fixed off by one error, changed min and max starting values and other general code cleaning

git-svn-id: http://svn.cactuscode.org/arrangements/CactusBase/LocalReduce/trunk@51 7daa882c-dc44-4453-834e-278d26b18e6a

31 files changed, 2735 insertions, 2735 deletions

diff --git a/src/AvgFunctions.c b/src/AvgFunctions.c
index 5228ef9..428598d 100644
--- a/src/AvgFunctions.c
+++ b/src/AvgFunctions.c
@@ -19,7 +19,7 @@ extern "C" {
 
 
 
-int LocalReduce_Avg_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -40,47 +40,47 @@ int LocalReduce_Avg_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -106,7 +106,7 @@ int LocalReduce_Avg_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -142,7 +142,7 @@ int LocalReduce_Avg_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -180,7 +180,7 @@ int LocalReduce_Avg_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -218,7 +218,7 @@ int LocalReduce_Avg_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -235,7 +235,7 @@ int LocalReduce_Avg_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
   return num_points;
 }
 
-int LocalReduce_Avg_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -256,47 +256,47 @@ int LocalReduce_Avg_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -321,7 +321,7 @@ int LocalReduce_Avg_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -357,7 +357,7 @@ int LocalReduce_Avg_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -394,7 +394,7 @@ int LocalReduce_Avg_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd16( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -431,7 +431,7 @@ int LocalReduce_Avg_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd32( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -449,7 +449,7 @@ int LocalReduce_Avg_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 }
 
 #ifdef CCTK_VARIABLE_INT1
-int LocalReduce_Avg_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -470,47 +470,47 @@ int LocalReduce_Avg_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT11
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT12
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT14
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT18
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -535,7 +535,7 @@ int LocalReduce_Avg_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -571,7 +571,7 @@ int LocalReduce_Avg_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -609,7 +609,7 @@ int LocalReduce_Avg_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -647,7 +647,7 @@ int LocalReduce_Avg_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -666,7 +666,7 @@ int LocalReduce_Avg_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif                                                              
 
 #ifdef CCTK_VARIABLE_INT2
-int LocalReduce_Avg_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -687,47 +687,47 @@ int LocalReduce_Avg_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT21
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT22
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT24
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT28
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -752,7 +752,7 @@ int LocalReduce_Avg_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -788,7 +788,7 @@ int LocalReduce_Avg_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -826,7 +826,7 @@ int LocalReduce_Avg_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -864,7 +864,7 @@ int LocalReduce_Avg_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -883,7 +883,7 @@ int LocalReduce_Avg_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif
 
 #ifdef CCTK_VARIABLE_INT4
-int LocalReduce_Avg_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -904,47 +904,47 @@ int LocalReduce_Avg_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -969,7 +969,7 @@ int LocalReduce_Avg_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1005,7 +1005,7 @@ int LocalReduce_Avg_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1043,7 +1043,7 @@ int LocalReduce_Avg_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1081,7 +1081,7 @@ int LocalReduce_Avg_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1100,7 +1100,7 @@ int LocalReduce_Avg_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif
 
 #ifdef CCTK_VARIABLE_INT8
-int LocalReduce_Avg_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1121,47 +1121,47 @@ int LocalReduce_Avg_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1186,7 +1186,7 @@ int LocalReduce_Avg_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1222,7 +1222,7 @@ int LocalReduce_Avg_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1260,7 +1260,7 @@ int LocalReduce_Avg_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1298,7 +1298,7 @@ int LocalReduce_Avg_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1316,7 +1316,7 @@ int LocalReduce_Avg_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 }
 #endif
 
-int LocalReduce_Avg_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1337,47 +1337,47 @@ int LocalReduce_Avg_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1402,7 +1402,7 @@ int LocalReduce_Avg_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1438,7 +1438,7 @@ int LocalReduce_Avg_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1476,7 +1476,7 @@ int LocalReduce_Avg_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1514,7 +1514,7 @@ int LocalReduce_Avg_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1532,7 +1532,7 @@ int LocalReduce_Avg_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 }
 
 #ifdef CCTK_VARIABLE_REAL4
-int LocalReduce_Avg_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1553,47 +1553,47 @@ int LocalReduce_Avg_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1618,7 +1618,7 @@ int LocalReduce_Avg_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1654,7 +1654,7 @@ int LocalReduce_Avg_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1692,7 +1692,7 @@ int LocalReduce_Avg_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1730,7 +1730,7 @@ int LocalReduce_Avg_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1749,7 +1749,7 @@ int LocalReduce_Avg_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #endif
 
 #ifdef CCTK_VARIABLE_REAL8
-int LocalReduce_Avg_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1770,47 +1770,47 @@ int LocalReduce_Avg_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1835,7 +1835,7 @@ int LocalReduce_Avg_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1871,7 +1871,7 @@ int LocalReduce_Avg_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1909,7 +1909,7 @@ int LocalReduce_Avg_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1947,7 +1947,7 @@ int LocalReduce_Avg_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1967,7 +1967,7 @@ int LocalReduce_Avg_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Avg_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1988,47 +1988,47 @@ int LocalReduce_Avg_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2053,7 +2053,7 @@ int LocalReduce_Avg_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2089,7 +2089,7 @@ int LocalReduce_Avg_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2127,7 +2127,7 @@ int LocalReduce_Avg_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2165,7 +2165,7 @@ int LocalReduce_Avg_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2183,7 +2183,7 @@ int LocalReduce_Avg_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 #endif
 
-int LocalReduce_Avg_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2215,47 +2215,47 @@ int LocalReduce_Avg_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2270,7 +2270,7 @@ int LocalReduce_Avg_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2307,7 +2307,7 @@ int LocalReduce_Avg_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2346,7 +2346,7 @@ int LocalReduce_Avg_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2385,7 +2385,7 @@ int LocalReduce_Avg_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2415,7 +2415,7 @@ int LocalReduce_Avg_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Avg_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2447,47 +2447,47 @@ int LocalReduce_Avg_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2502,7 +2502,7 @@ int LocalReduce_Avg_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2539,7 +2539,7 @@ int LocalReduce_Avg_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2578,7 +2578,7 @@ int LocalReduce_Avg_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2617,7 +2617,7 @@ int LocalReduce_Avg_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2647,7 +2647,7 @@ int LocalReduce_Avg_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Avg_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2679,47 +2679,47 @@ int LocalReduce_Avg_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2734,7 +2734,7 @@ int LocalReduce_Avg_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2771,7 +2771,7 @@ int LocalReduce_Avg_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2810,7 +2810,7 @@ int LocalReduce_Avg_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2849,7 +2849,7 @@ int LocalReduce_Avg_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2879,7 +2879,7 @@ int LocalReduce_Avg_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Avg_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Avg_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2911,47 +2911,47 @@ int LocalReduce_Avg_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2966,7 +2966,7 @@ int LocalReduce_Avg_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3003,7 +3003,7 @@ int LocalReduce_Avg_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(avg, scalar)   avg = CCTK_CmplxAdd8( avg, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3042,7 +3042,7 @@ int LocalReduce_Avg_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3081,7 +3081,7 @@ int LocalReduce_Avg_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
diff --git a/src/Avg_Functions.h b/src/Avg_Functions.h
index c183594..eece300 100644
--- a/src/Avg_Functions.h
+++ b/src/Avg_Functions.h
@@ -18,52 +18,52 @@ extern "C" {
 #endif
 
 /* Avg reduction functions */
-int LocalReduce_Avg_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
-int LocalReduce_Avg_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_INT1
-int LocalReduce_Avg_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Avg_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Avg_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Avg_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Avg_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_REAL4
-int LocalReduce_Avg_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Avg_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Avg_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Avg_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Avg_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Avg_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Avg_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef __cplusplus
diff --git a/src/CountFunctions.c b/src/CountFunctions.c
index 1f0a5a9..a86f057 100644
--- a/src/CountFunctions.c
+++ b/src/CountFunctions.c
@@ -17,7 +17,7 @@
 extern "C" {
 #endif
 
-int LocalReduce_Count_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -38,47 +38,47 @@ int LocalReduce_Count_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -104,7 +104,7 @@ int LocalReduce_Count_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -140,7 +140,7 @@ int LocalReduce_Count_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -178,7 +178,7 @@ int LocalReduce_Count_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -216,7 +216,7 @@ int LocalReduce_Count_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -233,7 +233,7 @@ int LocalReduce_Count_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
   return num_points;
 }
 
-int LocalReduce_Count_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -254,47 +254,47 @@ int LocalReduce_Count_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -319,7 +319,7 @@ int LocalReduce_Count_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -355,7 +355,7 @@ int LocalReduce_Count_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -392,7 +392,7 @@ int LocalReduce_Count_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -429,7 +429,7 @@ int LocalReduce_Count_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -447,7 +447,7 @@ int LocalReduce_Count_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 }
 
 #ifdef CCTK_INT1
-int LocalReduce_Count_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -468,47 +468,47 @@ int LocalReduce_Count_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT11
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT12
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT14
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT18
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -533,7 +533,7 @@ int LocalReduce_Count_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -569,7 +569,7 @@ int LocalReduce_Count_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -607,7 +607,7 @@ int LocalReduce_Count_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -645,7 +645,7 @@ int LocalReduce_Count_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -664,7 +664,7 @@ int LocalReduce_Count_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Count_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -685,47 +685,47 @@ int LocalReduce_Count_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT21
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT22
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT24
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT28
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -750,7 +750,7 @@ int LocalReduce_Count_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -786,7 +786,7 @@ int LocalReduce_Count_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -824,7 +824,7 @@ int LocalReduce_Count_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -862,7 +862,7 @@ int LocalReduce_Count_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -881,7 +881,7 @@ int LocalReduce_Count_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Count_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -902,47 +902,47 @@ int LocalReduce_Count_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -967,7 +967,7 @@ int LocalReduce_Count_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1003,7 +1003,7 @@ int LocalReduce_Count_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1041,7 +1041,7 @@ int LocalReduce_Count_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1079,7 +1079,7 @@ int LocalReduce_Count_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1098,7 +1098,7 @@ int LocalReduce_Count_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Count_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1119,47 +1119,47 @@ int LocalReduce_Count_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1184,7 +1184,7 @@ int LocalReduce_Count_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1220,7 +1220,7 @@ int LocalReduce_Count_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1258,7 +1258,7 @@ int LocalReduce_Count_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1296,7 +1296,7 @@ int LocalReduce_Count_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1314,7 +1314,7 @@ int LocalReduce_Count_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 #endif
 
-int LocalReduce_Count_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1335,47 +1335,47 @@ int LocalReduce_Count_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1400,7 +1400,7 @@ int LocalReduce_Count_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1436,7 +1436,7 @@ int LocalReduce_Count_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1474,7 +1474,7 @@ int LocalReduce_Count_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1512,7 +1512,7 @@ int LocalReduce_Count_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1530,7 +1530,7 @@ int LocalReduce_Count_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 
 #ifdef CCTK_REAL4
-int LocalReduce_Count_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1551,47 +1551,47 @@ int LocalReduce_Count_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1616,7 +1616,7 @@ int LocalReduce_Count_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1652,7 +1652,7 @@ int LocalReduce_Count_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1690,7 +1690,7 @@ int LocalReduce_Count_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1728,7 +1728,7 @@ int LocalReduce_Count_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1747,7 +1747,7 @@ int LocalReduce_Count_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Count_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1768,47 +1768,47 @@ int LocalReduce_Count_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1833,7 +1833,7 @@ int LocalReduce_Count_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1869,7 +1869,7 @@ int LocalReduce_Count_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1907,7 +1907,7 @@ int LocalReduce_Count_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1945,7 +1945,7 @@ int LocalReduce_Count_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1965,7 +1965,7 @@ int LocalReduce_Count_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Count_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1986,47 +1986,47 @@ int LocalReduce_Count_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2051,7 +2051,7 @@ int LocalReduce_Count_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2087,7 +2087,7 @@ int LocalReduce_Count_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2125,7 +2125,7 @@ int LocalReduce_Count_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2163,7 +2163,7 @@ int LocalReduce_Count_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2181,7 +2181,7 @@ int LocalReduce_Count_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 }
 #endif
 
-int LocalReduce_Count_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2213,47 +2213,47 @@ int LocalReduce_Count_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2268,7 +2268,7 @@ int LocalReduce_Count_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2305,7 +2305,7 @@ int LocalReduce_Count_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2344,7 +2344,7 @@ int LocalReduce_Count_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2383,7 +2383,7 @@ int LocalReduce_Count_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2413,7 +2413,7 @@ int LocalReduce_Count_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Count_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2445,47 +2445,47 @@ int LocalReduce_Count_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2500,7 +2500,7 @@ int LocalReduce_Count_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2537,7 +2537,7 @@ int LocalReduce_Count_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2576,7 +2576,7 @@ int LocalReduce_Count_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2615,7 +2615,7 @@ int LocalReduce_Count_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2645,7 +2645,7 @@ int LocalReduce_Count_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Count_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2677,47 +2677,47 @@ int LocalReduce_Count_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2732,7 +2732,7 @@ int LocalReduce_Count_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2769,7 +2769,7 @@ int LocalReduce_Count_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2808,7 +2808,7 @@ int LocalReduce_Count_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2847,7 +2847,7 @@ int LocalReduce_Count_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2877,7 +2877,7 @@ int LocalReduce_Count_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Count_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Count_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2909,47 +2909,47 @@ int LocalReduce_Count_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2964,7 +2964,7 @@ int LocalReduce_Count_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3001,7 +3001,7 @@ int LocalReduce_Count_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Count, scalar) (Count).Re = num_points;
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3040,7 +3040,7 @@ int LocalReduce_Count_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3079,7 +3079,7 @@ int LocalReduce_Count_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
diff --git a/src/Count_Functions.h b/src/Count_Functions.h
index 9df20f3..452ceb5 100644
--- a/src/Count_Functions.h
+++ b/src/Count_Functions.h
@@ -18,52 +18,52 @@ extern "C" {
 #endif
 
 /* Count reduction functions */
-int LocalReduce_Count_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
-int LocalReduce_Count_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_INT1
-int LocalReduce_Count_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Count_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Count_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Count_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Count_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_REAL4
-int LocalReduce_Count_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Count_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Count_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Count_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Count_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Count_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Count_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Count_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef __cplusplus
diff --git a/src/MaxFunctions.c b/src/MaxFunctions.c
index 92b0dc3..69446c9 100644
--- a/src/MaxFunctions.c
+++ b/src/MaxFunctions.c
@@ -18,7 +18,7 @@ extern "C" {
 #endif
 
 
-int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -32,54 +32,54 @@ int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -103,9 +103,9 @@ int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -114,7 +114,7 @@ int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -139,9 +139,9 @@ int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -150,7 +150,7 @@ int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -176,10 +176,10 @@ int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -188,7 +188,7 @@ int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -214,10 +214,10 @@ int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -226,7 +226,7 @@ int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -234,7 +234,7 @@ int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
   return num_points;
 }
 
-int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -248,54 +248,54 @@ int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -318,9 +318,9 @@ int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -329,7 +329,7 @@ int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -354,9 +354,9 @@ int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -365,7 +365,7 @@ int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -391,9 +391,9 @@ int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -402,7 +402,7 @@ int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -428,9 +428,9 @@ int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -439,7 +439,7 @@ int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -448,7 +448,7 @@ int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 }
 
 #ifdef CCTK_INT1
-int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -462,54 +462,54 @@ int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT11
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT12
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT14
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT18
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -532,9 +532,9 @@ int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -543,7 +543,7 @@ int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -568,9 +568,9 @@ int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -579,7 +579,7 @@ int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -605,10 +605,10 @@ int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -617,7 +617,7 @@ int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -643,10 +643,10 @@ int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -655,7 +655,7 @@ int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -665,7 +665,7 @@ int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -679,54 +679,54 @@ int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT21
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT22
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT24
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT28
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -749,9 +749,9 @@ int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -760,7 +760,7 @@ int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -785,9 +785,9 @@ int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -796,7 +796,7 @@ int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -822,10 +822,10 @@ int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -834,7 +834,7 @@ int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -860,10 +860,10 @@ int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -872,7 +872,7 @@ int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -882,7 +882,7 @@ int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -896,54 +896,54 @@ int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -966,9 +966,9 @@ int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -977,7 +977,7 @@ int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1002,9 +1002,9 @@ int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1013,7 +1013,7 @@ int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1039,10 +1039,10 @@ int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1051,7 +1051,7 @@ int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1077,10 +1077,10 @@ int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1089,7 +1089,7 @@ int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -1099,7 +1099,7 @@ int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1113,54 +1113,54 @@ int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1183,9 +1183,9 @@ int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1194,7 +1194,7 @@ int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1219,9 +1219,9 @@ int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1230,7 +1230,7 @@ int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1256,10 +1256,10 @@ int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1268,7 +1268,7 @@ int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1294,10 +1294,10 @@ int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1306,7 +1306,7 @@ int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -1315,7 +1315,7 @@ int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 }
 #endif
 
-int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1329,54 +1329,54 @@ int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1399,9 +1399,9 @@ int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1410,7 +1410,7 @@ int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1435,9 +1435,9 @@ int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1446,7 +1446,7 @@ int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1472,10 +1472,10 @@ int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1484,7 +1484,7 @@ int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1510,10 +1510,10 @@ int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1522,7 +1522,7 @@ int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -1531,7 +1531,7 @@ int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 }
 
 #ifdef CCTK_REAL4
-int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1545,54 +1545,54 @@ int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1615,9 +1615,9 @@ int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1626,7 +1626,7 @@ int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1651,9 +1651,9 @@ int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1662,7 +1662,7 @@ int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1688,10 +1688,10 @@ int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1700,7 +1700,7 @@ int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1726,10 +1726,10 @@ int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1738,7 +1738,7 @@ int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -1748,7 +1748,7 @@ int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1762,54 +1762,54 @@ int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1832,9 +1832,9 @@ int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1843,7 +1843,7 @@ int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1868,9 +1868,9 @@ int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1879,7 +1879,7 @@ int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1905,10 +1905,10 @@ int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1917,7 +1917,7 @@ int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1943,10 +1943,10 @@ int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1956,7 +1956,7 @@ int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval = (out_type) typed_vdata[sum_indices];
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -1966,7 +1966,7 @@ int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1980,54 +1980,54 @@ int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2050,9 +2050,9 @@ int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2061,7 +2061,7 @@ int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2086,9 +2086,9 @@ int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2097,7 +2097,7 @@ int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2123,10 +2123,10 @@ int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2135,7 +2135,7 @@ int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2161,10 +2161,10 @@ int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2173,7 +2173,7 @@ int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -2182,7 +2182,7 @@ int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 #endif
 
-int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2207,54 +2207,54 @@ int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2267,9 +2267,9 @@ int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2289,7 +2289,7 @@ int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2304,9 +2304,9 @@ int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL4) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL4) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2326,7 +2326,7 @@ int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2342,10 +2342,10 @@ int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL8) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL8) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2365,7 +2365,7 @@ int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2381,10 +2381,10 @@ int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL16) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL16) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2404,7 +2404,7 @@ int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -2414,7 +2414,7 @@ int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2439,54 +2439,54 @@ int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2499,9 +2499,9 @@ int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2521,7 +2521,7 @@ int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2536,9 +2536,9 @@ int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL4) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL4) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2558,7 +2558,7 @@ int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2574,10 +2574,10 @@ int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL8) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL8) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2597,7 +2597,7 @@ int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2613,10 +2613,10 @@ int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL16) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL16) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2636,7 +2636,7 @@ int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -2646,7 +2646,7 @@ int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2671,54 +2671,54 @@ int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2731,9 +2731,9 @@ int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2753,7 +2753,7 @@ int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2768,9 +2768,9 @@ int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL4) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL4) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2790,7 +2790,7 @@ int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2806,10 +2806,10 @@ int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL8) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL8) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2829,7 +2829,7 @@ int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2845,10 +2845,10 @@ int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL16) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL16) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2868,7 +2868,7 @@ int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -2878,7 +2878,7 @@ int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Max_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Max_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2903,54 +2903,54 @@ int LocalReduce_Max_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2963,9 +2963,9 @@ int LocalReduce_Max_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2985,7 +2985,7 @@ int LocalReduce_Max_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -3000,9 +3000,9 @@ int LocalReduce_Max_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL4) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL4) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3022,7 +3022,7 @@ int LocalReduce_Max_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -3038,10 +3038,10 @@ int LocalReduce_Max_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL8) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL8) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3061,7 +3061,7 @@ int LocalReduce_Max_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -3077,10 +3077,10 @@ int LocalReduce_Max_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL16) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL16) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Max, scalar)   if ((Max).Re < (scalar).Re) (Max).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3100,7 +3100,7 @@ int LocalReduce_Max_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #define  REDUCTION_OPERATION(Max, scalar)   if (Max < scalar) Max = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
     break;
     #endif
diff --git a/src/Max_Functions.h b/src/Max_Functions.h
index 275fe78..c6d2cbd 100644
--- a/src/Max_Functions.h
+++ b/src/Max_Functions.h
@@ -18,52 +18,52 @@ extern "C" {
 #endif
 
 /* Max reduction functions */
-int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
-int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_INT1
-int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_REAL4
-int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Max_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Max_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef __cplusplus
diff --git a/src/MinFunctions.c b/src/MinFunctions.c
index 1922ebf..9d77a3f 100644
--- a/src/MinFunctions.c
+++ b/src/MinFunctions.c
@@ -17,7 +17,7 @@
 extern "C" {
 #endif
 
-int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -30,55 +30,55 @@ int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -102,9 +102,9 @@ int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -112,8 +112,8 @@ int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -138,9 +138,9 @@ int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -148,8 +148,8 @@ int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -175,10 +175,10 @@ int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -186,8 +186,8 @@ int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -213,10 +213,10 @@ int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -224,8 +224,8 @@ int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -233,7 +233,7 @@ int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
   return num_points;
 }
 
-int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -246,55 +246,55 @@ int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -317,9 +317,9 @@ int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -327,8 +327,8 @@ int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -353,9 +353,9 @@ int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -363,8 +363,8 @@ int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -390,9 +390,9 @@ int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -400,8 +400,8 @@ int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -427,9 +427,9 @@ int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -437,8 +437,8 @@ int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -447,7 +447,7 @@ int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 }
 
 #ifdef CCTK_INT1
-int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -460,55 +460,55 @@ int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT11
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT12
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT14
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT18
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -531,9 +531,9 @@ int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -541,8 +541,8 @@ int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -567,9 +567,9 @@ int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -577,8 +577,8 @@ int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -604,10 +604,10 @@ int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -615,8 +615,8 @@ int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -642,10 +642,10 @@ int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -653,8 +653,8 @@ int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -664,7 +664,7 @@ int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -677,55 +677,55 @@ int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT21
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT22
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT24
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT28
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -748,9 +748,9 @@ int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -758,8 +758,8 @@ int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -784,9 +784,9 @@ int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -794,8 +794,8 @@ int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -821,10 +821,10 @@ int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -832,8 +832,8 @@ int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -859,10 +859,10 @@ int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -870,8 +870,8 @@ int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -881,7 +881,7 @@ int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -894,55 +894,55 @@ int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -965,9 +965,9 @@ int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -975,8 +975,8 @@ int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1001,9 +1001,9 @@ int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1011,8 +1011,8 @@ int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1038,10 +1038,10 @@ int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1049,8 +1049,8 @@ int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1076,10 +1076,10 @@ int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1087,8 +1087,8 @@ int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -1098,7 +1098,7 @@ int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1111,55 +1111,55 @@ int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1182,9 +1182,9 @@ int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1192,8 +1192,8 @@ int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1218,9 +1218,9 @@ int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1228,8 +1228,8 @@ int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1255,10 +1255,10 @@ int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1266,8 +1266,8 @@ int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1293,10 +1293,10 @@ int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1304,8 +1304,8 @@ int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -1314,7 +1314,7 @@ int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 }
 #endif
 
-int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1327,55 +1327,55 @@ int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1398,9 +1398,9 @@ int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1408,8 +1408,8 @@ int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1434,9 +1434,9 @@ int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1444,8 +1444,8 @@ int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1471,10 +1471,10 @@ int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1482,8 +1482,8 @@ int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1509,10 +1509,10 @@ int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1520,8 +1520,8 @@ int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -1530,7 +1530,7 @@ int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 }
 
 #ifdef CCTK_REAL4
-int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1543,55 +1543,55 @@ int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1614,9 +1614,9 @@ int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1624,8 +1624,8 @@ int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1650,9 +1650,9 @@ int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1660,8 +1660,8 @@ int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1687,10 +1687,10 @@ int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1698,8 +1698,8 @@ int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1725,10 +1725,10 @@ int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1736,8 +1736,8 @@ int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -1747,7 +1747,7 @@ int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1760,55 +1760,55 @@ int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1831,9 +1831,9 @@ int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1841,8 +1841,8 @@ int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1867,9 +1867,9 @@ int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1877,8 +1877,8 @@ int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1904,10 +1904,10 @@ int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1915,8 +1915,8 @@ int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -1942,10 +1942,10 @@ int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1954,8 +1954,8 @@ int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -1965,7 +1965,7 @@ int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1978,55 +1978,55 @@ int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2049,9 +2049,9 @@ int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2059,8 +2059,8 @@ int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2085,9 +2085,9 @@ int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2095,8 +2095,8 @@ int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2122,10 +2122,10 @@ int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2133,8 +2133,8 @@ int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2160,10 +2160,10 @@ int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
         inval.Re = (CCTK_REAL16) typed_vdata[sum_indices];
 #endif
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0]; (num).Im = 0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2171,8 +2171,8 @@ int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 
 #define REDUCTION_PREOP_CAST(inval, typed_vdata,sum_indices, out_type) \
         inval = (out_type) typed_vdata[sum_indices];
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -2181,7 +2181,7 @@ int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 #endif
 
-int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2205,55 +2205,55 @@ int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2266,9 +2266,9 @@ int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2287,8 +2287,8 @@ int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2303,9 +2303,9 @@ int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL4) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL4) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2324,8 +2324,8 @@ int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2341,10 +2341,10 @@ int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL8) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL8) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2363,8 +2363,8 @@ int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2380,10 +2380,10 @@ int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL16) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL16) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2402,8 +2402,8 @@ int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -2413,7 +2413,7 @@ int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2437,55 +2437,55 @@ int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2498,9 +2498,9 @@ int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2519,8 +2519,8 @@ int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2535,9 +2535,9 @@ int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL4) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL4) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2556,8 +2556,8 @@ int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2573,10 +2573,10 @@ int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL8) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL8) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2595,8 +2595,8 @@ int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2612,10 +2612,10 @@ int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
         (inval).Re = (CCTK_REAL16) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL16) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2634,8 +2634,8 @@ int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -2645,7 +2645,7 @@ int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2669,55 +2669,55 @@ int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2730,9 +2730,9 @@ int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2751,8 +2751,8 @@ int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2767,9 +2767,9 @@ int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL4) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL4) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2788,8 +2788,8 @@ int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2805,10 +2805,10 @@ int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL8) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL8) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2827,8 +2827,8 @@ int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2844,10 +2844,10 @@ int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL16) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL16) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2866,8 +2866,8 @@ int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
     break;
     #endif
@@ -2877,7 +2877,7 @@ int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Min_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Min_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2901,55 +2901,55 @@ int LocalReduce_Min_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
   switch (output_number_type_codes[i])
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2962,9 +2962,9 @@ int LocalReduce_Min_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2983,8 +2983,8 @@ int LocalReduce_Min_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0];
 #define EXTRA_STEP(a, b)
 
     break;
@@ -2999,9 +2999,9 @@ int LocalReduce_Min_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL4) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL4) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3020,8 +3020,8 @@ int LocalReduce_Min_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -3037,10 +3037,10 @@ int LocalReduce_Min_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL8) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL8) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3059,8 +3059,8 @@ int LocalReduce_Min_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
 
     break;
@@ -3076,10 +3076,10 @@ int LocalReduce_Min_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
         (inval).Re = (CCTK_REAL16) (typed_vdata[sum_indices]).Re; (inval).Im = (CCTK_REAL16) (typed_vdata[sum_indices]).Im;
 
 #define REDUCTION_OPERATION(Min, scalar)   if ((Min).Re > (scalar).Re) (Min).Re = (scalar).Re;
-#define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
+#define REDUCTION_INITIAL(num) (num).Re = typed_vdata[0].Re; (num).Im = typed_vdata[0].Im;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3098,8 +3098,8 @@ int LocalReduce_Min_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
         inval = (CCTK_REAL16) typed_vdata[sum_indices].Re;
 #endif
 
-#define  REDUCTION_OPERATION(Min, scalar)   if (Min < scalar) Min = scalar;
-#define REDUCTION_INITIAL(num) num = 0;
+#define  REDUCTION_OPERATION(Min, scalar)   if (Min > scalar) Min = scalar;
+#define REDUCTION_INITIAL(num) num = typed_vdata[0].Re;
 #define EXTRA_STEP(a, b)
     break;
     #endif
diff --git a/src/Min_Functions.h b/src/Min_Functions.h
index e690828..43f3483 100644
--- a/src/Min_Functions.h
+++ b/src/Min_Functions.h
@@ -18,52 +18,52 @@ extern "C" {
 #endif
 
 /* Min reduction functions */
-int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
-int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_INT1
-int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_REAL4
-int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Min_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Min_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef __cplusplus
diff --git a/src/Norm1Functions.c b/src/Norm1Functions.c
index 6330c9f..0fe5e7e 100644
--- a/src/Norm1Functions.c
+++ b/src/Norm1Functions.c
@@ -18,7 +18,7 @@ extern "C" {
 #endif
 
 
-int LocalReduce_Norm1_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -43,47 +43,47 @@ int LocalReduce_Norm1_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -113,7 +113,7 @@ int LocalReduce_Norm1_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -157,7 +157,7 @@ int LocalReduce_Norm1_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -203,7 +203,7 @@ int LocalReduce_Norm1_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -249,7 +249,7 @@ int LocalReduce_Norm1_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -270,7 +270,7 @@ int LocalReduce_Norm1_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
   return num_points;
 }
 
-int LocalReduce_Norm1_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -295,47 +295,47 @@ int LocalReduce_Norm1_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -364,7 +364,7 @@ int LocalReduce_Norm1_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -408,7 +408,7 @@ int LocalReduce_Norm1_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -453,7 +453,7 @@ int LocalReduce_Norm1_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd16( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -498,7 +498,7 @@ int LocalReduce_Norm1_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd32( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -520,7 +520,7 @@ int LocalReduce_Norm1_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 }
 
 #ifdef CCTK_INT1
-int LocalReduce_Norm1_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -545,47 +545,47 @@ int LocalReduce_Norm1_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT11
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT12
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT14
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT18
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -614,7 +614,7 @@ int LocalReduce_Norm1_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -658,7 +658,7 @@ int LocalReduce_Norm1_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -704,7 +704,7 @@ int LocalReduce_Norm1_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -750,7 +750,7 @@ int LocalReduce_Norm1_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -773,7 +773,7 @@ int LocalReduce_Norm1_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Norm1_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -798,47 +798,47 @@ int LocalReduce_Norm1_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT21
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT22
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT24
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT28
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -867,7 +867,7 @@ int LocalReduce_Norm1_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -911,7 +911,7 @@ int LocalReduce_Norm1_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -957,7 +957,7 @@ int LocalReduce_Norm1_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1003,7 +1003,7 @@ int LocalReduce_Norm1_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1026,7 +1026,7 @@ int LocalReduce_Norm1_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Norm1_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1051,47 +1051,47 @@ int LocalReduce_Norm1_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1120,7 +1120,7 @@ int LocalReduce_Norm1_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1164,7 +1164,7 @@ int LocalReduce_Norm1_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1210,7 +1210,7 @@ int LocalReduce_Norm1_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1256,7 +1256,7 @@ int LocalReduce_Norm1_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1279,7 +1279,7 @@ int LocalReduce_Norm1_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Norm1_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1304,47 +1304,47 @@ int LocalReduce_Norm1_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1373,7 +1373,7 @@ int LocalReduce_Norm1_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1417,7 +1417,7 @@ int LocalReduce_Norm1_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1463,7 +1463,7 @@ int LocalReduce_Norm1_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1509,7 +1509,7 @@ int LocalReduce_Norm1_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1531,7 +1531,7 @@ int LocalReduce_Norm1_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 #endif
 
-int LocalReduce_Norm1_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1556,47 +1556,47 @@ int LocalReduce_Norm1_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1625,7 +1625,7 @@ int LocalReduce_Norm1_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1669,7 +1669,7 @@ int LocalReduce_Norm1_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1715,7 +1715,7 @@ int LocalReduce_Norm1_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1761,7 +1761,7 @@ int LocalReduce_Norm1_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1783,7 +1783,7 @@ int LocalReduce_Norm1_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 
 #ifdef CCTK_REAL4
-int LocalReduce_Norm1_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1808,47 +1808,47 @@ int LocalReduce_Norm1_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1877,7 +1877,7 @@ int LocalReduce_Norm1_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1921,7 +1921,7 @@ int LocalReduce_Norm1_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1967,7 +1967,7 @@ int LocalReduce_Norm1_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2013,7 +2013,7 @@ int LocalReduce_Norm1_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2036,7 +2036,7 @@ int LocalReduce_Norm1_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Norm1_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2061,47 +2061,47 @@ int LocalReduce_Norm1_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2130,7 +2130,7 @@ int LocalReduce_Norm1_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2174,7 +2174,7 @@ int LocalReduce_Norm1_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2220,7 +2220,7 @@ int LocalReduce_Norm1_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2266,7 +2266,7 @@ int LocalReduce_Norm1_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2290,7 +2290,7 @@ int LocalReduce_Norm1_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Norm1_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2315,47 +2315,47 @@ int LocalReduce_Norm1_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2384,7 +2384,7 @@ int LocalReduce_Norm1_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2428,7 +2428,7 @@ int LocalReduce_Norm1_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2474,7 +2474,7 @@ int LocalReduce_Norm1_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2520,7 +2520,7 @@ int LocalReduce_Norm1_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2542,7 +2542,7 @@ int LocalReduce_Norm1_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 }
 #endif
 
-int LocalReduce_Norm1_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2574,47 +2574,47 @@ int LocalReduce_Norm1_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2633,7 +2633,7 @@ int LocalReduce_Norm1_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP 
@@ -2674,7 +2674,7 @@ int LocalReduce_Norm1_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2717,7 +2717,7 @@ int LocalReduce_Norm1_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2760,7 +2760,7 @@ int LocalReduce_Norm1_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2790,7 +2790,7 @@ int LocalReduce_Norm1_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Norm1_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2822,47 +2822,47 @@ int LocalReduce_Norm1_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2881,7 +2881,7 @@ int LocalReduce_Norm1_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2922,7 +2922,7 @@ int LocalReduce_Norm1_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2965,7 +2965,7 @@ int LocalReduce_Norm1_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3008,7 +3008,7 @@ int LocalReduce_Norm1_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3038,7 +3038,7 @@ int LocalReduce_Norm1_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Norm1_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -3070,47 +3070,47 @@ int LocalReduce_Norm1_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -3129,7 +3129,7 @@ int LocalReduce_Norm1_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3170,7 +3170,7 @@ int LocalReduce_Norm1_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3213,7 +3213,7 @@ int LocalReduce_Norm1_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3256,7 +3256,7 @@ int LocalReduce_Norm1_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3286,7 +3286,7 @@ int LocalReduce_Norm1_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Norm1_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm1_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -3318,47 +3318,47 @@ int LocalReduce_Norm1_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -3377,7 +3377,7 @@ int LocalReduce_Norm1_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);Norm1 = CCTK_CmplxAdd( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3418,7 +3418,7 @@ int LocalReduce_Norm1_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm1, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);Norm1 = CCTK_CmplxAdd8( Norm1, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3461,7 +3461,7 @@ int LocalReduce_Norm1_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3504,7 +3504,7 @@ int LocalReduce_Norm1_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
diff --git a/src/Norm1_Functions.h b/src/Norm1_Functions.h
index e6fd99b..54bd8e4 100644
--- a/src/Norm1_Functions.h
+++ b/src/Norm1_Functions.h
@@ -18,52 +18,52 @@ extern "C" {
 #endif
 
 /* Norm1 reduction functions */
-int LocalReduce_Norm1_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
-int LocalReduce_Norm1_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_INT1
-int LocalReduce_Norm1_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Norm1_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Norm1_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Norm1_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Norm1_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_REAL4
-int LocalReduce_Norm1_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Norm1_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Norm1_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Norm1_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Norm1_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Norm1_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Norm1_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm1_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef __cplusplus
diff --git a/src/Norm2Functions.c b/src/Norm2Functions.c
index 9caae45..e544d18 100644
--- a/src/Norm2Functions.c
+++ b/src/Norm2Functions.c
@@ -18,7 +18,7 @@ extern "C" {
 #endif
 
 
-int LocalReduce_Norm2_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -43,47 +43,47 @@ int LocalReduce_Norm2_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -113,7 +113,7 @@ int LocalReduce_Norm2_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -157,7 +157,7 @@ int LocalReduce_Norm2_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -203,7 +203,7 @@ int LocalReduce_Norm2_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -249,7 +249,7 @@ int LocalReduce_Norm2_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -270,7 +270,7 @@ int LocalReduce_Norm2_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
   return num_points;
 }
 
-int LocalReduce_Norm2_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -295,47 +295,47 @@ int LocalReduce_Norm2_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -364,7 +364,7 @@ int LocalReduce_Norm2_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -408,7 +408,7 @@ int LocalReduce_Norm2_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -453,7 +453,7 @@ int LocalReduce_Norm2_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd16( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -498,7 +498,7 @@ int LocalReduce_Norm2_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd32( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -520,7 +520,7 @@ int LocalReduce_Norm2_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 }
 
 #ifdef CCTK_INT1
-int LocalReduce_Norm2_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -545,47 +545,47 @@ int LocalReduce_Norm2_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT11
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT12
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT14
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT18
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -614,7 +614,7 @@ int LocalReduce_Norm2_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -658,7 +658,7 @@ int LocalReduce_Norm2_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -704,7 +704,7 @@ int LocalReduce_Norm2_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -750,7 +750,7 @@ int LocalReduce_Norm2_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -773,7 +773,7 @@ int LocalReduce_Norm2_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Norm2_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -798,47 +798,47 @@ int LocalReduce_Norm2_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT21
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT22
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT24
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT28
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -867,7 +867,7 @@ int LocalReduce_Norm2_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -911,7 +911,7 @@ int LocalReduce_Norm2_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -957,7 +957,7 @@ int LocalReduce_Norm2_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1003,7 +1003,7 @@ int LocalReduce_Norm2_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1026,7 +1026,7 @@ int LocalReduce_Norm2_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Norm2_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1051,47 +1051,47 @@ int LocalReduce_Norm2_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1120,7 +1120,7 @@ int LocalReduce_Norm2_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1164,7 +1164,7 @@ int LocalReduce_Norm2_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1210,7 +1210,7 @@ int LocalReduce_Norm2_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1256,7 +1256,7 @@ int LocalReduce_Norm2_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1279,7 +1279,7 @@ int LocalReduce_Norm2_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Norm2_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1304,47 +1304,47 @@ int LocalReduce_Norm2_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1373,7 +1373,7 @@ int LocalReduce_Norm2_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1417,7 +1417,7 @@ int LocalReduce_Norm2_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1463,7 +1463,7 @@ int LocalReduce_Norm2_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1509,7 +1509,7 @@ int LocalReduce_Norm2_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1531,7 +1531,7 @@ int LocalReduce_Norm2_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 #endif
 
-int LocalReduce_Norm2_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1556,47 +1556,47 @@ int LocalReduce_Norm2_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1625,7 +1625,7 @@ int LocalReduce_Norm2_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1669,7 +1669,7 @@ int LocalReduce_Norm2_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1715,7 +1715,7 @@ int LocalReduce_Norm2_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1761,7 +1761,7 @@ int LocalReduce_Norm2_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1783,7 +1783,7 @@ int LocalReduce_Norm2_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 
 #ifdef CCTK_REAL4
-int LocalReduce_Norm2_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1808,47 +1808,47 @@ int LocalReduce_Norm2_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1877,7 +1877,7 @@ int LocalReduce_Norm2_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1921,7 +1921,7 @@ int LocalReduce_Norm2_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1967,7 +1967,7 @@ int LocalReduce_Norm2_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2013,7 +2013,7 @@ int LocalReduce_Norm2_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2036,7 +2036,7 @@ int LocalReduce_Norm2_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Norm2_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2061,47 +2061,47 @@ int LocalReduce_Norm2_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2130,7 +2130,7 @@ int LocalReduce_Norm2_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2174,7 +2174,7 @@ int LocalReduce_Norm2_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2220,7 +2220,7 @@ int LocalReduce_Norm2_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2266,7 +2266,7 @@ int LocalReduce_Norm2_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2290,7 +2290,7 @@ int LocalReduce_Norm2_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Norm2_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2315,47 +2315,47 @@ int LocalReduce_Norm2_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2384,7 +2384,7 @@ int LocalReduce_Norm2_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2428,7 +2428,7 @@ int LocalReduce_Norm2_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm2, scalar)   (scalar).Re = SQR((scalar).Re); Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2474,7 +2474,7 @@ int LocalReduce_Norm2_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2520,7 +2520,7 @@ int LocalReduce_Norm2_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2542,7 +2542,7 @@ int LocalReduce_Norm2_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 }
 #endif
 
-int LocalReduce_Norm2_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2578,47 +2578,47 @@ int LocalReduce_Norm2_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2637,7 +2637,7 @@ int LocalReduce_Norm2_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Norm2, scalar)   scalar = CCTK_CmplxMul(scalar,scalar); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2682,7 +2682,7 @@ int LocalReduce_Norm2_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Norm2, scalar)   scalar = CCTK_CmplxMul8(scalar,scalar);Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2729,7 +2729,7 @@ int LocalReduce_Norm2_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2776,7 +2776,7 @@ int LocalReduce_Norm2_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2810,7 +2810,7 @@ int LocalReduce_Norm2_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Norm2_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2846,47 +2846,47 @@ int LocalReduce_Norm2_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2905,7 +2905,7 @@ int LocalReduce_Norm2_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_OPERATION(Norm2, scalar)   scalar = CCTK_CmplxMul(scalar,scalar); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2950,7 +2950,7 @@ int LocalReduce_Norm2_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_OPERATION(Norm2, scalar)   scalar = CCTK_CmplxMul8(scalar,scalar);Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2997,7 +2997,7 @@ int LocalReduce_Norm2_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3044,7 +3044,7 @@ int LocalReduce_Norm2_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3078,7 +3078,7 @@ int LocalReduce_Norm2_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Norm2_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -3114,47 +3114,47 @@ int LocalReduce_Norm2_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -3173,7 +3173,7 @@ int LocalReduce_Norm2_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm2, scalar)   scalar = CCTK_CmplxMul(scalar,scalar); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3218,7 +3218,7 @@ int LocalReduce_Norm2_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm2, scalar)   scalar = CCTK_CmplxMul8(scalar,scalar);Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3265,7 +3265,7 @@ int LocalReduce_Norm2_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3312,7 +3312,7 @@ int LocalReduce_Norm2_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3346,7 +3346,7 @@ int LocalReduce_Norm2_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Norm2_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm2_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -3382,47 +3382,47 @@ int LocalReduce_Norm2_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -3441,7 +3441,7 @@ int LocalReduce_Norm2_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm2, scalar)   scalar = CCTK_CmplxMul(scalar,scalar); Norm2 = CCTK_CmplxAdd( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3486,7 +3486,7 @@ int LocalReduce_Norm2_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm2, scalar)   scalar = CCTK_CmplxMul8(scalar,scalar);Norm2 = CCTK_CmplxAdd8( Norm2, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3533,7 +3533,7 @@ int LocalReduce_Norm2_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3580,7 +3580,7 @@ int LocalReduce_Norm2_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
diff --git a/src/Norm2_Functions.h b/src/Norm2_Functions.h
index 68edced..7e4c33c 100644
--- a/src/Norm2_Functions.h
+++ b/src/Norm2_Functions.h
@@ -18,52 +18,52 @@ extern "C" {
 #endif
 
 /* Norm2 reduction functions */
-int LocalReduce_Norm2_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
-int LocalReduce_Norm2_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_INT1
-int LocalReduce_Norm2_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Norm2_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Norm2_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Norm2_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Norm2_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_REAL4
-int LocalReduce_Norm2_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Norm2_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Norm2_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Norm2_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Norm2_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Norm2_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Norm2_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm2_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef __cplusplus
diff --git a/src/Norm3Functions.c b/src/Norm3Functions.c
index f12d68c..9b34c7a 100644
--- a/src/Norm3Functions.c
+++ b/src/Norm3Functions.c
@@ -23,7 +23,7 @@ extern "C" {
 #define ABS(x)      ((x) < 0 ? -(x) : (x))
 
 
-int LocalReduce_Norm3_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -48,47 +48,47 @@ int LocalReduce_Norm3_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -118,7 +118,7 @@ int LocalReduce_Norm3_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -162,7 +162,7 @@ int LocalReduce_Norm3_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -208,7 +208,7 @@ int LocalReduce_Norm3_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -254,7 +254,7 @@ int LocalReduce_Norm3_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -275,7 +275,7 @@ int LocalReduce_Norm3_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
   return num_points;
 }
 
-int LocalReduce_Norm3_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -300,47 +300,47 @@ int LocalReduce_Norm3_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -369,7 +369,7 @@ int LocalReduce_Norm3_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -413,7 +413,7 @@ int LocalReduce_Norm3_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -458,7 +458,7 @@ int LocalReduce_Norm3_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); Norm3 = CCTK_CmplxAdd16( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -503,7 +503,7 @@ int LocalReduce_Norm3_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); Norm3 = CCTK_CmplxAdd32( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -525,7 +525,7 @@ int LocalReduce_Norm3_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 }
 
 #ifdef CCTK_INT1
-int LocalReduce_Norm3_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -550,47 +550,47 @@ int LocalReduce_Norm3_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT11
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT12
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT14
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT18
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -619,7 +619,7 @@ int LocalReduce_Norm3_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -663,7 +663,7 @@ int LocalReduce_Norm3_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -709,7 +709,7 @@ int LocalReduce_Norm3_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -755,7 +755,7 @@ int LocalReduce_Norm3_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -778,7 +778,7 @@ int LocalReduce_Norm3_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Norm3_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -803,47 +803,47 @@ int LocalReduce_Norm3_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT21
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT22
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT24
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT28
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -872,7 +872,7 @@ int LocalReduce_Norm3_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -916,7 +916,7 @@ int LocalReduce_Norm3_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -962,7 +962,7 @@ int LocalReduce_Norm3_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1008,7 +1008,7 @@ int LocalReduce_Norm3_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1031,7 +1031,7 @@ int LocalReduce_Norm3_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Norm3_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1056,47 +1056,47 @@ int LocalReduce_Norm3_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1125,7 +1125,7 @@ int LocalReduce_Norm3_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1169,7 +1169,7 @@ int LocalReduce_Norm3_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1215,7 +1215,7 @@ int LocalReduce_Norm3_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1261,7 +1261,7 @@ int LocalReduce_Norm3_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1284,7 +1284,7 @@ int LocalReduce_Norm3_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Norm3_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1309,47 +1309,47 @@ int LocalReduce_Norm3_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1378,7 +1378,7 @@ int LocalReduce_Norm3_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1422,7 +1422,7 @@ int LocalReduce_Norm3_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1468,7 +1468,7 @@ int LocalReduce_Norm3_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1514,7 +1514,7 @@ int LocalReduce_Norm3_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1536,7 +1536,7 @@ int LocalReduce_Norm3_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 #endif
 
-int LocalReduce_Norm3_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1561,47 +1561,47 @@ int LocalReduce_Norm3_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1630,7 +1630,7 @@ int LocalReduce_Norm3_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1674,7 +1674,7 @@ int LocalReduce_Norm3_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1720,7 +1720,7 @@ int LocalReduce_Norm3_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1766,7 +1766,7 @@ int LocalReduce_Norm3_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1788,7 +1788,7 @@ int LocalReduce_Norm3_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 
 #ifdef CCTK_REAL4
-int LocalReduce_Norm3_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1813,47 +1813,47 @@ int LocalReduce_Norm3_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1882,7 +1882,7 @@ int LocalReduce_Norm3_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1926,7 +1926,7 @@ int LocalReduce_Norm3_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1972,7 +1972,7 @@ int LocalReduce_Norm3_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2018,7 +2018,7 @@ int LocalReduce_Norm3_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2041,7 +2041,7 @@ int LocalReduce_Norm3_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Norm3_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2066,47 +2066,47 @@ int LocalReduce_Norm3_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2135,7 +2135,7 @@ int LocalReduce_Norm3_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2179,7 +2179,7 @@ int LocalReduce_Norm3_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2225,7 +2225,7 @@ int LocalReduce_Norm3_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2271,7 +2271,7 @@ int LocalReduce_Norm3_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2295,7 +2295,7 @@ int LocalReduce_Norm3_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Norm3_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2320,47 +2320,47 @@ int LocalReduce_Norm3_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2389,7 +2389,7 @@ int LocalReduce_Norm3_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2433,7 +2433,7 @@ int LocalReduce_Norm3_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm3, scalar)   (scalar).Re = CUBE_ABS((scalar).Re); Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2479,7 +2479,7 @@ int LocalReduce_Norm3_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2525,7 +2525,7 @@ int LocalReduce_Norm3_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2547,7 +2547,7 @@ int LocalReduce_Norm3_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 }
 #endif
 
-int LocalReduce_Norm3_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2583,47 +2583,47 @@ int LocalReduce_Norm3_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2642,7 +2642,7 @@ int LocalReduce_Norm3_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Norm3, scalar)   scalar = CCTK_CmplxMul(scalar, CCTK_CmplxMul(scalar,scalar)); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2687,7 +2687,7 @@ int LocalReduce_Norm3_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Norm3, scalar)   scalar = CCTK_CmplxMul8(scalar, CCTK_CmplxMul8(scalar,scalar)); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2734,7 +2734,7 @@ int LocalReduce_Norm3_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2781,7 +2781,7 @@ int LocalReduce_Norm3_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2815,7 +2815,7 @@ int LocalReduce_Norm3_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Norm3_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2851,47 +2851,47 @@ int LocalReduce_Norm3_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2910,7 +2910,7 @@ int LocalReduce_Norm3_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_OPERATION(Norm3, scalar)   scalar = CCTK_CmplxMul(scalar, CCTK_CmplxMul(scalar,scalar)); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2955,7 +2955,7 @@ int LocalReduce_Norm3_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_OPERATION(Norm3, scalar)   scalar = CCTK_CmplxMul8(scalar, CCTK_CmplxMul8(scalar,scalar)); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3002,7 +3002,7 @@ int LocalReduce_Norm3_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3049,7 +3049,7 @@ int LocalReduce_Norm3_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3083,7 +3083,7 @@ int LocalReduce_Norm3_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Norm3_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -3119,47 +3119,47 @@ int LocalReduce_Norm3_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -3178,7 +3178,7 @@ int LocalReduce_Norm3_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm3, scalar)   scalar = CCTK_CmplxMul(scalar, CCTK_CmplxMul(scalar,scalar)); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3223,7 +3223,7 @@ int LocalReduce_Norm3_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm3, scalar)   scalar = CCTK_CmplxMul8(scalar, CCTK_CmplxMul8(scalar,scalar)); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3270,7 +3270,7 @@ int LocalReduce_Norm3_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3317,7 +3317,7 @@ int LocalReduce_Norm3_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3351,7 +3351,7 @@ int LocalReduce_Norm3_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Norm3_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm3_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -3387,47 +3387,47 @@ int LocalReduce_Norm3_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -3446,7 +3446,7 @@ int LocalReduce_Norm3_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm3, scalar)   scalar = CCTK_CmplxMul(scalar, CCTK_CmplxMul(scalar,scalar)); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3491,7 +3491,7 @@ int LocalReduce_Norm3_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm3, scalar)   scalar = CCTK_CmplxMul8(scalar, CCTK_CmplxMul8(scalar,scalar)); (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);  Norm3 = CCTK_CmplxAdd8( Norm3, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3538,7 +3538,7 @@ int LocalReduce_Norm3_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3585,7 +3585,7 @@ int LocalReduce_Norm3_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
diff --git a/src/Norm3_Functions.h b/src/Norm3_Functions.h
index 09a1aa2..2501e1d 100644
--- a/src/Norm3_Functions.h
+++ b/src/Norm3_Functions.h
@@ -18,52 +18,52 @@ extern "C" {
 #endif
 
 /* Norm3 reduction functions */
-int LocalReduce_Norm3_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
-int LocalReduce_Norm3_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_INT1
-int LocalReduce_Norm3_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Norm3_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Norm3_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Norm3_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Norm3_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_REAL4
-int LocalReduce_Norm3_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Norm3_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Norm3_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Norm3_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Norm3_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Norm3_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Norm3_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm3_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef __cplusplus
diff --git a/src/Norm4Functions.c b/src/Norm4Functions.c
index 64cc946..a191b10 100644
--- a/src/Norm4Functions.c
+++ b/src/Norm4Functions.c
@@ -18,7 +18,7 @@ extern "C" {
 #endif
 
 
-int LocalReduce_Norm4_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -43,47 +43,47 @@ int LocalReduce_Norm4_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -113,7 +113,7 @@ int LocalReduce_Norm4_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re);  Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -157,7 +157,7 @@ int LocalReduce_Norm4_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re); Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -203,7 +203,7 @@ int LocalReduce_Norm4_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -249,7 +249,7 @@ int LocalReduce_Norm4_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -270,7 +270,7 @@ int LocalReduce_Norm4_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, i
   return num_points;
 }
 
-int LocalReduce_Norm4_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -295,47 +295,47 @@ int LocalReduce_Norm4_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -364,7 +364,7 @@ int LocalReduce_Norm4_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re);  Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -408,7 +408,7 @@ int LocalReduce_Norm4_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re); Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -453,7 +453,7 @@ int LocalReduce_Norm4_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re); Norm4 = CCTK_CmplxAdd16( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -498,7 +498,7 @@ int LocalReduce_Norm4_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re); Norm4 = CCTK_CmplxAdd32( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -520,7 +520,7 @@ int LocalReduce_Norm4_INT(int i, int mask_on, CCTK_INT * input_array_offsets, in
 }
 
 #ifdef CCTK_INT1
-int LocalReduce_Norm4_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -545,47 +545,47 @@ int LocalReduce_Norm4_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT11
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT12
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT14
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT18
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -614,7 +614,7 @@ int LocalReduce_Norm4_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re);  Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -658,7 +658,7 @@ int LocalReduce_Norm4_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re); Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -704,7 +704,7 @@ int LocalReduce_Norm4_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -750,7 +750,7 @@ int LocalReduce_Norm4_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -773,7 +773,7 @@ int LocalReduce_Norm4_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Norm4_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -798,47 +798,47 @@ int LocalReduce_Norm4_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT21
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT22
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT24
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT28
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -867,7 +867,7 @@ int LocalReduce_Norm4_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re);  Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -911,7 +911,7 @@ int LocalReduce_Norm4_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re); Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -957,7 +957,7 @@ int LocalReduce_Norm4_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1003,7 +1003,7 @@ int LocalReduce_Norm4_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1026,7 +1026,7 @@ int LocalReduce_Norm4_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Norm4_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1051,47 +1051,47 @@ int LocalReduce_Norm4_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1120,7 +1120,7 @@ int LocalReduce_Norm4_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re);  Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1164,7 +1164,7 @@ int LocalReduce_Norm4_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re); Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1210,7 +1210,7 @@ int LocalReduce_Norm4_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1256,7 +1256,7 @@ int LocalReduce_Norm4_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1279,7 +1279,7 @@ int LocalReduce_Norm4_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Norm4_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1304,47 +1304,47 @@ int LocalReduce_Norm4_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1373,7 +1373,7 @@ int LocalReduce_Norm4_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re);  Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1417,7 +1417,7 @@ int LocalReduce_Norm4_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re); Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1463,7 +1463,7 @@ int LocalReduce_Norm4_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1509,7 +1509,7 @@ int LocalReduce_Norm4_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1531,7 +1531,7 @@ int LocalReduce_Norm4_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 #endif
 
-int LocalReduce_Norm4_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1556,47 +1556,47 @@ int LocalReduce_Norm4_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1625,7 +1625,7 @@ int LocalReduce_Norm4_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re);  Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1669,7 +1669,7 @@ int LocalReduce_Norm4_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re); Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1715,7 +1715,7 @@ int LocalReduce_Norm4_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1761,7 +1761,7 @@ int LocalReduce_Norm4_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1783,7 +1783,7 @@ int LocalReduce_Norm4_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 
 #ifdef CCTK_REAL4
-int LocalReduce_Norm4_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1808,47 +1808,47 @@ int LocalReduce_Norm4_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1877,7 +1877,7 @@ int LocalReduce_Norm4_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re);  Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1921,7 +1921,7 @@ int LocalReduce_Norm4_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re); Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1967,7 +1967,7 @@ int LocalReduce_Norm4_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2013,7 +2013,7 @@ int LocalReduce_Norm4_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2036,7 +2036,7 @@ int LocalReduce_Norm4_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Norm4_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2061,47 +2061,47 @@ int LocalReduce_Norm4_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2130,7 +2130,7 @@ int LocalReduce_Norm4_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re);  Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2174,7 +2174,7 @@ int LocalReduce_Norm4_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re); Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2220,7 +2220,7 @@ int LocalReduce_Norm4_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2266,7 +2266,7 @@ int LocalReduce_Norm4_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2290,7 +2290,7 @@ int LocalReduce_Norm4_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Norm4_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2315,47 +2315,47 @@ int LocalReduce_Norm4_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2384,7 +2384,7 @@ int LocalReduce_Norm4_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re);  Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2428,7 +2428,7 @@ int LocalReduce_Norm4_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Norm4, scalar)   (scalar).Re = POWER4((scalar).Re); Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2474,7 +2474,7 @@ int LocalReduce_Norm4_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2520,7 +2520,7 @@ int LocalReduce_Norm4_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2542,7 +2542,7 @@ int LocalReduce_Norm4_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets,
 }
 #endif
 
-int LocalReduce_Norm4_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2578,47 +2578,47 @@ int LocalReduce_Norm4_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2637,7 +2637,7 @@ int LocalReduce_Norm4_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Norm4, scalar)   scalar = CCTK_CmplxMul(scalar, scalar); CCTK_CmplxMul(scalar,scalar); Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2682,7 +2682,7 @@ int LocalReduce_Norm4_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Norm4, scalar)   scalar = CCTK_CmplxMul8(scalar,scalar); scalar = CCTK_CmplxMul8(scalar,scalar);Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2729,7 +2729,7 @@ int LocalReduce_Norm4_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2776,7 +2776,7 @@ int LocalReduce_Norm4_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2810,7 +2810,7 @@ int LocalReduce_Norm4_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets
 
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Norm4_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2846,47 +2846,47 @@ int LocalReduce_Norm4_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2905,7 +2905,7 @@ int LocalReduce_Norm4_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_OPERATION(Norm4, scalar)   scalar = CCTK_CmplxMul(scalar, scalar); CCTK_CmplxMul(scalar,scalar); Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2950,7 +2950,7 @@ int LocalReduce_Norm4_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_OPERATION(Norm4, scalar)   scalar = CCTK_CmplxMul8(scalar,scalar); scalar = CCTK_CmplxMul8(scalar,scalar);Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2997,7 +2997,7 @@ int LocalReduce_Norm4_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3044,7 +3044,7 @@ int LocalReduce_Norm4_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3078,7 +3078,7 @@ int LocalReduce_Norm4_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offset
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Norm4_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -3114,47 +3114,47 @@ int LocalReduce_Norm4_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -3173,7 +3173,7 @@ int LocalReduce_Norm4_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm4, scalar)   scalar = CCTK_CmplxMul(scalar, scalar); CCTK_CmplxMul(scalar,scalar); Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3218,7 +3218,7 @@ int LocalReduce_Norm4_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm4, scalar)   scalar = CCTK_CmplxMul8(scalar,scalar); scalar = CCTK_CmplxMul8(scalar,scalar);Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3265,7 +3265,7 @@ int LocalReduce_Norm4_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3312,7 +3312,7 @@ int LocalReduce_Norm4_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3346,7 +3346,7 @@ int LocalReduce_Norm4_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offse
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Norm4_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Norm4_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -3382,47 +3382,47 @@ int LocalReduce_Norm4_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -3441,7 +3441,7 @@ int LocalReduce_Norm4_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm4, scalar)   scalar = CCTK_CmplxMul(scalar, scalar); CCTK_CmplxMul(scalar,scalar); Norm4 = CCTK_CmplxAdd( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3486,7 +3486,7 @@ int LocalReduce_Norm4_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(Norm4, scalar)   scalar = CCTK_CmplxMul8(scalar,scalar); scalar = CCTK_CmplxMul8(scalar,scalar);Norm4 = CCTK_CmplxAdd8( Norm4, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3533,7 +3533,7 @@ int LocalReduce_Norm4_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3580,7 +3580,7 @@ int LocalReduce_Norm4_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
diff --git a/src/Norm4_Functions.h b/src/Norm4_Functions.h
index 8517704..930e18a 100644
--- a/src/Norm4_Functions.h
+++ b/src/Norm4_Functions.h
@@ -18,52 +18,52 @@ extern "C" {
 #endif
 
 /* Norm4 reduction functions */
-int LocalReduce_Norm4_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
-int LocalReduce_Norm4_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_INT1
-int LocalReduce_Norm4_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Norm4_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Norm4_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Norm4_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Norm4_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_REAL4
-int LocalReduce_Norm4_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Norm4_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Norm4_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Norm4_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Norm4_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Norm4_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Norm4_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Norm4_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef __cplusplus
diff --git a/src/NormInfFunctions.c b/src/NormInfFunctions.c
index f690619..099dba0 100644
--- a/src/NormInfFunctions.c
+++ b/src/NormInfFunctions.c
@@ -18,7 +18,7 @@ extern "C" {
 #endif
 
 
-int LocalReduce_NormInf_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -47,47 +47,47 @@ int LocalReduce_NormInf_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -121,7 +121,7 @@ int LocalReduce_NormInf_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   (NormInf).Re = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -173,7 +173,7 @@ int LocalReduce_NormInf_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   NormInf = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -227,7 +227,7 @@ int LocalReduce_NormInf_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -281,7 +281,7 @@ int LocalReduce_NormInf_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -306,7 +306,7 @@ int LocalReduce_NormInf_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets,
   return num_points;
 }
 
-int LocalReduce_NormInf_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -335,47 +335,47 @@ int LocalReduce_NormInf_INT(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -408,7 +408,7 @@ int LocalReduce_NormInf_INT(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   (NormInf).Re = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -460,7 +460,7 @@ int LocalReduce_NormInf_INT(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   NormInf = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -513,7 +513,7 @@ int LocalReduce_NormInf_INT(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   NormInf = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -566,7 +566,7 @@ int LocalReduce_NormInf_INT(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   NormInf = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -592,7 +592,7 @@ int LocalReduce_NormInf_INT(int i, int mask_on, CCTK_INT * input_array_offsets,
 }
 
 #ifdef CCTK_INT1
-int LocalReduce_NormInf_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -621,47 +621,47 @@ int LocalReduce_NormInf_INT1(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT11
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT12
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT14
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT18
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -694,7 +694,7 @@ int LocalReduce_NormInf_INT1(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   (NormInf).Re = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -746,7 +746,7 @@ int LocalReduce_NormInf_INT1(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   NormInf = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -800,7 +800,7 @@ int LocalReduce_NormInf_INT1(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -854,7 +854,7 @@ int LocalReduce_NormInf_INT1(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -881,7 +881,7 @@ int LocalReduce_NormInf_INT1(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_NormInf_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -910,47 +910,47 @@ int LocalReduce_NormInf_INT2(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT21
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT22
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT24
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT28
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -983,7 +983,7 @@ int LocalReduce_NormInf_INT2(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   (NormInf).Re = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1035,7 +1035,7 @@ int LocalReduce_NormInf_INT2(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   NormInf = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1089,7 +1089,7 @@ int LocalReduce_NormInf_INT2(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1143,7 +1143,7 @@ int LocalReduce_NormInf_INT2(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1170,7 +1170,7 @@ int LocalReduce_NormInf_INT2(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_NormInf_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1199,47 +1199,47 @@ int LocalReduce_NormInf_INT4(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1272,7 +1272,7 @@ int LocalReduce_NormInf_INT4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   (NormInf).Re = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1324,7 +1324,7 @@ int LocalReduce_NormInf_INT4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   NormInf = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1378,7 +1378,7 @@ int LocalReduce_NormInf_INT4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1432,7 +1432,7 @@ int LocalReduce_NormInf_INT4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1459,7 +1459,7 @@ int LocalReduce_NormInf_INT4(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_NormInf_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1488,47 +1488,47 @@ int LocalReduce_NormInf_INT8(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1561,7 +1561,7 @@ int LocalReduce_NormInf_INT8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   (NormInf).Re = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1613,7 +1613,7 @@ int LocalReduce_NormInf_INT8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   NormInf = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1667,7 +1667,7 @@ int LocalReduce_NormInf_INT8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1721,7 +1721,7 @@ int LocalReduce_NormInf_INT8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1747,7 +1747,7 @@ int LocalReduce_NormInf_INT8(int i, int mask_on, CCTK_INT * input_array_offsets,
 }
 #endif
 
-int LocalReduce_NormInf_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1776,47 +1776,47 @@ int LocalReduce_NormInf_REAL(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1849,7 +1849,7 @@ int LocalReduce_NormInf_REAL(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   (NormInf).Re = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1901,7 +1901,7 @@ int LocalReduce_NormInf_REAL(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(NormInf, scalar)   NormInf = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1955,7 +1955,7 @@ int LocalReduce_NormInf_REAL(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2009,7 +2009,7 @@ int LocalReduce_NormInf_REAL(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2035,7 +2035,7 @@ int LocalReduce_NormInf_REAL(int i, int mask_on, CCTK_INT * input_array_offsets,
 }
 
 #ifdef CCTK_REAL4
-int LocalReduce_NormInf_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2064,47 +2064,47 @@ int LocalReduce_NormInf_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2137,7 +2137,7 @@ int LocalReduce_NormInf_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(NormInf, scalar)   (NormInf).Re = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2189,7 +2189,7 @@ int LocalReduce_NormInf_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(NormInf, scalar)   NormInf = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2243,7 +2243,7 @@ int LocalReduce_NormInf_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2297,7 +2297,7 @@ int LocalReduce_NormInf_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2324,7 +2324,7 @@ int LocalReduce_NormInf_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_NormInf_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2353,47 +2353,47 @@ int LocalReduce_NormInf_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2426,7 +2426,7 @@ int LocalReduce_NormInf_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(NormInf, scalar)   (NormInf).Re = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2478,7 +2478,7 @@ int LocalReduce_NormInf_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(NormInf, scalar)   NormInf = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2532,7 +2532,7 @@ int LocalReduce_NormInf_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2586,7 +2586,7 @@ int LocalReduce_NormInf_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2614,7 +2614,7 @@ int LocalReduce_NormInf_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_NormInf_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2643,47 +2643,47 @@ int LocalReduce_NormInf_REAL16(int i, int mask_on, CCTK_INT * input_array_offset
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2716,7 +2716,7 @@ int LocalReduce_NormInf_REAL16(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_OPERATION(NormInf, scalar)   (NormInf).Re = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2768,7 +2768,7 @@ int LocalReduce_NormInf_REAL16(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_OPERATION(NormInf, scalar)   NormInf = MAX( (NormInf).Re, ABS((scalar).Re));
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2822,7 +2822,7 @@ int LocalReduce_NormInf_REAL16(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2876,7 +2876,7 @@ int LocalReduce_NormInf_REAL16(int i, int mask_on, CCTK_INT * input_array_offset
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2902,7 +2902,7 @@ int LocalReduce_NormInf_REAL16(int i, int mask_on, CCTK_INT * input_array_offset
 }
 #endif
 
-int LocalReduce_NormInf_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2934,47 +2934,47 @@ int LocalReduce_NormInf_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offse
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2997,7 +2997,7 @@ int LocalReduce_NormInf_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(NormInf, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);(NormInf).Re = MAX( (NormInf).Re, (scalar).Re); (NormInf).Im = MAX( (NormInf).Im, (scalar).Im); 
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP 
@@ -3042,7 +3042,7 @@ int LocalReduce_NormInf_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_OPERATION(NormInf, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);(NormInf).Re = MAX( (NormInf).Re, (scalar).Re); (NormInf).Im = MAX( (NormInf).Im, (scalar).Im); 
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3089,7 +3089,7 @@ int LocalReduce_NormInf_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3136,7 +3136,7 @@ int LocalReduce_NormInf_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offse
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3166,7 +3166,7 @@ int LocalReduce_NormInf_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offse
 
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_NormInf_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -3198,47 +3198,47 @@ int LocalReduce_NormInf_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offs
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -3261,7 +3261,7 @@ int LocalReduce_NormInf_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offs
 #define REDUCTION_OPERATION(NormInf, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);(NormInf).Re = MAX( (NormInf).Re, (scalar).Re); (NormInf).Im = MAX( (NormInf).Im, (scalar).Im); 
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3306,7 +3306,7 @@ int LocalReduce_NormInf_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offs
 #define REDUCTION_OPERATION(NormInf, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);(NormInf).Re = MAX( (NormInf).Re, (scalar).Re); (NormInf).Im = MAX( (NormInf).Im, (scalar).Im); 
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3353,7 +3353,7 @@ int LocalReduce_NormInf_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offs
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3400,7 +3400,7 @@ int LocalReduce_NormInf_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offs
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3430,7 +3430,7 @@ int LocalReduce_NormInf_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offs
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_NormInf_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -3462,47 +3462,47 @@ int LocalReduce_NormInf_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_off
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -3525,7 +3525,7 @@ int LocalReduce_NormInf_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_off
 #define REDUCTION_OPERATION(NormInf, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);(NormInf).Re = MAX( (NormInf).Re, (scalar).Re); (NormInf).Im = MAX( (NormInf).Im, (scalar).Im); 
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3570,7 +3570,7 @@ int LocalReduce_NormInf_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_off
 #define REDUCTION_OPERATION(NormInf, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);(NormInf).Re = MAX( (NormInf).Re, (scalar).Re); (NormInf).Im = MAX( (NormInf).Im, (scalar).Im); 
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3617,7 +3617,7 @@ int LocalReduce_NormInf_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_off
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3664,7 +3664,7 @@ int LocalReduce_NormInf_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_off
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3694,7 +3694,7 @@ int LocalReduce_NormInf_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_off
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_NormInf_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_NormInf_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -3726,47 +3726,47 @@ int LocalReduce_NormInf_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_off
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -3789,7 +3789,7 @@ int LocalReduce_NormInf_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_off
 #define REDUCTION_OPERATION(NormInf, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);(NormInf).Re = MAX( (NormInf).Re, (scalar).Re); (NormInf).Im = MAX( (NormInf).Im, (scalar).Im); 
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3834,7 +3834,7 @@ int LocalReduce_NormInf_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_off
 #define REDUCTION_OPERATION(NormInf, scalar)   (scalar).Re = ABS((scalar).Re); (scalar).Im = ABS((scalar).Im);(NormInf).Re = MAX( (NormInf).Re, (scalar).Re); (NormInf).Im = MAX( (NormInf).Im, (scalar).Im); 
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3881,7 +3881,7 @@ int LocalReduce_NormInf_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_off
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3928,7 +3928,7 @@ int LocalReduce_NormInf_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_off
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
diff --git a/src/NormInf_Functions.h b/src/NormInf_Functions.h
index 85bfc76..73ffa1d 100644
--- a/src/NormInf_Functions.h
+++ b/src/NormInf_Functions.h
@@ -18,52 +18,52 @@ extern "C" {
 #endif
 
 /* NormInf reduction functions */
-int LocalReduce_NormInf_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
-int LocalReduce_NormInf_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_INT1
-int LocalReduce_NormInf_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_NormInf_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_NormInf_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_NormInf_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_NormInf_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_REAL4
-int LocalReduce_NormInf_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_NormInf_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_NormInf_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_NormInf_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_NormInf_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_NormInf_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_NormInf_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef __cplusplus
diff --git a/src/ReductionAvg.c b/src/ReductionAvg.c
index 0b3eeda..57302ed 100644
--- a/src/ReductionAvg.c
+++ b/src/ReductionAvg.c
@@ -35,52 +35,52 @@ static int ReductionAvg (int N_dims, int operator_handle,
                         void * const output_numbers[]);
 
 /* Average (mean) reduction functions */
-int LocalReduce_Avg_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
-int LocalReduce_Avg_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_INT1
-int LocalReduce_Avg_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Avg_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Avg_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Avg_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Avg_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_REAL4
-int LocalReduce_Avg_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Avg_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Avg_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Avg_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Avg_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Avg_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Avg_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Avg_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 /*@@
@@ -305,65 +305,65 @@ static int ReductionAvg (int N_dims, int operator_handle,
     {
     /* in values type switches*/
       case CCTK_VARIABLE_CHAR:
-        num_points = LocalReduce_Avg_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       case CCTK_VARIABLE_INT:
-        num_points = LocalReduce_Avg_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_INT1
       case CCTK_VARIABLE_INT1:
-        num_points = LocalReduce_Avg_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif                                                              
       #ifdef CCTK_INT2
       case CCTK_VARIABLE_INT2:
-        num_points = LocalReduce_Avg_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT4
       case CCTK_VARIABLE_INT4:
-        num_points = LocalReduce_Avg_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT8
       case CCTK_VARIABLE_INT8:
-        num_points = LocalReduce_Avg_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;                                                                
       #endif
       case CCTK_VARIABLE_REAL:
-        num_points = LocalReduce_Avg_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_REAL4
       case CCTK_VARIABLE_REAL4:
-        num_points = LocalReduce_Avg_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL8
       case CCTK_VARIABLE_REAL8:
-        num_points = LocalReduce_Avg_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL16
       case CCTK_VARIABLE_REAL16:
-        num_points = LocalReduce_Avg_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif  
       case CCTK_VARIABLE_COMPLEX:
-        num_points = LocalReduce_Avg_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_COMPLEX8
       case CCTK_VARIABLE_COMPLEX8:
-        num_points = LocalReduce_Avg_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX16
       case CCTK_VARIABLE_COMPLEX16:
-        num_points = LocalReduce_Avg_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX32
       case CCTK_VARIABLE_COMPLEX32:
-        num_points = LocalReduce_Avg_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Avg_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
     }
diff --git a/src/ReductionCount.c b/src/ReductionCount.c
index e25e106..7be0d36 100644
--- a/src/ReductionCount.c
+++ b/src/ReductionCount.c
@@ -255,65 +255,65 @@ static int ReductionCount (int N_dims, int operator_handle,
     {
     /* in values type switches*/
       case CCTK_VARIABLE_CHAR:
-        num_points = LocalReduce_Count_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       case CCTK_VARIABLE_INT:
-        num_points = LocalReduce_Count_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_INT1
       case CCTK_VARIABLE_INT1:
-        num_points = LocalReduce_Count_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif                                                              
       #ifdef CCTK_INT2
       case CCTK_VARIABLE_INT2:
-        num_points = LocalReduce_Count_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT4
       case CCTK_VARIABLE_INT4:
-        num_points = LocalReduce_Count_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT8
       case CCTK_VARIABLE_INT8:
-        num_points = LocalReduce_Count_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;                                                                
       #endif
       case CCTK_VARIABLE_REAL:
-        num_points = LocalReduce_Count_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_REAL4
       case CCTK_VARIABLE_REAL4:
-        num_points = LocalReduce_Count_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL8
       case CCTK_VARIABLE_REAL8:
-        num_points = LocalReduce_Count_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL16
       case CCTK_VARIABLE_REAL16:
-        num_points = LocalReduce_Count_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif  
       case CCTK_VARIABLE_COMPLEX:
-        num_points = LocalReduce_Count_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_COMPLEX8
       case CCTK_VARIABLE_COMPLEX8:
-        num_points = LocalReduce_Count_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX16
       case CCTK_VARIABLE_COMPLEX16:
-        num_points = LocalReduce_Count_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX32
       case CCTK_VARIABLE_COMPLEX32:
-        num_points = LocalReduce_Count_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Count_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
     }
diff --git a/src/ReductionMax.c b/src/ReductionMax.c
index bd78caa..4b11890 100644
--- a/src/ReductionMax.c
+++ b/src/ReductionMax.c
@@ -247,65 +247,65 @@ static int ReductionMax (int N_dims, int operator_handle,
     {
       /* in values type switches*/
       case CCTK_VARIABLE_CHAR:
-        num_points = LocalReduce_Max_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       case CCTK_VARIABLE_INT:
-        num_points = LocalReduce_Max_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_INT1
       case CCTK_VARIABLE_INT1:
-        num_points = LocalReduce_Max_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif                                                              
       #ifdef CCTK_INT2
       case CCTK_VARIABLE_INT2:
-        num_points = LocalReduce_Max_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT4
       case CCTK_VARIABLE_INT4:
-        num_points = LocalReduce_Max_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT8
       case CCTK_VARIABLE_INT8:
-        num_points = LocalReduce_Max_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;                                                                
       #endif
       case CCTK_VARIABLE_REAL:
-        num_points = LocalReduce_Max_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_REAL4
       case CCTK_VARIABLE_REAL4:
-        num_points = LocalReduce_Max_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL8
       case CCTK_VARIABLE_REAL8:
-        num_points = LocalReduce_Max_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL16
       case CCTK_VARIABLE_REAL16:
-        num_points = LocalReduce_Max_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif  
       case CCTK_VARIABLE_COMPLEX:
-        num_points = LocalReduce_Max_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_COMPLEX8
       case CCTK_VARIABLE_COMPLEX8:
-        num_points = LocalReduce_Max_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX16
       case CCTK_VARIABLE_COMPLEX16:
-        num_points = LocalReduce_Max_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX32
       case CCTK_VARIABLE_COMPLEX32:
-        num_points = LocalReduce_Max_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Max_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
     }
diff --git a/src/ReductionMin.c b/src/ReductionMin.c
index ee42eb4..e630c38 100644
--- a/src/ReductionMin.c
+++ b/src/ReductionMin.c
@@ -256,65 +256,65 @@ static int ReductionMin (int N_dims, int operator_handle,
     {
     /* in values type switches*/
       case CCTK_VARIABLE_CHAR:
-        num_points = LocalReduce_Min_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       case CCTK_VARIABLE_INT:
-        num_points = LocalReduce_Min_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_INT1
       case CCTK_VARIABLE_INT1:
-        num_points = LocalReduce_Min_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif                                                              
       #ifdef CCTK_INT2
       case CCTK_VARIABLE_INT2:
-        num_points = LocalReduce_Min_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT4
       case CCTK_VARIABLE_INT4:
-        num_points = LocalReduce_Min_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT8
       case CCTK_VARIABLE_INT8:
-        num_points = LocalReduce_Min_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;                                                                
       #endif
       case CCTK_VARIABLE_REAL:
-        num_points = LocalReduce_Min_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_REAL4
       case CCTK_VARIABLE_REAL4:
-        num_points = LocalReduce_Min_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL8
       case CCTK_VARIABLE_REAL8:
-        num_points = LocalReduce_Min_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL16
       case CCTK_VARIABLE_REAL16:
-        num_points = LocalReduce_Min_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif  
       case CCTK_VARIABLE_COMPLEX:
-        num_points = LocalReduce_Min_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_COMPLEX8
       case CCTK_VARIABLE_COMPLEX8:
-        num_points = LocalReduce_Min_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX16
       case CCTK_VARIABLE_COMPLEX16:
-        num_points = LocalReduce_Min_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX32
       case CCTK_VARIABLE_COMPLEX32:
-        num_points = LocalReduce_Min_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Min_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
     }
diff --git a/src/ReductionNorm1.c b/src/ReductionNorm1.c
index 3108f4a..d47d2d0 100644
--- a/src/ReductionNorm1.c
+++ b/src/ReductionNorm1.c
@@ -260,65 +260,65 @@ static int ReductionL1 (int N_dims, int operator_handle,
     {
     /* in values type switches*/
       case CCTK_VARIABLE_CHAR:
-        num_points = LocalReduce_Norm1_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       case CCTK_VARIABLE_INT:
-        num_points = LocalReduce_Norm1_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_INT1
       case CCTK_VARIABLE_INT1:
-        num_points = LocalReduce_Norm1_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif                                                              
       #ifdef CCTK_INT2
       case CCTK_VARIABLE_INT2:
-        num_points = LocalReduce_Norm1_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT4
       case CCTK_VARIABLE_INT4:
-        num_points = LocalReduce_Norm1_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT8
       case CCTK_VARIABLE_INT8:
-        num_points = LocalReduce_Norm1_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;                                                                
       #endif
       case CCTK_VARIABLE_REAL:
-        num_points = LocalReduce_Norm1_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_REAL4
       case CCTK_VARIABLE_REAL4:
-        num_points = LocalReduce_Norm1_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL8
       case CCTK_VARIABLE_REAL8:
-        num_points = LocalReduce_Norm1_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL16
       case CCTK_VARIABLE_REAL16:
-        num_points = LocalReduce_Norm1_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif  
       case CCTK_VARIABLE_COMPLEX:
-        num_points = LocalReduce_Norm1_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_COMPLEX8
       case CCTK_VARIABLE_COMPLEX8:
-        num_points = LocalReduce_Norm1_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX16
       case CCTK_VARIABLE_COMPLEX16:
-        num_points = LocalReduce_Norm1_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX32
       case CCTK_VARIABLE_COMPLEX32:
-        num_points = LocalReduce_Norm1_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm1_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
     }
diff --git a/src/ReductionNorm2.c b/src/ReductionNorm2.c
index 1605d49..c6c1835 100644
--- a/src/ReductionNorm2.c
+++ b/src/ReductionNorm2.c
@@ -261,65 +261,65 @@ static int ReductionL2 (int N_dims, int operator_handle,
     {
     /* in values type switches*/
       case CCTK_VARIABLE_CHAR:
-        num_points = LocalReduce_Norm2_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       case CCTK_VARIABLE_INT:
-        num_points = LocalReduce_Norm2_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_INT1
       case CCTK_VARIABLE_INT1:
-        num_points = LocalReduce_Norm2_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif                                                              
       #ifdef CCTK_INT2
       case CCTK_VARIABLE_INT2:
-        num_points = LocalReduce_Norm2_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT4
       case CCTK_VARIABLE_INT4:
-        num_points = LocalReduce_Norm2_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT8
       case CCTK_VARIABLE_INT8:
-        num_points = LocalReduce_Norm2_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;                                                                
       #endif
       case CCTK_VARIABLE_REAL:
-        num_points = LocalReduce_Norm2_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_REAL4
       case CCTK_VARIABLE_REAL4:
-        num_points = LocalReduce_Norm2_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL8
       case CCTK_VARIABLE_REAL8:
-        num_points = LocalReduce_Norm2_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL16
       case CCTK_VARIABLE_REAL16:
-        num_points = LocalReduce_Norm2_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif  
       case CCTK_VARIABLE_COMPLEX:
-        num_points = LocalReduce_Norm2_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_COMPLEX8
       case CCTK_VARIABLE_COMPLEX8:
-        num_points = LocalReduce_Norm2_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX16
       case CCTK_VARIABLE_COMPLEX16:
-        num_points = LocalReduce_Norm2_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX32
       case CCTK_VARIABLE_COMPLEX32:
-        num_points = LocalReduce_Norm2_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm2_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
     }
diff --git a/src/ReductionNorm3.c b/src/ReductionNorm3.c
index 43a001b..cc3bcd5 100644
--- a/src/ReductionNorm3.c
+++ b/src/ReductionNorm3.c
@@ -258,65 +258,65 @@ static int ReductionL3 (int N_dims, int operator_handle,
     {
     /* in values type switches*/
       case CCTK_VARIABLE_CHAR:
-        num_points = LocalReduce_Norm3_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       case CCTK_VARIABLE_INT:
-        num_points = LocalReduce_Norm3_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_INT1
       case CCTK_VARIABLE_INT1:
-        num_points = LocalReduce_Norm3_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif                                                              
       #ifdef CCTK_INT2
       case CCTK_VARIABLE_INT2:
-        num_points = LocalReduce_Norm3_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT4
       case CCTK_VARIABLE_INT4:
-        num_points = LocalReduce_Norm3_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT8
       case CCTK_VARIABLE_INT8:
-        num_points = LocalReduce_Norm3_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;                                                                
       #endif
       case CCTK_VARIABLE_REAL:
-        num_points = LocalReduce_Norm3_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_REAL4
       case CCTK_VARIABLE_REAL4:
-        num_points = LocalReduce_Norm3_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL8
       case CCTK_VARIABLE_REAL8:
-        num_points = LocalReduce_Norm3_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL16
       case CCTK_VARIABLE_REAL16:
-        num_points = LocalReduce_Norm3_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif  
       case CCTK_VARIABLE_COMPLEX:
-        num_points = LocalReduce_Norm3_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_COMPLEX8
       case CCTK_VARIABLE_COMPLEX8:
-        num_points = LocalReduce_Norm3_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX16
       case CCTK_VARIABLE_COMPLEX16:
-        num_points = LocalReduce_Norm3_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX32
       case CCTK_VARIABLE_COMPLEX32:
-        num_points = LocalReduce_Norm3_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm3_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
     }
diff --git a/src/ReductionNorm4.c b/src/ReductionNorm4.c
index 5d4763e..9f895b9 100644
--- a/src/ReductionNorm4.c
+++ b/src/ReductionNorm4.c
@@ -258,65 +258,65 @@ static int ReductionL4 (int N_dims, int operator_handle,
     {
     /* in values type switches*/
       case CCTK_VARIABLE_CHAR:
-        num_points = LocalReduce_Norm4_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       case CCTK_VARIABLE_INT:
-        num_points = LocalReduce_Norm4_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_INT1
       case CCTK_VARIABLE_INT1:
-        num_points = LocalReduce_Norm4_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif                                                              
       #ifdef CCTK_INT2
       case CCTK_VARIABLE_INT2:
-        num_points = LocalReduce_Norm4_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT4
       case CCTK_VARIABLE_INT4:
-        num_points = LocalReduce_Norm4_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT8
       case CCTK_VARIABLE_INT8:
-        num_points = LocalReduce_Norm4_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;                                                                
       #endif
       case CCTK_VARIABLE_REAL:
-        num_points = LocalReduce_Norm4_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_REAL4
       case CCTK_VARIABLE_REAL4:
-        num_points = LocalReduce_Norm4_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL8
       case CCTK_VARIABLE_REAL8:
-        num_points = LocalReduce_Norm4_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL16
       case CCTK_VARIABLE_REAL16:
-        num_points = LocalReduce_Norm4_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif  
       case CCTK_VARIABLE_COMPLEX:
-        num_points = LocalReduce_Norm4_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_COMPLEX8
       case CCTK_VARIABLE_COMPLEX8:
-        num_points = LocalReduce_Norm4_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX16
       case CCTK_VARIABLE_COMPLEX16:
-        num_points = LocalReduce_Norm4_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX32
       case CCTK_VARIABLE_COMPLEX32:
-        num_points = LocalReduce_Norm4_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Norm4_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
     }
diff --git a/src/ReductionNormInf.c b/src/ReductionNormInf.c
index 307359a..e532027 100644
--- a/src/ReductionNormInf.c
+++ b/src/ReductionNormInf.c
@@ -35,52 +35,52 @@ static int ReductionLInf (int N_dims, int operator_handle,
                         void * const output_numbers[]);
 
 /* NormInf reduction functions */
-int LocalReduce_NormInf_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
-int LocalReduce_NormInf_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_INT1
-int LocalReduce_NormInf_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_NormInf_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_NormInf_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_NormInf_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_NormInf_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_REAL4
-int LocalReduce_NormInf_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_NormInf_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_NormInf_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_NormInf_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_NormInf_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_NormInf_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_NormInf_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_NormInf_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 /*@@
@@ -314,65 +314,65 @@ static int ReductionLInf (int N_dims, int operator_handle,
     {
     /* in values type switches*/
       case CCTK_VARIABLE_CHAR:
-        num_points = LocalReduce_NormInf_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       case CCTK_VARIABLE_INT:
-        num_points = LocalReduce_NormInf_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_INT1
       case CCTK_VARIABLE_INT1:
-        num_points = LocalReduce_NormInf_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif                                                              
       #ifdef CCTK_INT2
       case CCTK_VARIABLE_INT2:
-        num_points = LocalReduce_NormInf_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT4
       case CCTK_VARIABLE_INT4:
-        num_points = LocalReduce_NormInf_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT8
       case CCTK_VARIABLE_INT8:
-        num_points = LocalReduce_NormInf_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;                                                                
       #endif
       case CCTK_VARIABLE_REAL:
-        num_points = LocalReduce_NormInf_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_REAL4
       case CCTK_VARIABLE_REAL4:
-        num_points = LocalReduce_NormInf_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL8
       case CCTK_VARIABLE_REAL8:
-        num_points = LocalReduce_NormInf_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL16
       case CCTK_VARIABLE_REAL16:
-        num_points = LocalReduce_NormInf_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif  
       case CCTK_VARIABLE_COMPLEX:
-        num_points = LocalReduce_NormInf_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_COMPLEX8
       case CCTK_VARIABLE_COMPLEX8:
-        num_points = LocalReduce_NormInf_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX16
       case CCTK_VARIABLE_COMPLEX16:
-        num_points = LocalReduce_NormInf_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX32
       case CCTK_VARIABLE_COMPLEX32:
-        num_points = LocalReduce_NormInf_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_NormInf_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
     }
diff --git a/src/ReductionSum.c b/src/ReductionSum.c
index b8df856..95f957c 100644
--- a/src/ReductionSum.c
+++ b/src/ReductionSum.c
@@ -226,7 +226,7 @@ static int ReductionSum (int N_dims, int operator_handle,
   /* reduction maps an array to a single value of the same type */
   if (M_output_numbers != N_input_arrays)
   {
-    CCTK_WARN (1, "Average reduction returns a single value\n                 \
+    CCTK_WARN (1, "Sum reduction returns a single value\n                 \
                    for each input array\n");
     return (-1);
   }
@@ -237,7 +237,7 @@ static int ReductionSum (int N_dims, int operator_handle,
   {
     indices       [j]        = 0;
     actual_indices[j]        = input_array_min_subscripts[j];
-    actual_iters_per_dim [j] = (int) (input_array_max_subscripts[j] - input_array_min_subscripts[j]);
+    actual_iters_per_dim [j] = (int) (input_array_max_subscripts[j] -input_array_min_subscripts[j]);
     iters_per_dim [j]        = (int) ((input_array_max_subscripts[j] - input_array_min_subscripts[j])/input_array_strides[j]);
     max_iter *= iters_per_dim [j];
   }
@@ -257,65 +257,65 @@ static int ReductionSum (int N_dims, int operator_handle,
     {
     /* in values type switches*/
       case CCTK_VARIABLE_CHAR:
-        num_points = LocalReduce_Sum_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_CHAR(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       case CCTK_VARIABLE_INT:
-        num_points = LocalReduce_Sum_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_INT(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_INT1
       case CCTK_VARIABLE_INT1:
-        num_points = LocalReduce_Sum_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_INT1(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif                                                              
       #ifdef CCTK_INT2
       case CCTK_VARIABLE_INT2:
-        num_points = LocalReduce_Sum_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_INT2(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT4
       case CCTK_VARIABLE_INT4:
-        num_points = LocalReduce_Sum_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_INT4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_INT8
       case CCTK_VARIABLE_INT8:
-        num_points = LocalReduce_Sum_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_INT8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;                                                                
       #endif
       case CCTK_VARIABLE_REAL:
-        num_points = LocalReduce_Sum_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_REAL(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_REAL4
       case CCTK_VARIABLE_REAL4:
-        num_points = LocalReduce_Sum_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_REAL4(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL8
       case CCTK_VARIABLE_REAL8:
-        num_points = LocalReduce_Sum_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_REAL8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_REAL16
       case CCTK_VARIABLE_REAL16:
-        num_points = LocalReduce_Sum_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_REAL16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif  
       case CCTK_VARIABLE_COMPLEX:
-        num_points = LocalReduce_Sum_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_COMPLEX(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #ifdef CCTK_COMPLEX8
       case CCTK_VARIABLE_COMPLEX8:
-        num_points = LocalReduce_Sum_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_COMPLEX8(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX16
       case CCTK_VARIABLE_COMPLEX16:
-        num_points = LocalReduce_Sum_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_COMPLEX16(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
       #ifdef CCTK_COMPLEX32
       case CCTK_VARIABLE_COMPLEX32:
-        num_points = LocalReduce_Sum_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
+        num_points = LocalReduce_Sum_COMPLEX32(i, mask_on, input_array_offsets, indices, max_iter, actual_indices, input_array_strides, input_array_min_subscripts, input_array_dims, num_points, actual_iters_per_dim, iters_per_dim,   N_dims, input_arrays, output_number_type_codes, output_numbers);
       break;
       #endif
     }
diff --git a/src/SumFunctions.c b/src/SumFunctions.c
index b7960e0..b713dbd 100644
--- a/src/SumFunctions.c
+++ b/src/SumFunctions.c
@@ -18,7 +18,7 @@ extern "C" {
 #endif
 
 
-int LocalReduce_Sum_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -39,47 +39,47 @@ int LocalReduce_Sum_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -105,7 +105,7 @@ int LocalReduce_Sum_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -141,7 +141,7 @@ int LocalReduce_Sum_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -179,7 +179,7 @@ int LocalReduce_Sum_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -217,7 +217,7 @@ int LocalReduce_Sum_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_BYTE, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -234,7 +234,7 @@ int LocalReduce_Sum_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int
   return num_points;
 }
 
-int LocalReduce_Sum_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -255,47 +255,47 @@ int LocalReduce_Sum_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -320,7 +320,7 @@ int LocalReduce_Sum_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -356,7 +356,7 @@ int LocalReduce_Sum_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -393,7 +393,7 @@ int LocalReduce_Sum_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd16( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -430,7 +430,7 @@ int LocalReduce_Sum_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd32( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -448,7 +448,7 @@ int LocalReduce_Sum_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int
 }
 
 #ifdef CCTK_INT1
-int LocalReduce_Sum_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -469,47 +469,47 @@ int LocalReduce_Sum_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT11
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT12
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT14
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT18
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -534,7 +534,7 @@ int LocalReduce_Sum_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -570,7 +570,7 @@ int LocalReduce_Sum_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -608,7 +608,7 @@ int LocalReduce_Sum_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -646,7 +646,7 @@ int LocalReduce_Sum_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT1, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -665,7 +665,7 @@ int LocalReduce_Sum_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Sum_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -686,47 +686,47 @@ int LocalReduce_Sum_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT21
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT22
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT24
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT28
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -751,7 +751,7 @@ int LocalReduce_Sum_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -787,7 +787,7 @@ int LocalReduce_Sum_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -825,7 +825,7 @@ int LocalReduce_Sum_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -863,7 +863,7 @@ int LocalReduce_Sum_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT2, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -882,7 +882,7 @@ int LocalReduce_Sum_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Sum_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -903,47 +903,47 @@ int LocalReduce_Sum_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -968,7 +968,7 @@ int LocalReduce_Sum_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1004,7 +1004,7 @@ int LocalReduce_Sum_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1042,7 +1042,7 @@ int LocalReduce_Sum_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1080,7 +1080,7 @@ int LocalReduce_Sum_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1099,7 +1099,7 @@ int LocalReduce_Sum_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Sum_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1120,47 +1120,47 @@ int LocalReduce_Sum_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1185,7 +1185,7 @@ int LocalReduce_Sum_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1221,7 +1221,7 @@ int LocalReduce_Sum_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1259,7 +1259,7 @@ int LocalReduce_Sum_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1297,7 +1297,7 @@ int LocalReduce_Sum_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_INT8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1315,7 +1315,7 @@ int LocalReduce_Sum_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int
 }
 #endif
 
-int LocalReduce_Sum_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1336,47 +1336,47 @@ int LocalReduce_Sum_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1401,7 +1401,7 @@ int LocalReduce_Sum_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1437,7 +1437,7 @@ int LocalReduce_Sum_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1475,7 +1475,7 @@ int LocalReduce_Sum_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1513,7 +1513,7 @@ int LocalReduce_Sum_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1531,7 +1531,7 @@ int LocalReduce_Sum_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int
 }
 
 #ifdef CCTK_REAL4
-int LocalReduce_Sum_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1552,47 +1552,47 @@ int LocalReduce_Sum_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1617,7 +1617,7 @@ int LocalReduce_Sum_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1653,7 +1653,7 @@ int LocalReduce_Sum_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1691,7 +1691,7 @@ int LocalReduce_Sum_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1729,7 +1729,7 @@ int LocalReduce_Sum_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL4, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1748,7 +1748,7 @@ int LocalReduce_Sum_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Sum_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1769,47 +1769,47 @@ int LocalReduce_Sum_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -1834,7 +1834,7 @@ int LocalReduce_Sum_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1870,7 +1870,7 @@ int LocalReduce_Sum_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1908,7 +1908,7 @@ int LocalReduce_Sum_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1946,7 +1946,7 @@ int LocalReduce_Sum_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -1966,7 +1966,7 @@ int LocalReduce_Sum_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, in
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Sum_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -1987,47 +1987,47 @@ int LocalReduce_Sum_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2052,7 +2052,7 @@ int LocalReduce_Sum_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2088,7 +2088,7 @@ int LocalReduce_Sum_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2126,7 +2126,7 @@ int LocalReduce_Sum_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2164,7 +2164,7 @@ int LocalReduce_Sum_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_REAL16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2182,7 +2182,7 @@ int LocalReduce_Sum_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, i
 }
 #endif
 
-int LocalReduce_Sum_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2214,47 +2214,47 @@ int LocalReduce_Sum_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2269,7 +2269,7 @@ int LocalReduce_Sum_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2306,7 +2306,7 @@ int LocalReduce_Sum_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2345,7 +2345,7 @@ int LocalReduce_Sum_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2384,7 +2384,7 @@ int LocalReduce_Sum_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2414,7 +2414,7 @@ int LocalReduce_Sum_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets,
 
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Sum_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2446,47 +2446,47 @@ int LocalReduce_Sum_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2501,7 +2501,7 @@ int LocalReduce_Sum_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2538,7 +2538,7 @@ int LocalReduce_Sum_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2577,7 +2577,7 @@ int LocalReduce_Sum_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2616,7 +2616,7 @@ int LocalReduce_Sum_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX8, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2646,7 +2646,7 @@ int LocalReduce_Sum_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets,
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Sum_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2678,47 +2678,47 @@ int LocalReduce_Sum_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_BYTE, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2733,7 +2733,7 @@ int LocalReduce_Sum_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2770,7 +2770,7 @@ int LocalReduce_Sum_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2809,7 +2809,7 @@ int LocalReduce_Sum_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2848,7 +2848,7 @@ int LocalReduce_Sum_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX16, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -2878,7 +2878,7 @@ int LocalReduce_Sum_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Sum_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
+int LocalReduce_Sum_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[])
 {
   int iter = 0;
   int sum_indices = 0;
@@ -2910,47 +2910,47 @@ int LocalReduce_Sum_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
   {
     /* out values type switches*/
     case CCTK_VARIABLE_CHAR:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     case CCTK_VARIABLE_INT:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_INT1
     case CCTK_VARIABLE_INT1:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT1, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif                                                              
     #ifdef CCTK_INT2
     case CCTK_VARIABLE_INT2:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT2, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT4
     case CCTK_VARIABLE_INT4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_INT8
     case CCTK_VARIABLE_INT8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_INT8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;                                                                
     #endif
     case CCTK_VARIABLE_REAL:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #ifdef CCTK_REAL4
     case CCTK_VARIABLE_REAL4:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL4, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL8
     case CCTK_VARIABLE_REAL8:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     #ifdef CCTK_REAL16
     case CCTK_VARIABLE_REAL16:
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_REAL16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
     break;
     #endif
     case CCTK_VARIABLE_COMPLEX:
@@ -2965,7 +2965,7 @@ int LocalReduce_Sum_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3002,7 +3002,7 @@ int LocalReduce_Sum_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_OPERATION(Sum, scalar)   Sum = CCTK_CmplxAdd8( Sum, scalar);
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX8, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3041,7 +3041,7 @@ int LocalReduce_Sum_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX16, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
@@ -3080,7 +3080,7 @@ int LocalReduce_Sum_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets
 #define REDUCTION_INITIAL(num) (num).Re = 0.0; (num).Im = 0.0;
 #define EXTRA_STEP(a, b)
 
-      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)
+      ITERATE_ON_ARRAY(i,CCTK_COMPLEX32, input_arrays[i], CCTK_COMPLEX32, output_numbers[i], mask_on, input_array_offsets[i], indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)
 #undef REDUCTION_OPERATION
 #undef REDUCTION_INITIAL
 #undef EXTRA_STEP
diff --git a/src/Sum_Functions.h b/src/Sum_Functions.h
index 9d915ad..7419ecd 100644
--- a/src/Sum_Functions.h
+++ b/src/Sum_Functions.h
@@ -18,52 +18,52 @@ extern "C" {
 #endif
 
 /* Sum reduction functions */
-int LocalReduce_Sum_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_CHAR(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
-int LocalReduce_Sum_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_INT(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_INT1
-int LocalReduce_Sum_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_INT1(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif                                                              
 
 #ifdef CCTK_INT2
-int LocalReduce_Sum_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_INT2(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT4
-int LocalReduce_Sum_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_INT4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_INT8
-int LocalReduce_Sum_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_INT8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Sum_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_REAL(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_REAL4
-int LocalReduce_Sum_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_REAL4(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL8
-int LocalReduce_Sum_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_REAL8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_REAL16
-int LocalReduce_Sum_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_REAL16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
-int LocalReduce_Sum_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_COMPLEX(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 
 #ifdef CCTK_COMPLEX8
-int LocalReduce_Sum_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_COMPLEX8(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX16
-int LocalReduce_Sum_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_COMPLEX16(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef CCTK_COMPLEX32
-int LocalReduce_Sum_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
+int LocalReduce_Sum_COMPLEX32(int i, int mask_on, CCTK_INT * input_array_offsets, int * indices, int max_iter, int * actual_indices, CCTK_INT * input_array_strides, CCTK_INT * input_array_min_subscripts,const CCTK_INT * input_array_dims, int num_points, int * actual_iters_per_dim, int * iters_per_dim,    int N_dims, const void *const input_arrays[], const CCTK_INT output_number_type_codes[], void * const output_numbers[]);
 #endif
 
 #ifdef __cplusplus
diff --git a/src/local_reductions.h b/src/local_reductions.h
index 3e3eba8..2f6804b 100644
--- a/src/local_reductions.h
+++ b/src/local_reductions.h
@@ -23,7 +23,7 @@ extern "C" {
 #endif
 
 
-#define ITERATE_ON_ARRAY(i,cctk_type, in_data, out_type, out_num, mask_on,input_array_offset, indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,product)           \
+#define ITERATE_ON_ARRAY(i,cctk_type, in_data, out_type, out_num, mask_on,input_array_offset, indices, sum_indices, max_iter, iter, flag, actual_indices,input_array_strides, input_array_min_subscripts,input_array_dims,product)           \
 {                                                                                 \
   const cctk_type * typed_vdata = (const cctk_type *)(in_data);                   \
   out_type inval;                                                                 \
@@ -44,7 +44,7 @@ extern "C" {
           product = 1;                                                            \
           for (j=k-1;j>=0;j--)                                                    \
           {                                                                       \
-            product *= actual_iters_per_dim[j];                                   \
+            product *= input_array_dims[j];                           \
           }                                                                       \
           sum_indices += actual_indices[k]*product;                               \
         }                                                                         \
@@ -59,7 +59,7 @@ extern "C" {
           {                                                                       \
             if (flag==1)                                                          \
             {                                                                     \
-              actual_indices[k] += input_array_strides[k-1];                      \
+              actual_indices[k] += input_array_strides[k];                      \
               indices[k]++;                                                       \
               flag = 0;                                                           \
               break;                                                              \