LoopControl: Override the flesh CCTK_LOOP macros

Override the flesh CCTK_LOOP macros, so that one can use CCTK_LOOP instead of LC_LOOP.
Remove some outdated vectorisation code.
Ensure the LC_LOOP and LC_ENDLOOP macros match in pairs.

3 files changed, 27 insertions, 62 deletions

diff --git a/Carpet/LoopControl/src/loopcontrol.c b/Carpet/LoopControl/src/loopcontrol.c
index a0bbe564b..fa2761717 100644
--- a/Carpet/LoopControl/src/loopcontrol.c
+++ b/Carpet/LoopControl/src/loopcontrol.c
@@ -1,3 +1,7 @@
+#include <cctk.h>
+#include <cctk_Arguments.h>
+#include <cctk_Parameters.h>
+
 #include <assert.h>
 #include <errno.h>
 #include <float.h>
@@ -12,9 +16,6 @@
 #  include <omp.h>
 #endif
 
-#include <cctk.h>
-#include <cctk_Parameters.h>
-
 /* #ifdef HAVE_TGMATH_H */
 /* #  include <tgmath.h> */
 /* #endif */
@@ -854,12 +855,6 @@ lc_control_init (lc_control_t * restrict const lc,
   lc->jjjstep = (lc->jjjmax - lc->jjjmin + lt->jnthreads-1) / lt->jnthreads;
   lc->kkkstep = (lc->kkkmax - lc->kkkmin + lt->knthreads-1) / lt->knthreads;
   
-#if 0
-  /* Correct threading for vectorisation (cache line size) */
-  lc->iiistep =
-    (lc->iiistep + LC_VECTORSIZE - 1) / LC_VECTORSIZE * LC_VECTORSIZE;
-#endif
-  
   /* Find location of current thread */
   lc->thread_num  = omp_get_thread_num();
   int c = lc->thread_num;
@@ -886,11 +881,6 @@ lc_control_init (lc_control_t * restrict const lc,
   lc->jjstep = lt->jnpoints;
   lc->kkstep = lt->knpoints;
   
-#if 0
-  /* Correct tiling for vectorisation (cache line size) */
-  lc->iistep = (lc->iistep + LC_VECTORSIZE - 1) / LC_VECTORSIZE * LC_VECTORSIZE;
-#endif
-  
   
   
   /*** Inner threads **********************************************************/
diff --git a/Carpet/LoopControl/src/loopcontrol.h b/Carpet/LoopControl/src/loopcontrol.h
index e23806fcb..70d90ebb5 100644
--- a/Carpet/LoopControl/src/loopcontrol.h
+++ b/Carpet/LoopControl/src/loopcontrol.h
@@ -6,16 +6,14 @@
 
 #include <cctk.h>
 
-#ifdef CCODE
 
-#include <cctk_Arguments.h>
+
+#ifdef CCODE
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-
-
 #ifdef __cplusplus
 #  ifdef CCTK_CXX_RESTRICT
 #    define restrict CCTK_CXX_RESTRICT
@@ -24,36 +22,6 @@ extern "C" {
 
 
 
-#if 0
-/* The most simple implementation */
-
-#define LC_LOOP3(name, i,j,k, imin,jmin,kmin, imax,jmax,kmax, ilsh,jlsh,klsh) \
-  do {                                                                  \
-    int const lc_imin = (imin);                                         \
-    int const lc_jmin = (jmin);                                         \
-    int const lc_kmin = (kmin);                                         \
-    int const lc_imax = (imax);                                         \
-    int const lc_jmax = (jmax);                                         \
-    int const lc_kmax = (kmax);                                         \
-    for (int k = lc_kmin; k < lc_kmax; ++k)                             \
-      for (int j = lc_jmin; j < lc_jmax; ++j)                           \
-        for (int i = lc_imin; i < lc_imax; ++i)
-
-#define LC_ENDLOOP3(name)                       \
-  } while (0)
-
-#endif
-
-
-
-#if 0
-/* Vector size */
-#define LC_VECTORSIZE 2         /* Correct for double precision on
-                                   Intel */
-#endif
-
-
-
 /* A topology */
 typedef struct lc_topology_t {
   int nthreads[2][3];           /* [0:outer|1:inner][ijk] */
@@ -319,6 +287,7 @@ lc_control_finish (lc_control_t * restrict lc);
 
 #define LC_LOOP3VEC(name, i,j,k, imin_,jmin_,kmin_, imax_,jmax_,kmax_, ilsh_,jlsh_,klsh_, di_) \
   do {                                                                  \
+    typedef int lc_loop3vec_##name;                                     \
     static int lc_initialised = 0;                                      \
     static lc_statmap_t lc_lm;                                          \
     if (! lc_initialised) {                                             \
@@ -364,16 +333,17 @@ lc_control_finish (lc_control_t * restrict lc);
                 int const lc_ioffset = (lc_ipos & - lc_di) - lc_ipos;   \
                 for (int i = lc_imin + lc_ioffset; i < lc_imax; i += lc_di) {
 
-#define LC_ENDLOOP3VEC(name)                    \
-                }                               \
-              }                                 \
-              LC_POSTLOOP_STATEMENTS            \
-            }                                   \
-          }                                     \
-        }                                       \
-      }                                         \
-    }                                           \
-    lc_control_finish (& lc_lc);                \
+#define LC_ENDLOOP3VEC(name)                                    \
+                }                                               \
+              }                                                 \
+              LC_POSTLOOP_STATEMENTS                            \
+            }                                                   \
+          }                                                     \
+        }                                                       \
+      }                                                         \
+    }                                                           \
+    lc_control_finish (& lc_lc);                                \
+    typedef lc_loop3vec_##name lc_ensure_proper_nesting;        \
   } while (0)
 
 /* Pre- and post loop statements are inserted around the innermost
@@ -383,8 +353,13 @@ lc_control_finish (lc_control_t * restrict lc);
 
 
 
-void
-lc_printstats (CCTK_ARGUMENTS);
+/* Replace CCTK_LOOP macros */
+#undef CCTK_LOOP3
+#undef CCTK_ENDLOOP3
+#define CCTK_LOOP3    LC_LOOP3
+#define CCTK_ENDLOOP3 LC_ENDLOOP3
+
+
 
 #ifdef __cplusplus
 }
diff --git a/Carpet/LoopControl/src/wavetoy-loopcontrol.c b/Carpet/LoopControl/src/wavetoy-loopcontrol.c
index c5ea9fecb..cdf61d1f2 100644
--- a/Carpet/LoopControl/src/wavetoy-loopcontrol.c
+++ b/Carpet/LoopControl/src/wavetoy-loopcontrol.c
@@ -255,10 +255,10 @@ int lc_demo (void)
   double sum = 0.0;
 #pragma omp parallel reduction(+: sum)
   {
-    LC_LOOP3 (analysis, i,j,k, 1,1,1, NI-1,NJ-1,NK-1, NI,NJ,NK) {
+    CCTK_LOOP3 (analysis, i,j,k, 1,1,1, NI-1,NJ-1,NK-1, NI,NJ,NK) {
       int const ind = ind3d(i,j,k);
       sum += phi0[ind];
-    } LC_ENDLOOP3 (analysis);
+    } CCTK_ENDLOOP3 (analysis);
   }
   double const avg = sum / ((NI-2)*(NJ-2)*(NK-2));
   double const time_analysis_end = omp_get_wtime();