Provide more optimised versions for common cases.

git-svn-id: http://svn.cactuscode.org/arrangements/CactusNumerical/Slab/trunk@34 2e825fa2-fb71-486d-8b7f-a5ff3f0f6cb8

1 files changed, 241 insertions, 12 deletions

diff --git a/src/slab.c b/src/slab.c
index f4450e8..e1b4ada 100644
--- a/src/slab.c
+++ b/src/slab.c
@@ -996,7 +996,7 @@ int Slab_Transfer (cGH             const * restrict const cctkGH,
     if (info[0].xpose==0 && info[1].xpose==1 && info[2].xpose==2
         && srcdetail[n*SLAB_MAXDIM  ].str==1 && srcdetail[n*SLAB_MAXDIM+1].str==1 && srcdetail[n*SLAB_MAXDIM+2].str==1
         && srctype == CCTK_VARIABLE_REAL) {
-      /* Optimised version for a special case */
+      /* Optimised version for a special case: no transposing */
       
       int const srcoffi = info[0].src.local.off;
       int const srcoffj = info[1].src.local.off;
@@ -1014,9 +1014,9 @@ int Slab_Transfer (cGH             const * restrict const cctkGH,
       int const srcdetaillenj = srcdetail[n*SLAB_MAXDIM+1].len;
       int const srcdetaillenk = srcdetail[n*SLAB_MAXDIM+2].len;
       
-if (n==0) assert (srcoffset[n]==0);
-if (n<size-1) assert (srcoffset[n+1]==srcoffset[n]+srcdetailleni*srcdetaillenj*srcdetaillenk);
-
+      if (n==0) assert (srcoffset[n]==0);
+      if (n<size-1) assert (srcoffset[n+1]==srcoffset[n]+srcdetailleni*srcdetaillenj*srcdetaillenk);
+      
       for (k = 0; k < srcdetaillenk; ++k) {
         for (j = 0; j < srcdetaillenj; ++j) {
           for (i = 0; i < srcdetailleni; ++i) {
@@ -1033,6 +1033,86 @@ if (n<size-1) assert (srcoffset[n+1]==srcoffset[n]+srcdetailleni*srcdetaillenj*s
         }
       }
 
+    } else if (info[0].xpose==1 && info[1].xpose==0 && info[2].xpose==2
+               && srcdetail[n*SLAB_MAXDIM  ].str==1 && srcdetail[n*SLAB_MAXDIM+1].str==1 && srcdetail[n*SLAB_MAXDIM+2].str==1
+               && srctype == CCTK_VARIABLE_REAL) {
+      /* Optimised version for a special case: transpose x and y */
+      
+      int const srcoffi = info[0].src.local.off;
+      int const srcoffj = info[1].src.local.off;
+      int const srcoffk = info[2].src.local.off;
+      
+      int const srcleni = info[0].src.local.len;
+      int const srclenj = info[1].src.local.len;
+      int const srclenk = info[2].src.local.len;
+      
+      int const srcdetailoffi = srcdetail[n*SLAB_MAXDIM+0].off;
+      int const srcdetailoffj = srcdetail[n*SLAB_MAXDIM+1].off;
+      int const srcdetailoffk = srcdetail[n*SLAB_MAXDIM+2].off;
+      
+      int const srcdetailleni = srcdetail[n*SLAB_MAXDIM+0].len;
+      int const srcdetaillenj = srcdetail[n*SLAB_MAXDIM+1].len;
+      int const srcdetaillenk = srcdetail[n*SLAB_MAXDIM+2].len;
+      
+      if (n==0) assert (srcoffset[n]==0);
+      if (n<size-1) assert (srcoffset[n+1]==srcoffset[n]+srcdetailleni*srcdetaillenj*srcdetaillenk);
+      
+      for (k = 0; k < srcdetaillenk; ++k) {
+        for (j = 0; j < srcdetaillenj; ++j) {
+          for (i = 0; i < srcdetailleni; ++i) {
+            int const srcindi = srcdetailoffi + i - srcoffi;
+            int const srcindj = srcdetailoffj + j - srcoffj;
+            int const srcindk = srcdetailoffk + k - srcoffk;
+            size_t const srcind = srcindi + srcleni * (srcindj + srclenj * srcindk);
+            size_t const bufind = j + srcdetaillenj * (i + srcdetailleni * k);
+            ifcheck assert (srcindi>=0 && srcindi<srcleni);
+            ifcheck assert (srcindj>=0 && srcindj<srclenj);
+            ifcheck assert (srcindk>=0 && srcindk<srclenk);
+            ((CCTK_REAL*)srcdata)[srcoffset[n] + bufind] = ((const CCTK_REAL*)srcptr)[srcind];
+          }
+        }
+      }
+      
+    } else if (srcdetail[n*SLAB_MAXDIM  ].str==1 && srcdetail[n*SLAB_MAXDIM+1].str==1 && srcdetail[n*SLAB_MAXDIM+2].str==1
+               && srctype == CCTK_VARIABLE_REAL) {
+      /* Optimised version for CCTK_REAL and stride 1 */
+      
+      for (k = 0; k < srcdetail[n*SLAB_MAXDIM+info[2].xpose].len; ++k) {
+        for (j = 0; j < srcdetail[n*SLAB_MAXDIM+info[1].xpose].len; ++j) {
+          for (i = 0; i < srcdetail[n*SLAB_MAXDIM+info[0].xpose].len; ++i) {
+            int ipos[SLAB_MAXDIM];
+            int srcipos[SLAB_MAXDIM];
+            int bufipos[SLAB_MAXDIM];
+            size_t srcind;
+            size_t bufind;
+            ipos[0] = i;
+            ipos[1] = j;
+            ipos[2] = k;
+            for (d=0; d<SLAB_MAXDIM; ++d) {
+              int const c = info[d].xpose;
+              srcipos[c] = srcdetail[n*SLAB_MAXDIM+c].off + ipos[d];
+              assert (srcipos[c] >= info[c].src.local.off
+                      && srcipos[c] < info[c].src.local.off + info[c].src.local.len);
+              assert (srcipos[c] >= allinfo[n*SLAB_MAXDIM+c].src.slab.off
+                      && srcipos[c] <= allinfo[n*SLAB_MAXDIM+c].src.slab.off + (allinfo[n*SLAB_MAXDIM+c].src.slab.len - 1));
+              bufipos[d] = ipos[d];
+              assert (bufipos[d] >= 0 && bufipos[d] < srcdetail[n*SLAB_MAXDIM+c].len);
+            }
+            srcind = 0;
+            bufind = 0;
+            for (d=SLAB_MAXDIM-1; d>=0; --d) {
+              int const c = info[d].xpose;
+              srcind = srcind * info[d].src.local.len + srcipos[d] - info[d].src.local.off;
+              bufind = bufind * srcdetail[n*SLAB_MAXDIM+c].len + bufipos[d];
+            }
+            assert (srcind < srclentot);
+            assert (bufind < (size_t)srccount[n]);
+            ((CCTK_REAL*)srcdata)[srcoffset[n] + bufind]
+              = ((const CCTK_REAL*)srcptr)[srcind];
+          }
+        }
+      }
+      
     } else {
       /* Generic, unoptimised version */
       
@@ -1052,9 +1132,6 @@ if (n<size-1) assert (srcoffset[n+1]==srcoffset[n]+srcdetailleni*srcdetaillenj*s
               srcipos[c] = srcdetail[n*SLAB_MAXDIM+c].off + ipos[d] * srcdetail[n*SLAB_MAXDIM+c].str;
               assert (srcipos[c] >= info[c].src.local.off
                       && srcipos[c] < info[c].src.local.off + info[c].src.local.len);
-              if (! (srcipos[c] >= allinfo[n*SLAB_MAXDIM+c].src.slab.off
-                     && srcipos[c] <= allinfo[n*SLAB_MAXDIM+c].src.slab.off + (allinfo[n*SLAB_MAXDIM+c].src.slab.len - 1) * allinfo[n*SLAB_MAXDIM+c].src.slab.str)) {
-              }
               assert (srcipos[c] >= allinfo[n*SLAB_MAXDIM+c].src.slab.off
                       && srcipos[c] <= allinfo[n*SLAB_MAXDIM+c].src.slab.off + (allinfo[n*SLAB_MAXDIM+c].src.slab.len - 1) * allinfo[n*SLAB_MAXDIM+c].src.slab.str);
               assert ((srcipos[c] - allinfo[n*SLAB_MAXDIM+c].src.slab.off) % allinfo[n*SLAB_MAXDIM+c].src.slab.str == 0);
@@ -1120,11 +1197,10 @@ if (n<size-1) assert (srcoffset[n+1]==srcoffset[n]+srcdetailleni*srcdetaillenj*s
   for (n = 0; n < size; ++n) {
     assert (SLAB_MAXDIM == 3);
     
-    if (info[0].xpose==0 && info[1].xpose==1 && info[2].xpose==2
-        && info[0].flip==0 && info[1].flip==0 && info[2].flip==0
+    if (info[0].flip==0 && info[1].flip==0 && info[2].flip==0
         && dstdetail[n*SLAB_MAXDIM  ].str==1 && dstdetail[n*SLAB_MAXDIM+1].str==1 && dstdetail[n*SLAB_MAXDIM+2].str==1
         && dsttype == CCTK_VARIABLE_REAL) {
-      /* Optimised version for a special case */
+      /* Optimised version for a special case: no flipping */
       
       int const dstoffi = info[0].dst.local.off;
       int const dstoffj = info[1].dst.local.off;
@@ -1154,9 +1230,162 @@ if (n<size-1) assert (srcoffset[n+1]==srcoffset[n]+srcdetailleni*srcdetaillenj*s
             ifcheck assert (dstindj>=0 && dstindj<dstlenj);
             ifcheck assert (dstindk>=0 && dstindk<dstlenk);
             ((CCTK_REAL*)dstptr)[dstind] = ((const CCTK_REAL*)dstdata)[dstoffset[n] + bufind];
-	}
+          }
+        }
+      }
+    
+    } else if (info[0].flip==1 && info[1].flip==0 && info[2].flip==0
+               && dstdetail[n*SLAB_MAXDIM  ].str==1 && dstdetail[n*SLAB_MAXDIM+1].str==1 && dstdetail[n*SLAB_MAXDIM+2].str==1
+               && dsttype == CCTK_VARIABLE_REAL) {
+      /* Optimised version for a special case: flip in x direction */
+      
+      int const dstoffi = info[0].dst.local.off;
+      int const dstoffj = info[1].dst.local.off;
+      int const dstoffk = info[2].dst.local.off;
+      
+      int const dstleni = info[0].dst.local.len;
+      int const dstlenj = info[1].dst.local.len;
+      int const dstlenk = info[2].dst.local.len;
+      
+      int const dstdetailoffi = dstdetail[n*SLAB_MAXDIM+0].off;
+      int const dstdetailoffj = dstdetail[n*SLAB_MAXDIM+1].off;
+      int const dstdetailoffk = dstdetail[n*SLAB_MAXDIM+2].off;
+      
+      int const dstdetailleni = dstdetail[n*SLAB_MAXDIM+0].len;
+      int const dstdetaillenj = dstdetail[n*SLAB_MAXDIM+1].len;
+      int const dstdetaillenk = dstdetail[n*SLAB_MAXDIM+2].len;
+      
+      for (k = 0; k < dstdetaillenk; ++k) {
+        for (j = 0; j < dstdetaillenj; ++j) {
+          for (i = 0; i < dstdetailleni; ++i) {
+            size_t const bufind = (dstdetailleni - 1 - i) + dstdetailleni * (j + dstdetaillenj * k);
+            int const dstindi = dstdetailoffi + i - dstoffi;
+            int const dstindj = dstdetailoffj + j - dstoffj;
+            int const dstindk = dstdetailoffk + k - dstoffk;
+            size_t const dstind = dstindi + dstleni * (dstindj + dstlenj * dstindk);
+            ifcheck assert (dstindi>=0 && dstindi<dstleni);
+            ifcheck assert (dstindj>=0 && dstindj<dstlenj);
+            ifcheck assert (dstindk>=0 && dstindk<dstlenk);
+            ((CCTK_REAL*)dstptr)[dstind] = ((const CCTK_REAL*)dstdata)[dstoffset[n] + bufind];
+          }
+        }
+      }
+    
+    } else if (info[0].flip==0 && info[1].flip==1 && info[2].flip==0
+               && dstdetail[n*SLAB_MAXDIM  ].str==1 && dstdetail[n*SLAB_MAXDIM+1].str==1 && dstdetail[n*SLAB_MAXDIM+2].str==1
+               && dsttype == CCTK_VARIABLE_REAL) {
+      /* Optimised version for a special case: flip in y direction */
+      
+      int const dstoffi = info[0].dst.local.off;
+      int const dstoffj = info[1].dst.local.off;
+      int const dstoffk = info[2].dst.local.off;
+      
+      int const dstleni = info[0].dst.local.len;
+      int const dstlenj = info[1].dst.local.len;
+      int const dstlenk = info[2].dst.local.len;
+      
+      int const dstdetailoffi = dstdetail[n*SLAB_MAXDIM+0].off;
+      int const dstdetailoffj = dstdetail[n*SLAB_MAXDIM+1].off;
+      int const dstdetailoffk = dstdetail[n*SLAB_MAXDIM+2].off;
+      
+      int const dstdetailleni = dstdetail[n*SLAB_MAXDIM+0].len;
+      int const dstdetaillenj = dstdetail[n*SLAB_MAXDIM+1].len;
+      int const dstdetaillenk = dstdetail[n*SLAB_MAXDIM+2].len;
+      
+      for (k = 0; k < dstdetaillenk; ++k) {
+        for (j = 0; j < dstdetaillenj; ++j) {
+          for (i = 0; i < dstdetailleni; ++i) {
+            size_t const bufind = i + dstdetailleni * ((dstdetaillenj - 1 - j) + dstdetaillenj * k);
+            int const dstindi = dstdetailoffi + i - dstoffi;
+            int const dstindj = dstdetailoffj + j - dstoffj;
+            int const dstindk = dstdetailoffk + k - dstoffk;
+            size_t const dstind = dstindi + dstleni * (dstindj + dstlenj * dstindk);
+            ifcheck assert (dstindi>=0 && dstindi<dstleni);
+            ifcheck assert (dstindj>=0 && dstindj<dstlenj);
+            ifcheck assert (dstindk>=0 && dstindk<dstlenk);
+            ((CCTK_REAL*)dstptr)[dstind] = ((const CCTK_REAL*)dstdata)[dstoffset[n] + bufind];
+          }
+        }
+      }
+    
+    } else if (info[0].flip==1 && info[1].flip==1 && info[2].flip==0
+               && dstdetail[n*SLAB_MAXDIM  ].str==1 && dstdetail[n*SLAB_MAXDIM+1].str==1 && dstdetail[n*SLAB_MAXDIM+2].str==1
+               && dsttype == CCTK_VARIABLE_REAL) {
+      /* Optimised version for a special case: flip in y direction */
+      
+      int const dstoffi = info[0].dst.local.off;
+      int const dstoffj = info[1].dst.local.off;
+      int const dstoffk = info[2].dst.local.off;
+      
+      int const dstleni = info[0].dst.local.len;
+      int const dstlenj = info[1].dst.local.len;
+      int const dstlenk = info[2].dst.local.len;
+      
+      int const dstdetailoffi = dstdetail[n*SLAB_MAXDIM+0].off;
+      int const dstdetailoffj = dstdetail[n*SLAB_MAXDIM+1].off;
+      int const dstdetailoffk = dstdetail[n*SLAB_MAXDIM+2].off;
+      
+      int const dstdetailleni = dstdetail[n*SLAB_MAXDIM+0].len;
+      int const dstdetaillenj = dstdetail[n*SLAB_MAXDIM+1].len;
+      int const dstdetaillenk = dstdetail[n*SLAB_MAXDIM+2].len;
+      
+      for (k = 0; k < dstdetaillenk; ++k) {
+        for (j = 0; j < dstdetaillenj; ++j) {
+          for (i = 0; i < dstdetailleni; ++i) {
+            size_t const bufind = (dstdetailleni - 1 - i) + dstdetailleni * ((dstdetaillenj - 1 - j) + dstdetaillenj * k);
+            int const dstindi = dstdetailoffi + i - dstoffi;
+            int const dstindj = dstdetailoffj + j - dstoffj;
+            int const dstindk = dstdetailoffk + k - dstoffk;
+            size_t const dstind = dstindi + dstleni * (dstindj + dstlenj * dstindk);
+            ifcheck assert (dstindi>=0 && dstindi<dstleni);
+            ifcheck assert (dstindj>=0 && dstindj<dstlenj);
+            ifcheck assert (dstindk>=0 && dstindk<dstlenk);
+            ((CCTK_REAL*)dstptr)[dstind] = ((const CCTK_REAL*)dstdata)[dstoffset[n] + bufind];
+          }
+        }
+      }
+      
+    } else if (dstdetail[n*SLAB_MAXDIM  ].str==1 && dstdetail[n*SLAB_MAXDIM+1].str==1 && dstdetail[n*SLAB_MAXDIM+2].str==1
+        && dsttype == CCTK_VARIABLE_REAL) {
+      /* Optimised version for CCTK_REAL and stride 1 */
+      
+      for (k = 0; k < dstdetail[n*SLAB_MAXDIM+2].len; ++k) {
+        for (j = 0; j < dstdetail[n*SLAB_MAXDIM+1].len; ++j) {
+          for (i = 0; i < dstdetail[n*SLAB_MAXDIM+0].len; ++i) {
+            int ipos[SLAB_MAXDIM];
+            int bufipos[SLAB_MAXDIM];
+            int dstipos[SLAB_MAXDIM];
+            size_t bufind;
+            size_t dstind;
+            ipos[0] = i;
+            ipos[1] = j;
+            ipos[2] = k;
+            for (d=0; d<SLAB_MAXDIM; ++d) {
+              if (! info[d].flip) {
+                bufipos[d] = ipos[d];
+              } else {
+                bufipos[d] = dstdetail[n*SLAB_MAXDIM+d].len - 1 - ipos[d];
+              }
+              ifcheck assert (bufipos[d] >= 0 && bufipos[d] < dstdetail[n*SLAB_MAXDIM+d].len);
+              dstipos[d] = dstdetail[n*SLAB_MAXDIM+d].off + ipos[d];
+              ifcheck assert (dstipos[d] >= info[d].dst.local.off
+                              && dstipos[d] < info[d].dst.local.off + info[d].dst.local.len);
+              ifcheck assert (dstipos[d] >= info[d].dst.slab.off
+                              && dstipos[d] <= info[d].dst.slab.off + info[d].dst.slab.len - 1);
+            }
+            bufind = 0;
+            dstind = 0;
+            for (d=SLAB_MAXDIM-1; d>=0; --d) {
+              bufind = bufind * dstdetail[n*SLAB_MAXDIM+d].len + bufipos[d];
+              dstind = dstind * info[d].dst.local.len + dstipos[d] - info[d].dst.local.off;
+            }
+            ifcheck assert (bufind < (size_t)dstcount[n]);
+            ifcheck assert (dstind < dstlentot);
+            ((CCTK_REAL*)dstptr)[dstind]
+              = ((const CCTK_REAL*)dstdata)[dstoffset[n] + bufind];
+          }
+        }
       }
-    }
 
     } else {
       /* Generic, unoptimised version */