mixed_cg_merge: added orphaned version of update_backward_gauge for u…

…se in Hopping_Matrix_32
Marcogarofalo · Nov 11, 2015 · 71c55aa · 71c55aa
1 parent 8b8b447
commit 71c55aa
Show file tree

Hide file tree

Showing 2 changed files with 39 additions and 20 deletions.
diff --git a/update_backward_gauge.c b/update_backward_gauge.c
@@ -67,16 +67,12 @@ void update_backward_gauge(su3 ** const gf) {
   return;
 }
 
-void update_backward_gauge_32(su3_32 ** const gf) {
-#ifdef OMP
-#pragma omp parallel
-  {
-#endif
+void update_backward_gauge_32_orphaned(su3_32 ** const gf) {
 
   int ix=0, kb=0, iy=0;
 
 #ifdef OMP
-#pragma omp for
+#pragma omp for nowait
 #endif 
   for(ix = 0; ix < VOLUME/2; ix++) {
     iy = (VOLUME+RAND)/2+ix;
@@ -99,15 +95,30 @@ void update_backward_gauge_32(su3_32 ** const gf) {
     _su3_assign(g_gauge_field_copy_32[1][ix][3], gf[kb][3]);
   }
 
+// we use the implicit barrier at the end of the single section to catch all
+// threads, in the meantime, one of them modifies the global flag
 #ifdef OMP
-  } /* OpenMP closing brace */
+#pragma omp single
+  {
 #endif
+    g_update_gauge_copy_32 = 0;
+#ifdef OMP
+  } 
+#endif
+}
 
-  g_update_gauge_copy_32 = 0;
+void update_backward_gauge_32(su3_32 ** const gf) {
+#ifdef OMP
+#pragma omp parallel
+  {
+#endif
+  update_backward_gauge_32_orphaned(gf);
+#ifdef OMP
+  } /* OpenMP closing brace */
+#endif
   return;
 }
 
-
 #elif _USE_TSPLITPAR 
 
 void update_backward_gauge(su3 ** const gf) {
@@ -230,17 +241,11 @@ void update_backward_gauge(su3 ** const gf) {
   return;
 }
 
-
-void update_backward_gauge_32(su3_32 ** const gf) {
-#ifdef OMP
-#pragma omp parallel
-  {
-#endif
-
+void update_backward_gauge_32_orphaned(su3_32 ** const gf) {
   int ix=0, kb=0, kb2=0;
 
 #ifdef OMP
-#pragma omp for
+#pragma omp for nowait
 #endif
   for(ix = 0; ix < VOLUME/2; ix++) {
     kb2=g_eo2lexic[ix];
@@ -261,7 +266,7 @@ void update_backward_gauge_32(su3_32 ** const gf) {
     _su3_assign(g_gauge_field_copy_32[ix][7],gf[kb][3]);
   }
 #ifdef OMP
-#pragma omp for
+#pragma omp for nowait
 #endif
   for(ix = (VOLUME+RAND)/2; ix < (VOLUME+RAND)/2+VOLUME/2; ix++) {
     kb2=g_eo2lexic[ix];
@@ -281,12 +286,25 @@ void update_backward_gauge_32(su3_32 ** const gf) {
     kb=g_idn[g_eo2lexic[ix]][3];
     _su3_assign(g_gauge_field_copy_32[ix][7],gf[kb][3]);
   }
+// the threads are caught by the implicit barrier here
+#ifdef OMP
+#pragma omp single
+#endif
+  g_update_gauge_copy_32 = 0;
+#ifdef OMP
+ }
+#endif
+}
 
+void update_backward_gauge_32(su3_32 ** const gf) {
+#ifdef OMP
+#pragma omp parallel
+  {
+#endif
+  update_backward_gauge_32_orphaned(gf);
 #ifdef OMP
   } /* OpenMP closing brace */
 #endif
-
-  g_update_gauge_copy_32 = 0;
   return;
 }
 

diff --git a/update_backward_gauge.h b/update_backward_gauge.h
@@ -23,6 +23,7 @@
 #include "su3.h"
 
 void update_backward_gauge(su3 ** const gf);
+void update_backward_gauge_32_orphaned(su3_32 ** const gf);
 void update_backward_gauge_32(su3_32 ** const gf);
 
 #endif