Add a bunch of local_barrier()'s.

They are overkill but seem to fix the problems with the testcases, at least so far.
Theano · Jul 24, 2017 · 588b546 · 588b546
1 parent 9983c33
commit 588b546
Show file tree

Hide file tree

Showing 2 changed files with 49 additions and 46 deletions.
diff --git a/src/gpuarray_reduction.c b/src/gpuarray_reduction.c
@@ -2600,6 +2600,7 @@ static void       reduxGenSrcAppendDecode       (GpuReduction*     gr){
 			"    TK1* restrict const SHMEMK1 = (TK1*)(SHMEM + SHMEMK1Off);\n");
 		}
 		srcbAppends(&gr->srcGen,
+		"    local_barrier();\n"
 		"    INITREDUXSTATE(SHMEMK0[LID_0], SHMEMK1[LID_0]);\n"
 		"    if(D<LDIM_0 && LID_0+LDIM_0<H){\n"
 		"        INITREDUXSTATE(SHMEMK0[LID_0+LDIM_0], SHMEMK1[LID_0+LDIM_0]);\n"
@@ -2736,6 +2737,7 @@ static void        reduxGenSrcAppendIncrement     (GpuReduction*        gr,
 static void        reduxGenSrcAppendDstWrite      (GpuReduction*        gr,
                                                    uint32_t             selector,
                                                    int                  initial){
+	srcbAppends(&gr->srcGen, "                local_barrier();\n");
 	if (initial){
 		srcbAppends(&gr->srcGen, "                if(LID_0 < D){\n"
 		                         "                    SETREDUXSTATE(W0R[GID_0*D + LID_0],\n"
@@ -2762,6 +2764,7 @@ static void        reduxGenSrcAppendDstWrite      (GpuReduction*        gr,
 			                         "                }\n");
 		}
 	}
+	srcbAppends(&gr->srcGen, "                local_barrier();\n");
 }
 static void        reduxGenSrcAppendPhase1        (GpuReduction*        gr){
 	/**