Merge pull request #47 from TimmyLiu/develop

enhancement of tune tool for better sgemm performance

Author: Kent Knox

src/include/granulation.h

@@ -46,6 +46,8 @@ typedef struct PGranularity {
     unsigned int wfSize;
     /** Record number of work-groups spawned */
     unsigned int numWGSpawned[2];
+	/** max number of work group size */
+	unsigned int maxWorkGroupSize;
 } PGranularity;
 
 /**

src/library/blas/gens/gemm.c

@@ -1104,6 +1104,8 @@ blockCheckCalcDecomp(
     int check)
 {
     bool ret = true;
+	bool ret_multiple = false;
+	int i;
 
     DUMMY_ARG_USAGE(subdimsNum);
 
@@ -1114,7 +1116,12 @@ blockCheckCalcDecomp(
         minSize = (dtype == TYPE_COMPLEX_DOUBLE) ? 1 : 2;
         ret = decompSanityCheck(subdims, minSize, maxSize, 24, dtype, true);
         ret = ret && (subdims[0].bwidth == subdims[1].bwidth);
-        ret = ret && (pgran->wgSize[0] * pgran->wgSize[1] == 64);
+		for(i = 0; i < ( (pgran->maxWorkGroupSize) / (pgran->wfSize) ); i++)
+		{
+			// returns true if wgSize[0] * wgSize[1] is multiples of the 64 but not bigger than maxWorkGroupSize
+			ret_multiple = ret_multiple || ( pgran->wgSize[0] * pgran->wgSize[1] == pgran->wfSize * (i + 1) );
+		}
+		ret = ret && ret_multiple;
     }
     else {
         calcPgranDedicated(pgran, subdims, 1, 3);

src/library/tools/tune/subdim.c

@@ -213,8 +213,8 @@ initVector(SubDimInfo* sd)
         setVariable(sd, V_L1_BW, 6, &dim[0]);
     }
     else {
-        setVariable(sd, V_L0_X,  4, &dim[4]);
-        setVariable(sd, V_L0_Y,  4, &dim[4]);
+        setVariable(sd, V_L0_X,  4, &dim[5]);
+        setVariable(sd, V_L0_Y,  4, &dim[5]);
         setVariable(sd, V_L0_BW, 6, &dim[0]);
         setVariable(sd, V_L1_X,  6, &dim[0]);
         setVariable(sd, V_L1_Y,  6, &dim[0]);
@@ -258,7 +258,7 @@ calcPGranularity (SubDimInfo* sd)
 
     pgran->wgDim = 2;
     pgran->wfSize = 64;
-
+	pgran->maxWorkGroupSize = sd->workGroupSizes;
 
     // if pattern provides granularity calculation
     // call the pattern function