Merge pull request #133 from TimmyLiu/develop

fix the performance drop of SGEMM column major NT or row major TN when lda and ldb are big multiples of 1024 such as 4096, 5120, 6144, 7168, 8192

Author: Timmy

src/library/CMakeLists.txt

@@ -81,6 +81,7 @@ set(SRC_BLAS
 	blas/functor/gcn_dgemmSmallMatrices.cc
 	blas/functor/gcn_sgemmSmallMatrices.cc
 	blas/functor/hawaii_sgemmBranchKernel.cc
+	blas/functor/hawaii_sgemmBig1024Kernel.cc
 )
 
 set(SRC_BLAS_HEADERS
@@ -114,6 +115,7 @@ set(SRC_BLAS_HEADERS
 	blas/functor/include/gcn_dgemmSmallMatrices.h
 	blas/functor/include/gcn_sgemmSmallMatrices.h
 	blas/functor/include/hawaii_sgemmBranchKernel.h
+	blas/functor/include/hawaii_sgemmBig1024Kernel.h
 )
 
 set(SRC_BLAS_GENERIC
@@ -234,6 +236,7 @@ set (SRC_CL_TEMPLATES
     dtrsm_gpu192.cl
 	dgemm_gcn_SmallMatrices.cl
 	sgemm_gcn_SmallMatrices.cl
+	sgemm_gcn_bigMatrices.cl
 	sgemm_gcn.cl
 	zgemm_gcn.cl
 )
@@ -253,6 +256,9 @@ set(SRC_CL_TEMPLATES_GEN
 	sgemm_gcn_SmallMatrices.clHawaii_64.bin.cl
 	sgemm_gcn_SmallMatrices.clTahiti_64.bin.cl
     sgemm_gcn_SmallMatrices.clBonaire_64.bin.cl
+	sgemm_gcn_bigMatrices.clHawaii_64.bin.cl
+	sgemm_gcn_bigMatrices.clTahiti_64.bin.cl
+    sgemm_gcn_bigMatrices.clBonaire_64.bin.cl
 	sgemm_gcn.clHawaii_64.bin.cl
 	zgemm_gcn.clHawaii_64.bin.cl
     sgemm_gcn.clBonaire_64.bin.cl

src/library/bingen.cmake

@@ -15,6 +15,7 @@ ${CLTEMPLATE_PATH}/sgemm_hawaiiSplitKernel.cl
 ${CLTEMPLATE_PATH}/sgemm_gcn.cl
 ${CLTEMPLATE_PATH}/zgemm_gcn.cl
 ${CLTEMPLATE_PATH}/sgemm_gcn_SmallMatrices.cl
+${CLTEMPLATE_PATH}/sgemm_gcn_bigMatrices.cl
 ${CLTEMPLATE_PATH}/sgemm_hawaiiSplit64_32.cl
 ${CLTEMPLATE_PATH}/dtrsm_gpu192.cl
 )

src/library/blas/functor/hawaii.cc

@@ -27,6 +27,7 @@
 #include "hawaii_sgemmSplit64_32.h"
 #include "gcn_zgemm.h"
 #include "gpu_dtrsm192.h"
+#include "hawaii_sgemmBig1024Kernel.h"
 
 FunctorSelectorHawaii FunctorSelectorHawaii::instance ;
 
@@ -116,6 +117,24 @@ clblasSgemmFunctor * FunctorSelectorHawaii::select_sgemm_specific(clblasSgemmFun
   //if (functor)
   //  return functor;
 
+  if ((args.lda % 1024 == 0) && (args.ldb % 1024 == 0) && (args.K > args.lda / 4))
+  {
+	  if ((args.lda == args.ldb) && (args.lda >= 4096) && (args.lda <= 8192)) // between 4096 and 8192 for now
+	  {
+		  if (args.lda != 6144)// 6144 is handled by a special case split
+		  {
+			  // we are going to call 16 GEMMs with M=M/2, N=N/2, K=K/4
+			  // each GEMM requires M%128 == 0, N%128 == 0, K%16 == 0
+			  if (args.M % 256 == 0 && args.N % 256 == 0 && args.K % 64 == 0)
+			  {
+				  functor = clBlashawaiiSgemmBig1024KernelFunctor::provide(args, "Hawaii");
+				  if (functor)
+					  return functor;
+			  }
+		  }
+	  }
+  }
+
   if ((args.M >= 1184 && args.N >= 1184) && (args.M <= 3872 && args.N <= 3872) && (args.M % 64 != 0 && args.N % 64 != 0) && (args.M % 96 != 0 && args.N % 96 != 0) && (args.K % 16 == 0))
   {
 	  //all the mod32 sizes that is not mod64 or mod96 ranging from 1184 to 3872