added regression script

Author: AamirRaihan

Makefile

@@ -36,7 +36,7 @@ LDFLAGS = --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -I$(INC_D
 cutlass-test: $(obj)
 	$(CC) $(CFLAGS) -o $@ $^ $(LDFLAGS)
 
-clean: 
+clean:
 	rm cutlass-test
 	rm -rf gemm_tt_*
 	rm -rf host_results_*

gemm-test/dgemm_tests.h

@@ -1,19 +1,20 @@
+  #ifdef DGEMM_1
   typedef cutlass::gemm::DgemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                      cutlass::MatrixLayout::kRowMajor,
                                      cutlass::Shape<8, 32, 64> > DGemmTraits1;
   run_gemm<DGemmTraits1>(64, 32, 8);
+  #endif
 
-
-
+  #ifdef DGEMM_2
   typedef cutlass::gemm::DgemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                      cutlass::MatrixLayout::kRowMajor,
                                      cutlass::Shape<8, 32, 64> > DGemmTraits2;
   run_gemm<DGemmTraits2>(256, 128, 64);
+  #endif
 
-
-
+  #ifdef DGEMM_3
   typedef cutlass::gemm::DgemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                      cutlass::MatrixLayout::kRowMajor,
                                      cutlass::Shape<8, 64, 64> > DGemmTraits3;
   run_gemm<DGemmTraits3>(64, 64, 8);
-
+  #endif

gemm-test/gemm.h

@@ -83,14 +83,21 @@ static void run_gemm(
 
   Gemm::launch(params);
 
-  cudaError_t result = cudaDeviceSynchronize();
-  if(result==cudaSuccess){
-	printf("Successfully Launched\n");	
-  }
-  int save=0;
+
+  cudaError_t result;
+  do
+  { 
+	result = cudaDeviceSynchronize();
+  }while(result!=cudaSuccess);  
+  printf("Successfully Launched\n");	
+  
+  int save=1;
   int completedsuccessfully=testbed.verify_with_host(save,save);
   if (completedsuccessfully==1){
-	printf("Result Verified\n");	
+  	printf("Result Verified\n");	
+  }
+  else{
+  	printf("ERROR");	
   }
 
 }

gemm-test/sgemm_tests.h

@@ -1,16 +1,20 @@
+  #ifdef SGEMM_1
   typedef cutlass::gemm::SgemmTraits<cutlass::MatrixLayout::kColumnMajor,
     cutlass::MatrixLayout::kRowMajor, cutlass::Shape<8, 128, 128> >
     SgemmTraits1;
   run_gemm<SgemmTraits1>(1024, 512, 8);
+  #endif
 
-
+  #ifdef SGEMM_2
   typedef cutlass::gemm::SgemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                      cutlass::MatrixLayout::kRowMajor, cutlass::Shape<8, 128, 128> >
       SgemmTraits2;
   run_gemm<SgemmTraits2>(128, 81, 1);
+  #endif
 
-
+  #ifdef SGEMM_3
   typedef cutlass::gemm::SgemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                      cutlass::MatrixLayout::kRowMajor, cutlass::Shape<8, 128, 128> >
       SgemmTraits3;
   run_gemm<SgemmTraits3>(128, 112, 8);
+  #endif