generalized v4 completed;Note I have only added row major support for matrix b

4 files changed, 731 insertions, 61 deletions

diff --git a/cuda-kernels/genericMatrixMultiply.cu b/cuda-kernels/genericMatrixMultiply.cu
index 95cf021..6a5d33f 100644
--- a/cuda-kernels/genericMatrixMultiply.cu
+++ b/cuda-kernels/genericMatrixMultiply.cu
@@ -40,9 +40,9 @@ void cudaErrCheck_(cudaError_t stat, const char *file, int line) {
 using namespace nvcuda;
 
 // Must be multiples of 16 for wmma code to work
-#define MATRIX_M (16)
-#define MATRIX_N (16)
-#define MATRIX_K (16)
+#define MATRIX_M (32)
+#define MATRIX_N (32)
+#define MATRIX_K (32)
 
 
 
@@ -152,27 +152,27 @@ int main(int argc, char* argv[]) {
    b_host_wmma      = (float*)malloc(MATRIX_K * MATRIX_N * sizeof(float));
 
 
-   printf("a\n"); 
+   printf("INITIAL_MATRIX_A\n"); 
    for(int m=0;m<MATRIX_M;m++){
 	for(int n=0;n<MATRIX_K;n++){
 		a_host_wmma[m*MATRIX_K+n]= (rand()%3);///3.0; 
-		printf("%f ",a_host_wmma[m*MATRIX_K+n]);
+		printf("%.2f ",a_host_wmma[m*MATRIX_K+n]);
 	}
 	printf("\n");
    }
-   printf("b\n"); 
+   printf("INITIAL_MATRIX_B\n"); 
    for(int m=0;m<MATRIX_K;m++){
 	for(int n=0;n<MATRIX_N;n++){
 		b_host_wmma[m*MATRIX_N+n]=(rand()%3);///3.0;
-		printf("%f ",b_host_wmma[m*MATRIX_K+n]);
+		printf("%.2f ",b_host_wmma[m*MATRIX_K+n]);
 	}
 	printf("\n");
    }
-   printf("c\n"); 
+   printf("INITIAL_MATRIX_C\n"); 
    for(int m=0;m<MATRIX_M;m++){
 	for(int n=0;n<MATRIX_N;n++){
 		c_host_wmma[m*MATRIX_N+n]= (rand()%3);///3.0;
-		printf("%f ",c_host_wmma[m*MATRIX_K+n]);
+		printf("%.2f ",c_host_wmma[m*MATRIX_K+n]);
 	}
 	printf("\n");
    }
@@ -215,34 +215,8 @@ int main(int argc, char* argv[]) {
    printf("\nChecking results...\n");
    cudaErrCheck(cudaMemcpy(d_host_wmma, c_wmma, MATRIX_M * MATRIX_N * sizeof(float), cudaMemcpyDeviceToHost));
    
-   printf("wmma:d\n"); 
-   for(int m=0;m<MATRIX_M;m++){
-	for(int n=0;n<MATRIX_N;n++){
-		printf("%f ",d_host_wmma[m*MATRIX_K+n]);
-	}
-	printf("\n");
-   }
-   printf("wmma:d\n"); 
-   for(int m=0;m<MATRIX_M;m++){
-	for(int n=0;n<MATRIX_N;n++){
-		printf("%f ",d_host_wmma[m*MATRIX_K+n]);
-	}
-	printf("\n");
-   }
-   printf("wmma:d\n"); 
-   for(int m=0;m<MATRIX_M;m++){
-	for(int n=0;n<MATRIX_N;n++){
-		printf("%f ",d_host_wmma[m*MATRIX_K+n]);
-	}
-	printf("\n");
-   }
-   printf("wmma:d\n"); 
-   for(int m=0;m<MATRIX_M;m++){
-	for(int n=0;n<MATRIX_N;n++){
-		printf("%f ",d_host_wmma[m*MATRIX_K+n]);
-	}
-	printf("\n");
-   }
+   int t=200000000;
+   while(t-->0);
    for(int m=0;m<MATRIX_M;m++){
 	for(int n=0;n<MATRIX_N;n++){
 		d_cal_host_wmma[n*MATRIX_N+m]=0;
diff --git a/cuda-kernels/genericMatrixMultiplyRow.cu b/cuda-kernels/genericMatrixMultiplyRow.cu
new file mode 100644
index 0000000..6194492
--- /dev/null
+++ b/cuda-kernels/genericMatrixMultiplyRow.cu
@@ -0,0 +1,289 @@
+/* Copyright (c) 1993-2017, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+// Define some error checking macros.
+#define cudaErrCheck(stat) { cudaErrCheck_((stat), __FILE__, __LINE__); }
+void cudaErrCheck_(cudaError_t stat, const char *file, int line) {
+   if (stat != cudaSuccess) {
+      fprintf(stderr, "CUDA Error: %s %s %d\n", cudaGetErrorString(stat), file, line);
+   }
+}
+
+
+#include <mma.h>
+using namespace nvcuda;
+
+// Must be multiples of 16 for wmma code to work
+#define MATRIX_M (32)
+#define MATRIX_N (32)
+#define MATRIX_K (32)
+
+
+
+// The only dimensions currently supported by WMMA
+const int WMMA_M = 16;
+const int WMMA_N = 16;
+const int WMMA_K = 16;
+
+__global__ void wmma_example(half *a, half *b, float *c, int M, int N, int K, float alpha, float beta) {
+   // Leading dimensions. Packed with no transpositions.
+   int lda = M;
+   int ldb = K;
+   int ldc = M;
+
+   // Tile using a 2D grid
+   int warpM = (blockIdx.x * blockDim.x + threadIdx.x) / warpSize;
+   int warpN = (blockIdx.y * blockDim.y + threadIdx.y);
+ 
+   // Declare the fragments
+   wmma::fragment<wmma::matrix_a, WMMA_M, WMMA_N, WMMA_K, half, wmma::row_major> a_frag;
+   wmma::fragment<wmma::matrix_b, WMMA_M, WMMA_N, WMMA_K, half, wmma::row_major> b_frag;
+   wmma::fragment<wmma::accumulator, WMMA_M, WMMA_N, WMMA_K, float> acc_frag;
+   wmma::fragment<wmma::accumulator, WMMA_M, WMMA_N, WMMA_K, float> c_frag;
+
+   wmma::fill_fragment(acc_frag, 0.0f);
+
+   // Loop over k
+   for (int i = 0; i < K; i += WMMA_K) {
+      int aRow = warpM * WMMA_M;
+      int aCol = i;
+
+      int bRow = i;
+      int bCol = warpN * WMMA_N;
+
+      // Bounds checking
+      if (aRow < M && aCol < K && bRow < K && bCol < N) {
+         // Load the inputs
+         wmma::load_matrix_sync(a_frag, a + aRow * lda+ aCol , lda);
+         wmma::load_matrix_sync(b_frag, b + bRow * ldb+ bCol , ldb);
+ 
+         // Perform the matrix multiplication
+         wmma::mma_sync(acc_frag, a_frag, b_frag, acc_frag);
+
+      }
+   }
+
+   // Load in the current value of c, scale it by beta, and add this our result scaled by alpha
+   int cRow = warpM * WMMA_M;
+   int cCol = warpN * WMMA_N;
+
+   if (cRow < M && cCol < N) {
+      wmma::load_matrix_sync(c_frag, c + cRow*ldc + cCol , ldc, wmma::mem_row_major);
+
+
+      for(int i=0; i < c_frag.num_elements; i++) {
+         c_frag.x[i] = alpha * acc_frag.x[i] + beta * c_frag.x[i];
+      }
+
+      // Store the output
+      wmma::store_matrix_sync(c + cRow *ldc + cCol , c_frag, ldc, wmma::mem_row_major);
+   }
+}
+
+__global__ void convertFp32ToFp16 (half *out, float *in, int n) {
+   int idx = blockDim.x * blockIdx.x + threadIdx.x;
+   if (idx < n) {
+      out[idx] = in[idx];
+   }
+}
+
+int main(int argc, char* argv[]) {
+   float *a_fp32;
+   float *b_fp32;
+   half *a_fp16;
+   half *b_fp16;
+
+   float *c;
+   float *c_wmma;
+
+   float *d_host_wmma;
+   float *d_cal_host_wmma;
+   float *a_host_wmma;
+   float *b_host_wmma;
+   float *c_host_wmma;
+   
+   
+   cudaEvent_t startWMMA;
+   cudaEvent_t stopWMMA;
+   
+   
+   cudaErrCheck(cudaEventCreate(&startWMMA));
+   cudaErrCheck(cudaEventCreate(&stopWMMA));
+   
+   // Use tensor cores
+   cudaErrCheck(cudaMalloc((void**)&a_fp32, MATRIX_M * MATRIX_K * sizeof(float)));
+   cudaErrCheck(cudaMalloc((void**)&b_fp32, MATRIX_K * MATRIX_N * sizeof(float)));
+   cudaErrCheck(cudaMalloc((void**)&a_fp16, MATRIX_M * MATRIX_K * sizeof(half)));
+   cudaErrCheck(cudaMalloc((void**)&b_fp16, MATRIX_K * MATRIX_N * sizeof(half)));
+
+   cudaErrCheck(cudaMalloc((void**)&c, MATRIX_M * MATRIX_N * sizeof(float)));
+   cudaErrCheck(cudaMalloc((void**)&c_wmma, MATRIX_M * MATRIX_N * sizeof(float)));
+
+   d_host_wmma      = (float*)malloc(MATRIX_M * MATRIX_N * sizeof(float));
+   d_cal_host_wmma      = (float*)malloc(MATRIX_M * MATRIX_N * sizeof(float));
+   c_host_wmma = (float*)malloc(MATRIX_M * MATRIX_N * sizeof(float));
+   a_host_wmma      = (float*)malloc(MATRIX_M * MATRIX_K * sizeof(float));
+   b_host_wmma      = (float*)malloc(MATRIX_K * MATRIX_N * sizeof(float));
+
+
+   printf("INITIAL_MATRIX_A\n"); 
+   for(int m=0;m<MATRIX_M;m++){
+	for(int n=0;n<MATRIX_K;n++){
+		a_host_wmma[m*MATRIX_K+n]= (rand()%3);///3.0; 
+		printf("%.2f ",a_host_wmma[m*MATRIX_K+n]);
+	}
+	printf("\n");
+   }
+   printf("INITIAL_MATRIX_B\n"); 
+   for(int m=0;m<MATRIX_K;m++){
+	for(int n=0;n<MATRIX_N;n++){
+		b_host_wmma[m*MATRIX_N+n]=(rand()%3);///3.0;
+		printf("%.2f ",b_host_wmma[m*MATRIX_K+n]);
+	}
+	printf("\n");
+   }
+   printf("INITIAL_MATRIX_C\n"); 
+   for(int m=0;m<MATRIX_M;m++){
+	for(int n=0;n<MATRIX_N;n++){
+		c_host_wmma[m*MATRIX_N+n]= (rand()%3);///3.0;
+		printf("%.2f ",c_host_wmma[m*MATRIX_K+n]);
+	}
+	printf("\n");
+   }
+
+   cudaErrCheck(cudaMemcpy(a_fp32,a_host_wmma,  MATRIX_M * MATRIX_K * sizeof(float), cudaMemcpyHostToDevice));
+   cudaErrCheck(cudaMemcpy(b_fp32,b_host_wmma,  MATRIX_K * MATRIX_N * sizeof(float), cudaMemcpyHostToDevice));
+
+   convertFp32ToFp16 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (a_fp16, a_fp32, MATRIX_M * MATRIX_K);
+   convertFp32ToFp16 <<< (MATRIX_K * MATRIX_N + 255) / 256, 256 >>> (b_fp16, b_fp32, MATRIX_K * MATRIX_N);
+
+   cudaErrCheck(cudaMemcpy(c, c_host_wmma,  MATRIX_M * MATRIX_N * sizeof(float), cudaMemcpyHostToDevice));
+   cudaErrCheck(cudaMemcpy(c_wmma, c, MATRIX_M * MATRIX_N * sizeof(float), cudaMemcpyDeviceToDevice));
+
+   float alpha = 1.0f;
+   float beta = 1.0f;
+
+
+   printf("\nM = %d, N = %d, K = %d. alpha = %f, beta = %f\n\n", MATRIX_M, MATRIX_N, MATRIX_K, alpha, beta);
+   
+   // First: using WMMA
+   dim3 gridDim;
+   dim3 blockDim;
+ 
+   // blockDim.x must be a multple of warpSize
+   // 128x4 means we have 16 warps and a block computes a 64x64 output tile
+   blockDim.x = 64;
+   blockDim.y = 2;
+
+   gridDim.x = (MATRIX_M + (WMMA_M * blockDim.x / 32 - 1)) / (WMMA_M * blockDim.x / 32);
+   gridDim.y = (MATRIX_N + WMMA_N * blockDim.y - 1) / (WMMA_N * blockDim.y);
+   printf("GRID:X=%d,Y=%d\n",gridDim.x,gridDim.y);
+   printf("BLOCK:X=%d,Y=%d\n",blockDim.x,blockDim.y);
+   
+   printf("Running with wmma...\n");
+   cudaErrCheck(cudaEventRecord(startWMMA));
+   wmma_example <<< gridDim, blockDim >>> (a_fp16, b_fp16, c_wmma, MATRIX_M, MATRIX_N, MATRIX_K, alpha, beta);
+   cudaErrCheck(cudaEventRecord(stopWMMA));
+   cudaErrCheck(cudaEventSynchronize(stopWMMA));
+
+   printf("\nChecking results...\n");
+   cudaErrCheck(cudaMemcpy(d_host_wmma, c_wmma, MATRIX_M * MATRIX_N * sizeof(float), cudaMemcpyDeviceToHost));
+   
+   int t=200000000;  
+   while(t-->0);
+ 
+   for(int m=0;m<MATRIX_M;m++){
+	for(int n=0;n<MATRIX_N;n++){
+		d_cal_host_wmma[m*MATRIX_N+n]=0;
+		for(int k=0;k<MATRIX_K;k++){
+			d_cal_host_wmma[m*MATRIX_N+n]+=	a_host_wmma[m*MATRIX_K+k]*b_host_wmma[k*MATRIX_K+n];
+		}
+		d_cal_host_wmma[m*MATRIX_N+n]+=c_host_wmma[m*MATRIX_N+n];
+	}
+   }
+   printf("cal:d\n"); 
+   for(int m=0;m<MATRIX_M;m++){
+	for(int n=0;n<MATRIX_N;n++){
+		printf("%.2f ",d_cal_host_wmma[m*MATRIX_K+n]);
+	}
+	printf("\n");
+  }
+   printf("wmma:d\n"); 
+   for(int m=0;m<MATRIX_M;m++){
+	for(int n=0;n<MATRIX_N;n++){
+		printf("%.2f ",d_host_wmma[m*MATRIX_K+n]);
+	}
+	printf("\n");
+   }
+   int suc=1;
+   float relative_error;
+   for(int m=0;m<MATRIX_M;m++){
+	for(int n=0;n<MATRIX_N;n++){
+ 		relative_error=100*abs(d_cal_host_wmma[m*MATRIX_N+n]-d_host_wmma[m*MATRIX_N+n])/d_host_wmma[m*MATRIX_N+n];
+		printf("relative_error=%f\n",relative_error);
+		if((int)relative_error>1)
+		{	
+			printf("ERROR:\n");
+			suc=0;
+			printf("ROW=%d,COL=%d:cpu=%f,gpgpusim=%f\n",m,n,d_cal_host_wmma[m*MATRIX_N+n],d_host_wmma[m*MATRIX_N+n]);
+		}
+	}
+   }
+   if(suc==1)
+	printf("COMPLETED_SUCCESSFULLY\n");
+   
+   //int errors = 0;
+   //for (int i = 0; i < MATRIX_M * MATRIX_N; i++) {
+   //   float v1 = c_host_wmma[i];
+   //   float v2 = c_host_cublas[i];
+   //   if (v1 / v2 > 1.0001 || v2 / v1 > 1.0001 || abs(v1 - v2) > 1e-5) {
+   //      errors++;
+   //      if (errors < 10) printf("%f %f\n", v1, v2);
+   //   }
+   //}
+   
+   float wmmaTime;
+   cudaErrCheck(cudaEventElapsedTime(&wmmaTime, startWMMA, stopWMMA));
+   printf("wmma took %fms\n", wmmaTime);
+   
+   cudaErrCheck(cudaEventDestroy(startWMMA));
+   cudaErrCheck(cudaEventDestroy(stopWMMA));
+   cudaErrCheck(cudaFree(a_fp32));
+   cudaErrCheck(cudaFree(b_fp32));
+   cudaErrCheck(cudaFree(a_fp16));
+   cudaErrCheck(cudaFree(b_fp16));
+   cudaErrCheck(cudaFree(c));
+   cudaErrCheck(cudaFree(c_wmma));
+   free(d_host_wmma);
+   free(c_host_wmma);
+   cudaErrCheck(cudaDeviceReset());
+   return 0;
+}
+
+
diff --git a/cuda-kernels/v4p_genericMatrixMultiply.cu b/cuda-kernels/v4p_genericMatrixMultiply.cu
new file mode 100644
index 0000000..a08903b
--- /dev/null
+++ b/cuda-kernels/v4p_genericMatrixMultiply.cu
@@ -0,0 +1,384 @@
+#include <stdio.h>
+#include <curand.h>
+
+// Define some error checking macros.
+#define cudaErrCheck(stat) { cudaErrCheck_((stat), __FILE__, __LINE__); }
+void cudaErrCheck_(cudaError_t stat, const char *file, int line) {
+   if (stat != cudaSuccess) {
+      fprintf(stderr, "CUDA Error: %s %s %d\n", cudaGetErrorString(stat), file, line);
+   }
+}
+
+#define curandErrCheck(stat) { curandErrCheck_((stat), __FILE__, __LINE__); }
+void curandErrCheck_(curandStatus_t stat, const char *file, int line) {
+   if (stat != CURAND_STATUS_SUCCESS) {
+      fprintf(stderr, "cuRand Error: %d %s %d\n", stat, file, line);
+   }
+}
+
+#include <mma.h>
+using namespace nvcuda;
+
+// Must be multiples of 16 for wmma code to work
+#define MATRIX_M (32)
+#define MATRIX_N (32)
+#define MATRIX_K (32)
+
+
+// The only dimensions currently supported by WMMA
+const int WMMA_M = 16;
+const int WMMA_N = 16;
+const int WMMA_K = 16;
+
+
+
+__global__ void vp_example(int *a, int *b, int *c, int M, int N, int K ) {
+   // Leading dimensions. Packed with no transpositions.
+   int lda = M;
+   int ldb = K;
+   int ldc = M;
+
+   // Tile using a 2D grid
+   int warpM = (blockIdx.x * blockDim.x + threadIdx.x) / warpSize;
+   int warpN = (blockIdx.y * blockDim.y + threadIdx.y);
+ 
+   // Declare the fragments
+   int a_frag[8];
+   int b_frag[8];
+   int c_frag[8];
+   int acc_frag[8];
+   
+   acc_frag[0]=0;
+   acc_frag[1]=0;
+   acc_frag[2]=0;
+   acc_frag[3]=0;
+   acc_frag[4]=0;
+   acc_frag[5]=0;
+   acc_frag[6]=0;
+   acc_frag[7]=0;
+
+   // Loop over k
+   for (int i = 0; i < K; i += WMMA_K) {
+      int aRow = warpM * WMMA_M;
+      int aCol = i;
+
+      int bRow = i;
+      int bCol = warpN * WMMA_N;
+
+      // Bounds checking
+      if (aRow < M && aCol < K && bRow < K && bCol < N) {
+         // Load the inputs
+        // vp::load_matrix_sync(a_frag, a + aRow * lda+ aCol , lda);
+	asm("/*");
+	asm("CPTX_BEGIN");
+	asm("vp.load.a.sync.row.m16n16k16.s32 {%0,%1,%2,%3,%4,%5,%6,%7},[%8],%9;" : 
+	"=r"(a_frag[0]), "=r"(a_frag[1]),"=r"(a_frag[2]),"=r"(a_frag[3]),
+	"=r"(a_frag[4]),"=r"(a_frag[5]),"=r"(a_frag[6]),"=r"(a_frag[7]):
+	"l"(a+aRow*lda+aCol),"r"(lda)
+	);
+	asm("CPTX_END");
+	asm("*/");
+        //vp::load_matrix_sync(b_frag, b + bRow * ldb+ bCol , ldb);
+	asm("/*");
+	asm("CPTX_BEGIN");
+	asm("vp.load.b4.sync.row.m16n16k16.s32 {%0},[%1],%2;" : 
+	"=r"(b_frag[0]):
+	"l"(b+bRow*ldb/8+bCol),"r"(ldb/8)
+	);
+	asm("CPTX_END");
+	asm("*/");
+ 
+         // Perform the matrix multiplication
+         //vp::mma_sync(acc_frag, a_frag, b_frag, acc_frag);
+	asm("/*");
+	asm("CPTX_BEGIN");
+	asm("vp.mma.sync.row.row.m16n16k16.s32 {%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9, %10, %11, %12, %13, %14, %15}, {%16}, {%17, %18, %19, %20, %21, %22, %23, %24};" : 
+	"=r"(acc_frag[0]), "=r"(acc_frag[1]),"=r"(acc_frag[2]),"=r"(acc_frag[3]),
+	"=r"(acc_frag[4]),"=r"(acc_frag[5]),"=r"(acc_frag[6]),"=r"(acc_frag[7]):
+	"r"(a_frag[0]),"r"(a_frag[1]),"r"(a_frag[2]),"r"(a_frag[3]),
+	"r"(a_frag[4]),"r"(a_frag[5]),"r"(a_frag[6]),"r"(a_frag[7]),
+	"r"(b_frag[0]),
+	"r"(acc_frag[0]),"r"(acc_frag[1]),"r"(acc_frag[2]),"r"(acc_frag[3]),
+	"r"(acc_frag[4]),"r"(acc_frag[5]),"r"(acc_frag[6]),"r"(acc_frag[7])
+	);
+	asm("CPTX_END");
+	asm("*/");
+
+      }
+   }
+
+   // Load in the current value of c, scale it by beta, and add this our result scaled by alpha
+   int cRow = warpM * WMMA_M;
+   int cCol = warpN * WMMA_N;
+
+   if (cRow < M && cCol < N) {
+        //vp::load_matrix_sync(c_frag, c + cRow*ldc + cCol , ldc, wmma::mem_row_major);
+	asm("/*");
+	asm("CPTX_BEGIN");
+	asm("vp.load.c.sync.row.m16n16k16.s32 {%0,%1,%2,%3,%4,%5,%6,%7},[%8],%9;" : 
+	"=r"(c_frag[0]), "=r"(c_frag[1]),"=r"(c_frag[2]),"=r"(c_frag[3]),
+	"=r"(c_frag[4]),"=r"(c_frag[5]),"=r"(c_frag[6]),"=r"(c_frag[7]):
+	"l"(c+cRow*ldc),"r"(ldc)
+	);
+	asm("CPTX_END");
+	asm("*/");
+
+
+      for(int i=0; i < 8; i++) {
+         c_frag[i] =  acc_frag[i] + c_frag[i];
+      }
+
+      // Store the output
+      //vp::store_matrix_sync(c + cRow *ldc + cCol , c_frag, ldc, wmma::mem_row_major);
+	asm("/*");
+	asm("CPTX_BEGIN");
+	asm("vp.store.d.sync.row.m16n16k16.s32 [%0], {%1,%2,%3,%4,%5,%6,%7,%8},%9;" : 
+	:"l"(c+cRow*ldc+cCol),
+	"r"(c_frag[0]), "r"(c_frag[1]),"r"(c_frag[2]),"r"(c_frag[3]),
+	"r"(c_frag[4]),"r"(c_frag[5]),"r"(c_frag[6]),"r"(c_frag[7]),
+	"r"(ldc)
+	);
+	asm("CPTX_END");
+	asm("*/");
+   }
+}
+
+__global__ void convertFp32ToFp16 (half *out, float *in, int n) {
+   int idx = blockDim.x * blockIdx.x + threadIdx.x;
+   if (idx < n) {
+      out[idx] = in[idx];
+   }
+}
+__global__ void convertFp16ToFp32 (float *out, half *in, int n) {
+   int idx = blockDim.x * blockIdx.x + threadIdx.x;
+   if (idx < n) {
+      out[idx] = in[idx];
+   }
+}
+
+__global__ void convertInt32ToInt4 (int *out, int *in, int n) {
+   int idx = blockDim.x * blockIdx.x + threadIdx.x;
+   if (idx < n/8) {
+      		out[idx] =(in[8*idx]&0xf)|(in[8*idx+1]&0xf)<<4|(in[8*idx+2]&0xf)<<8|(in[8*idx+3]&0xf)<<12|
+      			  (in[8*idx+4]&0xf)<<16|(in[8*idx+5]&0xf)<<20|(in[8*idx+6]&0xf)<<24|(in[8*idx+7]&0xf)<<28;
+   }
+}
+__global__ void convertInt32ToInt8 (int *out, int *in, int n) {
+   int idx = blockDim.x * blockIdx.x + threadIdx.x;
+   if (idx < n/4) {
+      		out[idx] =(in[4*idx]&0xff)|(in[4*idx+1]&0xff)<<8|(in[4*idx+2]&0xff)<<16|(in[4*idx+3]&0xff)<<24;
+   }
+}
+__global__ void convertInt32ToInt16 (int *out, int *in, int n) {
+   int idx = blockDim.x * blockIdx.x + threadIdx.x;
+   if (idx < n/2) {
+      		out[idx] =(in[2*idx]&0xffff)|(in[2*idx+1]&0xffff)<<16;
+   }
+}
+
+__global__ void convertInt4ToInt32 (int *out, int *in, int n) {
+   int idx = blockDim.x * blockIdx.x + threadIdx.x;
+   int shft_amt=4*(idx%8);
+   int shft_mask=0xf<<shft_amt;
+   if (idx < n) {
+      	out[idx]= (in[idx/8]&shft_mask)>>shft_amt;
+   }
+}
+__global__ void convertInt8ToInt32 (int *out, int *in, int n) {
+   int idx = blockDim.x * blockIdx.x + threadIdx.x;
+   int shft_amt=8*(idx%4);
+   int shft_mask=0xff<<shft_amt;
+   if (idx < n) {
+      	out[idx]= (in[idx/4]&shft_mask)>>shft_amt;
+   }
+}
+__global__ void convertInt16ToInt32 (int *out, int *in, int n) {
+   int idx = blockDim.x * blockIdx.x + threadIdx.x;
+   int shft_amt=16*(idx%2);
+   int shft_mask=0xffff<<shft_amt;
+   if (idx < n) {
+      	out[idx]= (in[idx/2]&shft_mask)>>shft_amt;
+   }
+}
+
+int main(int argc, char* argv[]) {
+   int *a_int32;
+   int *b_int32;
+   int *c_int32;
+   int *d_int32;
+
+   int *a_int4;
+   int *b_int4;
+   int *a_int8;
+   int *b_int8;
+   int *a_int16;
+   int *b_int16;
+   
+   int  *a_host_wmma;
+   int  *b_host_wmma;
+   int  *c_host_wmma;
+   int  *d_host_wmma;
+   int  *d_cal_host_wmma;
+
+   cudaEvent_t startWMMA;
+   cudaEvent_t stopWMMA;
+   
+   
+   cudaErrCheck(cudaEventCreate(&startWMMA));
+   cudaErrCheck(cudaEventCreate(&stopWMMA));
+   
+   // Use tensor cores
+   cudaErrCheck(cudaMalloc((void**)&a_int32, MATRIX_M * MATRIX_K * sizeof(int)));
+   cudaErrCheck(cudaMalloc((void**)&b_int32, MATRIX_K * MATRIX_N * sizeof(int)));
+   cudaErrCheck(cudaMalloc((void**)&c_int32, MATRIX_K * MATRIX_N * sizeof(int)));
+   cudaErrCheck(cudaMalloc((void**)&d_int32, MATRIX_K * MATRIX_N * sizeof(int)));
+   cudaErrCheck(cudaMalloc((void**)&a_int4, MATRIX_M * MATRIX_K * sizeof(int)/8));
+   cudaErrCheck(cudaMalloc((void**)&b_int4, MATRIX_K * MATRIX_N * sizeof(int)/8));
+   cudaErrCheck(cudaMalloc((void**)&a_int8, MATRIX_M * MATRIX_K * sizeof(int)/4));
+   cudaErrCheck(cudaMalloc((void**)&b_int8, MATRIX_K * MATRIX_N * sizeof(int)/4));
+   cudaErrCheck(cudaMalloc((void**)&a_int16, MATRIX_M * MATRIX_K * sizeof(int)/2));
+   cudaErrCheck(cudaMalloc((void**)&b_int16, MATRIX_K * MATRIX_N * sizeof(int)/2));
+
+
+   a_host_wmma      = (int *)malloc(MATRIX_M * MATRIX_K * sizeof(int));
+   b_host_wmma      = (int *)malloc(MATRIX_K * MATRIX_N * sizeof(int));
+   c_host_wmma      = (int *)malloc(MATRIX_M * MATRIX_N * sizeof(int));
+   d_host_wmma      = (int *)malloc(MATRIX_M * MATRIX_N * sizeof(int));
+   d_cal_host_wmma  = (int *)malloc(MATRIX_M * MATRIX_N * sizeof(int));
+
+   printf("a_int32\n");
+   for(int m=0;m<MATRIX_M;m++){
+	for(int n=0;n<MATRIX_K;n++){
+		a_host_wmma[m*MATRIX_K+n]=(m*MATRIX_K+n)%1024;
+		printf("%d ",a_host_wmma[m*MATRIX_K+n]);
+	}
+	printf(";\n");
+   }
+  
+   printf("b_int32\n");
+   for(int m=0;m<MATRIX_K;m++){
+	for(int n=0;n<MATRIX_N;n++){
+		b_host_wmma[m*MATRIX_N+n]=(m*MATRIX_N+n)%4;
+		printf("%d ",b_host_wmma[m*MATRIX_N+n]);
+	}
+		printf(";\n");
+   }
+   
+   printf("c_int32\n");
+   for(int m=0;m<MATRIX_M;m++){
+	for(int n=0;n<MATRIX_N;n++){
+		c_host_wmma[m*MATRIX_N+n]=(m*MATRIX_N+n)%4;
+		d_cal_host_wmma[m*MATRIX_N+n]=0;
+		printf("%d ",c_host_wmma[m*MATRIX_N+n]);
+	}
+		printf(";\n");
+   }
+   for(int m=0;m<MATRIX_M;m++){
+	for(int n=0;n<MATRIX_N;n++){
+		for(int k=0;k<MATRIX_K;k++){
+			d_cal_host_wmma[m*MATRIX_N+n]+=	a_host_wmma[m*MATRIX_K+k]*b_host_wmma[k*MATRIX_K+n];
+		}
+		d_cal_host_wmma[m*MATRIX_N+n]+=c_host_wmma[m*MATRIX_N+n];
+	}
+   }
+
+
+   cudaErrCheck(cudaMemcpy(a_int32,a_host_wmma,  MATRIX_M * MATRIX_K * sizeof(int), cudaMemcpyHostToDevice));
+   cudaErrCheck(cudaMemcpy(b_int32,b_host_wmma,  MATRIX_K * MATRIX_N * sizeof(int), cudaMemcpyHostToDevice));
+   cudaErrCheck(cudaMemcpy(c_int32,c_host_wmma,  MATRIX_M * MATRIX_N * sizeof(int), cudaMemcpyHostToDevice));
+	
+   #ifdef TEST16
+   	convertInt32ToInt16 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (a_int16, a_int32, MATRIX_M * MATRIX_K);
+   	convertInt16ToInt32 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (d_int32, a_int16, MATRIX_M * MATRIX_K);
+   	cudaErrCheck(cudaMemcpy(d_host_wmma, d_int32, MATRIX_M * MATRIX_N * sizeof(int), cudaMemcpyDeviceToHost));
+   #endif
+   #ifdef TEST8
+   	convertInt32ToInt8 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (a_int8, a_int32, MATRIX_M * MATRIX_K);
+  	convertInt8ToInt32 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (d_int32, a_int8, MATRIX_M * MATRIX_K);
+ 	cudaErrCheck(cudaMemcpy(d_host_wmma, d_int32, MATRIX_M * MATRIX_N * sizeof(int), cudaMemcpyDeviceToHost));
+   #endif
+   #ifdef TEST4
+   	convertInt32ToInt4 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (b_int4, b_int32, MATRIX_M * MATRIX_K);
+  	convertInt4ToInt32 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (d_int32, b_int4, MATRIX_M * MATRIX_K);
+ 	cudaErrCheck(cudaMemcpy(d_host_wmma, d_int32, MATRIX_M * MATRIX_N * sizeof(int), cudaMemcpyDeviceToHost));
+   #endif
+   convertInt32ToInt4 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (b_int4, b_int32, MATRIX_M * MATRIX_K);
+
+   dim3 gridDim;
+   dim3 blockDim;
+ 
+   // blockDim.x must be a multple of warpSize
+   // 128x4 means we have 16 warps and a block computes a 64x64 output tile
+   blockDim.x = 64;
+   blockDim.y = 2;
+
+   gridDim.x = (MATRIX_M + (WMMA_M * blockDim.x / 32 - 1)) / (WMMA_M * blockDim.x / 32);
+   gridDim.y = (MATRIX_N + WMMA_N * blockDim.y - 1) / (WMMA_N * blockDim.y);
+   printf("GRID:X=%d,Y=%d\n",gridDim.x,gridDim.y);
+   printf("BLOCK:X=%d,Y=%d\n",blockDim.x,blockDim.y);
+   
+  
+   printf("Running with wmma...\n");
+   cudaErrCheck(cudaEventRecord(startWMMA));
+   vp_example <<< gridDim, blockDim >>> (a_int32, b_int4, c_int32, MATRIX_M, MATRIX_N, MATRIX_K);
+   cudaErrCheck(cudaEventRecord(stopWMMA));
+   cudaErrCheck(cudaEventSynchronize(stopWMMA));
+
+   // Error checking
+   printf("\nChecking results...\n");
+   cudaErrCheck(cudaMemcpy(d_host_wmma, c_int32, MATRIX_M * MATRIX_N * sizeof(float), cudaMemcpyDeviceToHost));
+   
+   float wmmaTime;
+   cudaErrCheck(cudaEventElapsedTime(&wmmaTime, startWMMA, stopWMMA));
+   printf("wmma took %fms\n", wmmaTime);
+   
+   cudaErrCheck(cudaEventDestroy(startWMMA));
+   cudaErrCheck(cudaEventDestroy(stopWMMA));
+
+   int t=200000;
+   while(t-->0);
+   printf("D_CALCULATED\n");
+
+   for(int m=0;m<MATRIX_M;m++){
+	for(int n=0;n<MATRIX_N;n++){
+		printf("%d,",d_cal_host_wmma[m*MATRIX_N+n]);
+	}
+	printf("\n");
+   }
+   printf("D_WMMA\n");
+   for(int m=0;m<MATRIX_M;m++){
+	for(int n=0;n<MATRIX_N;n++){
+		printf("%d,",d_host_wmma[m*MATRIX_N+n]);
+	}
+	printf("\n");
+    }
+   int suc=1;
+   for(int m=0;m<MATRIX_M;m++){
+	for(int n=0;n<MATRIX_N;n++){
+ 		if(abs(d_cal_host_wmma[m*MATRIX_N+n]-d_host_wmma[m*MATRIX_N+n])) 
+		{	
+			printf("ERROR:\n");
+			suc=0;
+		}
+	}
+   }
+   if(suc==1)
+	printf("COMPLETED_SUCCESSFULLY\n");
+   
+   
+   cudaErrCheck(cudaFree(a_int32));
+   cudaErrCheck(cudaFree(b_int32));
+   cudaErrCheck(cudaFree(c_int32));
+   cudaErrCheck(cudaFree(d_int32));
+   cudaErrCheck(cudaFree(a_int8));
+   cudaErrCheck(cudaFree(b_int8));
+
+   free(a_host_wmma);
+   free(b_host_wmma);
+   free(c_host_wmma);
+   free(d_host_wmma);
+   cudaErrCheck(cudaDeviceReset());
+   return 0;
+}
+
+
diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc
index 1e84425..b26adfd 100644
--- a/src/cuda-sim/instructions.cc
+++ b/src/cuda-sim/instructions.cc
@@ -133,7 +133,7 @@ unsigned thread_group_offset(int thread,unsigned  wmma_type,unsigned wmma_layout
 				if(wmma_layout==ROW)
 					offset=load_c_float_row[thread_group]+16*in_tg_index;
 				else	
-					offset=load_c_float_col[thread_group]+16*in_tg_index;
+					offset=load_c_float_col[thread_group]+in_tg_index;
 			break;	
 
          	default:
@@ -1810,7 +1810,7 @@ void vp_mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
 	printf("MATRIX_B\n");
 	for (i=0;i<16;i++){
 		for(j=0;j<16;j++){
-			printf("%x ",matrix_b[i][j].s32);
+			printf("%d ",matrix_b[i][j].s32);
 		}
 		printf("\n");
 	}	
@@ -1922,7 +1922,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
 				for(k=0;k<2*nelem;k++){
 					temp=nw_v[k].f16;
 					if(g_debug_instruction)
-						printf("%f ",temp);
+						printf("%.2f ",temp);
 				}
 				if(g_debug_instruction)
 					printf("\n");
@@ -1930,7 +1930,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
 			else{
 				if(g_debug_instruction){
 					for(k=0;k<8;k++){
-						printf("%f ",v[k].f32);
+						printf("%.2f ",v[k].f32);
 					}
 					printf("\n");
 				}
@@ -1977,7 +1977,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
 		for (i=0;i<16;i++){
 			for(j=0;j<16;j++){
 				temp=matrix_a[i][j].f16;
-				printf("%f ",temp);
+				printf("%.2f ",temp);
 			}
 			printf("\n");
 		}
@@ -1985,7 +1985,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
 		for (i=0;i<16;i++){
 			for(j=0;j<16;j++){
 				temp=matrix_b[i][j].f16;
-				printf("%f ",temp);
+				printf("%.2f ",temp);
 			}
 			printf("\n");
 		}	
@@ -1994,10 +1994,10 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
 			for(j=0;j<16;j++){
 				if(type2==F16_TYPE){
 					temp=matrix_c[i][j].f16;
-					printf("%f ",temp);
+					printf("%.2f ",temp);
 				}
 				else
-					printf("%f ",matrix_c[i][j].f32);
+					printf("%.2f ",matrix_c[i][j].f32);
 			}
 			printf("\n");
 		}	
@@ -2038,7 +2038,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
 			for(j=0;j<16;j++){
 				if(type==F16_TYPE){
 					temp=matrix_d[i][j].f16;
-					printf("%f ",temp);
+					printf("%.2f ",temp);
 				}
 				else
 					printf("%.2f ",matrix_d[i][j].f32);
@@ -2064,7 +2064,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
 			{
 				printf("thread%d:",thrd);
 				for(k=0;k<8;k++){
-					printf("%f ",matrix_d[row_t[k]][col_t[k]].f32);
+					printf("%.2f ",matrix_d[row_t[k]][col_t[k]].f32);
 				}
 				printf("\n");
 			}
@@ -2074,7 +2074,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
 				printf("thread%d:",thrd);
 				for(k=0;k<8;k++){
 					temp=matrix_d[row_t[k]][col_t[k]].f16;
-					printf("%f ",temp);
+					printf("%.2f ",temp);
 				}
 				printf("\n");
 
@@ -3162,10 +3162,14 @@ void vp_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
 	addr_t new_addr = addr+thread_group_offset(thrd,VP_MMA,wmma_layout,type,stride)*size/8;  
 
 	if(g_debug_instruction){
-		printf("vp:store:thread%d=%d,%d,%d,%d,%d,%d,%d,%d\n",thrd,v[0].s32,v[1].s32,v[2].s32,v[3].s32,v[4].s32,v[5].s32,v[6].s32,v[7].s32);   
+		printf("vp_st:thread%d=%d,%d,%d,%d,%d,%d,%d,%d\n",thrd,v[0].s32,v[1].s32,v[2].s32,v[3].s32,v[4].s32,v[5].s32,v[6].s32,v[7].s32);   
 	}
+	
 	for(k=0;k<8;k++){
-       		mem->write(new_addr+4*k,size/8,&v[k].s64,thread,pI);
+		if(wmma_layout==ROW)
+       			mem->write(new_addr+4*k,size/8,&v[k].s64,thread,pI);
+		else if(wmma_layout==COL)
+       			mem->write(new_addr+k*4*stride,size/8,&v[k].s64,thread,pI);
 	}
    	
    	delete [] v;
@@ -3236,7 +3240,7 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
 				printf("thread=%d:",thrd);
 				for(l=0;l<8;l++){
 					temp=v[l].f32;
-					printf("%f",temp);	
+					printf("%.2f",temp);	
 				}
 				printf("\n");
 			}
@@ -3292,6 +3296,7 @@ void vp_ld_impl(const ptx_instruction *pI, core_t *core, warp_inst_t inst)
 	
 	ptx_reg_t data[8];
 	addr_t new_addr;	
+	
 	//note we are using distribution of VP_MMA for every type of load!
 	if(wmma_type==LOAD_A||wmma_type==LOAD_C){
 		 new_addr = addr+thread_group_offset(thrd,VP_MMA,wmma_layout,type,stride)*size/8;  
@@ -3311,22 +3316,40 @@ void vp_ld_impl(const ptx_instruction *pI, core_t *core, warp_inst_t inst)
 
 	if(wmma_type==LOAD_A||wmma_type==LOAD_C){
 		for(i=0;i<8;i++){
-			mem->read(new_addr+4*i,size/8,&data[i].s64);
+			if(wmma_layout==ROW)
+				mem->read(new_addr+4*i,size/8,&data[i].s64);
+			else if(wmma_layout==COL)
+				mem->read(new_addr+4*stride*i,size/8,&data[i].s64);
 		}
 	}
 	else if(wmma_type==LOAD_B4){
-			mem->read(new_addr,size/8,&data[0].s64);
+			if(wmma_layout==ROW){
+				mem->read(new_addr,size/8,&data[0].s64);
+			}
+			else if(wmma_layout==COL){
+	
+			}
 	}
 	else if(wmma_type==LOAD_B8){
-			mem->read(new_addr,size/8,&data[0].s64);
-			mem->read(new_addr+4,size/8,&data[1].s64);
+			if(wmma_layout==ROW){
+				mem->read(new_addr,size/8,&data[0].s64);
+				mem->read(new_addr+4,size/8,&data[1].s64);
+			}
+			else if(wmma_layout==COL){
+	
+			}
 	}
 	else if(wmma_type==LOAD_B16){
 			printf("LOADB16_MODE");
-			mem->read(new_addr,size/8,&data[0].s64);
-			mem->read(new_addr+4,size/8,&data[1].s64);
-			mem->read(new_addr+8,size/8,&data[2].s64);
-			mem->read(new_addr+12,size/8,&data[3].s64);
+			if(wmma_layout==ROW){
+				mem->read(new_addr,size/8,&data[0].s64);
+				mem->read(new_addr+4,size/8,&data[1].s64);
+				mem->read(new_addr+8,size/8,&data[2].s64);
+				mem->read(new_addr+12,size/8,&data[3].s64);
+			}
+			else if(wmma_layout==COL){
+	
+			}
 	}
 	else{
 		printf("wrong vp_load type\n");;
@@ -3490,14 +3513,14 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
 			float temp;
 			for(i=0;i<16;i++){
 				temp=data[i].f16;
-				printf("%f ",temp);
+				printf("%.2f ",temp);
 			}
 			printf("\n");
 		}
 		else{
 			printf("\nmma_ld:thread%d= ",thrd);
 			for(i=0;i<8;i++){
-				printf("%f ",data[i].f32);
+				printf("%.2f ",data[i].f32);
 			}
 			printf("\n");
 			printf("\nmma_ld:thread%d= ",thrd);