diff options
| author | aamir <[email protected]> | 2018-09-22 15:28:40 -0700 |
|---|---|---|
| committer | aamir <[email protected]> | 2018-09-22 15:28:40 -0700 |
| commit | 09e6092ace5213a5d5a49bf80b052802c06a4268 (patch) | |
| tree | 4cef1b01b7015d5b7c730ef7b4d9589db28ffd7d /cuda-kernels/v16p_kernel.cu | |
| parent | 3949357047a621a06b2e7fb4fd6099cce1469d27 (diff) | |
debug prints
Diffstat (limited to 'cuda-kernels/v16p_kernel.cu')
| -rw-r--r-- | cuda-kernels/v16p_kernel.cu | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/cuda-kernels/v16p_kernel.cu b/cuda-kernels/v16p_kernel.cu index 864597d..d84328f 100644 --- a/cuda-kernels/v16p_kernel.cu +++ b/cuda-kernels/v16p_kernel.cu @@ -31,7 +31,7 @@ const int WMMA_M = 16; const int WMMA_N = 16; const int WMMA_K = 16; -__global__ void v4p_example(int *a_int32, int *b_int4, int *c,int *d_int32, int M, int N, int K) { +__global__ void v16p_example(int *a_int32, int *b_int4, int *c,int *d_int32, int M, int N, int K) { int registers_a[8]; int registers_b[8]; @@ -294,7 +294,7 @@ int main(int argc, char* argv[]) { //AAMIR //AAMIR printf("Running with wmma...\n"); cudaErrCheck(cudaEventRecord(startWMMA)); - v4p_example <<< 1, 32>>> (a_int32, b_int16, c_int32, d_int32, MATRIX_M, MATRIX_N, MATRIX_K); + v16p_example <<< 1, 32>>> (a_int32, b_int16, c_int32, d_int32, MATRIX_M, MATRIX_N, MATRIX_K); cudaErrCheck(cudaEventRecord(stopWMMA)); cudaErrCheck(cudaEventSynchronize(stopWMMA)); |