diff options
| author | Amruth <[email protected]> | 2018-04-05 01:09:57 +0000 |
|---|---|---|
| committer | Tor Aamodt <[email protected]> | 2018-04-05 01:09:57 +0000 |
| commit | a17c466a9578f983c48704495566df5a373ff6ba (patch) | |
| tree | 9856a74f2124d669bbd35daac4a346eacc526cd2 /libcuda | |
| parent | deee9038d3d67e60f106776be3dd0a846dd11df9 (diff) | |
| parent | 69c57b077d6799f46ad43b99cc0e0049b7c43775 (diff) | |
Merged in amruth_s/apr2_gpgpusim (pull request #1)
Support for CDP
Diffstat (limited to 'libcuda')
| -rw-r--r-- | libcuda/cuda_runtime_api.cc | 19 |
1 files changed, 10 insertions, 9 deletions
diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 6cf21dd..c103244 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1085,8 +1085,8 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetStreamPriorityRange(int* leastPriori return cudaSuccess; } -__host__ __device__ cudaError_t CUDARTAPI cudaStreamCreateWithFlags(cudaStream_t *pStream, unsigned int flags) { - return cudaStreamCreate(pStream); +__host__ __device__ cudaError_t CUDARTAPI cudaStreamCreateWithFlags(cudaStream_t *stream, unsigned int flags) { + return cudaStreamCreate(stream); } __host__ cudaError_t CUDARTAPI cudaStreamDestroy(cudaStream_t stream) @@ -1499,6 +1499,12 @@ void extract_code_using_cuobjdump(){ no_of_ptx = no_of_ptx + 1; fclose(fp); } + if(no_of_ptx==0){ + printf("WARNING: Number of ptx in the executable file are 0. One of the reasons might be\n"); + printf("\t1. CDP is enabled\n"); + printf("\t2. cuobjdump -lptx doesnt recognize sm_%u\n",forced_max_capability); + printf("\t3. the application was not compiled with nvcc flag sm_%u\n",forced_max_capability); + } } if(!g_cdp_enabled) { //based on the list above, dump ptx files individually. Format of dumped ptx file is prog_name.unique_no.sm_<>.ptx @@ -1520,15 +1526,10 @@ void extract_code_using_cuobjdump(){ snprintf(fname,1024,"_cuobjdump_complete_output_XXXXXX"); int fd=mkstemp(fname); close(fd); - if(!g_cdp_enabled) { -#if (CUDART_VERSION >= 6000) - snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass -arch=sm_%u %s > %s", forced_max_capability, app_binary.c_str(), fname); -#else + if(!g_cdp_enabled) snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass %s > %s", app_binary.c_str(), fname); -#endif - } else { + else snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass -all %s > %s", app_binary.c_str(), fname); - } bool parse_output = true; result = system(command); if(result) { |