diff options
Diffstat (limited to 'src/cuda-sim/ptx_loader.cc')
| -rw-r--r-- | src/cuda-sim/ptx_loader.cc | 50 |
1 files changed, 33 insertions, 17 deletions
diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc index f5a1c77..8c6f361 100644 --- a/src/cuda-sim/ptx_loader.cc +++ b/src/cuda-sim/ptx_loader.cc @@ -55,6 +55,7 @@ extern int ptxinfo_debug; extern FILE *ptxinfo_in; static bool g_save_embedded_ptx; +static int g_occupancy_sm_number; bool g_keep_intermediate_files; bool m_ptx_save_converted_ptxplus; @@ -72,6 +73,10 @@ void ptx_reg_options(option_parser_t opp) &m_ptx_save_converted_ptxplus, "Saved converted ptxplus to a file", "0"); + option_parser_register(opp, "-gpgpu_occupancy_sm_number", OPT_INT32, &g_occupancy_sm_number, + "The SM number to pass to ptxas when getting register usage for computing GPU occupancy. " + "This parameter is required in the config.", + "0"); } void print_ptx_file( const char *p, unsigned source_num, const char *filename ) @@ -121,7 +126,7 @@ char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptxfilenam fflush(stdout); printf("GPGPU-Sim PTX: calling cuobjdump_to_ptxplus\ncommandline: %s\n", commandline); result = system(commandline); - if(result){printf("GPGPU-Sim PTX: ERROR ** could not execute %s\n", commandline); exit(1);} + if(result){fprintf(stderr, "GPGPU-Sim PTX: ERROR ** could not execute %s\n", commandline); exit(1);} // Get ptxplus from file @@ -143,7 +148,7 @@ char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptxfilenam printf("GPGPU-Sim PTX: removing temporary files using \"%s\"\n", rm_commandline); int rm_result = system(rm_commandline); if( rm_result != 0 ) { - printf("GPGPU-Sim PTX: ERROR ** while removing temporary files %d\n", rm_result); + fprintf(stderr, "GPGPU-Sim PTX: ERROR ** while removing temporary files %d\n", rm_result); exit(1); } } @@ -152,6 +157,7 @@ char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptxfilenam return ptxplus_str; } + symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ) { char buf[1024]; @@ -200,7 +206,7 @@ void fix_duplicate_errors(char fname2[1024]) { printf("Running: %s\n", commandline); int result = system(commandline); if (result != 0) { - printf("GPGPU-Sim PTX: ERROR ** while changing filename from %s to %s", fname2, tempfile); + fprintf(stderr, "GPGPU-Sim PTX: ERROR ** while changing filename from %s to %s", fname2, tempfile); exit(1); } @@ -289,11 +295,12 @@ void fix_duplicate_errors(char fname2[1024]) { printf("Running: %s\n", commandline); result = system(commandline); if (result != 0) { - printf("GPGPU-Sim PTX: ERROR ** while deleting %s", tempfile); + fprintf(stderr, "GPGPU-Sim PTX: ERROR ** while deleting %s", tempfile); exit(1); } } + //we need the application name here too. char* get_app_binary_name(){ char exe_path[1025]; @@ -323,7 +330,7 @@ char* get_app_binary_name(){ void gpgpu_ptx_info_load_from_filename( const char *filename, unsigned sm_version) { std::string ptxas_filename(std::string(filename) + "as"); - char buff[1024], extra_flags[256]; + char buff[1024], extra_flags[1024]; extra_flags[0]=0; extern bool g_cdp_enabled; if(!g_cdp_enabled) @@ -382,19 +389,28 @@ void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num char extra_flags[1024]; extra_flags[0]=0; - #if CUDART_VERSION >= 3000 - if (sm_version == 0) sm_version = 20; - snprintf(extra_flags,1024,"--gpu-name=sm_%u",sm_version); - #endif +#if CUDART_VERSION >= 3000 + if ( g_occupancy_sm_number == 0 ) { + fprintf( stderr, "gpgpusim.config must specify the sm version for the GPU that you use to compute occupancy \"-gpgpu_occupancy_sm_number XX\".\n" + "The register file size is specifically tied to the sm version used to querry ptxas for register usage.\n" + "A register size/SM mismatch may result in occupancy differences." ); + exit(1); + } + extern bool g_cdp_enabled; + if(!g_cdp_enabled) + snprintf(extra_flags,1024,"--gpu-name=sm_%u", g_occupancy_sm_number); + else + snprintf(extra_flags,1024,"--compile-only --gpu-name=sm_%u",g_occupancy_sm_number); +#endif - snprintf(commandline,1024,"$CUDA_INSTALL_PATH/bin/ptxas %s -v %s --output-file /dev/null 2> %s", + snprintf(commandline,1024,"$PTXAS_CUDA_INSTALL_PATH/bin/ptxas %s -v %s --output-file /dev/null 2> %s", extra_flags, fname2, tempfile_ptxinfo); - printf("GPGPU-Sim PTX: generating ptxinfo using \"%s\"\n", commandline); - result = system(commandline); - if( result != 0 ) { - // 65280 = duplicate errors - if (result == 65280) { - ptxinfo_in = fopen(tempfile_ptxinfo,"r"); + printf("GPGPU-Sim PTX: generating ptxinfo using \"%s\"\n", commandline); + result = system(commandline); + if( result != 0 ) { + // 65280 = duplicate errors + if (result == 65280) { + ptxinfo_in = fopen(tempfile_ptxinfo,"r"); g_ptxinfo_filename = tempfile_ptxinfo; ptxinfo_parse(); @@ -499,4 +515,4 @@ void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num exit(1); } } -} +}
\ No newline at end of file |