diff options
| author | Wilson Fung <[email protected]> | 2012-08-13 04:19:58 -0800 |
|---|---|---|
| committer | Andrew Boktor <[email protected]> | 2014-08-14 13:48:53 -0700 |
| commit | 8f2b4a6f61ef30d86b9bae7c92ed62f223550a76 (patch) | |
| tree | a1cbb71be63ae88e4fd6ef1b27fd9231e0f97c19 | |
| parent | 5009c035ab4b201e46af58feb4a231483a0c948e (diff) | |
Extended andrew's fatbin hack to support CUDA 4.1 and 4.2 as well.
[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 13744]
| -rw-r--r-- | libcuda/cuda_runtime_api.cc | 9 |
1 files changed, 9 insertions, 0 deletions
diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index c08bb0a..6bbd2f2 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1492,7 +1492,16 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) static unsigned next_fat_bin_handle = 1; if(context->get_device()->get_gpgpu()->get_config().use_cuobjdump()) { char * s1 = *((char**)((char*)fatCubin+8)); +#if (CUDART_VERSION < 4010) char * filename = (s1+72); +#else + // for CUDA 4.1 and 4.2, the source file name can be found at offset 80, + // or offset 96 if the source file has been compiled with a ptxas option + char * filename = (s1+80); + if (strncmp(filename, "-dlcm=", 6) == 0) { + filename += 16; + } +#endif unsigned fat_cubin_handle = next_fat_bin_handle; next_fat_bin_handle++; printf("GPGPU-Sim PTX: __cudaRegisterFatBinary, fat_cubin_handle = %u, filename=%s\n", fat_cubin_handle, filename); |