Integration from TM-311 branch.

- Updated PTX parser to support CUDA 4.1 and 4.2.
  - Revised fatbin workaround to a more robust version (with comments explaining it).
  - Added print_simulation_time() in gpgpu_sim_thread_concurrent().

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 13789]

1 files changed, 30 insertions, 12 deletions

diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc
index 6bbd2f2..58ffd4d 100644
--- a/libcuda/cuda_runtime_api.cc
+++ b/libcuda/cuda_runtime_api.cc
@@ -1387,7 +1387,14 @@ void cuobjdumpRegisterFatBinary(unsigned int handle, char* filename){
 	cuobjdumpPTXSection* ptx = findptxsection(cuobjdumpSectionList, fname);
 
 	symbol_table *symtab;
-	char *ptxcode = readfile(ptx->getPTXfilename());
+	char *ptxcode;
+	const char *override_ptx_name = getenv("PTX_SIM_KERNELFILE"); 
+   if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL) {
+		ptxcode = readfile(ptx->getPTXfilename());
+	} else {
+		printf("GPGPU-Sim PTX: overriding embedded ptx with '%s' (PTX_SIM_USE_PTX_FILE is set)\n", override_ptx_name);
+		ptxcode = readfile(override_ptx_name);
+	}
 	if(context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() ) {
 		cuobjdumpELFSection* elfsection = findelfsection(cuobjdumpSectionList, ptx->getIdentifier());
 		assert (elfsection!= NULL);
@@ -1491,17 +1498,28 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin )
 	CUctx_st *context = GPGPUSim_Context();
 	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
+		// The following workaround has only been verified on 64-bit systems. 
+		if (sizeof(void*) == 4) 
+			printf("GPGPU-Sim PTX: FatBin file name extraction has not been tested on 32-bit system.\n"); 
+
+		// FatBin handle from the .fatbin.c file (one of the intermediate files generated by NVCC)
+		typedef struct {int m; int v; const unsigned long long* d; char* f;} __fatDeviceText __attribute__ ((aligned (8))); 
+		__fatDeviceText * fatDeviceText = (__fatDeviceText *) fatCubin;
+
+		// Extract the source code file name that generate the given FatBin. 
+		// - Obtains the pointer to the actual fatbin structure from the FatBin handle (fatCubin).
+		// - An integer inside the fatbin structure contains the relative offset to the source code file name.
+		// - This offset differs among different CUDA and GCC versions. 
+		char * pfatbin = (char*) fatDeviceText->d; 
+		int offset = *((int*)(pfatbin+48)); 
+		char * filename = (pfatbin+16+offset); 
+
+		// The extracted file name is associated with a fat_cubin_handle passed
+		// into cudaLaunch().  Inside cudaLaunch(), the associated file name is
+		// used to find the PTX/SASS section from cuobjdump, which contains the
+		// PTX/SASS code for the launched kernel function.  
+		// This allows us to work around the fact that cuobjdump only outputs the
+		// file name associated with each section. 
 		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);