Merge pull request #17 from ElTantawy/dev

Initial support for 7.5 - tested on simple sdk benchmaks (e.g., template, vectorAdd, ...)

1 files changed, 31 insertions, 10 deletions

diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc
index eed4017..910bebd 100644
--- a/libcuda/cuda_runtime_api.cc
+++ b/libcuda/cuda_runtime_api.cc
@@ -1457,18 +1457,22 @@ std::list<cuobjdumpSection*> pruneSectionList(std::list<cuobjdumpSection*> cuobj
 
 	std::list<cuobjdumpSection*> prunedList;
 
-	//Find the highest capability (that is lower than the forces maximum) for each cubin file
+	//Find the highest capability (that is lower than the forced maximum) for each cubin file
 	//and set it in cuobjdumpSectionMap. Do this only for ptx sections
 	std::map<std::string, unsigned> cuobjdumpSectionMap;
+	int min_ptx_capability_found=0;
 	for (	std::list<cuobjdumpSection*>::iterator iter = cuobjdumpSectionList.begin();
 			iter != cuobjdumpSectionList.end();
 			iter++){
 		unsigned capability = (*iter)->getArch();
-		if(dynamic_cast<cuobjdumpPTXSection*>(*iter) != NULL &&
-				(capability <= forced_max_capability ||
-						forced_max_capability==0)) {
-			if(cuobjdumpSectionMap[(*iter)->getIdentifier()] < capability)
-				cuobjdumpSectionMap[(*iter)->getIdentifier()] = capability;
+		if(dynamic_cast<cuobjdumpPTXSection*>(*iter) != NULL){
+			if(capability<min_ptx_capability_found || min_ptx_capability_found==0)
+				min_ptx_capability_found=capability;
+			if (capability <= forced_max_capability ||	forced_max_capability==0) {
+				if((cuobjdumpSectionMap.find((*iter)->getIdentifier())==cuobjdumpSectionMap.end())
+						|| (cuobjdumpSectionMap[(*iter)->getIdentifier()] < capability))
+						cuobjdumpSectionMap[(*iter)->getIdentifier()] = capability;
+			}
 		}
 	}
 
@@ -1484,6 +1488,10 @@ std::list<cuobjdumpSection*> pruneSectionList(std::list<cuobjdumpSection*> cuobj
 			delete *iter;
 		}
 	}
+	if(prunedList.empty()){
+		printf("Error: No PTX sections found with sm capability that is lower than current forced maximum capability \n minimum ptx capability found = %u, maximum forced ptx capability = %u \n User might want to change either the forced maximum capability from gpgpusim configuration or update the compilation to generate the required PTX version\n",min_ptx_capability_found,forced_max_capability);
+		abort();
+	}
 	return prunedList;
 }
 
@@ -1567,6 +1575,16 @@ void cuobjdumpParseBinary(unsigned int handle){
 	fatbin_registered[handle] = true;
 	CUctx_st *context = GPGPUSim_Context();
 
+	unsigned max_capability = 0;
+	for (	std::list<cuobjdumpSection*>::iterator iter = cuobjdumpSectionList.begin();
+			iter != cuobjdumpSectionList.end();
+			iter++){
+		unsigned capability = (*iter)->getArch();
+		if (capability > max_capability) max_capability = capability;
+	}
+
+	if (max_capability > 20) printf("WARNING: No guarantee that PTX will be parsed for SM version %u\n", max_capability);
+
 	std::string fname = fatbinmap[handle];
 	cuobjdumpPTXSection* ptx = findPTXSection(fname);
 
@@ -1589,13 +1607,13 @@ void cuobjdumpParseBinary(unsigned int handle){
 		symtab=gpgpu_ptx_sim_load_ptx_from_string(ptxplus_str, handle);
 		printf("Adding %s with cubin handle %u\n", ptx->getPTXfilename().c_str(), handle);
 		context->add_binary(symtab, handle);
-		gpgpu_ptxinfo_load_from_string( ptxcode, handle);
+		gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability );
 		delete[] ptxplus_str;
 	} else {
 		symtab=gpgpu_ptx_sim_load_ptx_from_string(ptxcode, handle);
 		printf("Adding %s with cubin handle %u\n", ptx->getPTXfilename().c_str(), handle);
 		context->add_binary(symtab, handle);
-		gpgpu_ptxinfo_load_from_string( ptxcode, handle);
+		gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability );
 	}
 	load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu());
 	load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu());
@@ -1624,10 +1642,13 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin )
 		// - 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. 
+		#if (CUDART_VERSION <= 6000)
 		char * pfatbin = (char*) fatDeviceText->d; 
 		int offset = *((int*)(pfatbin+48)); 
 		char * filename = (pfatbin+16+offset); 
-
+		#else
+		char * filename = "default";
+		#endif
 		// 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
@@ -1693,7 +1714,7 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin )
 			} else {
 				symtab=gpgpu_ptx_sim_load_ptx_from_string(ptx,source_num);
 				context->add_binary(symtab,fat_cubin_handle);
-				gpgpu_ptxinfo_load_from_string( ptx, source_num );
+				gpgpu_ptxinfo_load_from_string( ptx, source_num, max_capability );
 			}
 			source_num++;
 			load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu());