Merge pull request #25 from echoedit/dev

Dev

3 files changed, 31 insertions, 11 deletions

diff --git a/libcuda/cuda_api_object.h b/libcuda/cuda_api_object.h
index db5e6a4..51416f2 100644
--- a/libcuda/cuda_api_object.h
+++ b/libcuda/cuda_api_object.h
@@ -199,6 +199,7 @@ class cuda_runtime_api {
 	std::list<cuobjdumpSection*> mergeSections();
 	cuobjdumpELFSection* findELFSection(const std::string identifier);
 	cuobjdumpPTXSection* findPTXSection(const std::string identifier);
+	cuobjdumpPTXSection* findPTXSectionInList(std::list<cuobjdumpSection*> &sectionlist, const std::string identifier);
 	void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename, CUctx_st *context);
 	kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *kernel_key,
 		gpgpu_ptx_sim_arg_list_t args,
diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc
index 59d2a60..10a651a 100644
--- a/libcuda/cuda_runtime_api.cc
+++ b/libcuda/cuda_runtime_api.cc
@@ -2421,7 +2421,6 @@ __host__ cudaError_t CUDARTAPI cudaGetExportTable(const void **ppExportTable, co
 
 //extracts all ptx files from binary and dumps into prog_name.unique_no.sm_<>.ptx files
 void cuda_runtime_api::extract_ptx_files_using_cuobjdump(CUctx_st *context){
-    extern bool g_cdp_enabled;
     char command[1000];
     char *pytorch_bin = getenv("PYTORCH_BIN");
     std::string app_binary = get_app_binary(); 
@@ -2442,7 +2441,7 @@ void cuda_runtime_api::extract_ptx_files_using_cuobjdump(CUctx_st *context){
         printf("WARNING: Failed to execute cuobjdump to get list of ptx files \n");
         exit(0);
     }   
-    if(!g_cdp_enabled) {
+    if(!gpgpu_ctx->device_runtime->g_cdp_enabled) {
         //based on the list above, dump ptx files individually. Format of dumped ptx file is prog_name.unique_no.sm_<>.ptx
 
        std::ifstream infile(ptx_list_file_name);
@@ -2515,7 +2514,6 @@ void cuda_runtime_api::extract_code_using_cuobjdump(){
     }
     // Running cuobjdump using dynamic link to current process
     // Needs the option '-all' to extract PTX from CDP-enabled binary 
-    extern bool g_cdp_enabled;
 
     //dump ptx for all individial ptx files into sepearte files which is later used by ptxas.
     int result=0;
@@ -2530,7 +2528,7 @@ void cuda_runtime_api::extract_code_using_cuobjdump(){
 	snprintf(fname,1024,"_cuobjdump_complete_output_XXXXXX");
 	int fd=mkstemp(fname);
 	close(fd);
-	if(!g_cdp_enabled)
+	if(!gpgpu_ctx->device_runtime->g_cdp_enabled)
             snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass %s > %s", app_binary.c_str(), fname);
 	else
             snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass -all %s > %s", app_binary.c_str(), fname);
@@ -2822,7 +2820,7 @@ cuobjdumpELFSection* cuda_runtime_api::findELFSection(const std::string identifi
 }
 
 //! Within the section list, find the PTX section corresponding to a given identifier
-cuobjdumpPTXSection* findPTXSectionInList(std::list<cuobjdumpSection*> &sectionlist, const std::string identifier){
+cuobjdumpPTXSection* cuda_runtime_api::findPTXSectionInList(std::list<cuobjdumpSection*> &sectionlist, const std::string identifier){
 	std::list<cuobjdumpSection*>::iterator iter;
 	for (	iter = sectionlist.begin();
 			iter != sectionlist.end();
@@ -2833,8 +2831,7 @@ cuobjdumpPTXSection* findPTXSectionInList(std::list<cuobjdumpSection*> &sectionl
 			if(ptxsection->getIdentifier() == identifier)
 				return ptxsection;
 			else {
-				extern bool g_cdp_enabled;
-				if(g_cdp_enabled) {
+				if(gpgpu_ctx->device_runtime->g_cdp_enabled) {
 					printf("Warning: __cudaRegisterFatBinary needs %s, but find PTX section with %s\n",
 						identifier.c_str(), ptxsection->getIdentifier().c_str());
 					return ptxsection;
diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h
index 3c9f87c..45c5cdd 100644
--- a/libcuda/gpgpu_context.h
+++ b/libcuda/gpgpu_context.h
@@ -6,32 +6,51 @@
 #include "../src/gpgpusim_entrypoint.h"
 #include "../src/cuda-sim/cuda-sim.h"
 #include "../src/cuda-sim/cuda_device_runtime.h"
+#include "../src/cuda-sim/ptx-stats.h"
 
 class gpgpu_context {
     public:
 	gpgpu_context() {
 	    g_global_allfiles_symbol_table = NULL;
+	    sm_next_access_uid=0;
+	    warp_inst_sm_next_uid=0;
+	    operand_info_sm_next_uid = 1;
+	    kernel_info_m_next_uid = 1;
+	    g_num_ptx_inst_uid = 0;
+	    g_ptx_cta_info_uid = 1;
+	    symbol_sm_next_uid = 1;
+	    function_info_sm_next_uid = 1;
+	    debug_tensorcore = 0;
 	    api = new cuda_runtime_api(this);
 	    ptxinfo = new ptxinfo_data(this);
 	    ptx_parser = new ptx_recognizer(this);
 	    the_gpgpusim = new GPGPUsim_ctx(this);
 	    func_sim = new cuda_sim(this);
-#if (CUDART_VERSION >= 5000)
 	    device_runtime = new cuda_device_runtime(this);
-#endif
+	    stats = new ptx_stats(this);
 	}
 	// global list
 	symbol_table *g_global_allfiles_symbol_table;
 	const char *g_filename;
+	unsigned sm_next_access_uid;
+	unsigned warp_inst_sm_next_uid;
+	unsigned operand_info_sm_next_uid;//uid for operand_info
+	unsigned kernel_info_m_next_uid;//uid for kernel_info_t
+	unsigned g_num_ptx_inst_uid; //uid for ptx inst inside ptx_instruction
+	unsigned long long g_ptx_cta_info_uid;
+	unsigned symbol_sm_next_uid; //uid for symbol
+	unsigned function_info_sm_next_uid;
+	std::vector<ptx_instruction*> s_g_pc_to_insn; // a direct mapping from PC to instruction
+	bool debug_tensorcore;
+
 	// objects pointers for each file
 	cuda_runtime_api* api;
 	ptxinfo_data* ptxinfo;
 	ptx_recognizer* ptx_parser;
 	GPGPUsim_ctx* the_gpgpusim;
 	cuda_sim* func_sim;
-#if (CUDART_VERSION >= 5000)
 	cuda_device_runtime* device_runtime;
-#endif
+	ptx_stats* stats;
 	// member function list
 	void cuobjdumpParseBinary(unsigned int handle);
 	class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num );
@@ -43,6 +62,9 @@ class gpgpu_context {
 	class gpgpu_sim *gpgpu_ptx_sim_init_perf();
 	struct _cuda_device_id *GPGPUSim_Init();
 	void ptx_reg_options(option_parser_t opp);
+	const ptx_instruction* pc_to_instruction(unsigned pc); 
+	const warp_inst_t *ptx_fetch_inst( address_type pc );
+	unsigned translate_pc_to_ptxlineno(unsigned pc);
 };
 gpgpu_context* GPGPU_Context();