Supporting the option for removing temporary files

  . the option -gpgpu_ptx_save_converted_ptxplus allows keeping the ptxplus file
  . the option -gpgpu_keep allows keeping intermediate files used to communicate with other programs (e.g. cuobjdump)

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 12304]

5 files changed, 46 insertions, 27 deletions

diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc
index e9744d0..85664a7 100644
--- a/libcuda/cuda_runtime_api.cc
+++ b/libcuda/cuda_runtime_api.cc
@@ -1159,10 +1159,15 @@ extern "C" FILE *cuobjdump_in;
 /*!
  *	This Function extract the whole PTX (for all the files) using cuobjdump
  * */
-void extract_ptx(){
+void extract_code_using_cuobjdump(){
 	char command[1000];
+
+	char fname[1024];
+	snprintf(fname,1024,"_cuobjdump_complete_output_XXXXXX");
+	int fd=mkstemp(fname);
+	close(fd);
+
 	char* whole_code;
-	char fname[1024]="_ptx_whole_code_WESWW";
 	//! Running CUobjdump using dynamic link to current process
 	snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass /proc/%d/exe > %s",getpid(),fname);
 	printf("Running cuobjdump using \"%s\"\n", command);
@@ -1177,10 +1182,10 @@ void extract_ptx(){
 	printf("Done parsing!!!\n");
 }
 
-//! Return proper ptx version
+//! Find the proper sm version to be used
 /*!
- * This function return newest ptx version inside the argument
- * which is is not newer than forced_max_capability
+ * The function finds the highest sm version lower than the option
+ * force_max_capability
  */
 unsigned get_best_version(std::list<cuobjdumpSection> sectionlist, CUctx_st *context){
 
@@ -1209,9 +1214,10 @@ unsigned get_best_version(std::list<cuobjdumpSection> sectionlist, CUctx_st *con
 	return max_capability;
 }
 
+//! Find number of files with a certain sm version
 /*!
- * Return number of diffrent ptx files generated in first argument
- *  which have sm version equal to selected_capability
+ * Once the sm verison to be used is known (using get_best_version), this
+ * function counts the number of files that use this sm version
  */
 unsigned get_number_of_ptx(std::list<cuobjdumpSection> sectionlist, unsigned selected_capability){
 	int result = 0;
@@ -1275,7 +1281,7 @@ void useCuobjdump() {
 	CUctx_st *context = GPGPUSim_Context();
 	unsigned source_num=1;
 	char *sass, *elf;
-	extract_ptx(); //extract all the output of cuobjdump to _cuobjdump_*.*
+	extract_code_using_cuobjdump(); //extract all the output of cuobjdump to _cuobjdump_*.*
 	unsigned selected_capability = get_best_version(cuobjdump, context); //Find max capability less than forced_max_capability
 	unsigned total_ptx_files = get_number_of_ptx(cuobjdump,selected_capability); //Count ptx files for the given capability
 
@@ -1311,6 +1317,19 @@ void useCuobjdump() {
 			load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu());
 		}
 	}
+	if(!keep_intermediate_files()){
+		char rm_commandline[1024];
+
+		snprintf(rm_commandline,1024,"rm -f _cuobjdump_*");
+
+		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);
+			exit(1);
+		}
+	}
+
 }
 
 void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin )
@@ -1326,8 +1345,8 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin )
 		next_fat_bin_handle++;
 		printf("GPGPU-Sim PTX: __cudaRegisterFatBinary, fat_cubin_handle = %u\n", fat_cubin_handle);
 		/*!
-		 * This function extracting all data from all files in first call
-		 * then for next calls, only return the appropriate number
+		 * This function extracts all data from all files in first call
+		 * then for next calls, only returns the appropriate number
 		 */
 		assert(fat_cubin_handle >= 1);
 		if(fat_cubin_handle == 1)
diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc
index 6457edd..01899d5 100644
--- a/src/abstract_hardware_model.cc
+++ b/src/abstract_hardware_model.cc
@@ -62,10 +62,6 @@ void gpgpu_functional_sim_config::reg_options(class OptionParser * opp)
                  &m_ptx_convert_to_ptxplus,
                  "Convert SASS (native ISA) to ptxplus and run ptxplus",
                  "0");
-    option_parser_register(opp, "-gpgpu_ptx_save_converted_ptxplus", OPT_BOOL,
-                 &m_ptx_save_converted_ptxplus,
-                 "Saved converted ptxplus to a file",
-                 "0");
     option_parser_register(opp, "-gpgpu_ptx_force_max_capability", OPT_UINT32,
                  &m_ptx_force_max_capability,
                  "Force maximum compute capability",
diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index 28dc6cb..7a710f8 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -338,7 +338,6 @@ public:
     unsigned get_forced_max_capability() const { return m_ptx_force_max_capability; }
     bool convert_to_ptxplus() const { return m_ptx_convert_to_ptxplus; }
     bool use_cuobjdump() const { return m_ptx_use_cuobjdump; }
-    bool saved_converted_ptxplus() const { return m_ptx_save_converted_ptxplus; }
 
     int         get_ptx_inst_debug_to_file() const { return g_ptx_inst_debug_to_file; }
     const char* get_ptx_inst_debug_file() const  { return g_ptx_inst_debug_file; }
@@ -348,7 +347,6 @@ public:
 private:
     // PTX options
     int m_ptx_convert_to_ptxplus;
-    int m_ptx_save_converted_ptxplus;
     int m_ptx_use_cuobjdump;
     unsigned m_ptx_force_max_capability;
 
diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc
index cbfa9f4..d405288 100644
--- a/src/cuda-sim/ptx_loader.cc
+++ b/src/cuda-sim/ptx_loader.cc
@@ -53,6 +53,9 @@ extern "C" FILE *ptxinfo_in;
 
 static bool g_save_embedded_ptx;
 bool g_keep_intermediate_files;
+bool m_ptx_save_converted_ptxplus;
+
+bool keep_intermediate_files() {return g_keep_intermediate_files;}
 
 void ptx_reg_options(option_parser_t opp)
 {
@@ -62,6 +65,10 @@ void ptx_reg_options(option_parser_t opp)
    option_parser_register(opp, "-keep", OPT_BOOL, &g_keep_intermediate_files, 
                 "keep intermediate files created by GPGPU-Sim when interfacing with external programs",
                 "0");
+   option_parser_register(opp, "-gpgpu_ptx_save_converted_ptxplus", OPT_BOOL,
+                &m_ptx_save_converted_ptxplus,
+                "Saved converted ptxplus to a file",
+                "0");
 }
 
 void print_ptx_file( const char *p, unsigned source_num, const char *filename )
@@ -125,20 +132,18 @@ char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const char *ptxfilename, con
 	char* ptxplus_str = new char [strlen(text.c_str())+1];
 	strcpy(ptxplus_str, text.c_str());
 
-	// Remove temporary files
-	// TODO: Fix deleting/keeping PTXPlus files
-	/*
-	char rm_commandline[1024];
+	if (!m_ptx_save_converted_ptxplus){
+		char rm_commandline[1024];
 
-	snprintf(rm_commandline,1024,"rm -f %s", fname_ptxplus);
+		snprintf(rm_commandline,1024,"rm -f %s", fname_ptxplus);
 
-	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);
-	   exit(1);
+		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);
+			exit(1);
+		}
 	}
-	*/
 	printf("GPGPU-Sim PTX: DONE converting EMBEDDED .ptx file to ptxplus \n");
 
 	return ptxplus_str;
diff --git a/src/cuda-sim/ptx_loader.h b/src/cuda-sim/ptx_loader.h
index 9227d85..21974e6 100644
--- a/src/cuda-sim/ptx_loader.h
+++ b/src/cuda-sim/ptx_loader.h
@@ -33,5 +33,6 @@ extern bool g_override_embedded_ptx;
 class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num );
 void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num );
 char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const char *ptx_str, const char *sass_str, const char *elf_str);
+bool keep_intermediate_files();
 
 #endif