ptx parse only

3 files changed, 29 insertions, 61 deletions

diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc
index d03caf7..4efbd68 100644
--- a/libcuda/cuda_runtime_api.cc
+++ b/libcuda/cuda_runtime_api.cc
@@ -112,7 +112,6 @@
 #include <regex>
 #include <sstream>
 #include <fstream>
-#include <queue>
 #ifdef OPENGL_SUPPORT
 #define GL_GLEXT_PROTOTYPES
 #ifdef __APPLE__
@@ -149,7 +148,7 @@
 std::map<void *,void **> pinned_memory; //support for pinned memories added
 std::map<void *, size_t> pinned_memory_size;
 int no_of_ptx=0;
-char ptx_list_file_name[1024];
+std::map<int, std::set<std::string> > version_filename;
 std::map<int, std::string> fatbinmap;
 std::map<int, bool>fatbin_registered;
 std::map<std::string, symbol_table*> name_symtab;
@@ -1510,6 +1509,7 @@ void extract_ptx_files_using_cuobjdump(bool g_cdp_enabled){
     char command[1000];
     std::string app_binary = get_app_binary(); 
 
+    char ptx_list_file_name[1024];
     snprintf(ptx_list_file_name,1024,"_cuobjdump_list_ptx_XXXXXX");
     int fd2=mkstemp(ptx_list_file_name);
     close(fd2);
@@ -1542,6 +1542,27 @@ void extract_ptx_files_using_cuobjdump(bool g_cdp_enabled){
 	 	printf("WARNING: Number of ptx in the executable file are 0. One of the reasons might be\n");
 	 	printf("\t1. CDP is enabled\n");
 	 }
+
+    std::ifstream infile(ptx_list_file_name);
+    std::string line;
+    while (std::getline(infile, line))
+    {
+         //int pos = line.find(std::string(get_app_binary_name(app_binary)));
+         const char *ptx_file = line.c_str();
+         int pos1 = line.find("sm_");
+         int pos2 = line.find_last_of(".");
+         if (pos1==std::string::npos&&pos2==std::string::npos){
+             printf("ERROR: PTX list is not in correct format");
+             exit(0);
+         }
+         std::string vstr = line.substr(pos1+3,pos2-pos1-3);
+         int version = atoi(vstr.c_str());
+         if (version_filename.find(version)==version_filename.end()){
+            version_filename[version] = std::set<std::string>();
+         }
+         version_filename[version].insert(line);
+    }
+
 }
 
 void cuobjdumpParseBinary(unsigned int handle){
@@ -1557,28 +1578,6 @@ void cuobjdumpParseBinary(unsigned int handle){
 		return;
 	}
 
-   std::map<int, std::set<std::string> > version_filename;
-
-   std::ifstream infile(ptx_list_file_name);
-   std::string line;
-   while (std::getline(infile, line))
-   {
-        //int pos = line.find(std::string(get_app_binary_name(app_binary)));
-        const char *ptx_file = line.c_str();
-        int pos1 = line.find("sm_");
-        int pos2 = line.find_last_of(".");
-        if (pos1==std::string::npos&&pos2==std::string::npos){
-            printf("ERROR: PTX list is not in correct format");
-            exit(0);
-        }
-        std::string vstr = line.substr(pos1+3,pos2-pos1-3);
-        int version = atoi(vstr.c_str());
-        if (version_filename.find(version)==version_filename.end()){
-           version_filename[version] = std::set<std::string>();
-        }
-        version_filename[version].insert(line);
-   }
-
 	symbol_table *symtab;
    //loops through all ptx files from smallest sm version to largest
    std::map<int,std::set<std::string> >::iterator itr_m;
@@ -1671,7 +1670,6 @@ void extract_code_using_cuobjdump(){
     int result=0;
 #if (CUDART_VERSION >= 6000)
     extract_ptx_files_using_cuobjdump(g_cdp_enabled);
-    cuobjdumpParseBinary(1);
 #endif
     //TODO: redundant to dump twice. how can it be prevented?
     //dump only for specific arch
diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc
index 0348af0..97d5773 100644
--- a/src/cuda-sim/ptx_loader.cc
+++ b/src/cuda-sim/ptx_loader.cc
@@ -152,39 +152,6 @@ char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptxfilenam
 	return ptxplus_str;
 }
 
-
-symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num )
-{
-    char buf[1024];
-    snprintf(buf,1024,"_%u.ptx", source_num );
-    if( g_save_embedded_ptx ) {
-       FILE *fp = fopen(buf,"w");
-       fprintf(fp,"%s",p);
-       fclose(fp);
-    }
-    symbol_table *symtab=init_parser(buf);
-    ptx__scan_string(p);
-    int errors = ptx_parse ();
-    if ( errors ) {
-        char fname[1024];
-        snprintf(fname,1024,"_ptx_errors_XXXXXX");
-        int fd=mkstemp(fname); 
-        close(fd);
-        printf("GPGPU-Sim PTX: parser error detected, exiting... but first extracting .ptx to \"%s\"\n", fname);
-        FILE *ptxfile = fopen(fname,"w");
-        fprintf(ptxfile,"%s", p );
-        fclose(ptxfile);
-        abort();
-        exit(40);
-    }
-
-    if ( g_debug_execution >= 100 ) 
-       print_ptx_file(p,source_num,buf);
-
-    printf("GPGPU-Sim PTX: finished parsing EMBEDDED .ptx file %s\n",buf);
-    return symtab;
-}
-
 symbol_table *gpgpu_ptx_sim_load_ptx_from_filename( const char *filename )
 {
     symbol_table *symtab=init_parser(filename);
@@ -321,6 +288,10 @@ char* get_app_binary_name(){
    return self_exe_path;
 }
 
+void gpgpu_ptx_info_load_from_filename( const char *filename )
+{
+}
+
 void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version )
 {
     //do ptxas for individual files instead of one big embedded ptx. This prevents the duplicate defs and declarations.
diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc
index 06ca870..c418fac 100644
--- a/src/cuda-sim/ptx_parser.cc
+++ b/src/cuda-sim/ptx_parser.cc
@@ -142,9 +142,7 @@ symbol_table *init_parser( const char *ptx_filename )
        g_global_allfiles_symbol_table = new symbol_table("global_allfiles", 0, NULL);
        g_global_symbol_table = g_current_symbol_table = g_global_allfiles_symbol_table;
    }
-//   else {
-//       g_global_symbol_table = g_current_symbol_table = new symbol_table("global",0,g_global_allfiles_symbol_table);
-//   }
+
    ptx_lineno = 1;
 
 #define DEF(X,Y) g_ptx_token_decode[X] = Y;
@@ -164,6 +162,7 @@ symbol_table *init_parser( const char *ptx_filename )
    g_ptx_token_decode[global_space] = "global_space";
    g_ptx_token_decode[generic_space] = "generic_space";
    g_ptx_token_decode[instruction_space] = "instruction_space";
+
    init_directive_state();
    init_instruction_state();