Merge pull request #17 from ElTantawy/dev

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

3 files changed, 101 insertions, 31 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());
diff --git a/libcuda/cuobjdump.l b/libcuda/cuobjdump.l
index d7086ad..f63ee73 100644
--- a/libcuda/cuobjdump.l
+++ b/libcuda/cuobjdump.l
@@ -52,8 +52,12 @@ void cuobjdump_error(const char*);
 %s ptxcode
 %s sasscode
 %s elfcode
-%s header
+%s elfheader
+%s ptxheader
 %s endheader
+%x identifier
+%x conidentifier
+%s endidentifier
 
 alpha		[a-zA-Z]
 numeric		[0-9]
@@ -67,16 +71,20 @@ newlines	{newline}+
 "ptxasOptions"{notnewline}*{newline}
 [1-9]{numeric}* yylval.string_value = strdup(yytext); return DECIMAL;
 
+"has debug info"{newline}
+
 
 "Fatbin ptx code:"{newline}	{
 	yy_push_state(ptxcode);
-	yy_push_state(header);
+	yy_push_state(identifier);
+	yy_push_state(ptxheader);
 	yylval.string_value = strdup(yytext);
 	return PTXHEADER;
 }
 "Fatbin elf code:"{newline}	{
 	yy_push_state(elfcode);
-	yy_push_state(header);
+	yy_push_state(identifier);
+	yy_push_state(elfheader);
 	yylval.string_value = strdup(yytext);
 	return ELFHEADER;
 }
@@ -94,22 +102,50 @@ newlines	{newline}+
 	/*SASS code tokens*/
 <sasscode>{notnewline}*{newline} yylval.string_value = strdup(yytext); return SASSLINE;
 
+<identifier>{newline}"compressed"{newline}	BEGIN(conidentifier); return H_COMPRESSED;
+<identifier>{newline}"identifier = "	BEGIN(endidentifier); return H_IDENTIFIER;
+<identifier>{newline}{newline} yy_pop_state();
+
+
+<conidentifier>"identifier = "	BEGIN(endidentifier); return H_IDENTIFIER;
+<conidentifier>{newline} yy_pop_state();
+
+
+<endidentifier>{notnewline}+	yylval.string_value = strdup(yytext); yy_pop_state(); return FILENAME;
+
+
+
 	/*Header tokens*/
-<header>[[:alnum:]_]+ yylval.string_value = strdup(yytext); return IDENTIFIER;
-<header>"================"	return H_SEPARATOR;
-<header>"arch = "		return H_ARCH;
-<header>"code version = "	return H_CODEVERSION;
-<header>\[{numeric},{numeric}\]	return CODEVERSION;
-<header>"producer = "		return H_PRODUCER;
-<header>"host = "			return H_HOST;
-<header>"compile_size = "	return H_COMPILESIZE;
-<header>"compressed"{newline}
-<header>"identifier = "		BEGIN(endheader); return H_IDENTIFIER;
-<header>"has debug info"{newline}
+<elfheader>[[:alnum:]_]+ yylval.string_value = strdup(yytext); return IDENTIFIER;
+<elfheader>"================"	return H_SEPARATOR;
+<elfheader>"arch = "		return H_ARCH;
+<elfheader>"code version = "	return H_CODEVERSION;
+<elfheader>\[{numeric},{numeric}\]	return CODEVERSION;
+<elfheader>"producer = "		return H_PRODUCER;
+<elfheader>"<unknown>"			return H_UNKNOWN;
+<elfheader>"host = "			return H_HOST;
+<elfheader>"compile_size = "	BEGIN(endheader); return H_COMPILESIZE;
+
+
+	/*Header tokens*/
+<ptxheader>[[:alnum:]_]+ yylval.string_value = strdup(yytext); return IDENTIFIER;
+<ptxheader>"================"	return H_SEPARATOR;
+<ptxheader>"arch = "		return H_ARCH;
+<ptxheader>"code version = "	return H_CODEVERSION;
+<ptxheader>\[{numeric},{numeric}\]	return CODEVERSION;
+<ptxheader>"producer = "		return H_PRODUCER;
+<ptxheader>"host = "			return H_HOST;
+<ptxheader>"compile_size = "	BEGIN(endheader); return H_COMPILESIZE;
+
+
+
+
+
 
 	/* Looking for the identifier (filename) then the header is done */
 	/* <endheader>[[:alnum:]_\./]+	yylval.string_value = strdup(yytext); yy_pop_state(); return FILENAME; */
-<endheader>{notnewline}+	yylval.string_value = strdup(yytext); yy_pop_state(); return FILENAME;
+<endheader>{notnewline}+	yylval.string_value = strdup(yytext); yy_pop_state(); return IDENTIFIER;
+
 
 
 {newline}	return NEWLINE;
diff --git a/libcuda/cuobjdump.y b/libcuda/cuobjdump.y
index ec97283..dce7e3d 100644
--- a/libcuda/cuobjdump.y
+++ b/libcuda/cuobjdump.y
@@ -47,7 +47,7 @@ char filename [1024];
 %union {
 	char* string_value;
 }
-%token <string_value> H_SEPARATOR H_ARCH H_CODEVERSION H_PRODUCER H_HOST H_COMPILESIZE H_IDENTIFIER
+%token <string_value> H_SEPARATOR H_ARCH H_CODEVERSION H_PRODUCER H_HOST H_COMPILESIZE H_IDENTIFIER H_UNKNOWN H_COMPRESSED
 %token <string_value> CODEVERSION
 %token <string_value> STRING
 %token <string_value> FILENAME
@@ -61,6 +61,7 @@ char filename [1024];
 
 %%
 
+
 program :	{printf("######### cuobjdump parser ########\n");}
 			emptylines section
 		|	program section;
@@ -73,7 +74,7 @@ section :	PTXHEADER {
 				snprintf(filename, 1024, "_cuobjdump_%d.ptx", ptxserial++);
 				ptxfile = fopen(filename, "w");
 				setCuobjdumpptxfilename(filename);
-			} headerinfo ptxcode {
+			} headerinfo compressedkeyword identifier ptxcode {
 				fclose(ptxfile);
 			}
 		|	ELFHEADER {
@@ -81,7 +82,8 @@ section :	PTXHEADER {
 				snprintf(filename, 1024, "_cuobjdump_%d.elf", elfserial);
 				elffile = fopen(filename, "w");
 				setCuobjdumpelffilename(filename);
-			} headerinfo elfcode { 
+			} headerinfo identifier{
+			} elfcode {
 				fclose(elffile);
 				snprintf(filename, 1024, "_cuobjdump_%d.sass", elfserial++);
 				sassfile = fopen(filename, "w");
@@ -93,15 +95,26 @@ section :	PTXHEADER {
 headerinfo :	H_SEPARATOR NEWLINE
 				H_ARCH IDENTIFIER NEWLINE
 				H_CODEVERSION CODEVERSION NEWLINE
+				H_PRODUCER H_UNKNOWN NEWLINE
+				H_HOST IDENTIFIER NEWLINE
+				H_COMPILESIZE IDENTIFIER  {setCuobjdumparch($4);};
+			|   H_SEPARATOR NEWLINE
+				H_ARCH IDENTIFIER NEWLINE
+				H_CODEVERSION CODEVERSION NEWLINE
 				H_PRODUCER IDENTIFIER NEWLINE
 				H_HOST IDENTIFIER NEWLINE
-				H_COMPILESIZE IDENTIFIER NEWLINE
-				H_IDENTIFIER FILENAME emptylines {setCuobjdumparch($4); setCuobjdumpidentifier($19);};
+				H_COMPILESIZE IDENTIFIER {setCuobjdumparch($4);};
+
+identifier : H_IDENTIFIER FILENAME emptylines {setCuobjdumpidentifier($2);}
+			 |	{setCuobjdumpidentifier("default");};
+
+compressedkeyword : H_COMPRESSED emptylines
+                    | ;
 
 ptxcode :	ptxcode PTXLINE {fprintf(ptxfile, "%s", $2);}
 		|	;
 
-elfcode :	elfcode ELFLINE {fprintf(elffile, "%s", $2);}
+elfcode :	elfcode ELFLINE {printf(elffile, "%s", $2);}
 		|	;
 
 sasscode :	sasscode SASSLINE {fprintf(sassfile, "%s", $2);}