Integration from TM-311 branch.

- Updated PTX parser to support CUDA 4.1 and 4.2.
  - Revised fatbin workaround to a more robust version (with comments explaining it).
  - Added print_simulation_time() in gpgpu_sim_thread_concurrent().

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 13789]

6 files changed, 49 insertions, 23 deletions

diff --git a/CHANGES b/CHANGES
index 33ceee2..f077145 100644
--- a/CHANGES
+++ b/CHANGES
@@ -1,20 +1,25 @@
 CHANGE LOG:
 
 Version 3.1.1+edits (development branch) versus 3.1.1
+- Added support for CUDA 4.1 and 4.2 (PTX 3.0). 
 - Added an improved and less brittle support for associating fatbins to
   sections from the output of cuobjdump
+- Added raw simulation time printout back for CUDA.  
+- Revived support for PTX file override via environment variables for CUDA 4.0
+  onwards. 
 - Bug Fixes:
     - Fixed a compile error that happens with newer gcc/g++ versions (4.7.1)
-    - Fixed a bug with the association between cuobjdump output and cubin 
+    - Fixed a bug with the association between cuobjdump output and cubin
       handles (Bug #7 external)
     - Adding support for atomic operations with generic memory space.  Before
       this, atomic operations can only work on global memory space.  (Bug #14
       external)
-    - Fixed a bug in conversion to ptxplus that was causing local memory store 
+    - Fixed a bug in conversion to ptxplus that was causing local memory store
       or load instruction to be ignored.
     - Adding support for direct addressing using immediate values for the load
       and store instructions.
-    - Fixed a bug that was causing inconsistency in local memory address calculations
+    - Fixed a bug that was causing inconsistency in local memory address
+      calculations
 
 Version 3.1.1 versus 3.1.0
 - Add checks to top level makefile to ensure setup_environment is run and checks to
diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc
index 6bbd2f2..58ffd4d 100644
--- a/libcuda/cuda_runtime_api.cc
+++ b/libcuda/cuda_runtime_api.cc
@@ -1387,7 +1387,14 @@ void cuobjdumpRegisterFatBinary(unsigned int handle, char* filename){
 	cuobjdumpPTXSection* ptx = findptxsection(cuobjdumpSectionList, fname);
 
 	symbol_table *symtab;
-	char *ptxcode = readfile(ptx->getPTXfilename());
+	char *ptxcode;
+	const char *override_ptx_name = getenv("PTX_SIM_KERNELFILE"); 
+   if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL) {
+		ptxcode = readfile(ptx->getPTXfilename());
+	} else {
+		printf("GPGPU-Sim PTX: overriding embedded ptx with '%s' (PTX_SIM_USE_PTX_FILE is set)\n", override_ptx_name);
+		ptxcode = readfile(override_ptx_name);
+	}
 	if(context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() ) {
 		cuobjdumpELFSection* elfsection = findelfsection(cuobjdumpSectionList, ptx->getIdentifier());
 		assert (elfsection!= NULL);
@@ -1491,17 +1498,28 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin )
 	CUctx_st *context = GPGPUSim_Context();
 	static unsigned next_fat_bin_handle = 1;
 	if(context->get_device()->get_gpgpu()->get_config().use_cuobjdump()) {
-		char * s1 = *((char**)((char*)fatCubin+8));
-#if (CUDART_VERSION < 4010)
-		char * filename = (s1+72);
-#else
-		// for CUDA 4.1 and 4.2, the source file name can be found at offset 80, 
-		// or offset 96 if the source file has been compiled with a ptxas option 
-		char * filename = (s1+80);
-		if (strncmp(filename, "-dlcm=", 6) == 0) {
-			filename += 16; 
-		}
-#endif
+		// The following workaround has only been verified on 64-bit systems. 
+		if (sizeof(void*) == 4) 
+			printf("GPGPU-Sim PTX: FatBin file name extraction has not been tested on 32-bit system.\n"); 
+
+		// FatBin handle from the .fatbin.c file (one of the intermediate files generated by NVCC)
+		typedef struct {int m; int v; const unsigned long long* d; char* f;} __fatDeviceText __attribute__ ((aligned (8))); 
+		__fatDeviceText * fatDeviceText = (__fatDeviceText *) fatCubin;
+
+		// Extract the source code file name that generate the given FatBin. 
+		// - 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. 
+		char * pfatbin = (char*) fatDeviceText->d; 
+		int offset = *((int*)(pfatbin+48)); 
+		char * filename = (pfatbin+16+offset); 
+
+		// 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
+		// PTX/SASS code for the launched kernel function.  
+		// This allows us to work around the fact that cuobjdump only outputs the
+		// file name associated with each section. 
 		unsigned fat_cubin_handle = next_fat_bin_handle;
 		next_fat_bin_handle++;
 		printf("GPGPU-Sim PTX: __cudaRegisterFatBinary, fat_cubin_handle = %u, filename=%s\n", fat_cubin_handle, filename);
diff --git a/src/cuda-sim/ptx.l b/src/cuda-sim/ptx.l
index 4747181..2d0e065 100644
--- a/src/cuda-sim/ptx.l
+++ b/src/cuda-sim/ptx.l
@@ -66,8 +66,8 @@ bfind   TC; ptx_lval.int_value = BFIND_OP; return OPCODE;
 bra     TC; ptx_lval.int_value = BRA_OP; return OPCODE;
 brev    TC; ptx_lval.int_value = BREV_OP; return OPCODE;
 brkpt   TC; ptx_lval.int_value = BRKPT_OP; return OPCODE;
-call	TC; ptx_lval.int_value = CALL_OP; return OPCODE;
-callp    TC; ptx_lval.int_value = CALLP_OP; return OPCODE;
+call	TC; BEGIN(NOT_OPCODE); ptx_lval.int_value = CALL_OP; return OPCODE; // blocking opcode token in case the callee has the same name as an opcode
+callp    TC; BEGIN(NOT_OPCODE); ptx_lval.int_value = CALLP_OP; return OPCODE;
 clz	TC; ptx_lval.int_value = CLZ_OP; return OPCODE;
 cnot	TC; ptx_lval.int_value = CNOT_OP; return OPCODE;
 cos	TC; ptx_lval.int_value = COS_OP; return OPCODE;
@@ -155,7 +155,7 @@ breakaddr  TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE;
 \.entry TC; return ENTRY_DIRECTIVE;
 \.extern TC; return EXTERN_DIRECTIVE;
 \.file	 TC; BEGIN(INITIAL); return FILE_DIRECTIVE;
-\.func   TC; BEGIN(NOT_OPCODE); return FUNC_DIRECTIVE;
+\.func   TC; BEGIN(NOT_OPCODE); return FUNC_DIRECTIVE; // blocking opcode parsing in case the function has the same name as an opcode (e.g. sin(), cos())
 \.global TC; return GLOBAL_DIRECTIVE;
 \.local  TC; return LOCAL_DIRECTIVE;
 \.loc    TC; return LOC_DIRECTIVE;
@@ -352,7 +352,7 @@ breakaddr  TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE;
 ";"	TC; BEGIN(INITIAL); return SEMI_COLON;
 "!"	TC; return EXCLAMATION;
 "=" 	TC; return EQUALS;
-"{"	TC; return LEFT_BRACE;
+"{"	TC; BEGIN(INITIAL); return LEFT_BRACE; // starting a statement block (allow next token to be parsed as an opcode)
 "}"	TC; return RIGHT_BRACE;
 \.	TC; return PERIOD;
 "/"	TC; return BACKSLASH;
diff --git a/src/cuda-sim/ptx.y b/src/cuda-sim/ptx.y
index 2c01a70..5b69975 100644
--- a/src/cuda-sim/ptx.y
+++ b/src/cuda-sim/ptx.y
@@ -259,6 +259,7 @@ directive_statement: variable_declaration SEMI_COLON
 	| FILE_DIRECTIVE INT_OPERAND STRING { add_file($2,$3); } 
 	| LOC_DIRECTIVE INT_OPERAND INT_OPERAND INT_OPERAND 
 	| PRAGMA_DIRECTIVE STRING SEMI_COLON { add_pragma($2); }
+	| function_decl SEMI_COLON {/*Do nothing*/}
 	;
 
 variable_declaration: variable_spec identifier_list { add_variables(); }
diff --git a/src/cuda-sim/ptxinfo.y b/src/cuda-sim/ptxinfo.y
index 8bc1306..e4ba6b2 100644
--- a/src/cuda-sim/ptxinfo.y
+++ b/src/cuda-sim/ptxinfo.y
@@ -51,7 +51,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 %token SEMICOLON
 %token QUOTE
 %token LINE
-%token WARNING
+%token <string_value> WARNING
 %token FOR
 
 %{
@@ -78,6 +78,7 @@ input:	/* empty */
 
 line: 	HEADER INFO COLON line_info
 	| HEADER IDENTIFIER COMMA LINE INT_OPERAND SEMICOLON WARNING
+	| HEADER WARNING { printf("GPGPU-Sim: ptxas %s\n", $2); }
 	;
 
 line_info: function_name
diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc
index 006cecd..6271f79 100644
--- a/src/gpgpusim_entrypoint.cc
+++ b/src/gpgpusim_entrypoint.cc
@@ -132,9 +132,10 @@ void *gpgpu_sim_thread_concurrent(void*)
            printf("GPGPU-Sim: ** STOP simulation thread (no work) **\n");
            fflush(stdout);
         }
-		if(sim_cycles) {
-			g_the_gpu->update_stats();
-		}
+        if(sim_cycles) {
+            g_the_gpu->update_stats();
+            print_simulation_time();
+        }
         pthread_mutex_lock(&g_sim_lock);
         g_sim_active = false;
         pthread_mutex_unlock(&g_sim_lock);