1. integrating Inder's changes

2. edit to 3dfd to enable execution on Quadro
3. update script to not pass device directly to 3dfd on command line (which it does not support)

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 7810]

6 files changed, 52 insertions, 17 deletions

diff --git a/configs/QuadroFX5800/gpgpusim.config b/configs/QuadroFX5800/gpgpusim.config
index 8c14ea3..560796d 100644
--- a/configs/QuadroFX5800/gpgpusim.config
+++ b/configs/QuadroFX5800/gpgpusim.config
@@ -1,6 +1,7 @@
 # functional simulator specification
 -gpgpu_ptx_instruction_classification 0
 -gpgpu_ptx_sim_mode 0
+-gpgpu_ptx_force_max_capability 10
 
 # high level architecture configuration
 -gpgpu_sm_uarch GT200
@@ -59,3 +60,5 @@
 -gpgpu_operand_collector 1
 -gpgpu_operand_collector_num_units 6
 -gpgpu_operand_collector_num_units_sfu 8
+
+-visualizer_enabled 0
diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc
index c7d19e0..d8092e8 100644
--- a/libcuda/cuda_runtime_api.cc
+++ b/libcuda/cuda_runtime_api.cc
@@ -1042,22 +1042,18 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin )
    __cudaFatCudaBinary *info =   (__cudaFatCudaBinary *)fatCubin;
    assert( info->version >= 3 );
    unsigned num_ptx_versions=0;
-   unsigned max_capability=0;
-   unsigned selected_capability_offset=(unsigned)-1;
+   //unsigned max_capability=0;
+   //unsigned selected_capability_offset=(unsigned)-1;
    while( info->ptx[num_ptx_versions].gpuProfileName != NULL ) {
       unsigned capability=0;
       sscanf(info->ptx[num_ptx_versions].gpuProfileName,"compute_%u",&capability);
-      if( capability > max_capability ) {
-         max_capability = capability;
-         selected_capability_offset=num_ptx_versions;
-      }
+
+      printf("GPGPU-Sim PTX: __cudaRegisterFatBinary found PTX versions for '%s', ", info->ident);
+	  printf("capability = %s\n", info->ptx[num_ptx_versions].gpuProfileName );
+	  gpgpu_ptx_sim_add_ptxstring( fat_cubin_handle, info->ptx[num_ptx_versions].ptx, info->cubin[num_ptx_versions].cubin, info->ident, capability );
+
       num_ptx_versions++;
    }
-   if ( selected_capability_offset != (unsigned)-1 ) {
-      printf("GPGPU-Sim PTX: __cudaRegisterFatBinary found %u PTX versions for '%s', ", num_ptx_versions, info->ident);
-      printf("selected = %s\n", info->ptx[selected_capability_offset].gpuProfileName );
-      gpgpu_ptx_sim_add_ptxstring( fat_cubin_handle, info->ptx[selected_capability_offset].ptx, info->cubin[selected_capability_offset].cubin, info->ident );
-   }
 #endif
    return (void**)fat_cubin_handle;
 }
diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc
index db915fe..fd51de5 100644
--- a/src/cuda-sim/ptx_ir.cc
+++ b/src/cuda-sim/ptx_ir.cc
@@ -1219,8 +1219,8 @@ void gpgpu_ptx_assemble( std::string kname, void *kinfo )
 {
     function_info *func_info = (function_info *)kinfo;
     if((function_info *)kinfo == NULL) {
-       printf("GPGPU-Sim PTX: Error - missing function definition \'%s\'\n", kname.c_str());
-       abort();
+       printf("GPGPU-Sim PTX: Warning - missing function definition \'%s\'\n", kname.c_str());
+       return;
     }
     if( func_info->is_extern() ) {
        printf("GPGPU-Sim PTX: skipping assembly for extern declared function \'%s\'\n", func_info->get_name().c_str() );
diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc
index cea9912..c836fac 100644
--- a/src/cuda-sim/ptx_loader.cc
+++ b/src/cuda-sim/ptx_loader.cc
@@ -83,6 +83,7 @@ struct ptx_info_t {
     char *cubin_str;
     char *fname;
     ptx_info_t *next;
+    unsigned capability;
 };
 
 /// extern prototypes
@@ -98,6 +99,7 @@ extern "C" FILE *ptxinfo_in;
 
 extern int g_ptx_convert_to_ptxplus;
 extern int g_ptx_save_converted_ptxplus;
+extern unsigned g_ptx_force_max_capability;
 
 /// static functions
 
@@ -209,6 +211,7 @@ static int ptx_file_filter(
 // global functions
 
 static ptx_info_t *g_ptx_source_array = NULL;
+static unsigned g_ptx_max_capability = 0;
 
 void gpgpu_ptx_sim_load_gpu_kernels()
 {
@@ -226,9 +229,33 @@ void gpgpu_ptx_sim_load_gpu_kernels()
         load_constants(symtab,STATIC_ALLOC_LIMIT);
     } else {
         if (!g_override_embedded_ptx) {
-            printf("GPGPU-Sim PTX: USING EMBEDDED .ptx files...\n"); 
+            printf("GPGPU-Sim PTX: USING EMBEDDED .ptx files...\n");
+            unsigned selected_capability = 0;
+
+            if(g_ptx_force_max_capability == 0) {
+            	// No forced max capability, selected the highest capability
+            	//assert(g_ptx_max_capability > 0);
+            	selected_capability = g_ptx_max_capability;
+            	printf("GPGPU-Sim PTX: Loading PTX, selected capability: compute_%u\n", selected_capability);
+            } else {
+            	// Forced max capability, select the highest capability less than or equal to forced capability
+            	ptx_info_t *pti;
+            	for( pti=g_ptx_source_array; pti!=NULL; pti=pti->next ){
+            		if(selected_capability < pti->capability && pti->capability <= g_ptx_force_max_capability)
+            			selected_capability = pti->capability;
+            	}
+            	//assert(selected_capability > 0);
+            	printf("GPGPU-Sim PTX: Loading PTX, max forced capability: compute_%u, selected capability: compute_%u\n",
+            			g_ptx_force_max_capability, selected_capability);
+            }
+
             ptx_info_t *s;
             for ( s=g_ptx_source_array; s!=NULL; s=s->next ) {
+            	 if(s->capability != selected_capability)
+            		 continue;
+
+            	 printf("GPGPU-Sim PTX: Loading PTX for %s, capability = compute_%u\n", s->fname, s->capability);
+
             	 symbol_table *symtab;
             	 source_num++;
             	 if(g_ptx_convert_to_ptxplus) {
@@ -293,7 +320,7 @@ void gpgpu_ptx_sim_load_gpu_kernels()
    }
 }
 
-void gpgpu_ptx_sim_add_ptxstring( unsigned fat_cubin_handle, const char *ptx_string, const char *cubin_string, const char *sourcefname )
+void gpgpu_ptx_sim_add_ptxstring( unsigned fat_cubin_handle, const char *ptx_string, const char *cubin_string, const char *sourcefname, unsigned capability )
 {
     ptx_info_t *t = new ptx_info_t;
     t->next = NULL;
@@ -311,6 +338,11 @@ void gpgpu_ptx_sim_add_ptxstring( unsigned fat_cubin_handle, const char *ptx_str
        t->cubin_str = NULL; 
     }
     t->fname = strdup(sourcefname);
+    t->capability = capability;
+
+    // Set overall max capability
+    if(g_ptx_max_capability < capability)
+    	g_ptx_max_capability = capability;
 
     // put ptx source into a fifo
     if (g_ptx_source_array == NULL) {
diff --git a/src/cuda-sim/ptx_loader.h b/src/cuda-sim/ptx_loader.h
index 432ee10..6b9d2f8 100644
--- a/src/cuda-sim/ptx_loader.h
+++ b/src/cuda-sim/ptx_loader.h
@@ -74,8 +74,7 @@ extern memory_space *g_surf_mem;
 extern memory_space *g_param_mem;
 extern bool g_override_embedded_ptx;
  
-void gpgpu_ptx_sim_load_gpu_kernels();
-void gpgpu_ptx_sim_add_ptxstring( unsigned cubin_handle, const char *ptx_string, const char *cubin_string,  const char *sourcefname );
+void gpgpu_ptx_sim_add_ptxstring( unsigned fat_cubin_handle, const char *ptx_string, const char *cubin_string,  const char *sourcefname, unsigned capability );
 class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, const char *p_for_info, unsigned source_num );
 char* gpgpu_ptx_sim_convert_ptx_to_ptxplus(const char *ptx_str, const char *cubin_str, unsigned source_num);
 
diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc
index 563dbae..bfa4fda 100644
--- a/src/gpgpu-sim/gpu-sim.cc
+++ b/src/gpgpu-sim/gpu-sim.cc
@@ -188,6 +188,7 @@ int   g_ptx_inst_debug_thread_uid;
 
 int   g_ptx_convert_to_ptxplus;
 int   g_ptx_save_converted_ptxplus;
+unsigned g_ptx_force_max_capability;
 
 void visualizer_options(option_parser_t opp);
 
@@ -417,6 +418,10 @@ void gpgpu_sim::reg_options(option_parser_t opp)
                 &g_ptx_save_converted_ptxplus,
                 "Saved converted ptxplus to a file",
                 "0");
+   option_parser_register(opp, "-gpgpu_ptx_force_max_capability", OPT_UINT32,
+                &g_ptx_force_max_capability,
+                "Force maximum compute capability",
+                "0");
    option_parser_register(opp, "-gpgpu_operand_collector", OPT_BOOL, &m_shader_config->gpgpu_operand_collector,
                "Enable operand collector model (default = off)",
                "0");