summaryrefslogtreecommitdiff
path: root/src/cuda-sim/cuda-sim.cc
diff options
context:
space:
mode:
Diffstat (limited to 'src/cuda-sim/cuda-sim.cc')
-rw-r--r--src/cuda-sim/cuda-sim.cc378
1 files changed, 3 insertions, 375 deletions
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc
index 74b18bc..f5c283e 100644
--- a/src/cuda-sim/cuda-sim.cc
+++ b/src/cuda-sim/cuda-sim.cc
@@ -68,7 +68,6 @@
#include "ptx_ir.h"
#include "ptx_sim.h"
#include <stdio.h>
-#include <dirent.h>
#include "opcodes.h"
#include "../intersim/statwraper.h"
@@ -78,6 +77,7 @@
#include "../abstract_hardware_model.h"
#include "memory.h"
#include "ptx-stats.h"
+#include "ptx_loader.h"
extern bool g_interactive_debugger_enabled;
@@ -85,30 +85,15 @@ int gpgpu_ptx_instruction_classification=0;
void ** g_inst_classification_stat = NULL;
void ** g_inst_op_classification_stat= NULL;
int g_ptx_kernel_count = -1; // used for classification stat collection purposes
-extern "C" int ptx_parse();
-extern "C" int ptx__scan_string(const char*);
-extern "C" int ptx_debug;
-extern "C" FILE *ptx_in;
-
-extern "C" const char *g_ptxinfo_filename = NULL;
-extern "C" int ptxinfo_parse();
-extern "C" int ptxinfo_debug;
-extern "C" FILE *ptxinfo_in;
int g_debug_execution = 0;
int g_debug_thread_uid = 0;
addr_t g_debug_pc = 0xBEEF1518;
-int g_override_embedded_ptx = false;
const char *g_filename;
bool g_debug_ir_generation = false;
unsigned g_ptx_sim_num_insn = 0;
-extern memory_space *g_global_mem;
-extern memory_space *g_param_mem;
-extern memory_space *g_tex_mem;
-extern memory_space *g_surf_mem;
-
std::map<const struct textureReference*,const struct cudaArray*> TextureToArrayMap; // texture bindings
std::map<const struct textureReference*, const struct textureInfo*> TextureToInfoMap;
std::map<std::string, const struct textureReference*> NameToTextureMap;
@@ -309,100 +294,6 @@ void gpgpu_ptx_sim_init_memory()
}
}
-int load_static_globals( symbol_table *symtab, unsigned min_gaddr, unsigned max_gaddr)
-{
- printf( "GPGPU-Sim PTX: loading globals with explicit initializers... \n" );
- fflush(stdout);
- int ng_bytes=0;
- symbol_table::iterator g=symtab->global_iterator_begin();
-
- for ( ; g!=symtab->global_iterator_end(); g++) {
- symbol *global = *g;
- if ( global->has_initializer() ) {
- printf( "GPGPU-Sim PTX: initializing '%s' ... ", global->name().c_str() );
- unsigned addr=global->get_address();
- const type_info *type = global->type();
- type_info_key ti=type->get_key();
- size_t size;
- int t;
- ti.type_decode(size,t);
- int nbytes = size/8;
- int offset=0;
- std::list<operand_info> init_list = global->get_initializer();
- for ( std::list<operand_info>::iterator i=init_list.begin(); i!=init_list.end(); i++ ) {
- operand_info op = *i;
- ptx_reg_t value = op.get_literal_value();
- assert( (addr+offset+nbytes) < min_gaddr ); // min_gaddr is start of "heap" for cudaMalloc
- g_global_mem->write(addr+offset,nbytes,&value,NULL,NULL); // assuming little endian here
- offset+=nbytes;
- ng_bytes+=nbytes;
- }
- printf(" wrote %u bytes\n", offset );
- }
- }
- printf( "GPGPU-Sim PTX: finished loading globals (%u bytes total).\n", ng_bytes );
- fflush(stdout);
- return ng_bytes;
-}
-
-int load_constants( symbol_table *symtab, addr_t min_gaddr )
-{
- printf( "GPGPU-Sim PTX: loading constants with explicit initializers... " );
- fflush(stdout);
- int nc_bytes = 0;
- symbol_table::iterator g=symtab->const_iterator_begin();
-
- for ( ; g!=symtab->const_iterator_end(); g++) {
- symbol *constant = *g;
- if ( constant->is_const() && constant->has_initializer() ) {
-
- // get the constant element data size
- int basic_type;
- size_t num_bits;
- constant->type()->get_key().type_decode(num_bits,basic_type);
-
- std::list<operand_info> init_list = constant->get_initializer();
- int nbytes_written = 0;
- for ( std::list<operand_info>::iterator i=init_list.begin(); i!=init_list.end(); i++ ) {
- operand_info op = *i;
- ptx_reg_t value = op.get_literal_value();
- int nbytes = num_bits/8;
- switch ( op.get_type() ) {
- case int_t: assert(nbytes >= 1); break;
- case float_op_t: assert(nbytes == 4); break;
- case double_op_t: assert(nbytes >= 4); break; // account for double DEMOTING
- default:
- abort();
- }
- unsigned addr=constant->get_address() + nbytes_written;
- assert( addr+nbytes < min_gaddr );
-
- g_global_mem->write(addr,nbytes,&value,NULL,NULL); // assume little endian (so u8 is the first byte in u32)
- nc_bytes+=nbytes;
- nbytes_written += nbytes;
- }
- }
- }
- printf( " done.\n");
- fflush(stdout);
- return nc_bytes;
-}
-
-#define GLOBAL_HEAP_START 0x10000000
- // start allocating from this address (lower values used for allocating globals in .ptx file)
-
-#define SHARED_MEM_SIZE_MAX (64*1024)
-#define LOCAL_MEM_SIZE_MAX 1024
-#define MAX_STREAMING_MULTIPROCESSORS 64
-#define MAX_THREAD_PER_SM 1024
-#define TOTAL_LOCAL_MEM_PER_SM (MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX)
-#define TOTAL_SHARED_MEM (MAX_STREAMING_MULTIPROCESSORS*SHARED_MEM_SIZE_MAX)
-#define TOTAL_LOCAL_MEM (MAX_STREAMING_MULTIPROCESSORS*MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX)
-#define SHARED_GENERIC_START (GLOBAL_HEAP_START-TOTAL_SHARED_MEM)
-#define LOCAL_GENERIC_START (SHARED_GENERIC_START-TOTAL_LOCAL_MEM)
-
-#define STATIC_ALLOC_LIMIT (GLOBAL_HEAP_START - (TOTAL_LOCAL_MEM+TOTAL_SHARED_MEM))
-
addr_t shared_to_generic( unsigned smid, addr_t addr )
{
assert( addr < SHARED_MEM_SIZE_MAX );
@@ -1340,7 +1231,6 @@ std::map<const void*,std::string> g_const_name_lookup; // indexed by hostVar
std::map<const void*,std::string> g_global_name_lookup; // indexed by hostVar
std::set<std::string> g_globals;
std::set<std::string> g_constants;
-extern std::map<std::string,symbol_table*> g_sym_name_to_symbol_table;
void gpgpu_ptx_sim_register_const_variable(void *hostVar, const char *deviceName, size_t size )
{
@@ -1432,20 +1322,9 @@ int g_ptx_sim_mode=0;
// used by libcuda.a if non-zer cudaLaunch() will call gpgpu_ptx_sim_main_func()
// if zero it calls gpgpu_ptx_sim_main_perf()
-#if defined(__APPLE__)
-int ptx_file_filter(struct dirent *de )
-#else
-int ptx_file_filter(const struct dirent *de )
-#endif
-{
- const char *tmp = strstr(de->d_name,".ptx");
- if ( tmp != NULL && tmp[4] == 0 ) {
- return 1;
- }
- return 0;
-}
+extern "C" int ptx_debug;
-void read_environment_variables()
+void read_sim_environment_variables()
{
ptx_debug = 0;
g_debug_execution = 0;
@@ -1503,261 +1382,10 @@ void read_environment_variables()
}
}
-static FILE *open_ptxinfo (const char* ptx_filename)
-{
- const int ptx_fnamelen = strlen(ptx_filename);
- char *ptxi_fname = new char[ptx_fnamelen+5];
- strcpy (ptxi_fname, ptx_filename);
- strcpy (ptxi_fname+ptx_fnamelen, "info");
-
- //ptxinfo_debug=1;
- g_ptxinfo_filename = ptxi_fname;
- FILE *f = fopen (ptxi_fname, "rt");
- return f;
-}
-struct ptx_info_t {
- char *str;
- char *fname;
- ptx_info_t *next;
-};
-ptx_info_t *g_ptx_source_array = NULL;
-unsigned g_used_embedded_ptx_files;
-extern double g_ptx_version;
-void gpgpu_ptx_sim_add_ptxstring( const char *ptx_string, const char *sourcefname )
-{
- ptx_info_t *t = new ptx_info_t;
- t->next = NULL;
- t->str = strdup(ptx_string);
- t->fname = strdup(sourcefname);
-
- // put ptx source into a fifo
- if (g_ptx_source_array == NULL) {
- // first ptx source
- g_ptx_source_array = t;
- } else {
- // insert subsequent ptx source at the end of queue
- ptx_info_t *l_ptx_source = g_ptx_source_array;
- while (l_ptx_source->next != NULL) {
- l_ptx_source = l_ptx_source->next;
- }
- l_ptx_source->next = t;
- }
-}
-const ptx_instruction *ptx_instruction_lookup( const char *filename, unsigned linenumber );
-
-void print_ptx_file( const char *p, unsigned source_num, const char *filename )
-{
- printf("\nGPGPU-Sim PTX: file _%u.ptx contents:\n\n", source_num );
- char *s = strdup(p);
- char *t = s;
- unsigned n=1;
- while ( *t != '\0' ) {
- char *u = t;
- while ( (*u != '\n') && (*u != '\0') ) u++;
- unsigned last = (*u == '\0');
- *u = '\0';
- const ptx_instruction *pI = ptx_instruction_lookup(filename,n);
- char pc[64];
- if( pI && pI->get_PC() )
- snprintf(pc,64,"%4u", pI->get_PC() );
- else
- snprintf(pc,64," ");
- printf(" _%u.ptx %4u (pc=%s): %s\n", source_num, n, pc, t );
- if ( last ) break;
- t = u+1;
- n++;
- }
- free(s);
- fflush(stdout);
-}
-
-extern int g_save_embedded_ptx;
-
-void 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);
- }
- g_filename = strdup(buf);
- init_parser();
- ptx__scan_string(p);
- int errors = ptx_parse ();
- if ( errors ) {
- char fname[1024];
- snprintf(fname,1024,"_ptx_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,g_filename);
-
- printf("GPGPU-Sim PTX: finished parsing EMBEDDED .ptx file %s\n",g_filename);
-
- char fname[1024];
- snprintf(fname,1024,"_ptx_XXXXXX");
- int fd=mkstemp(fname);
- close(fd);
-
- printf("GPGPU-Sim PTX: extracting embedded .ptx to temporary file \"%s\"\n", fname);
- FILE *ptxfile = fopen(fname,"w");
- fprintf(ptxfile,"%s",p);
- fclose(ptxfile);
-
- char fname2[1024];
- snprintf(fname2,1024,"_ptx2_XXXXXX");
- fd=mkstemp(fname2);
- close(fd);
- char commandline2[4096];
- snprintf(commandline2,4096,"cat %s | sed 's/.version 1.5/.version 1.4/' | sed 's/, texmode_independent//' | sed 's/\\(\\.extern \\.const\\[1\\] .b8 \\w\\+\\)\\[\\]/\\1\\[1\\]/' | sed 's/const\\[.\\]/const\\[0\\]/g' > %s", fname, fname2);
- int result = system(commandline2);
- if( result != 0 ) {
- printf("GPGPU-Sim PTX: ERROR ** while loading PTX (a) %d\n", result);
- printf(" Ensure you have write access to simulation directory\n");
- printf(" and have \'cat\' and \'sed\' in your path.\n");
- exit(1);
- }
-
- char tempfile_ptxinfo[1024];
- snprintf(tempfile_ptxinfo,1024,"%sinfo",fname);
- char commandline[1024];
- char extra_flags[1024];
- extra_flags[0]=0;
-#if CUDART_VERSION >= 3000
- snprintf(extra_flags,1024,"--gpu-name=sm_20");
-#endif
- snprintf(commandline,1024,"ptxas %s -v %s --output-file /dev/null 2> %s",
- extra_flags, fname2, tempfile_ptxinfo);
- printf("GPGPU-Sim PTX: generating ptxinfo using \"%s\"\n", commandline);
- result = system(commandline);
- if( result != 0 ) {
- printf("GPGPU-Sim PTX: ERROR ** while loading PTX (b) %d\n", result);
- printf(" Ensure ptxas is in your path.\n");
- exit(1);
- }
-
- ptxinfo_in = fopen(tempfile_ptxinfo,"r");
- g_ptxinfo_filename = tempfile_ptxinfo;
- ptxinfo_parse();
- snprintf(commandline,1024,"rm -f %s %s %s", fname, fname2, tempfile_ptxinfo);
- printf("GPGPU-Sim PTX: removing ptxinfo using \"%s\"\n", commandline);
- result = system(commandline);
- if( result != 0 ) {
- printf("GPGPU-Sim PTX: ERROR ** while loading PTX (c) %d\n", result);
- exit(1);
- }
- g_filename = NULL;
-}
-
-void gpgpu_ptx_assemble( std::string kname, void *kinfo )
-{
- function_info *func_info = (function_info *)kinfo;
- if( func_info->is_extern() ) {
- printf("GPGPU-Sim PTX: skipping assembly for extern declared function \'%s\'\n", func_info->get_name().c_str() );
- return;
- }
- if( g_kernel_name_to_symtab_lookup[ kname ] == NULL ) {
- printf("\nGPGPU-Sim PTX: ERROR no information for kernel \'%s\'\n"
- " this can happen for kernels contained in CUDA\n"
- " libraries (such as CUBLAS, CUFFT, or CUDPP)\n", kname.c_str() );
- exit(1);
- }
- g_func_info = func_info;
- func_info->ptx_assemble();
-}
-
-void gpgpu_ptx_sim_load_gpu_kernels()
-{
- ptx_in = NULL;
- if ( g_filename )
- ptx_in = fopen( g_filename, "r" );
- static unsigned source_num = 0;
- gpgpu_ptx_sim_init_memory();
- if (ptx_in) {
- init_parser();
- ptx_parse();
- ptxinfo_in = open_ptxinfo(g_filename);
- ptxinfo_parse();
- load_static_globals(g_global_symbol_table,STATIC_ALLOC_LIMIT,0xFFFFFFFF);
- load_constants(g_global_symbol_table,STATIC_ALLOC_LIMIT);
- } else {
- if (!g_override_embedded_ptx) {
- g_used_embedded_ptx_files=1;
- printf("GPGPU-Sim PTX: USING EMBEDDED .ptx files...\n");
- ptx_info_t *s;
- for ( s=g_ptx_source_array; s!=NULL; s=s->next ) {
- gpgpu_ptx_sim_load_ptx_from_string(s->str, ++source_num);
- load_static_globals(g_global_symbol_table,STATIC_ALLOC_LIMIT,0xFFFFFFFF);
- load_constants(g_global_symbol_table,STATIC_ALLOC_LIMIT);
- }
- } else {
- g_filename = NULL;
- struct dirent **namelist;
- int n;
-
- n = scandir(".", &namelist, ptx_file_filter, alphasort);
- if (n < 0)
- perror("scandir");
- else {
- while (n--) {
- if ( g_filename != NULL ) {
- printf("Loader error: support for multiple .ptx files not yet enabled\n");
- abort();
- }
- g_filename = strdup(namelist[n]->d_name);
- printf("Parsing %s..\n", g_filename);
- ptx_in = fopen( g_filename, "r" );
- free(namelist[n]);
- init_parser();
- ptx_parse ();
- ptxinfo_in = open_ptxinfo(g_filename);
- ptxinfo_parse();
- g_filename = NULL;
- load_static_globals(g_global_symbol_table,STATIC_ALLOC_LIMIT,0xFFFFFFFF);
- load_constants(g_global_symbol_table,STATIC_ALLOC_LIMIT);
- }
- free(namelist);
- }
- }
- }
-
- if ( ptx_in == NULL && g_override_embedded_ptx ) {
- printf("GPGPU-Sim PTX Simulator error: Could find/open .ptx file for reading\n");
- printf(" This means there are no .ptx files in the current directory.\n");
- printf(" Either place a .ptx file in the current directory, or ensure\n" );
- printf(" the PTX_SIM_KERNELFILE environment variable points to .ptx file.\n");
- printf(" PTX_SIM_KERNELFILE=\"%s\"\n", g_filename );
- exit(1);
- }
-
- if ( g_error_detected ) {
- printf( "GPGPU-Sim PTX: Program parsing completed: Errors detected.\n" );
- exit(1);
- } else {
- printf( "GPGPU-Sim PTX: Program parsing completed (max %u registers used per thread).\n", g_max_regs_per_thread );
- }
-
- if ( g_kernel_name_to_function_lookup ) {
- for ( std::map<std::string,function_info*>::iterator f=g_kernel_name_to_function_lookup->begin();
- f != g_kernel_name_to_function_lookup->end(); f++ ) {
- gpgpu_ptx_assemble(f->first,f->second);
- }
- }
-}
extern time_t simulation_starttime;
generated by cgit v1.3 (git 2.53.0) at 2026-07-11 00:32:33 +0000