aboutsummaryrefslogtreecommitdiff
path: root/src
diff options
context:
space:
mode:
Diffstat (limited to 'src')
-rw-r--r--src/abstract_hardware_model.h3
-rw-r--r--src/cuda-sim/cuda-math.h3
-rw-r--r--src/cuda-sim/cuda-sim.cc27
-rw-r--r--src/cuda-sim/cuda_device_runtime.cc14
-rw-r--r--src/cuda-sim/instructions.cc18
-rw-r--r--src/cuda-sim/ptx.l7
-rw-r--r--src/cuda-sim/ptx_ir.cc41
-rw-r--r--src/cuda-sim/ptx_ir.h21
-rw-r--r--src/cuda-sim/ptx_loader.cc201
-rw-r--r--src/cuda-sim/ptx_loader.h1
-rw-r--r--src/gpgpu-sim/gpu-sim.cc8
-rw-r--r--src/gpgpu-sim/gpu-sim.h1
-rw-r--r--src/gpgpu-sim/shader.h2
13 files changed, 285 insertions, 62 deletions
diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index 7125b6b..f5708bc 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -182,6 +182,9 @@ void increment_x_then_y_then_z( dim3 &i, const dim3 &bound);
class stream_manager;
struct CUstream_st;
extern stream_manager * g_stream_manager;
+//support for pinned memories added
+extern std::map<void *,void **> pinned_memory;
+extern std::map<void *, size_t> pinned_memory_size;
class kernel_info_t {
public:
diff --git a/src/cuda-sim/cuda-math.h b/src/cuda-sim/cuda-math.h
index a3db0df..afac330 100644
--- a/src/cuda-sim/cuda-math.h
+++ b/src/cuda-sim/cuda-math.h
@@ -67,6 +67,8 @@
#ifndef CUDA_MATH
#define CUDA_MATH
+#include <cmath>
+
// cuda math implementations
#undef max
#undef min
@@ -148,6 +150,7 @@ float __ll2float_rd(long long int a) {
#include <device_types.h>
#include <fenv.h>
+
// 32-bit integer to float
float __int2float_rn(int a) {
int orig_rnd_mode = fegetround();
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc
index a34b99b..946043a 100644
--- a/src/cuda-sim/cuda-sim.cc
+++ b/src/cuda-sim/cuda-sim.cc
@@ -212,7 +212,9 @@ void function_info::ptx_assemble()
m_start_PC = PC;
addr_t n=0; // offset in m_instr_mem
- s_g_pc_to_insn.reserve(s_g_pc_to_insn.size() + MAX_INST_SIZE*m_instructions.size());
+ //Why s_g_pc_to_insn.size() is needed to reserve additional memory for insts? reserve is cumulative.
+ //s_g_pc_to_insn.reserve(s_g_pc_to_insn.size() + MAX_INST_SIZE*m_instructions.size());
+ s_g_pc_to_insn.reserve(MAX_INST_SIZE*m_instructions.size());
for ( i=m_instructions.begin(); i != m_instructions.end(); i++ ) {
ptx_instruction *pI = *i;
if ( pI->is_label() ) {
@@ -250,11 +252,13 @@ void function_info::ptx_assemble()
target.set_type(label_t);
}
}
-
+ m_n = n;
printf(" done.\n");
fflush(stdout);
- printf("GPGPU-Sim PTX: finding reconvergence points for \'%s\'...\n", m_name.c_str() );
+ //disable pdom analysis here and do it at runtime
+#if 0
+ printf("GPGPU-Sim PTX: finding reconvergence points for \'%s\'...\n", m_name.c_str() );
create_basic_blocks();
connect_basic_blocks();
bool modified = false;
@@ -288,6 +292,7 @@ void function_info::ptx_assemble()
fflush(stdout);
m_assembled = true;
+#endif
}
addr_t shared_to_generic( unsigned smid, addr_t addr )
@@ -1134,8 +1139,13 @@ void function_info::finalize( memory_space *param_mem )
}
// copy the parameter over word-by-word so that parameter that crosses a memory page can be copied over
//Jin: copy parameter using aligned rules
+ const type_info *paramtype = param->type();
+ int align_amount = paramtype->get_key().get_alignment_spec();
+ align_amount = (align_amount == -1) ? size : align_amount;
+ param_address = (param_address + align_amount - 1) / align_amount * align_amount; //aligned
+
const size_t word_size = 4;
- param_address = (param_address + size - 1) / size * size; //aligned with size
+ //param_address = (param_address + size - 1) / size * size; //aligned with size
for (size_t idx = 0; idx < size; idx += word_size) {
const char *pdata = reinterpret_cast<const char*>(param_value.pdata) + idx; // cast to char * for ptr arithmetic
param_mem->write(param_address + idx, word_size, pdata,NULL,NULL);
@@ -1799,6 +1809,15 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL )
//using a shader core object for book keeping, it is not needed but as most function built for performance simulation need it we use it here
extern gpgpu_sim *g_the_gpu;
+ //before we execute, we should do PDOM analysis for functional simulation scenario.
+ function_info *kernel_func_info = kernel.entry();
+ if (kernel_func_info->is_pdom_set()) {
+ printf("GPGPU-Sim PTX: PDOM analysis already done for %s \n", kernel.name().c_str() );
+ } else {
+ printf("GPGPU-Sim PTX: finding reconvergence points for \'%s\'...\n", kernel.name().c_str() );
+ kernel_func_info->do_pdom();
+ kernel_func_info->set_pdom();
+ }
//we excute the kernel one CTA (Block) at the time, as synchronization functions work block wise
while(!kernel.no_more_ctas_to_run()){
diff --git a/src/cuda-sim/cuda_device_runtime.cc b/src/cuda-sim/cuda_device_runtime.cc
index 4a8ffe5..b399133 100644
--- a/src/cuda-sim/cuda_device_runtime.cc
+++ b/src/cuda-sim/cuda_device_runtime.cc
@@ -177,6 +177,20 @@ void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info *
//device_grid = op.grid;
device_kernel_entry = config.entry;
DEV_RUNTIME_REPORT("find device kernel " << device_kernel_entry->get_name());
+
+ //PDOM analysis is done for Parent kernel but not for child kernel.
+ if (device_kernel_entry->is_pdom_set()) {
+ printf("GPGPU-Sim PTX: PDOM analysis already done for %s \n", device_kernel_entry->get_name().c_str() );
+ } else {
+ printf("GPGPU-Sim PTX: finding reconvergence points for \'%s\'...\n", device_kernel_entry->get_name().c_str() );
+ /*
+ * Some of the instructions like printf() gives the gpgpusim the wrong impression that it is a function call.
+ * As printf() doesnt have a body like functions do, doing pdom analysis for printf() causes a crash.
+ */
+ if (device_kernel_entry->get_function_size() >0)
+ device_kernel_entry->do_pdom();
+ device_kernel_entry->set_pdom();
+ }
//copy data in parameter_buffer to device kernel param memory
unsigned device_kernel_arg_size = device_kernel_entry->get_args_aligned_size();
diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc
index 71286c9..e53aaab 100644
--- a/src/cuda-sim/instructions.cc
+++ b/src/cuda-sim/instructions.cc
@@ -154,7 +154,9 @@ ptx_reg_t ptx_thread_info::get_operand_value( const operand_info &op, operand_in
} else if ( op.is_local() ) {
result.u64 = op.get_symbol()->get_address();
} else if ( op.is_function_address() ) {
- result.u64 = (size_t)op.get_symbol()->get_pc();
+ result.u64 = (size_t)op.get_symbol()->get_pc();
+ } else if ( op.is_param_kernel()) {
+ result.u64 = op.get_symbol()->get_address();
} else {
const char *name = op.name().c_str();
printf("GPGPU-Sim PTX: ERROR ** get_operand_value : unknown operand type for %s\n", name );
@@ -1466,7 +1468,19 @@ void call_impl( const ptx_instruction *pI, ptx_thread_info *thread )
const operand_info &target = pI->func_addr();
assert( target.is_function_address() );
const symbol *func_addr = target.get_symbol();
- const function_info *target_func = func_addr->get_pc();
+ function_info *target_func = func_addr->get_pc();
+ if (target_func->is_pdom_set()) {
+ printf("GPGPU-Sim PTX: PDOM analysis already done for %s \n", target_func->get_name().c_str() );
+ } else {
+ printf("GPGPU-Sim PTX: finding reconvergence points for \'%s\'...\n", target_func->get_name().c_str() );
+ /*
+ * Some of the instructions like printf() gives the gpgpusim the wrong impression that it is a function call.
+ * As printf() doesnt have a body like functions do, doing pdom analysis for printf() causes a crash.
+ */
+ if (target_func->get_function_size() >0)
+ target_func->do_pdom();
+ target_func->set_pdom();
+ }
// check that number of args and return match function requirements
if( pI->has_return() ^ target_func->has_return() ) {
diff --git a/src/cuda-sim/ptx.l b/src/cuda-sim/ptx.l
index 5471d6f..908c5be 100644
--- a/src/cuda-sim/ptx.l
+++ b/src/cuda-sim/ptx.l
@@ -36,7 +36,8 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "ptx.tab.h"
#include <string.h>
-char linebuf[1024];
+#define LINEBUF_SIZE (64*1024)
+char linebuf[LINEBUF_SIZE];
unsigned col = 0;
#define TC col+=strlen(ptx_text);
#define CHECK_UNSIGNED \
@@ -162,6 +163,7 @@ breakaddr TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE;
\.file TC; BEGIN(INITIAL); return FILE_DIRECTIVE;
\.func TC; BEGIN(IN_FUNC_DECL); 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;
+\.global.volatile TC; return GLOBAL_DIRECTIVE; //TODO: fix this!
\.local TC; return LOCAL_DIRECTIVE;
\.loc TC; return LOC_DIRECTIVE;
\.maxnctapersm TC; return MAXNCTAPERSM_DIRECTIVE;
@@ -233,6 +235,7 @@ breakaddr TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE;
\.u32 TC; return U32_TYPE;
\.u64 TC; return U64_TYPE;
\.f16 TC; return F16_TYPE;
+\.f16x2 TC; return F16_TYPE; /* TODO: figure out what this should really be */
\.f32 TC; return F32_TYPE;
\.f64 TC; return F64_TYPE;
\.ff64 TC; return FF64_TYPE;
@@ -382,7 +385,7 @@ breakaddr TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE;
"//"[^\n]* TC; // eat single
-\n.* col=0; strncpy(linebuf, yytext + 1, 1024); yyless( 1 );
+\n.* col=0; strncpy(linebuf, yytext + 1, LINEBUF_SIZE); yyless( 1 );
" " TC;
"\t" TC;
diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc
index 8ebdcf8..be25dbe 100644
--- a/src/cuda-sim/ptx_ir.cc
+++ b/src/cuda-sim/ptx_ir.cc
@@ -280,8 +280,10 @@ type_info *symbol_table::get_array_type( type_info *base_type, unsigned array_di
{
type_info_key t = base_type->get_key();
t.set_array_dim(array_dim);
- type_info *pt;
- pt = m_types[t] = new type_info(this,t);
+ type_info *pt = new type_info(this,t);
+ //Where else is m_types being used? As of now, I dont find any use of it and causing seg fault. So disabling m_types.
+ //TODO: find where m_types can be used in future and solve the seg fault.
+ //pt = m_types[t] = new type_info(this,t);
return pt;
}
@@ -575,6 +577,40 @@ bool function_info::connect_break_targets() //connecting break instructions with
return modified;
}
+void function_info::do_pdom()
+{
+ create_basic_blocks();
+ connect_basic_blocks();
+ bool modified = false;
+ do {
+ find_dominators();
+ find_idominators();
+ modified = connect_break_targets();
+ } while (modified == true);
+
+ if ( g_debug_execution>=50 ) {
+ print_basic_blocks();
+ print_basic_block_links();
+ print_basic_block_dot();
+ }
+ if ( g_debug_execution>=2 ) {
+ print_dominators();
+ }
+ find_postdominators();
+ find_ipostdominators();
+ if ( g_debug_execution>=50 ) {
+ print_postdominators();
+ print_ipostdominators();
+ }
+ printf("GPGPU-Sim PTX: pre-decoding instructions for \'%s\'...\n", m_name.c_str() );
+ for ( unsigned ii=0; ii < m_n; ii += m_instr_mem[ii]->inst_size() ) { // handle branch instructions
+ ptx_instruction *pI = m_instr_mem[ii];
+ pI->pre_decode();
+ }
+ printf("GPGPU-Sim PTX: ... done pre-decoding instructions for \'%s\'.\n", m_name.c_str() );
+ fflush(stdout);
+ m_assembled = true;
+}
void intersect( std::set<int> &A, const std::set<int> &B )
{
// return intersection of A and B in A
@@ -1305,6 +1341,7 @@ function_info::function_info(int entry_point )
m_kernel_info.smem = 0;
m_local_mem_framesize = 0;
m_args_aligned_size = -1;
+ pdom_done = false; //initialize it to false
}
unsigned function_info::print_insn( unsigned pc, FILE * fp ) const
diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h
index 9ad1571..58d5f49 100644
--- a/src/cuda-sim/ptx_ir.h
+++ b/src/cuda-sim/ptx_ir.h
@@ -91,6 +91,7 @@ public:
bool is_tex() const { return m_space_spec == tex_space;}
bool is_func_addr() const { return m_is_function?true:false; }
int scalar_type() const { return m_scalar_type_spec;}
+ int get_alignment_spec() const { return m_alignment_spec;}
unsigned type_decode( size_t &size, int &t ) const;
static unsigned type_decode( int type, size_t &size, int &t );
memory_space_t get_memory_space() const { return m_space_spec; }
@@ -163,6 +164,7 @@ public:
m_is_global = false;
m_is_local = false;
m_is_param_local = false;
+ m_is_param_kernel = false;
m_is_tex = false;
m_is_func_addr = false;
m_reg_num_valid = false;
@@ -176,6 +178,7 @@ public:
if ( type ) m_is_global = type->get_key().is_global();
if ( type ) m_is_local = type->get_key().is_local();
if ( type ) m_is_param_local = type->get_key().is_param_local();
+ if ( type ) m_is_param_kernel = type->get_key().is_param_kernel();
if ( type ) m_is_tex = type->get_key().is_tex();
if ( type ) m_is_func_addr = type->get_key().is_func_addr();
}
@@ -226,6 +229,7 @@ public:
bool is_global() const { return m_is_global;}
bool is_local() const { return m_is_local;}
bool is_param_local() const { return m_is_param_local; }
+ bool is_param_kernel() const { return m_is_param_kernel; }
bool is_tex() const { return m_is_tex;}
bool is_func_addr() const { return m_is_func_addr; }
bool is_reg() const
@@ -283,6 +287,7 @@ private:
bool m_is_global;
bool m_is_local;
bool m_is_param_local;
+ bool m_is_param_kernel;
bool m_is_tex;
bool m_is_func_addr;
unsigned m_reg_num;
@@ -399,6 +404,8 @@ public:
m_type = symbolic_t;
} else if ( addr->is_param_local() ) {
m_type = symbolic_t;
+ } else if ( addr->is_param_kernel() ) {
+ m_type = symbolic_t;
} else if ( addr->is_tex() ) {
m_type = symbolic_t;
} else if ( addr->is_func_addr() ) {
@@ -675,6 +682,13 @@ public:
return m_value.m_symbolic->type()->get_key().is_param_local();
}
+ bool is_param_kernel() const
+ {
+ if ( m_type != symbolic_t )
+ return false;
+ return m_value.m_symbolic->type()->get_key().is_param_kernel();
+ }
+
bool is_vector() const
{
if ( m_vector) return true;
@@ -1178,7 +1192,7 @@ public:
//Muchnick's Adv. Compiler Design & Implemmntation Fig 7.15
void find_ipostdominators( );
void print_ipostdominators();
-
+ void do_pdom(); //function to call pdom analysis
unsigned get_num_reconvergence_pairs();
@@ -1274,6 +1288,8 @@ public:
m_local_mem_framesize = sz;
}
bool is_entry_point() const { return m_entry_point; }
+ bool is_pdom_set() const { return pdom_done; } //return pdom flag
+ void set_pdom() { pdom_done = true; } //set pdom flag
private:
unsigned m_uid;
@@ -1281,6 +1297,7 @@ private:
bool m_entry_point;
bool m_extern;
bool m_assembled;
+ bool pdom_done; //flag to check whether pdom is completed or not
std::string m_name;
ptx_instruction **m_instr_mem;
unsigned m_start_PC;
@@ -1305,6 +1322,8 @@ private:
//parameter size for device kernels
int m_args_aligned_size;
+
+ addr_t m_n; // offset in m_instr_mem (used in do_pdom)
};
class arg_buffer_t {
diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc
index 6c1b595..33a4260 100644
--- a/src/cuda-sim/ptx_loader.cc
+++ b/src/cuda-sim/ptx_loader.cc
@@ -32,6 +32,7 @@
#include <unistd.h>
#include <dirent.h>
#include <fstream>
+#include <sstream>
/// globals
@@ -287,56 +288,82 @@ void fix_duplicate_errors(char fname2[1024]) {
}
}
+//we need the application name here too.
+char* get_app_binary_name(){
+ char exe_path[1025];
+ char *self_exe_path;
+#ifdef __APPLE__
+ //AMRUTH: get apple device and check the result.
+ printf("WARNING: not tested for Apple-mac devices \n");
+ abort();
+#else
+ std::stringstream exec_link;
+ exec_link << "/proc/self/exe";
+ ssize_t path_length = readlink(exec_link.str().c_str(), exe_path, 1024);
+ assert(path_length != -1);
+ exe_path[path_length] = '\0';
+
+ char *token = strtok(exe_path, "/");
+ while(token !=NULL){
+ self_exe_path = token;
+ token = strtok(NULL,"/");
+ }
+#endif
+ self_exe_path = strtok(self_exe_path, ".");
+ printf("self exe links to: %s\n", self_exe_path);
+ return self_exe_path;
+}
+
void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version )
{
- 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_for_info);
- fclose(ptxfile);
+ //do ptxas for individual files instead of one big embedded ptx. This prevents the duplicate defs and declarations.
+ char ptx_file[1000];
+ char *name=get_app_binary_name();
+ char commandline[4096], fname[1024], fname2[1024], final_tempfile_ptxinfo[1024], tempfile_ptxinfo[1024];
+ for (int index=1; index <= no_of_ptx; index++){
+ snprintf(ptx_file, 1000, "%s.%d.sm_%u.ptx", name, index, sm_version);
+ snprintf(fname,1024,"_ptx_XXXXXX");
+ int fd=mkstemp(fname);
+ close(fd);
- 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);
- printf("Running: %s\n", commandline2);
- 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);
- }
+ printf("GPGPU-Sim PTX: extracting embedded .ptx to temporary file \"%s\"\n", fname);
+ snprintf(commandline,4096,"cat %s > %s",ptx_file, fname);
+ if (system(commandline) !=0) {
+ printf("ERROR: %s command failed\n", commandline);
+ exit(0);
+ }
+
+ 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);
+ printf("Running: %s\n", commandline2);
+ 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;
+ snprintf(tempfile_ptxinfo,1024,"%sinfo",fname);
+ char extra_flags[1024];
+ extra_flags[0]=0;
-#if CUDART_VERSION >= 3000
- if (sm_version == 0) sm_version = 20;
- extern bool g_cdp_enabled;
- if(!g_cdp_enabled)
+ #if CUDART_VERSION >= 3000
+ if (sm_version == 0) sm_version = 20;
snprintf(extra_flags,1024,"--gpu-name=sm_%u",sm_version);
- else
- snprintf(extra_flags,1024,"--compile-only --gpu-name=sm_%u",sm_version);
-#endif
+ #endif
- snprintf(commandline,1024,"$CUDA_INSTALL_PATH/bin/ptxas %s -v %s --output-file /dev/null 2> %s",
+ snprintf(commandline,1024,"$CUDA_INSTALL_PATH/bin/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 ) {
- // 65280 = duplicate errors
- if (result == 65280) {
- ptxinfo_in = fopen(tempfile_ptxinfo,"r");
+ printf("GPGPU-Sim PTX: generating ptxinfo using \"%s\"\n", commandline);
+ result = system(commandline);
+ if( result != 0 ) {
+ // 65280 = duplicate errors
+ if (result == 65280) {
+ ptxinfo_in = fopen(tempfile_ptxinfo,"r");
g_ptxinfo_filename = tempfile_ptxinfo;
ptxinfo_parse();
@@ -345,26 +372,100 @@ void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num
extra_flags, fname2, tempfile_ptxinfo);
printf("GPGPU-Sim PTX: regenerating 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);
+ }
+ }
+ }
+
+ //TODO: duplicate code! move it into a function so that it can be reused!
+ if(no_of_ptx==0) {
+ //For CDP, we dump everything. So no_of_ptx will be 0.
+ 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_for_info);
+ fclose(ptxfile);
+
+ 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);
+ printf("Running: %s\n", commandline2);
+ 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);
}
- if (result != 0) {
+ //char tempfile_ptxinfo[1024];
+ snprintf(tempfile_ptxinfo,1024,"%sinfo",fname);
+ char extra_flags[1024];
+ extra_flags[0]=0;
+
+ #if CUDART_VERSION >= 3000
+ if (sm_version == 0) sm_version = 20;
+ extern bool g_cdp_enabled;
+ if(!g_cdp_enabled)
+ snprintf(extra_flags,1024,"--gpu-name=sm_%u",sm_version);
+ else
+ snprintf(extra_flags,1024,"--compile-only --gpu-name=sm_%u",sm_version);
+ #endif
+
+ snprintf(commandline,1024,"$CUDA_INSTALL_PATH/bin/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);
+ fflush(stdout);
+ 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;
+ //Now that we got resource usage per kernel in a ptx file, we dump all into one file and pass it to rest of the code as usual.
+ if(no_of_ptx>0){
+ char commandline3[4096];
+ snprintf(final_tempfile_ptxinfo,1024,"f_tempfile_ptx");
+ snprintf(commandline3,4096, "cat *info > %s", final_tempfile_ptxinfo);
+ if (system(commandline3)!=0) {
+ printf("ERROR: Either we dont have info files or cat is not working \n");
+ printf("ERROR: %s command failed\n",commandline3);
+ exit(1);
+ }
+ }
+
+ if(no_of_ptx>0)
+ g_ptxinfo_filename = final_tempfile_ptxinfo;
+ else
+ g_ptxinfo_filename = tempfile_ptxinfo;
+ ptxinfo_in = fopen(g_ptxinfo_filename,"r");
+
ptxinfo_parse();
+ snprintf(commandline,1024,"rm -f *info");
+ if( system(commandline) != 0 ) {
+ printf("GPGPU-Sim PTX: ERROR ** while removing temporary info files\n");
+ exit(1);
+ }
if( ! g_save_embedded_ptx ) {
- snprintf(commandline,1024,"rm -f %s %s %s", fname, fname2, tempfile_ptxinfo);
+ if(no_of_ptx>0)
+ snprintf(commandline,1024,"rm -f %s %s %s", fname, fname2, final_tempfile_ptxinfo);
+ else
+ 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);
+ if( system(commandline) != 0 ) {
+ printf("GPGPU-Sim PTX: ERROR ** while removing temporary files\n");
exit(1);
}
}
}
-
diff --git a/src/cuda-sim/ptx_loader.h b/src/cuda-sim/ptx_loader.h
index d3d0c92..a8ecda3 100644
--- a/src/cuda-sim/ptx_loader.h
+++ b/src/cuda-sim/ptx_loader.h
@@ -30,6 +30,7 @@
#include <string>
extern bool g_override_embedded_ptx;
+extern int no_of_ptx; //counter to track number of ptx files to be extracted in an application.
class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num );
void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version=20 );
diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc
index bd4c00c..892b245 100644
--- a/src/gpgpu-sim/gpu-sim.cc
+++ b/src/gpgpu-sim/gpu-sim.cc
@@ -274,6 +274,9 @@ void shader_core_config::reg_options(class OptionParser * opp)
option_parser_register(opp, "-gpgpu_n_ldst_response_buffer_size", OPT_UINT32, &ldst_unit_response_queue_size,
"number of response packets in ld/st unit ejection buffer",
"2");
+ option_parser_register(opp, "-gpgpu_shmem_per_block", OPT_UINT32, &gpgpu_shmem_per_block,
+ "Size of shared memory per thread block or CTA (default 48kB)",
+ "49152");
option_parser_register(opp, "-gpgpu_shmem_size", OPT_UINT32, &gpgpu_shmem_size,
"Size of shared memory per shader core (default 16kB)",
"16384");
@@ -670,6 +673,11 @@ int gpgpu_sim::shared_mem_size() const
return m_shader_config->gpgpu_shmem_size;
}
+int gpgpu_sim::shared_mem_per_block() const
+{
+ return m_shader_config->gpgpu_shmem_per_block;
+}
+
int gpgpu_sim::num_registers_per_core() const
{
return m_shader_config->gpgpu_shader_registers;
diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h
index 7d92c66..8d1c4fc 100644
--- a/src/gpgpu-sim/gpu-sim.h
+++ b/src/gpgpu-sim/gpu-sim.h
@@ -389,6 +389,7 @@ public:
void get_pdom_stack_top_info( unsigned sid, unsigned tid, unsigned *pc, unsigned *rpc );
int shared_mem_size() const;
+ int shared_mem_per_block() const;
int num_registers_per_core() const;
int wrp_size() const;
int shader_clock() const;
diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h
index bdd8dbe..fbddd18 100644
--- a/src/gpgpu-sim/shader.h
+++ b/src/gpgpu-sim/shader.h
@@ -1285,7 +1285,7 @@ struct shader_core_config : public core_config
unsigned max_cta_per_core; //Limit on number of concurrent CTAs in shader core
unsigned max_barriers_per_cta;
char * gpgpu_scheduler_string;
-
+ unsigned gpgpu_shmem_per_block;
char* pipeline_widths_string;
int pipe_widths[N_PIPELINE_STAGES];
generated by cgit v1.3 (git 2.53.0) at 2026-07-19 17:59:27 +0000