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-sim.cc16
-rw-r--r--src/cuda-sim/ptx.l2
-rw-r--r--src/cuda-sim/ptx_ir.cc32
-rw-r--r--src/cuda-sim/ptx_ir.h5
-rw-r--r--src/cuda-sim/ptx_loader.cc194
-rw-r--r--src/cuda-sim/ptx_loader.h1
7 files changed, 198 insertions, 55 deletions
diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index aaa4b00..67b36c7 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-sim.cc b/src/cuda-sim/cuda-sim.cc
index d4ace76..39a04dd 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() ) {
@@ -255,6 +257,8 @@ void function_info::ptx_assemble()
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
create_basic_blocks();
connect_basic_blocks();
bool modified = false;
@@ -278,6 +282,7 @@ void function_info::ptx_assemble()
print_postdominators();
print_ipostdominators();
}
+#endif
printf("GPGPU-Sim PTX: pre-decoding instructions for \'%s\'...\n", m_name.c_str() );
for ( unsigned ii=0; ii < n; ii += m_instr_mem[ii]->inst_size() ) { // handle branch instructions
@@ -1799,6 +1804,15 @@ 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/ptx.l b/src/cuda-sim/ptx.l
index 5471d6f..1b5d7f6 100644
--- a/src/cuda-sim/ptx.l
+++ b/src/cuda-sim/ptx.l
@@ -162,6 +162,7 @@ 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 +234,7 @@ 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;
diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc
index 8ebdcf8..17e91df 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 +577,31 @@ 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();
+ }
+}
void intersect( std::set<int> &A, const std::set<int> &B )
{
// return intersection of A and B in A
@@ -1305,6 +1332,7 @@ 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..26a2839 100644
--- a/src/cuda-sim/ptx_ir.h
+++ b/src/cuda-sim/ptx_ir.h
@@ -1178,7 +1178,7 @@ 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 +1274,8 @@ 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 +1283,7 @@ 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;
diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc
index 6c1b595..9ff0859 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 +288,86 @@ 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)
- snprintf(extra_flags,1024,"--gpu-name=sm_%u",sm_version);
- else
- snprintf(extra_flags,1024,"--compile-only --gpu-name=sm_%u",sm_version);
-#endif
+ #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,"$CUDA_INSTALL_PATH/bin/ptxas %s -v %s --output-file /dev/null 2> %s",
+ snprintf(commandline,1024,"$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 +376,87 @@ 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
+ snprintf(extra_flags,1024,"--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);
+ 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);
+ }
+ }
+
+ ptxinfo_in = fopen(final_tempfile_ptxinfo,"r");
+ if(no_of_ptx>0)
+ g_ptxinfo_filename = final_tempfile_ptxinfo;
+ else
+ g_ptxinfo_filename = tempfile_ptxinfo;
ptxinfo_parse();
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 *info", fname, fname2, final_tempfile_ptxinfo);
+ else
+ snprintf(commandline,1024,"rm -f %s %s %s *info", 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 );
generated by cgit v1.3 (git 2.53.0) at 2026-07-19 19:36:56 +0000