dynamic pdom analysis at runtime

4 files changed, 57 insertions, 5 deletions

diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc
index e7952e2..948d81d 100644
--- a/libcuda/cuda_runtime_api.cc
+++ b/libcuda/cuda_runtime_api.cc
@@ -1013,7 +1013,16 @@ __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun )
 	printf("\nGPGPU-Sim PTX: cudaLaunch for 0x%p (mode=%s) on stream %u\n", hostFun,
 			g_ptx_sim_mode?"functional simulation":"performance simulation", stream?stream->get_uid():0 );
 	kernel_info_t *grid = gpgpu_cuda_ptx_sim_init_grid(hostFun,config.get_args(),config.grid_dim(),config.block_dim(),context);
+        //do dynamic PDOM analysis for performance simulation scenario
 	std::string kname = grid->name();
+	function_info *kernel_func_info = grid->entry();
+	if (kernel_func_info->is_pdom_set()) {
+    		printf("GPGPU-Sim PTX: PDOM analysis already done for %s \n", kname.c_str() );
+    	} else {
+    		printf("GPGPU-Sim PTX: finding reconvergence points for \'%s\'...\n", kname.c_str() );
+		kernel_func_info->do_pdom();
+		kernel_func_info->set_pdom();
+    	} 
 	dim3 gridDim = config.grid_dim();
 	dim3 blockDim = config.block_dim();
 	printf("GPGPU-Sim PTX: pushing kernel \'%s\' to stream %u, gridDim= (%u,%u,%u) blockDim = (%u,%u,%u) \n",
@@ -1400,7 +1409,7 @@ void extract_code_using_cuobjdump(){
     const char *override_cuobjdump = getenv("CUOBJDUMP_SIM_FILE"); 
     
     char fname[1024];
-    if (override_cuobjdump == NULL) {
+    if ((override_cuobjdump == NULL) || (strlen(override_cuobjdump)==0)) {
 	char command[1000];
    	std::string app_binary = get_app_binary(); 
 
@@ -1733,7 +1742,8 @@ cuobjdumpPTXSection* findPTXSection(const std::string identifier){
 void cuobjdumpInit(){
 	CUctx_st *context = GPGPUSim_Context();
 	extract_code_using_cuobjdump(); //extract all the output of cuobjdump to _cuobjdump_*.*
-	if (getenv("CUOBJDUMP_SIM_FILE")==NULL){
+	const char* pre_load = getenv("CUOBJDUMP_SIM_FILE");
+	if (pre_load ==NULL || strlen(pre_load)==0){
 		cuobjdumpSectionList = pruneSectionList(cuobjdumpSectionList, context);
 		cuobjdumpSectionList = mergeSections(cuobjdumpSectionList);
 	}
@@ -1772,12 +1782,13 @@ void cuobjdumpParseBinary(unsigned int handle){
 	if (max_capability > 20) printf("WARNING: No guarantee that PTX will be parsed for SM version %u\n", max_capability);
 
 	cuobjdumpPTXSection* ptx = NULL;
-	if(getenv("CUOBJDUMP_SIM_FILE")==NULL)
+	const char* pre_load = getenv("CUOBJDUMP_SIM_FILE");
+	if(pre_load==NULL || strlen(pre_load)==0)
 		ptx = findPTXSection(fname);
 	symbol_table *symtab;
 	char *ptxcode;
 	const char *override_ptx_name = getenv("PTX_SIM_KERNELFILE"); 
-	if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL) {
+	if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL or strlen(getenv("PTX_SIM_USE_PTX_FILE"))==0) {
 		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);
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc
index f51f57d..39a04dd 100644
--- a/src/cuda-sim/cuda-sim.cc
+++ b/src/cuda-sim/cuda-sim.cc
@@ -257,6 +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; 
@@ -280,6 +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
@@ -1801,6 +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_ir.cc b/src/cuda-sim/ptx_ir.cc
index 8ebdcf8..6a17eaf 100644
--- a/src/cuda-sim/ptx_ir.cc
+++ b/src/cuda-sim/ptx_ir.cc
@@ -575,6 +575,31 @@ 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 +1330,7 @@ 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;