diff options
Diffstat (limited to 'src/trace-driven/trace_driven.cc')
| -rw-r--r-- | src/trace-driven/trace_driven.cc | 707 |
1 files changed, 707 insertions, 0 deletions
diff --git a/src/trace-driven/trace_driven.cc b/src/trace-driven/trace_driven.cc new file mode 100644 index 0000000..76eb7ca --- /dev/null +++ b/src/trace-driven/trace_driven.cc @@ -0,0 +1,707 @@ +//developed by Mahmoud Khairy, Purdue Univ + +#include <time.h> +#include <stdio.h> +#include <vector> +#include <iostream> +#include <fstream> +#include <string> +#include <sstream> +#include <math.h> +#include <bits/stdc++.h> + +#include "../abstract_hardware_model.h" +#include "../option_parser.h" +#include "../cuda-sim/cuda-sim.h" +#include "../cuda-sim/ptx_ir.h" +#include "../cuda-sim/ptx_parser.h" +#include "../gpgpu-sim/gpu-sim.h" +#include "../../libcuda/gpgpu_context.h" +#include "trace_driven.h" +#include "ISA_Def/trace_opcode.h" +#include "ISA_Def/volta_opcode.h" +#include "ISA_Def/turing_opcode.h" +#include "ISA_Def/pascal_opcode.h" +#include "ISA_Def/kepler_opcode.h" +#include "../gpgpusim_entrypoint.h" + + +bool is_number(const std::string& s) +{ + std::string::const_iterator it = s.begin(); + while (it != s.end() && std::isdigit(*it)) ++it; + return !s.empty() && it == s.end(); +} + +void split(const std::string& str, std::vector<std::string>& cont, char delimi = ' ') +{ + std::stringstream ss(str); + std::string token; + while (std::getline(ss, token, delimi)) { + cont.push_back(token); + } +} + +trace_parser::trace_parser(const char* kernellist_filepath, gpgpu_sim * m_gpgpu_sim, gpgpu_context* m_gpgpu_context) +{ + + this->m_gpgpu_sim = m_gpgpu_sim; + this->m_gpgpu_context = m_gpgpu_context; + kernellist_filename = kernellist_filepath; +} + +std::vector<std::string> trace_parser::parse_kernellist_file() { + + ifs.open(kernellist_filename); + + if (!ifs.is_open()) { + std::cout << "Unable to open file: " <<kernellist_filename<<std::endl; + exit(1); + } + + std::string directory(kernellist_filename); + const size_t last_slash_idx = directory.rfind('/'); + if (std::string::npos != last_slash_idx) + { + directory = directory.substr(0, last_slash_idx); + } + + std::string line, filepath; + std::vector<std::string> kernellist; + while(!ifs.eof()) { + getline(ifs, line); + if(line.empty()) + continue; + else if(line.substr(0,6) == "Memcpy") { + kernellist.push_back(line); + } + else if(line.substr(0,6) == "kernel") { + filepath = directory+"/"+line; + kernellist.push_back(filepath); + } + } + + ifs.close(); + return kernellist; +} + +void trace_parser::parse_memcpy_info(const std::string& memcpy_command, size_t& address, size_t& count) { + std::vector<std::string> params; + split(memcpy_command, params, ','); + assert(params.size() == 3); + std::stringstream ss; + ss.str(params[1]); + ss>>std::hex>>address; + ss.clear(); + ss.str(params[2]); + ss>>std::dec>>count; +} + +trace_kernel_info_t* trace_parser::parse_kernel_info(const std::string& kerneltraces_filepath, trace_config* config) { + + ifs.open(kerneltraces_filepath.c_str()); + + if (!ifs.is_open()) { + std::cout << "Unable to open file: " <<kerneltraces_filepath<<std::endl; + exit(1); + } + + std::cout << "Processing kernel " <<kerneltraces_filepath<<std::endl; + + unsigned grid_dim_x=0, grid_dim_y=0, grid_dim_z=0, tb_dim_x=0, tb_dim_y=0, tb_dim_z=0; + unsigned shmem=0, nregs=0, cuda_stream_id=0, kernel_id=0, binary_verion=0; + std::string line; + std::stringstream ss; + std::string string1, string2; + std::string kernel_name; + + while(!ifs.eof()) { + getline(ifs, line); + + if (line.length() == 0) { + continue; + } + else if(line[0] == '#'){ + //the trace format, ignore this and assume fixed format for now + break; //the begin of the instruction stream + } + else if(line[0] == '-') { + ss.str(line); + ss.ignore(); + ss>>string1>>string2; + if(string1 == "kernel" && string2 == "name") { + const size_t equal_idx = line.find('='); + kernel_name = line.substr(equal_idx+1); + } + else if(string1 == "kernel" && string2 == "id") { + sscanf(line.c_str(), "-kernel id = %d", &kernel_id); + } + else if(string1 == "grid" && string2 == "dim") { + sscanf(line.c_str(), "-grid dim = (%d,%d,%d)", &grid_dim_x, &grid_dim_y, &grid_dim_z); + } + else if (string1 == "block" && string2 == "dim") { + sscanf(line.c_str(), "-block dim = (%d,%d,%d)", &tb_dim_x, &tb_dim_y, &tb_dim_z); + } + else if (string1 == "shmem") { + sscanf(line.c_str(), "-shmem = %d", &shmem); + } + else if (string1 == "nregs") { + sscanf(line.c_str(), "-nregs = %d", &nregs); + } + else if (string1 == "cuda" && string2 == "stream") { + sscanf(line.c_str(), "-cuda stream id = %d", &cuda_stream_id); + } + else if (string1 == "binary" && string2 == "version") { + sscanf(line.c_str(), "-binary version = %d", &binary_verion); + } + std::cout << line << std::endl; + continue; + } + } + + gpgpu_ptx_sim_info info; + info.smem = shmem; + info.regs = nregs; + dim3 gridDim(grid_dim_x, grid_dim_y, grid_dim_z); + dim3 blockDim(tb_dim_x, tb_dim_y, tb_dim_z); + trace_function_info* function_info = new trace_function_info(info, m_gpgpu_context); + function_info->set_name(kernel_name.c_str()); + trace_kernel_info_t* kernel_info = new trace_kernel_info_t(gridDim, blockDim, binary_verion, function_info, &ifs, m_gpgpu_sim, m_gpgpu_context, config); + + return kernel_info; +} + + +void trace_parser::kernel_finalizer(trace_kernel_info_t* kernel_info){ + if (ifs.is_open()) + ifs.close(); + + delete kernel_info->entry(); + delete kernel_info; +} + +const trace_warp_inst_t* trace_shd_warp_t::get_next_inst(){ + if(trace_pc < warp_traces.size()) + { + return &warp_traces[trace_pc++]; + } + else + return NULL; +} + +void trace_shd_warp_t::clear() { + trace_pc=0; + warp_traces.clear(); +} + +//functional_done +bool trace_shd_warp_t::trace_done() { + return trace_pc==(warp_traces.size()); +} + +address_type trace_shd_warp_t::get_start_pc(){ + assert(warp_traces.size() > 0); + return warp_traces[0].pc; +} + +address_type trace_shd_warp_t::get_pc(){ + assert(warp_traces.size() > 0 ); + assert(trace_pc < warp_traces.size()); + return warp_traces[trace_pc].pc; +} + +trace_kernel_info_t::trace_kernel_info_t(dim3 gridDim, dim3 blockDim, unsigned m_binary_verion, trace_function_info* m_function_info, std::ifstream* inputstream, gpgpu_sim * gpgpu_sim, gpgpu_context* gpgpu_context, class trace_config* config):kernel_info_t(gridDim, blockDim, m_function_info) { + ifs = inputstream; + m_gpgpu_sim = gpgpu_sim; + m_gpgpu_context = gpgpu_context; + binary_verion = m_binary_verion; + m_tconfig = config; + + //resolve the binary version + if(m_binary_verion == VOLTA_BINART_VERSION) + OpcodeMap = &Volta_OpcodeMap; + else if(m_binary_verion == PASCAL_TITANX_BINART_VERSION || m_binary_verion == PASCAL_P100_BINART_VERSION) + OpcodeMap = &Pascal_OpcodeMap; + else if(m_binary_verion == KEPLER_BINART_VERSION) + OpcodeMap = &Kepler_OpcodeMap; + else if(m_binary_verion == TURING_BINART_VERSION) + OpcodeMap = &Turing_OpcodeMap; + else + assert(0 && "unsupported binary version"); +} + +bool trace_kernel_info_t::get_next_threadblock_traces(std::vector<std::vector<trace_warp_inst_t>*> threadblock_traces) { + + for(unsigned i=0; i<threadblock_traces.size(); ++i) { + threadblock_traces[i]->clear(); + } + + unsigned block_id_x=0, block_id_y=0, block_id_z=0; + unsigned warp_id=0; + unsigned insts_num=0; + + + bool start_of_tb_stream_found = false; + + while(!ifs->eof()) { + std::string line; + std::stringstream ss; + std::string string1, string2; + + getline(*ifs, line); + + if (line.length() == 0) { + continue; + } + else { + ss.str(line); + ss>>string1>>string2; + if (string1 == "#BEGIN_TB") { + if(!start_of_tb_stream_found) + { + start_of_tb_stream_found=true; + } + else + assert(0 && "Parsing error: thread block start before the previous one finish"); + } + else if (string1 == "#END_TB") { + assert(start_of_tb_stream_found); + break; //end of TB stream + } + else if(string1 == "thread" && string2 == "block") { + assert(start_of_tb_stream_found); + sscanf(line.c_str(), "thread block = %d,%d,%d", &block_id_x, &block_id_y, &block_id_z); + std::cout << line << std::endl; + } + else if (string1 == "warp") { + //the start of new warp stream + assert(start_of_tb_stream_found); + sscanf(line.c_str(), "warp = %d", &warp_id); + } + else if (string1 == "insts") { + assert(start_of_tb_stream_found); + sscanf(line.c_str(), "insts = %d", &insts_num); + threadblock_traces[warp_id]->reserve(insts_num); + } + else { + assert(start_of_tb_stream_found); + trace_warp_inst_t inst(m_gpgpu_sim->getShaderCoreConfig(), m_gpgpu_context, m_tconfig); + inst.parse_from_string(line, OpcodeMap); + threadblock_traces[warp_id]->push_back(inst); + } + } + } + + return true; +} + +bool trace_warp_inst_t::check_opcode_contain(const std::vector<std::string>& opcode, std::string param) +{ + for(unsigned i=0; i<opcode.size(); ++i) + if(opcode[i] == param) + return true; + + return false; +} + + + +unsigned trace_warp_inst_t::get_datawidth_from_opcode(const std::vector<std::string>& opcode) +{ + for(unsigned i=0; i<opcode.size(); ++i) { + if(is_number(opcode[i])){ + return (std::stoi(opcode[i],NULL)/8); + } + else if(opcode[i][0] == 'U' && is_number(opcode[i].substr(1))){ + //handle the U* case + unsigned bits; + sscanf(opcode[i].c_str(), "U%u",&bits); + return bits/8; + } + } + + return 4; //default is 4 bytes +} + +bool trace_warp_inst_t::parse_from_string(std::string trace, const std::unordered_map<std::string,OpcodeChar>* OpcodeMap){ + + std::stringstream ss; + ss.str(trace); + + std::string temp; + unsigned threadblock_x=0, threadblock_y=0, threadblock_z=0, warpid_tb=0, sm_id=0, warpid_sm=0; + unsigned long long m_pc=0; + unsigned mask=0; + unsigned reg_dest[4]; + std::string opcode; + unsigned reg_dsts_num=0; + unsigned reg_srcs_num=0; + unsigned reg_srcs[4]; + unsigned mem_width=0; + unsigned long long mem_addresses[warp_size()]; + unsigned address_mode=0; + unsigned long long base_address=0; + int stride=0; + + //Start Parsing + ss>>std::dec>>threadblock_x>>threadblock_y>>threadblock_z>>warpid_tb; + + //ignore core id + //ss>>std::dec>>sm_id>>warpid_sm; + + ss>>std::hex>>m_pc; + ss>>std::hex>>mask; + + std::bitset<MAX_WARP_SIZE> mask_bits(mask); + + ss>>std::dec>>reg_dsts_num; + for(unsigned i=0; i<reg_dsts_num; ++i) { + ss>>std::dec>>temp; + sscanf(temp.c_str(), "R%d", ®_dest[i]); + } + + ss>>opcode; + + ss>>reg_srcs_num; + for(unsigned i=0; i<reg_srcs_num; ++i) { + ss>>temp; + sscanf(temp.c_str(), "R%d", ®_srcs[i]); + + } + + ss>>mem_width; + + if(mem_width > 0) //then it is a memory inst + { + ss>>std::dec>>address_mode; + if(address_mode==0){ + //read addresses one by one from the file + for (int s = 0; s < warp_size(); s++) { + if(mask_bits.test(s)) + ss>>std::hex>>mem_addresses[s]; + else + mem_addresses[s]=0; + } + } + else if(address_mode==1){ + //read addresses as base address and stride + ss>>std::hex>>base_address; + ss>>std::dec>>stride; + bool first_bit1_found=false; + bool last_bit1_found=false; + unsigned long long addra=base_address; + for (int s = 0; s < warp_size(); s++) { + if(mask_bits.test(s) && !first_bit1_found){ + first_bit1_found=true; + mem_addresses[s]=base_address; + } else if(first_bit1_found && !last_bit1_found) { + if(mask_bits.test(s)) { + addra += stride; + mem_addresses[s] = addra; + } else + last_bit1_found=true; + } + else + mem_addresses[s]=0; + } + } + } + //Finish Parsing + //After parsing, fill the inst_t and warp_inst_t params + + //fill active mask + active_mask_t active_mask = mask_bits; + set_active( active_mask ); + + //get the opcode + std::istringstream iss(opcode); + std::vector<std::string> opcode_tokens; + std::string token; + while (std::getline(iss, token, '.')) { + if (!token.empty()) + opcode_tokens.push_back(token); + } + + std::string opcode1 = opcode_tokens[0]; + + //fill and initialize common params + m_decoded = true; + pc = (address_type)m_pc; //we will lose the high 32 bits from casting long to unsigned, it should be okay! + + isize = 16; //starting from MAXWELL isize=16 bytes (including the control bytes) + for(unsigned i=0; i<MAX_OUTPUT_VALUES; i++) { + out[i] = 0; + } + for(unsigned i=0; i<MAX_INPUT_VALUES; i++) { + in[i] = 0; + } + + is_vectorin = 0; + is_vectorout = 0; + ar1 = 0; + ar2 = 0; + memory_op = no_memory_op; + data_size = 0; + op = ALU_OP; + mem_op= NOT_TEX; + + std::unordered_map<std::string,OpcodeChar>::const_iterator it= OpcodeMap->find(opcode1); + if (it != OpcodeMap->end()) { + m_opcode = it->second.opcode; + op = (op_type)(it->second.opcode_category); + } + else { + std::cout<<"ERROR: undefined instruction : "<<opcode<<" Opcode: "<<opcode1<<std::endl; + assert(0 && "undefined instruction"); + } + + //fill regs information + num_regs = reg_srcs_num+reg_dsts_num; + num_operands = num_regs; + outcount=reg_dsts_num; + for(unsigned m=0; m<reg_dsts_num; ++m){ + out[m]=reg_dest[m]+1; //Increment by one because GPGPU-sim starts from R1, while SASS starts from R0 + arch_reg.dst[m]=reg_dest[m]+1; + } + + incount=reg_srcs_num; + for(unsigned m=0; m<reg_srcs_num; ++m){ + in[m]=reg_srcs[m]+1; //Increment by one because GPGPU-sim starts from R1, while SASS starts from R0 + arch_reg.src[m]=reg_srcs[m]+1; + } + //TO DO: handle: vector, store insts have no output, double inst and hmma, and 64 bit address + //remove redundant registers + + //fill latency and initl + m_tconfig->set_latency(op, latency, initiation_interval); + + //fill addresses + if(mem_width > 0) { + for(unsigned i=0; i<warp_size(); ++i) + set_addr(i, mem_addresses[i]); + } + + + //handle special cases and fill memory space + switch(m_opcode){ + case OP_LDG: + case OP_LDL: + assert(mem_width>0); + //Nvbit reports incorrect data width, and we have to parse the opcode to get the correct data width + data_size = get_datawidth_from_opcode(opcode_tokens); + memory_op = memory_load; + cache_op = CACHE_ALL; + if(m_opcode == OP_LDL) + space.set_type(local_space); + else + space.set_type(global_space); + //check the cache scope, if its strong GPU, then bypass L1 + if(check_opcode_contain( opcode_tokens , "STRONG") && check_opcode_contain( opcode_tokens , "GPU")) { + cache_op = CACHE_GLOBAL; + } + break; + case OP_STG: + case OP_STL: + case OP_ATOMG: + case OP_RED: + case OP_ATOM: + assert(mem_width>0); + data_size = get_datawidth_from_opcode(opcode_tokens); + memory_op = memory_store; + cache_op = CACHE_ALL; + if(m_opcode == OP_STL) + space.set_type(local_space); + else + space.set_type(global_space); + + if(m_opcode == OP_ATOMG || m_opcode == OP_ATOM || m_opcode == OP_RED){ + m_isatomic = true; + memory_op = memory_load; + op=LOAD_OP; + cache_op = CACHE_GLOBAL; + + //ATOMIC writes to the first operand, we missed that in the trace so we fixed it here. TO be fixed in tracer + outcount=reg_dsts_num+1; + out[0]=in[0]; //Increment by one because GPGPU-sim starts from R1, while SASS starts from R0 + arch_reg.dst[0]=reg_srcs[0]; + num_regs = reg_srcs_num+reg_dsts_num+1; + num_operands = num_regs; + + } + + break; + case OP_LDS: + case OP_STS: + case OP_ATOMS: + assert(mem_width>0); + data_size = mem_width; + space.set_type(shared_space); + if(m_opcode == OP_ATOMS || m_opcode == OP_LDS) { + //m_isatomic = true; + op=LOAD_OP; + memory_op = memory_load; + } + break; + case OP_ST: + case OP_LD: + //TO DO: set generic load based on the address + //right now, we consider all loads are shared. + assert(mem_width>0); + data_size = get_datawidth_from_opcode(opcode_tokens); + space.set_type(shared_space); + if(m_opcode == OP_LD) + memory_op = memory_load; + else + memory_op = memory_store; + break; + case OP_BAR: + //TO DO: fill this correctly + bar_id = 0; + bar_count = (unsigned)-1; + bar_type = SYNC; + //TO DO + //if bar_type = RED; + //set bar_type + // barrier_type bar_type; + // reduction_type red_type; + break; + case OP_HADD2: + case OP_HADD2_32I: + case OP_HFMA2: + case OP_HFMA2_32I: + case OP_HMUL2_32I: + case OP_HSET2: + case OP_HSETP2: + initiation_interval = initiation_interval/2; //FP16 has 2X throughput than FP32 + break; + default: + break; + } + + return true; +} + +trace_config::trace_config(gpgpu_sim* m_gpgpu_sim){ + + this->m_gpgpu_sim=m_gpgpu_sim; + parse_config(); +} + +void trace_config::parse_config() +{ + + sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_int, "%u,%u", + &int_latency,&int_init); + sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_sp, "%u,%u", + &fp_latency,&fp_init); + sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_dp, "%u,%u", + &dp_latency,&dp_init); + sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_sfu, "%u,%u", + &sfu_latency,&sfu_init); + sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_tensor, "%u,%u", + &tensor_latency,&tensor_init); + +} +void trace_config::set_latency(unsigned category, unsigned& latency, unsigned& initiation_interval) +{ + initiation_interval = latency = 1; + + switch(category){ + case ALU_OP: + case INTP_OP: + case BRANCH_OP: + case CALL_OPS: + case RET_OPS: + latency = int_latency; + initiation_interval = int_init; + break; + case SP_OP: + latency = fp_latency; + initiation_interval = fp_init; + break; + case DP_OP: + latency = dp_latency; + initiation_interval = dp_init; + break; + case SFU_OP: + latency = sfu_latency; + initiation_interval = sfu_init; + break; + case TENSOR_CORE_OP: + latency = tensor_latency; + initiation_interval = tensor_init; + break; + default: + break; + } + +} + +unsigned trace_shader_core_ctx::trace_sim_inc_thread( kernel_info_t &kernel) +{ + + if ( kernel.no_more_ctas_to_run() ) { + return 0; //finished! + } + + if( kernel.more_threads_in_cta() ) { + kernel.increment_thread_id(); + } + + if( !kernel.more_threads_in_cta() ) + kernel.increment_cta_id(); + + return 1; +} + +void trace_shader_core_ctx::init_traces( unsigned start_warp, unsigned end_warp, kernel_info_t &kernel ) { + + std::vector<std::vector<trace_warp_inst_t>*> threadblock_traces; + for (unsigned i = start_warp; i < end_warp; ++i) { + m_trace_warp[i].clear(); + threadblock_traces.push_back(&(m_trace_warp[i].warp_traces)); + } + trace_kernel_info_t& trace_kernel = static_cast<trace_kernel_info_t&> (kernel); + trace_kernel.get_next_threadblock_traces(threadblock_traces); + + //set pc + for (unsigned i = start_warp; i < end_warp; ++i) { + m_warp[i].set_next_pc(m_trace_warp[i].get_start_pc()); + } +} + + +void trace_shader_core_ctx::checkExecutionStatusAndUpdate(warp_inst_t &inst, unsigned t, unsigned tid) +{ + if(inst.isatomic()) + m_warp[inst.warp_id()].inc_n_atomic(); + + if ( inst.op == EXIT_OPS ) { + m_warp[inst.warp_id()].set_completed(t); + } + +} + +void trace_shader_core_ctx::func_exec_inst( warp_inst_t &inst ) +{ + //here, we generate memory acessess and set the status if thread (done?) + if( inst.is_load() || inst.is_store() ) + { + inst.generate_mem_accesses(); + } + for ( unsigned t=0; t < m_warp_size; t++ ) { + if( inst.active(t) ) { + unsigned warpId = inst.warp_id(); + unsigned tid=m_warp_size*warpId+t; + + //virtual function + checkExecutionStatusAndUpdate(inst,t,tid); + } + } + if(m_trace_warp[inst.warp_id()].trace_done() && m_warp[inst.warp_id()].functional_done()) { + m_warp[inst.warp_id()].ibuffer_flush(); + m_barriers.warp_exit( inst.warp_id() ); + } +} + |