diff options
Diffstat (limited to 'src/cuda-sim/cuda-sim.cc')
| -rw-r--r-- | src/cuda-sim/cuda-sim.cc | 42 |
1 files changed, 29 insertions, 13 deletions
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index d4ace76..b5b79e7 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -69,9 +69,9 @@ unsigned cdp_latency[5]; void ptx_opcocde_latency_options (option_parser_t opp) { option_parser_register(opp, "-ptx_opcode_latency_int", OPT_CSTR, &opcode_latency_int, - "Opcode latencies for integers <ADD,MAX,MUL,MAD,DIV>" - "Default 1,1,19,25,145", - "1,1,19,25,145"); + "Opcode latencies for integers <ADD,MAX,MUL,MAD,DIV,BSMAD>" + "Default 1,1,19,25,145,1", + "1,1,19,25,145,1"); option_parser_register(opp, "-ptx_opcode_latency_fp", OPT_CSTR, &opcode_latency_fp, "Opcode latencies for single precision floating points <ADD,MAX,MUL,MAD,DIV>" "Default 1,1,1,1,30", @@ -81,9 +81,9 @@ void ptx_opcocde_latency_options (option_parser_t opp) { "Default 8,8,8,8,335", "8,8,8,8,335"); option_parser_register(opp, "-ptx_opcode_initiation_int", OPT_CSTR, &opcode_initiation_int, - "Opcode initiation intervals for integers <ADD,MAX,MUL,MAD,DIV>" - "Default 1,1,4,4,32", - "1,1,4,4,32"); + "Opcode initiation intervals for integers <ADD,MAX,MUL,MAD,DIV,BSMAD>" + "Default 1,1,4,4,32,1", + "1,1,4,4,32,1"); option_parser_register(opp, "-ptx_opcode_initiation_fp", OPT_CSTR, &opcode_initiation_fp, "Opcode initiation intervals for single precision floating points <ADD,MAX,MUL,MAD,DIV>" "Default 1,1,1,1,5", @@ -581,15 +581,18 @@ void ptx_instruction::set_bar_type() abort(); } } + else if(m_opcode==SST_OP) { + bar_type = SYNC; + } } void ptx_instruction::set_opcode_and_latency() { - unsigned int_latency[5]; + unsigned int_latency[6]; unsigned fp_latency[5]; unsigned dp_latency[5]; - unsigned int_init[5]; + unsigned int_init[6]; unsigned fp_init[5]; unsigned dp_init[5]; /* @@ -598,19 +601,20 @@ void ptx_instruction::set_opcode_and_latency() * [2] MUL * [3] MAD * [4] DIV + * [5] BSMAD */ - sscanf(opcode_latency_int, "%u,%u,%u,%u,%u", + sscanf(opcode_latency_int, "%u,%u,%u,%u,%u,%u", &int_latency[0],&int_latency[1],&int_latency[2], - &int_latency[3],&int_latency[4]); + &int_latency[3],&int_latency[4],&int_latency[5]); sscanf(opcode_latency_fp, "%u,%u,%u,%u,%u", &fp_latency[0],&fp_latency[1],&fp_latency[2], &fp_latency[3],&fp_latency[4]); sscanf(opcode_latency_dp, "%u,%u,%u,%u,%u", &dp_latency[0],&dp_latency[1],&dp_latency[2], &dp_latency[3],&dp_latency[4]); - sscanf(opcode_initiation_int, "%u,%u,%u,%u,%u", + sscanf(opcode_initiation_int, "%u,%u,%u,%u,%u,%u", &int_init[0],&int_init[1],&int_init[2], - &int_init[3],&int_init[4]); + &int_init[3],&int_init[4],&int_init[5]); sscanf(opcode_initiation_fp, "%u,%u,%u,%u,%u", &fp_init[0],&fp_init[1],&fp_init[2], &fp_init[3],&fp_init[4]); @@ -647,6 +651,7 @@ void ptx_instruction::set_opcode_and_latency() case TEX_OP: op = LOAD_OP; mem_op=TEX; break; case ATOM_OP: op = LOAD_OP; break; case BAR_OP: op = BARRIER_OP; break; + case SST_OP: op = BARRIER_OP; break; case MEMBAR_OP: op = MEMORY_BARRIER_OP; break; case CALL_OP: { @@ -783,6 +788,10 @@ void ptx_instruction::set_opcode_and_latency() initiation_interval = dp_init[2]; op = SFU_OP; break; + case BSMAD_OP: + latency = int_latency[5]; + initiation_interval = int_init[5]; + break; case SHFL_OP: latency = 32; initiation_interval = 15; @@ -1459,6 +1468,7 @@ unsigned ptx_sim_init_thread( kernel_info_t &kernel, std::list<ptx_thread_info *> &active_threads = kernel.active_threads(); static std::map<unsigned,memory_space*> shared_memory_lookup; + static std::map<unsigned,memory_space*> sstarr_memory_lookup; static std::map<unsigned,ptx_cta_info*> ptx_cta_lookup; static std::map<unsigned,ptx_warp_info*> ptx_warp_lookup; static std::map<unsigned,std::map<unsigned,memory_space*> > local_memory_lookup; @@ -1503,6 +1513,7 @@ unsigned ptx_sim_init_thread( kernel_info_t &kernel, //initializing new CTA ptx_cta_info *cta_info = NULL; memory_space *shared_mem = NULL; + memory_space *sstarr_mem = NULL; unsigned cta_size = kernel.threads_per_cta(); unsigned max_cta_per_sm = num_threads/cta_size; // e.g., 256 / 48 = 5 @@ -1520,6 +1531,9 @@ unsigned ptx_sim_init_thread( kernel_info_t &kernel, snprintf(buf,512,"shared_%u", sid); shared_mem = new memory_space_impl<16*1024>(buf,4); shared_memory_lookup[sm_idx] = shared_mem; + snprintf(buf,512,"sstarr_%u", sid); + sstarr_mem = new memory_space_impl<16*1024>(buf,4); + sstarr_memory_lookup[sm_idx] = sstarr_mem; cta_info = new ptx_cta_info(sm_idx); ptx_cta_lookup[sm_idx] = cta_info; } else { @@ -1528,6 +1542,7 @@ unsigned ptx_sim_init_thread( kernel_info_t &kernel, sm_idx, sid, max_cta_per_sm ); } shared_mem = shared_memory_lookup[sm_idx]; + sstarr_mem = sstarr_memory_lookup[sm_idx]; cta_info = ptx_cta_lookup[sm_idx]; cta_info->check_cta_thread_status_and_reset(); } @@ -1540,7 +1555,6 @@ unsigned ptx_sim_init_thread( kernel_info_t &kernel, kernel.increment_thread_id(); new_tid += tid; ptx_thread_info *thd = new ptx_thread_info(kernel); - ptx_warp_info *warp_info = NULL; if ( ptx_warp_lookup.find(hw_warp_id) == ptx_warp_lookup.end() ) { warp_info = new ptx_warp_info(); @@ -1569,9 +1583,11 @@ unsigned ptx_sim_init_thread( kernel_info_t &kernel, thd->cpy_tid_to_reg(tid3d); thd->set_valid(); thd->m_shared_mem = shared_mem; + thd->m_sstarr_mem = sstarr_mem; function_info *finfo = thd->func_info(); symbol_table *st = finfo->get_symtab(); thd->func_info()->param_to_shared(thd->m_shared_mem,st); + thd->func_info()->param_to_shared(thd->m_sstarr_mem,st); thd->m_cta_info = cta_info; cta_info->add_thread(thd); thd->m_local_mem = local_mem; |