diff options
Diffstat (limited to 'src/gpgpu-sim/gpu-sim.cc')
| -rw-r--r-- | src/gpgpu-sim/gpu-sim.cc | 91 |
1 files changed, 85 insertions, 6 deletions
diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 8bbadf5..ec570bf 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -58,6 +58,7 @@ #include "../debug.h" #include "../gpgpusim_entrypoint.h" #include "../cuda-sim/cuda-sim.h" +#include "../cuda-sim/ptx_ir.h" #include "../trace.h" #include "mem_latency_stat.h" #include "power_stat.h" @@ -284,6 +285,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shader_registers", OPT_UINT32, &gpgpu_shader_registers, "Number of registers per shader core. Limits number of concurrent CTAs. (default 8192)", "8192"); + option_parser_register(opp, "-gpgpu_registers_per_block", OPT_UINT32, &gpgpu_registers_per_block, + "Maximum number of registers per CTA. (default 8192)", + "8192"); option_parser_register(opp, "-gpgpu_ignore_resources_limitation", OPT_BOOL, &gpgpu_ignore_resources_limitation, "gpgpu_ignore_resources_limitation (default 0)", "0"); @@ -305,6 +309,9 @@ 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"); @@ -329,6 +336,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shmem_warp_parts", OPT_INT32, &mem_warp_parts, "Number of portions a warp is divided into for shared memory bank conflict check ", "2"); + option_parser_register(opp, "-mem_unit_ports", OPT_INT32, &mem_unit_ports, + "The number of memory transactions allowed per core cycle", + "1"); option_parser_register(opp, "-gpgpu_shmem_warp_parts", OPT_INT32, &mem_warp_parts, "Number of portions a warp is divided into for shared memory bank conflict check ", "2"); @@ -347,8 +357,14 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_reg_bank_use_warp_id", OPT_BOOL, &gpgpu_reg_bank_use_warp_id, "Use warp ID in mapping registers to banks (default = off)", "0"); + option_parser_register(opp, "-sub_core_model", OPT_BOOL, &sub_core_model, + "Sub Core Volta/Pascal model (default = off)", + "0"); + option_parser_register(opp, "-enable_specialized_operand_collector", OPT_BOOL, &enable_specialized_operand_collector, + "enable_specialized_operand_collector", + "1"); option_parser_register(opp, "-gpgpu_operand_collector_num_units_sp", OPT_INT32, &gpgpu_operand_collector_num_units_sp, - "number of collector units (default = 4)", + "number of collector units (default = 4)", "4"); option_parser_register(opp, "-gpgpu_operand_collector_num_units_dp", OPT_INT32, &gpgpu_operand_collector_num_units_dp, "number of collector units (default = 0)", @@ -356,6 +372,12 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_operand_collector_num_units_sfu", OPT_INT32, &gpgpu_operand_collector_num_units_sfu, "number of collector units (default = 4)", "4"); + option_parser_register(opp, "-gpgpu_operand_collector_num_units_int", OPT_INT32, &gpgpu_operand_collector_num_units_int, + "number of collector units (default = 0)", + "0"); + option_parser_register(opp, "-gpgpu_operand_collector_num_units_tensor_core", OPT_INT32, &gpgpu_operand_collector_num_units_tensor_core, + "number of collector units (default = 4)", + "4"); option_parser_register(opp, "-gpgpu_operand_collector_num_units_mem", OPT_INT32, &gpgpu_operand_collector_num_units_mem, "number of collector units (default = 2)", "2"); @@ -371,6 +393,12 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_operand_collector_num_in_ports_sfu", OPT_INT32, &gpgpu_operand_collector_num_in_ports_sfu, "number of collector unit in ports (default = 1)", "1"); + option_parser_register(opp, "-gpgpu_operand_collector_num_in_ports_int", OPT_INT32, &gpgpu_operand_collector_num_in_ports_int, + "number of collector unit in ports (default = 0)", + "0"); + option_parser_register(opp, "-gpgpu_operand_collector_num_in_ports_tensor_core", OPT_INT32, &gpgpu_operand_collector_num_in_ports_tensor_core, + "number of collector unit in ports (default = 1)", + "1"); option_parser_register(opp, "-gpgpu_operand_collector_num_in_ports_mem", OPT_INT32, &gpgpu_operand_collector_num_in_ports_mem, "number of collector unit in ports (default = 1)", "1"); @@ -386,6 +414,12 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_operand_collector_num_out_ports_sfu", OPT_INT32, &gpgpu_operand_collector_num_out_ports_sfu, "number of collector unit in ports (default = 1)", "1"); + option_parser_register(opp, "-gpgpu_operand_collector_num_out_ports_int", OPT_INT32, &gpgpu_operand_collector_num_out_ports_int, + "number of collector unit in ports (default = 0)", + "0"); + option_parser_register(opp, "-gpgpu_operand_collector_num_out_ports_tensor_core", OPT_INT32, &gpgpu_operand_collector_num_out_ports_tensor_core, + "number of collector unit in ports (default = 1)", + "1"); option_parser_register(opp, "-gpgpu_operand_collector_num_out_ports_mem", OPT_INT32, &gpgpu_operand_collector_num_out_ports_mem, "number of collector unit in ports (default = 1)", "1"); @@ -409,17 +443,26 @@ void shader_core_config::reg_options(class OptionParser * opp) "1"); option_parser_register(opp, "-gpgpu_pipeline_widths", OPT_CSTR, &pipeline_widths_string, "Pipeline widths " - "ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB", - "1,1,1,1,1,1,1,1,1" ); + "ID_OC_SP,ID_OC_DP,ID_OC_INT,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_INT,OC_EX_SFU,OC_EX_MEM,EX_WB,ID_OC_TENSOR_CORE,OC_EX_TENSOR_CORE", + "1,1,1,1,1,1,1,1,1,1,1,1,1" ); + option_parser_register(opp, "-gpgpu_tensor_core_avail", OPT_INT32, &gpgpu_tensor_core_avail, + "Tensor Core Available (default=0)", + "0"); option_parser_register(opp, "-gpgpu_num_sp_units", OPT_INT32, &gpgpu_num_sp_units, "Number of SP units (default=1)", "1"); option_parser_register(opp, "-gpgpu_num_dp_units", OPT_INT32, &gpgpu_num_dp_units, "Number of DP units (default=0)", "0"); + option_parser_register(opp, "-gpgpu_num_int_units", OPT_INT32, &gpgpu_num_int_units, + "Number of INT units (default=0)", + "0"); option_parser_register(opp, "-gpgpu_num_sfu_units", OPT_INT32, &gpgpu_num_sfu_units, "Number of SF units (default=1)", "1"); + option_parser_register(opp, "-gpgpu_num_tensor_core_units", OPT_INT32, &gpgpu_num_tensor_core_units, + "Number of tensor_core units (default=1)", + "1"); option_parser_register(opp, "-gpgpu_num_mem_units", OPT_INT32, &gpgpu_num_mem_units, "Number if ldst units (default=1) WARNING: not hooked up to anything", "1"); @@ -721,11 +764,21 @@ 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; } +int gpgpu_sim::num_registers_per_block() const +{ + return m_shader_config->gpgpu_registers_per_block; +} + int gpgpu_sim::wrp_size() const { return m_shader_config->warp_size; @@ -1178,6 +1231,9 @@ void shader_core_ctx::mem_instruction_stats(const warp_inst_t &inst) case shared_space: m_stats->gpgpu_n_shmem_insn += active_count; break; + case sstarr_space: + m_stats->gpgpu_n_sstarr_insn += active_count; + break; case const_space: m_stats->gpgpu_n_const_insn += active_count; break; @@ -1379,22 +1435,44 @@ void shader_core_ctx::issue_block2core( kernel_info_t &kernel ) // bind functional simulation state of threads to hardware resources (simulation) warp_set_t warps; unsigned nthreads_in_block= 0; + function_info *kernel_func_info = kernel.entry(); + symbol_table * symtab= kernel_func_info->get_symtab(); + unsigned ctaid= kernel.get_next_cta_id_single(); + checkpoint *g_checkpoint= new checkpoint(); for (unsigned i = start_thread; i<end_thread; i++) { m_threadState[i].m_cta_id = free_cta_hw_id; unsigned warp_id = i/m_config->warp_size; nthreads_in_block += ptx_sim_init_thread(kernel,&m_thread[i],m_sid,i,cta_size-(i-start_thread),m_config->n_thread_per_shader,this,free_cta_hw_id,warp_id,m_cluster->get_gpu()); m_threadState[i].m_active = true; + // load thread local memory and register file + if(m_gpu->resume_option==1 && kernel.get_uid()==m_gpu->resume_kernel && ctaid>=m_gpu->resume_CTA && ctaid<m_gpu->checkpoint_CTA_t ) + { + char fname[2048]; + snprintf(fname,2048,"checkpoint_files/thread_%d_%d_reg.txt",i%cta_size,ctaid ); + m_thread[i]->resume_reg_thread(fname,symtab); + char f1name[2048]; + snprintf(f1name,2048,"checkpoint_files/local_mem_thread_%d_%d_reg.txt",i%cta_size,ctaid); + g_checkpoint->load_global_mem(m_thread[i]->m_local_mem, f1name); + } + // warps.set( warp_id ); } assert( nthreads_in_block > 0 && nthreads_in_block <= m_config->n_thread_per_shader); // should be at least one, but less than max m_cta_status[free_cta_hw_id]=nthreads_in_block; + if(m_gpu->resume_option==1 && kernel.get_uid()==m_gpu->resume_kernel && ctaid>=m_gpu->resume_CTA && ctaid<m_gpu->checkpoint_CTA_t ) + { + char f1name[2048]; + snprintf(f1name,2048,"checkpoint_files/shared_mem_%d.txt", ctaid); + + g_checkpoint->load_global_mem(m_thread[start_thread]->m_shared_mem, f1name); + } // now that we know which warps are used in this CTA, we can allocate // resources for use in CTA-wide barrier operations m_barriers.allocate_barrier(free_cta_hw_id,warps); // initialize the SIMT stacks and fetch hardware - init_warps( free_cta_hw_id, start_thread, end_thread); + init_warps( free_cta_hw_id, start_thread, end_thread, ctaid, cta_size, kernel.get_uid()); m_n_active_cta++; shader_CTA_count_log(m_sid, 1); @@ -1553,7 +1631,8 @@ void gpgpu_sim::cycle() if( g_single_step && ((gpu_sim_cycle+gpu_tot_sim_cycle) >= g_single_step) ) { raise(SIGTRAP); // Debug breakpoint } - gpu_sim_cycle++; + gpu_sim_cycle++; + if( g_interactive_debugger_enabled ) gpgpu_debug(); @@ -1643,7 +1722,7 @@ void gpgpu_sim::cycle() } } - if (!(gpu_sim_cycle % 20000)) { + if (!(gpu_sim_cycle % 50000)) { // deadlock detection if (m_config.gpu_deadlock_detect && gpu_sim_insn == last_gpu_sim_insn) { gpu_deadlock = true; |