From ab0d5d829d3f825736c5f91a365e755b2c2752c9 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Tue, 23 Oct 2018 11:14:22 -0400 Subject: Moving configs to folders with better names --- .../tested-cfgs/SM2_GTX480/config_fermi_islip.icnt | 70 +++ configs/tested-cfgs/SM2_GTX480/gpgpusim.config | 153 ++++++ .../tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml | 538 +++++++++++++++++++++ 3 files changed, 761 insertions(+) create mode 100644 configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt create mode 100644 configs/tested-cfgs/SM2_GTX480/gpgpusim.config create mode 100755 configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml (limited to 'configs/tested-cfgs/SM2_GTX480') diff --git a/configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt b/configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt new file mode 100644 index 0000000..c399db9 --- /dev/null +++ b/configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 27; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 64; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config new file mode 100644 index 0000000..4096b09 --- /dev/null +++ b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config @@ -0,0 +1,153 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 15 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 700.0:700.0:700.0:924.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 20 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# 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 +#For Fermi, DP unit =0, DP inst is executed on SFU +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,S:64:8,8 +-gpgpu_shmem_size 49152 +-icnt_flit_size 40 +-gmem_skip_L1D 0 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 35 +-smem_latency 26 +-gpgpu_flush_l1_cache 1 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,S:64:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 S:64:128:8,L:B:m:L:L,A:256:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 1 +-memory_partition_indexing 0 + +-gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,S:2:32,4 +-gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,T:128:4,128:2 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,S:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 20 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml b/configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + -- cgit v1.3 From 111cca2a061fe4f247be930cb44fdcdaec2b59f5 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 7 Nov 2018 20:30:02 -0500 Subject: Adding INT unit, fixing tensor core latency, updating config files --- configs/tested-cfgs/SM2_GTX480/gpgpusim.config | 7 +- configs/tested-cfgs/SM6_TITANX/gpgpusim.config | 7 +- configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 8 +- src/abstract_hardware_model.h | 3 + src/cuda-sim/cuda-sim.cc | 35 ++++-- src/gpgpu-sim/gpu-sim.cc | 16 ++- src/gpgpu-sim/shader.cc | 165 +++++++++++++++++++------ src/gpgpu-sim/shader.h | 57 +++++++-- 8 files changed, 232 insertions(+), 66 deletions(-) (limited to 'configs/tested-cfgs/SM2_GTX480') diff --git a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config index 4096b09..d71b2fd 100644 --- a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config @@ -30,9 +30,10 @@ -gpgpu_simd_model 1 # Pipeline widths and number of FUs -# 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 -#For Fermi, DP unit =0, DP inst is executed on SFU --gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 +# 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 +# For Fermi, DP unit =0, DP inst is executed on SFU +# For Fermi, INT unit =0, INT inst is executed on SP +-gpgpu_pipeline_widths 2,0,0,1,1,2,0,0,1,1,2 -gpgpu_num_sp_units 2 -gpgpu_num_sfu_units 1 -gpgpu_num_dp_units 0 diff --git a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config index ed35531..cb23ab3 100644 --- a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config @@ -35,10 +35,11 @@ -gpgpu_simd_model 1 # Pipeline widths and number of FUs -# 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 -## Pascal GP102 has 4 SP SIMD units and 4 SFU units +# 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 +## Pascal GP102 has 4 SP SIMD units and 4 SFU units per SM. In this config, we split SM into two shader cores, each has 2 SPs and 2 SFUs +# There is no int unit in Pascal ## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 +-gpgpu_pipeline_widths 2,1,0,2,1,2,1,0,2,1,5 -gpgpu_num_sp_units 2 -gpgpu_num_sfu_units 2 -gpgpu_num_dp_units 1 diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index 7b0369a..e143f31 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -41,13 +41,14 @@ -gpgpu_simd_model 1 # Pipeline widths and number of FUs -# 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,ID_OC_TENSOR_CORE,OC_EX_TENSOR_CORE +# 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 ## Volta GV100 has 4 SP SIMD units, 4 SFU units, 4 DP units per core ## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,4,4,4,4,4,4,8,4,4 +-gpgpu_pipeline_widths 4,4,4,4,4,4,4,4,4,4,8,4,4 -gpgpu_num_sp_units 4 -gpgpu_num_sfu_units 4 -gpgpu_num_dp_units 4 +-gpgpu_num_int_units 4 -gpgpu_tensor_core_avail 1 -gpgpu_num_tensor_core_units 4 @@ -64,7 +65,8 @@ -ptx_opcode_initiation_dp 4,4,4,4,130 -ptx_opcode_latency_sfu 100 -ptx_opcode_initiation_sfu 8 - +-ptx_opcode_latency_tesnor 64 +-ptx_opcode_initiation_tensor 64 # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index edd48a1..b65df81 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -83,6 +83,8 @@ enum uarch_op_t { SFU_OP, TENSOR_CORE_OP, DP_OP, + SP_OP, + INTP_OP, ALU_SFU_OP, LOAD_OP, TENSOR_CORE_LOAD_OP, @@ -141,6 +143,7 @@ enum operation_pipeline_t { UNKOWN_OP, SP__OP, DP__OP, + INTP__OP, SFU__OP, TENSOR_CORE__OP, MEM__OP diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index c9ac8cb..7dc20cd 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -62,8 +62,8 @@ addr_t g_debug_pc = 0xBEEF1518; unsigned g_ptx_sim_num_insn = 0; unsigned gpgpu_param_num_shaders = 0; -char *opcode_latency_int, *opcode_latency_fp, *opcode_latency_dp,*opcode_latency_sfu; -char *opcode_initiation_int, *opcode_initiation_fp, *opcode_initiation_dp,*opcode_initiation_sfu; +char *opcode_latency_int, *opcode_latency_fp, *opcode_latency_dp,*opcode_latency_sfu,*opcode_latency_tensor; +char *opcode_initiation_int, *opcode_initiation_fp, *opcode_initiation_dp,*opcode_initiation_sfu,*opcode_initiation_tensor; char *cdp_latency_str; unsigned cdp_latency[5]; @@ -84,6 +84,10 @@ void ptx_opcocde_latency_options (option_parser_t opp) { "Opcode latencies for SFU instructions" "Default 8", "8"); + option_parser_register(opp, "-ptx_opcode_latency_tesnor", OPT_CSTR, &opcode_latency_tensor, + "Opcode latencies for Tensor instructions" + "Default 64", + "64"); option_parser_register(opp, "-ptx_opcode_initiation_int", OPT_CSTR, &opcode_initiation_int, "Opcode initiation intervals for integers " "Default 1,1,4,4,32", @@ -100,6 +104,10 @@ void ptx_opcocde_latency_options (option_parser_t opp) { "Opcode initiation intervals for sfu instructions" "Default 8", "8"); + option_parser_register(opp, "-ptx_opcode_initiation_tensor", OPT_CSTR, &opcode_initiation_tensor, + "Opcode initiation intervals for tensor instructions" + "Default 64", + "64"); option_parser_register(opp, "-cdp_latency", OPT_CSTR, &cdp_latency_str, "CDP API latency sub_core_model) { //in subcore model, each scheduler should has its own issue register, so num scheduler = reg width assert(m_config->gpgpu_num_sched_per_core == m_pipeline_reg[ID_OC_SP].get_size() ); - if(m_config->gpgpu_num_dp_units > 0) - assert(m_config->gpgpu_num_sched_per_core == m_pipeline_reg[ID_OC_DP].get_size() ); assert(m_config->gpgpu_num_sched_per_core == m_pipeline_reg[ID_OC_SFU].get_size() ); assert(m_config->gpgpu_num_sched_per_core == m_pipeline_reg[ID_OC_MEM].get_size() ); - assert(m_config->gpgpu_num_sched_per_core == m_pipeline_reg[ID_OC_TENSOR_CORE].get_size() ); + if(m_config->gpgpu_tensor_core_avail) + assert(m_config->gpgpu_num_sched_per_core == m_pipeline_reg[ID_OC_TENSOR_CORE].get_size() ); + if(m_config->gpgpu_num_dp_units > 0) + assert(m_config->gpgpu_num_sched_per_core == m_pipeline_reg[ID_OC_DP].get_size() ); + if(m_config->gpgpu_num_int_units > 0) + assert(m_config->gpgpu_num_sched_per_core == m_pipeline_reg[ID_OC_INT].get_size() ); } m_threadState = (thread_ctx_t*) calloc(sizeof(thread_ctx_t), config->n_thread_per_shader); @@ -161,6 +164,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, &m_pipeline_reg[ID_OC_SP], &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], + &m_pipeline_reg[ID_OC_INT], &m_pipeline_reg[ID_OC_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], i @@ -177,6 +181,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, &m_pipeline_reg[ID_OC_SP], &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], + &m_pipeline_reg[ID_OC_INT], &m_pipeline_reg[ID_OC_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], i, @@ -194,6 +199,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, &m_pipeline_reg[ID_OC_SP], &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], + &m_pipeline_reg[ID_OC_INT], &m_pipeline_reg[ID_OC_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], i @@ -210,7 +216,8 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, &m_pipeline_reg[ID_OC_SP], &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], - &m_pipeline_reg[ID_OC_TENSOR_CORE], + &m_pipeline_reg[ID_OC_INT], + &m_pipeline_reg[ID_OC_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], i ) @@ -226,6 +233,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, &m_pipeline_reg[ID_OC_SP], &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], + &m_pipeline_reg[ID_OC_INT], &m_pipeline_reg[ID_OC_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], i, @@ -248,7 +256,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, //op collector configuration - enum { SP_CUS, DP_CUS, SFU_CUS, TENSOR_CORE_CUS, MEM_CUS, GEN_CUS }; + enum { SP_CUS, DP_CUS, SFU_CUS, TENSOR_CORE_CUS, INT_CUS, MEM_CUS, GEN_CUS }; opndcoll_rfu_t::port_vector_t in_ports; opndcoll_rfu_t::port_vector_t out_ports; @@ -260,16 +268,22 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, for (unsigned i = 0; i < m_config->gpgpu_operand_collector_num_in_ports_gen; i++) { in_ports.push_back(&m_pipeline_reg[ID_OC_SP]); in_ports.push_back(&m_pipeline_reg[ID_OC_SFU]); - in_ports.push_back(&m_pipeline_reg[ID_OC_TENSOR_CORE]); in_ports.push_back(&m_pipeline_reg[ID_OC_MEM]); out_ports.push_back(&m_pipeline_reg[OC_EX_SP]); out_ports.push_back(&m_pipeline_reg[OC_EX_SFU]); - out_ports.push_back(&m_pipeline_reg[OC_EX_TENSOR_CORE]); out_ports.push_back(&m_pipeline_reg[OC_EX_MEM]); + if(m_config->gpgpu_tensor_core_avail) { + in_ports.push_back(&m_pipeline_reg[ID_OC_TENSOR_CORE]); + out_ports.push_back(&m_pipeline_reg[OC_EX_TENSOR_CORE]); + } if(m_config->gpgpu_num_dp_units > 0) { in_ports.push_back(&m_pipeline_reg[ID_OC_DP]); out_ports.push_back(&m_pipeline_reg[OC_EX_DP]); } + if(m_config->gpgpu_num_int_units > 0) { + in_ports.push_back(&m_pipeline_reg[ID_OC_INT]); + out_ports.push_back(&m_pipeline_reg[OC_EX_INT]); + } cu_sets.push_back((unsigned)GEN_CUS); m_operand_collector.add_port(in_ports,out_ports,cu_sets); in_ports.clear(),out_ports.clear(),cu_sets.clear(); @@ -281,6 +295,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_operand_collector.add_cu_set(TENSOR_CORE_CUS, config->gpgpu_operand_collector_num_units_tensor_core, config->gpgpu_operand_collector_num_out_ports_tensor_core); m_operand_collector.add_cu_set(SFU_CUS, m_config->gpgpu_operand_collector_num_units_sfu, m_config->gpgpu_operand_collector_num_out_ports_sfu); m_operand_collector.add_cu_set(MEM_CUS, m_config->gpgpu_operand_collector_num_units_mem, m_config->gpgpu_operand_collector_num_out_ports_mem); + m_operand_collector.add_cu_set(INT_CUS, m_config->gpgpu_operand_collector_num_units_int, m_config->gpgpu_operand_collector_num_out_ports_int); for (unsigned i = 0; i < m_config->gpgpu_operand_collector_num_in_ports_sp; i++) { in_ports.push_back(&m_pipeline_reg[ID_OC_SP]); @@ -326,11 +341,20 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_operand_collector.add_port(in_ports,out_ports,cu_sets); in_ports.clear(),out_ports.clear(),cu_sets.clear(); } + + for (unsigned i = 0; i < m_config->gpgpu_operand_collector_num_in_ports_int; i++) { + in_ports.push_back(&m_pipeline_reg[ID_OC_INT]); + out_ports.push_back(&m_pipeline_reg[OC_EX_INT]); + cu_sets.push_back((unsigned)INT_CUS); + cu_sets.push_back((unsigned)GEN_CUS); + m_operand_collector.add_port(in_ports,out_ports,cu_sets); + in_ports.clear(),out_ports.clear(),cu_sets.clear(); + } } m_operand_collector.init( m_config->gpgpu_num_reg_banks, this ); - m_num_function_units = m_config->gpgpu_num_sp_units + m_config->gpgpu_num_dp_units + m_config->gpgpu_num_sfu_units + m_config->gpgpu_num_tensor_core_units + 1; // sp_unit, sfu, ldst_unit + m_num_function_units = m_config->gpgpu_num_sp_units + m_config->gpgpu_num_dp_units + m_config->gpgpu_num_sfu_units + m_config->gpgpu_num_tensor_core_units + m_config->gpgpu_num_int_units + 1; // sp_unit, sfu, dp, tensor, int, ldst_unit //m_dispatch_port = new enum pipeline_stage_name_t[ m_num_function_units ]; //m_issue_port = new enum pipeline_stage_name_t[ m_num_function_units ]; @@ -347,6 +371,11 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_dispatch_port.push_back(ID_OC_DP); m_issue_port.push_back(OC_EX_DP); } + for (int k = 0; k < m_config->gpgpu_num_int_units; k++) { + m_fu.push_back(new int_unit( &m_pipeline_reg[EX_WB], m_config, this )); + m_dispatch_port.push_back(ID_OC_INT); + m_issue_port.push_back(OC_EX_INT); + } for (int k = 0; k < m_config->gpgpu_num_sfu_units; k++) { m_fu.push_back(new sfu( &m_pipeline_reg[EX_WB], m_config, this )); @@ -359,6 +388,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_dispatch_port.push_back(ID_OC_TENSOR_CORE); m_issue_port.push_back(OC_EX_TENSOR_CORE); } + m_ldst_unit = new ldst_unit( m_icnt, m_mem_fetch_allocator, this, &m_operand_collector, m_scoreboard, config, mem_config, stats, shader_id, tpc_id ); m_fu.push_back(m_ldst_unit); m_dispatch_port.push_back(ID_OC_MEM); @@ -958,8 +988,8 @@ void scheduler_unit::cycle() unsigned checked=0; unsigned issued=0; exec_unit_type_t previous_issued_inst_exec_type = exec_unit_type_t::NONE; - unsigned max_issue = m_shader->m_config->gpgpu_max_insn_issue_per_warp; - bool diff_exec_units = m_shader->m_config->gpgpu_dual_issue_diff_exec_units; + unsigned max_issue = m_shader->m_config->gpgpu_max_insn_issue_per_warp; + bool diff_exec_units = m_shader->m_config->gpgpu_dual_issue_diff_exec_units; //In tis mode, we only allow dual issue to diff execution units (as in Maxwell and Pascal) while( !warp(warp_id).waiting() && !warp(warp_id).ibuffer_empty() && (checked < max_issue) && (checked <= issued) && (issued < max_issue) ) { const warp_inst_t *pI = warp(warp_id).ibuffer_next_inst(); @@ -1008,33 +1038,62 @@ void scheduler_unit::cycle() bool sfu_pipe_avail = m_sfu_out->has_free(m_shader->m_config->sub_core_model, m_id); bool tensor_core_pipe_avail = m_tensor_core_out->has_free(m_shader->m_config->sub_core_model, m_id); bool dp_pipe_avail = m_dp_out->has_free(m_shader->m_config->sub_core_model, m_id); - if( sp_pipe_avail && (pI->op != TENSOR_CORE_OP) && (pI->op != SFU_OP && pI->op != DP_OP) && (!diff_exec_units || previous_issued_inst_exec_type != exec_unit_type_t::SP)) { - - //Jin: special for CDP api - if(pI->m_is_cdp && !warp(warp_id).m_cdp_dummy) { - assert(warp(warp_id).m_cdp_latency == 0); - - extern unsigned cdp_latency[5]; - if(pI->m_is_cdp == 1) - warp(warp_id).m_cdp_latency = cdp_latency[pI->m_is_cdp - 1]; - else //cudaLaunchDeviceV2 and cudaGetParameterBufferV2 - warp(warp_id).m_cdp_latency = cdp_latency[pI->m_is_cdp - 1] - + cdp_latency[pI->m_is_cdp] * active_mask.count(); - warp(warp_id).m_cdp_dummy = true; - break; - } - else if(pI->m_is_cdp && warp(warp_id).m_cdp_dummy) { - assert(warp(warp_id).m_cdp_latency == 0); - warp(warp_id).m_cdp_dummy = false; - } + bool int_pipe_avail = m_int_out->has_free(m_shader->m_config->sub_core_model, m_id); - // always prefer SP pipe for operations that can use both SP and SFU pipelines - m_shader->issue_warp(*m_sp_out,pI,active_mask,warp_id,m_id); - issued++; - issued_inst=true; - warp_inst_issued = true; - previous_issued_inst_exec_type = exec_unit_type_t::SP; - } else if ( (m_shader->m_config->gpgpu_num_dp_units != 0) && (pI->op == DP_OP) && (!diff_exec_units || previous_issued_inst_exec_type != exec_unit_type_t::DP)) { + //This code need to be refactored + if(pI->op != TENSOR_CORE_OP && pI->op != SFU_OP && pI->op != DP_OP) { + + bool execute_on_SP = false; + bool execute_on_INT = false; + + //if INT unit pipline exist, then execute ALU and INT operations on INT unit and SP-FPU on SP unit (like in Volta) + //if INT unit pipline does not exist, then execute all ALU, INT and SP operations on SP unit (as in Fermi, Pascal GPUs) + if(m_shader->m_config->gpgpu_num_int_units > 0 && + int_pipe_avail && + pI->op != SP_OP && + !(diff_exec_units && previous_issued_inst_exec_type == exec_unit_type_t::INT)) + execute_on_INT = true; + else if (sp_pipe_avail && + (m_shader->m_config->gpgpu_num_int_units == 0 || + (m_shader->m_config->gpgpu_num_int_units > 0 && pI->op == SP_OP)) && + !(diff_exec_units && previous_issued_inst_exec_type == exec_unit_type_t::SP) ) + execute_on_SP = true; + + + if(execute_on_INT || execute_on_SP) { + //Jin: special for CDP api + if(pI->m_is_cdp && !warp(warp_id).m_cdp_dummy) { + assert(warp(warp_id).m_cdp_latency == 0); + + extern unsigned cdp_latency[5]; + if(pI->m_is_cdp == 1) + warp(warp_id).m_cdp_latency = cdp_latency[pI->m_is_cdp - 1]; + else //cudaLaunchDeviceV2 and cudaGetParameterBufferV2 + warp(warp_id).m_cdp_latency = cdp_latency[pI->m_is_cdp - 1] + + cdp_latency[pI->m_is_cdp] * active_mask.count(); + warp(warp_id).m_cdp_dummy = true; + break; + } + else if(pI->m_is_cdp && warp(warp_id).m_cdp_dummy) { + assert(warp(warp_id).m_cdp_latency == 0); + warp(warp_id).m_cdp_dummy = false; + } + } + + if(execute_on_SP) { + m_shader->issue_warp(*m_sp_out,pI,active_mask,warp_id,m_id); + issued++; + issued_inst=true; + warp_inst_issued = true; + previous_issued_inst_exec_type = exec_unit_type_t::SP; + } else if (execute_on_INT) { + m_shader->issue_warp(*m_int_out,pI,active_mask,warp_id,m_id); + issued++; + issued_inst=true; + warp_inst_issued = true; + previous_issued_inst_exec_type = exec_unit_type_t::INT; + } + } else if ( (m_shader->m_config->gpgpu_num_dp_units > 0) && (pI->op == DP_OP) && !(diff_exec_units && previous_issued_inst_exec_type == exec_unit_type_t::DP)) { if( dp_pipe_avail ) { m_shader->issue_warp(*m_dp_out,pI,active_mask,warp_id,m_id); issued++; @@ -1042,8 +1101,8 @@ void scheduler_unit::cycle() warp_inst_issued = true; previous_issued_inst_exec_type = exec_unit_type_t::DP; } - } //If the DP units = 0 (like in Fermi archi), then change DP inst to SFU inst - else if ( ((m_shader->m_config->gpgpu_num_dp_units == 0 && pI->op == DP_OP) || (pI->op == SFU_OP) || (pI->op == ALU_SFU_OP)) && (!diff_exec_units || previous_issued_inst_exec_type != exec_unit_type_t::SFU)) { + } //If the DP units = 0 (like in Fermi archi), then execute DP inst on SFU unit + else if ( ((m_shader->m_config->gpgpu_num_dp_units == 0 && pI->op == DP_OP) || (pI->op == SFU_OP) || (pI->op == ALU_SFU_OP)) && !(diff_exec_units && previous_issued_inst_exec_type == exec_unit_type_t::SFU)) { if( sfu_pipe_avail ) { m_shader->issue_warp(*m_sfu_out,pI,active_mask,warp_id,m_id); issued++; @@ -1052,12 +1111,13 @@ void scheduler_unit::cycle() previous_issued_inst_exec_type = exec_unit_type_t::SFU; } } - else if ( (pI->op == TENSOR_CORE_OP) ) { + else if ( (pI->op == TENSOR_CORE_OP) && !(diff_exec_units && previous_issued_inst_exec_type == exec_unit_type_t::SP) ) { if( tensor_core_pipe_avail ) { m_shader->issue_warp(*m_tensor_core_out,pI,active_mask,warp_id,m_id); issued++; issued_inst=true; warp_inst_issued = true; + previous_issued_inst_exec_type = exec_unit_type_t::TENSOR; } } }//end of else @@ -1245,11 +1305,12 @@ swl_scheduler::swl_scheduler ( shader_core_stats* stats, shader_core_ctx* shader register_set* sp_out, register_set* dp_out, register_set* sfu_out, + register_set* int_out, register_set* tensor_core_out, register_set* mem_out, int id, char* config_string ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ) + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, int_out, tensor_core_out, mem_out, id ) { unsigned m_prioritization_readin; int ret = sscanf( config_string, @@ -1822,6 +1883,7 @@ void ldst_unit::active_lanes_in_pipeline(){ assert(active_count<=m_core->get_config()->warp_size); m_core->incfumemactivelanes_stat(active_count); } + void sp_unit::active_lanes_in_pipeline(){ unsigned active_count=pipelined_simd_unit::get_active_lanes_in_pipeline(); assert(active_count<=m_core->get_config()->warp_size); @@ -1837,6 +1899,13 @@ void dp_unit::active_lanes_in_pipeline(){ m_core->incfumemactivelanes_stat(active_count); } +void int_unit::active_lanes_in_pipeline(){ + unsigned active_count=pipelined_simd_unit::get_active_lanes_in_pipeline(); + assert(active_count<=m_core->get_config()->warp_size); + m_core->incspactivelanes_stat(active_count); + m_core->incfuactivelanes_stat(active_count); + m_core->incfumemactivelanes_stat(active_count); +} void sfu::active_lanes_in_pipeline(){ unsigned active_count=pipelined_simd_unit::get_active_lanes_in_pipeline(); assert(active_count<=m_core->get_config()->warp_size); @@ -1866,6 +1935,12 @@ dp_unit::dp_unit( register_set* result_port, const shader_core_config *config,sh m_name = "DP "; } +int_unit::int_unit( register_set* result_port, const shader_core_config *config,shader_core_ctx *core) + : pipelined_simd_unit(result_port,config,config->max_sp_latency,core) +{ + m_name = "INT "; +} + void sp_unit :: issue(register_set& source_reg) { warp_inst_t** ready_reg = source_reg.get_ready(); @@ -1884,6 +1959,15 @@ void dp_unit :: issue(register_set& source_reg) pipelined_simd_unit::issue(source_reg); } +void int_unit :: issue(register_set& source_reg) +{ + warp_inst_t** ready_reg = source_reg.get_ready(); + //m_core->incexecstat((*ready_reg)); + (*ready_reg)->op_pipe=INTP__OP; + m_core->incsp_stat(m_core->get_config()->warp_size,(*ready_reg)->latency); + pipelined_simd_unit::issue(source_reg); +} + pipelined_simd_unit::pipelined_simd_unit( register_set* result_port, const shader_core_config *config, unsigned max_latency,shader_core_ctx *core ) : simd_function_unit(config) { @@ -2006,6 +2090,7 @@ ldst_unit::ldst_unit( mem_fetch_interface *icnt, l1_latency_queue.push_back((mem_fetch*)NULL); } } + m_name = "MEM "; } ldst_unit::ldst_unit( mem_fetch_interface *icnt, diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 7b33c14..45cf8e4 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -76,7 +76,9 @@ enum exec_unit_type_t SP = 1, SFU = 2, MEM = 3, - DP = 4 + DP = 4, + INT = 5, + TENSOR = 6 }; class thread_ctx_t { @@ -328,12 +330,13 @@ public: register_set* sp_out, register_set* dp_out, register_set* sfu_out, + register_set* int_out, register_set* tensor_core_out, register_set* mem_out, int id) : m_supervised_warps(), m_stats(stats), m_shader(shader), m_scoreboard(scoreboard), m_simt_stack(simt), /*m_pipeline_reg(pipe_regs),*/ m_warp(warp), - m_sp_out(sp_out),m_dp_out(dp_out),m_sfu_out(sfu_out),m_tensor_core_out(tensor_core_out),m_mem_out(mem_out), m_id(id){} + m_sp_out(sp_out),m_dp_out(dp_out),m_sfu_out(sfu_out),m_int_out(int_out),m_tensor_core_out(tensor_core_out),m_mem_out(mem_out), m_id(id){} virtual ~scheduler_unit(){} virtual void add_supervised_warp_id(int i) { m_supervised_warps.push_back(&warp(i)); @@ -409,6 +412,7 @@ protected: register_set* m_sp_out; register_set* m_dp_out; register_set* m_sfu_out; + register_set* m_int_out; register_set* m_tensor_core_out; register_set* m_mem_out; @@ -423,10 +427,11 @@ public: register_set* sp_out, register_set* dp_out, register_set* sfu_out, + register_set* int_out, register_set* tensor_core_out, register_set* mem_out, int id ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ){} + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, int_out, tensor_core_out, mem_out, id ){} virtual ~lrr_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -442,10 +447,11 @@ public: register_set* sp_out, register_set* dp_out, register_set* sfu_out, + register_set* int_out, register_set* tensor_core_out, register_set* mem_out, int id ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ){} + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, int_out, tensor_core_out, mem_out, id ){} virtual ~gto_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -462,10 +468,11 @@ public: register_set* sp_out, register_set* dp_out, register_set* sfu_out, - register_set* tensor_core_out, + register_set* int_out, + register_set* tensor_core_out, register_set* mem_out, int id ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ){} + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, int_out, tensor_core_out, mem_out, id ){} virtual ~oldest_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -482,11 +489,12 @@ public: register_set* sp_out, register_set* dp_out, register_set* sfu_out, + register_set* int_out, register_set* tensor_core_out, register_set* mem_out, int id, char* config_str ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ), + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, int_out, tensor_core_out, mem_out, id ), m_pending_warps() { unsigned inner_level_readin; @@ -534,6 +542,7 @@ public: register_set* sp_out, register_set* dp_out, register_set* sfu_out, + register_set* int_out, register_set* tensor_core_out, register_set* mem_out, int id, @@ -1046,6 +1055,9 @@ public: fprintf(fp,"%s dispatch= ", m_name.c_str() ); m_dispatch_reg->print(fp); } + const char* get_name() { + return m_name.c_str(); + } protected: std::string m_name; const shader_core_config *m_config; @@ -1154,6 +1166,29 @@ public: }; +class int_unit : public pipelined_simd_unit +{ +public: + int_unit( register_set* result_port, const shader_core_config *config, shader_core_ctx *core ); + virtual bool can_issue( const warp_inst_t &inst ) const + { + switch(inst.op) { + case SFU_OP: return false; + case LOAD_OP: return false; + case TENSOR_CORE_LOAD_OP: return false; + case STORE_OP: return false; + case TENSOR_CORE_STORE_OP: return false; + case MEMORY_BARRIER_OP: return false; + case SP_OP: return false; + case DP_OP: return false; + default: break; + } + return pipelined_simd_unit::can_issue(inst); + } + virtual void active_lanes_in_pipeline(); + virtual void issue( register_set& source_reg ); +}; + class sp_unit : public pipelined_simd_unit { public: @@ -1303,10 +1338,12 @@ protected: enum pipeline_stage_name_t { ID_OC_SP=0, 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, @@ -1318,10 +1355,12 @@ enum pipeline_stage_name_t { const char* const pipeline_stage_name_decode[] = { "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", @@ -1429,6 +1468,7 @@ struct shader_core_config : public core_config int gpgpu_operand_collector_num_units_tensor_core; int gpgpu_operand_collector_num_units_mem; int gpgpu_operand_collector_num_units_gen; + int gpgpu_operand_collector_num_units_int; unsigned int gpgpu_operand_collector_num_in_ports_sp; unsigned int gpgpu_operand_collector_num_in_ports_dp; @@ -1436,6 +1476,7 @@ struct shader_core_config : public core_config unsigned int gpgpu_operand_collector_num_in_ports_tensor_core; unsigned int gpgpu_operand_collector_num_in_ports_mem; unsigned int gpgpu_operand_collector_num_in_ports_gen; + unsigned int gpgpu_operand_collector_num_in_ports_int; unsigned int gpgpu_operand_collector_num_out_ports_sp; unsigned int gpgpu_operand_collector_num_out_ports_dp; @@ -1443,6 +1484,7 @@ struct shader_core_config : public core_config unsigned int gpgpu_operand_collector_num_out_ports_tensor_core; unsigned int gpgpu_operand_collector_num_out_ports_mem; unsigned int gpgpu_operand_collector_num_out_ports_gen; + unsigned int gpgpu_operand_collector_num_out_ports_int; int gpgpu_num_sp_units; int gpgpu_tensor_core_avail; @@ -1450,6 +1492,7 @@ struct shader_core_config : public core_config int gpgpu_num_sfu_units; int gpgpu_num_tensor_core_units; int gpgpu_num_mem_units; + int gpgpu_num_int_units; //Shader core resources unsigned gpgpu_shader_registers; -- cgit v1.3 From 74289a429633176e37116353347f60c75f2b644a Mon Sep 17 00:00:00 2001 From: Suchita Pati Date: Wed, 17 Apr 2019 23:30:20 -0500 Subject: Modified config files to add device limit configurations --- configs/tested-cfgs/SM2_GTX480/gpgpusim.config | 3 +++ configs/tested-cfgs/SM6_TITANX/gpgpusim.config | 6 +++++ configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 6 +++++ libcuda/cuda_runtime_api.cc | 35 ++++++++++++++++++-------- 4 files changed, 40 insertions(+), 10 deletions(-) (limited to 'configs/tested-cfgs/SM2_GTX480') diff --git a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config index d71b2fd..8d54d8b 100644 --- a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config @@ -3,6 +3,9 @@ -gpgpu_ptx_sim_mode 0 -gpgpu_ptx_force_max_capability 20 +# Device Limits +-gpgpu_stack_size_limit 1024 +-gpgpu_heap_size_limit 8388608 # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 diff --git a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config index cb23ab3..3371186 100644 --- a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config @@ -8,6 +8,12 @@ -gpgpu_ptx_force_max_capability 61 -gpgpu_ignore_resources_limitation 1 +# Device Limits +-gpgpu_stack_size_limit 1024 +-gpgpu_heap_size_limit 8388608 +-gpgpu_runtime_sync_depth_limit 2 +-gpgpu_runtime_pending_launch_count_limit 2048 + # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 -gpgpu_ptx_save_converted_ptxplus 0 diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index 8ed4cd0..1af64ac 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -13,6 +13,12 @@ -gpgpu_ptx_sim_mode 0 -gpgpu_ptx_force_max_capability 70 +# Device Limits +-gpgpu_stack_size_limit 1024 +-gpgpu_heap_size_limit 8388608 +-gpgpu_runtime_sync_depth_limit 2 +-gpgpu_runtime_pending_launch_count_limit 2048 + # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 -gpgpu_ptx_save_converted_ptxplus 0 diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 86435b9..6f1ab08 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1121,10 +1121,10 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetAttribute(int *value, enum cudaDevic *value= 0; break; case 75: - *value= 7 ; + *value= 7 ; //cudaDevAttrComputeCapabilityMajor for Volta architecture break; case 76: - *value= 0 ; + *value= 0 ; //cudaDevAttrComputeCapabilityMinor for Volta architecture break; case 78: *value= 0 ; //TODO: as of now, we dont support stream priorities. @@ -1206,20 +1206,35 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetLimit ( size_t* pValue, cudaLimit li announce_call(__my_func__); } _cuda_device_id *dev = GPGPUSim_Init(); + const struct cudaDeviceProp *prop = dev->get_prop(); const gpgpu_sim_config& config=dev->get_gpgpu()->get_config(); switch(limit) { case 0: // cudaLimitStackSize - *pValue=config.stack_limit(); - break; + *pValue=config.stack_limit(); + break; case 2: // cudaLimitMallocHeapSize - *pValue=config.heap_limit(); - break; + *pValue=config.heap_limit(); + break; +#if (CUDART_VERSION > 5050) case 3: // cudaLimitDevRuntimeSyncDepth - *pValue=config.sync_depth_limit(); - break; + if(prop->major > 2){ + *pValue=config.sync_depth_limit(); + break; + } + else{ + printf("ERROR:Limit %s is not supported on this architecture \n",limit); + abort(); + } case 4: // cudaLimitDevRuntimePendingLaunchCount - *pValue=config.pending_launch_count_limit(); - break; + if(prop->major > 2){ + *pValue=config.pending_launch_count_limit(); + break; + } + else{ + printf("ERROR:Limit %s is not supported on this architecture \n",limit); + abort(); + } +#endif default: printf("ERROR:Limit %s unimplemented \n",limit); abort(); -- cgit v1.3 From 537b342ab102b75a6da694b968b2226719b267c3 Mon Sep 17 00:00:00 2001 From: Preyesh Dalmia Date: Thu, 18 Apr 2019 17:13:06 -0500 Subject: changes for compute capability major/minor --- configs/tested-cfgs/SM2_GTX480/gpgpusim.config | 4 ++++ configs/tested-cfgs/SM6_TITANX/gpgpusim.config | 4 ++++ configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 6 ++++++ libcuda/cuda_runtime_api.cc | 12 ++++++------ src/gpgpu-sim/gpu-sim.cc | 17 +++++++++++++++++ src/gpgpu-sim/gpu-sim.h | 8 ++++++++ 6 files changed, 45 insertions(+), 6 deletions(-) (limited to 'configs/tested-cfgs/SM2_GTX480') diff --git a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config index 8d54d8b..cf3627b 100644 --- a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config @@ -7,6 +7,10 @@ -gpgpu_stack_size_limit 1024 -gpgpu_heap_size_limit 8388608 +# Compute Capability +-gpgpu_compute_capability_major 2 +-gpgpu_compute_capability_minor 0 + # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 -gpgpu_ptx_save_converted_ptxplus 0 diff --git a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config index 3371186..23d044c 100644 --- a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config @@ -14,6 +14,10 @@ -gpgpu_runtime_sync_depth_limit 2 -gpgpu_runtime_pending_launch_count_limit 2048 +# Compute Capability +-gpgpu_compute_capability_major 6 +-gpgpu_compute_capability_minor 1 + # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 -gpgpu_ptx_save_converted_ptxplus 0 diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index 1af64ac..a5b427f 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -13,12 +13,18 @@ -gpgpu_ptx_sim_mode 0 -gpgpu_ptx_force_max_capability 70 + # Device Limits -gpgpu_stack_size_limit 1024 -gpgpu_heap_size_limit 8388608 -gpgpu_runtime_sync_depth_limit 2 -gpgpu_runtime_pending_launch_count_limit 2048 +# Compute Capability +-gpgpu_compute_capability_major 7 +-gpgpu_compute_capability_minor 0 + + # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 -gpgpu_ptx_save_converted_ptxplus 0 diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 6f1ab08..9068700 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -354,8 +354,8 @@ struct _cuda_device_id *GPGPUSim_Init() cudaDeviceProp *prop = (cudaDeviceProp *) calloc(sizeof(cudaDeviceProp),1); snprintf(prop->name,256,"GPGPU-Sim_v%s", g_gpgpusim_version_string ); - prop->major = 5; - prop->minor = 2; + prop->major = the_gpu->compute_capability_major(); + prop->minor = the_gpu->compute_capability_minor(); prop->totalGlobalMem = 0x80000000 /* 2 GB */; prop->memPitch = 0; if(prop->major >= 2) { @@ -1120,11 +1120,11 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetAttribute(int *value, enum cudaDevic case 41: *value= 0; break; - case 75: - *value= 7 ; //cudaDevAttrComputeCapabilityMajor for Volta architecture + case 75:// Major compute capability version number + *value= prop->major ; break; - case 76: - *value= 0 ; //cudaDevAttrComputeCapabilityMinor for Volta architecture + case 76:// Minor compute capability version number + *value= prop->minor ; break; case 78: *value= 0 ; //TODO: as of now, we dont support stream priorities. diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 2a862a4..358c0f7 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -499,6 +499,12 @@ void gpgpu_sim_config::reg_options(option_parser_t opp) option_parser_register(opp, "-liveness_message_freq", OPT_INT64, &liveness_message_freq, "Minimum number of seconds between simulation liveness messages (0 = always print)", "1"); + option_parser_register(opp, "-gpgpu_compute_capability_major", OPT_UINT32, &gpgpu_compute_capability_major, + "Major compute capability version number", + "7"); + option_parser_register(opp, "-gpgpu_compute_capability_minor", OPT_UINT32, &gpgpu_compute_capability_minor, + "Minor compute capability version number", + "0"); option_parser_register(opp, "-gpgpu_flush_l1_cache", OPT_BOOL, &gpgpu_flush_l1_cache, "Flush L1 cache at the end of each kernel call", "0"); @@ -802,6 +808,17 @@ void gpgpu_sim::set_prop( cudaDeviceProp *prop ) m_cuda_properties = prop; } +int gpgpu_sim::compute_capability_major() const +{ + printf("Everything works %d\n", m_config.gpgpu_compute_capability_major); + return m_config.gpgpu_compute_capability_major; +} + +int gpgpu_sim::compute_capability_minor() const +{ + return m_config.gpgpu_compute_capability_minor; +} + const struct cudaDeviceProp *gpgpu_sim::get_prop() const { return m_cuda_properties; diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 11ffe49..39e1d9a 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -388,6 +388,10 @@ private: size_t runtime_sync_depth_limit; size_t runtime_pending_launch_count_limit; + //gpu compute capability options + unsigned int gpgpu_compute_capability_major; + unsigned int gpgpu_compute_capability_minor; + unsigned long long liveness_message_freq; friend class gpgpu_sim; @@ -447,6 +451,10 @@ public: int shared_mem_size() const; int shared_mem_per_block() const; + int compute_capability_major() const; + int compute_capability_minor() const; + + int num_registers_per_core() const; int num_registers_per_block() const; int wrp_size() const; -- cgit v1.3 From 17815dfdd7b6e9b558997bc5a9ba157a0493da16 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 12 Jul 2019 21:31:54 -0400 Subject: fixing device and function parameters config --- configs/tested-cfgs/SM2_GTX480/gpgpusim.config | 1 + configs/tested-cfgs/SM6_TITANX/gpgpusim.config | 1 + configs/tested-cfgs/SM7_QV100/gpgpusim.config | 11 +++++++---- configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 3 +++ libcuda/cuda_runtime_api.cc | 21 ++++++++++++--------- src/gpgpu-sim/gpu-sim.cc | 4 ++-- 6 files changed, 26 insertions(+), 15 deletions(-) (limited to 'configs/tested-cfgs/SM2_GTX480') diff --git a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config index cf3627b..4a7a3c3 100644 --- a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config @@ -61,6 +61,7 @@ # Note: Hashing set index function (H) only applies to a set size of 32 or 64. -gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,S:64:8,8 -gpgpu_shmem_size 49152 +-gpgpu_shmem_sizeDefault 49152 -icnt_flit_size 40 -gmem_skip_L1D 0 -gpgpu_n_cluster_ejection_buffer_size 32 diff --git a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config index 2fe898a..e6d8f1d 100644 --- a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config @@ -81,6 +81,7 @@ -gpgpu_cache:dl1PrefL1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 -gpgpu_cache:dl1PrefShared S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 -gpgpu_shmem_size 49152 +-gpgpu_shmem_sizeDefault 49152 -gpgpu_shmem_size_PrefL1 49152 -gpgpu_shmem_size_PrefShared 49152 # By default, L1 cache is disabled in Pascal P102. diff --git a/configs/tested-cfgs/SM7_QV100/gpgpusim.config b/configs/tested-cfgs/SM7_QV100/gpgpusim.config index 1a34d0f..5f64908 100644 --- a/configs/tested-cfgs/SM7_QV100/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100/gpgpusim.config @@ -11,7 +11,7 @@ # functional simulator specification -gpgpu_ptx_instruction_classification 0 -gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 60 +-gpgpu_ptx_force_max_capability 70 # Device Limits @@ -21,7 +21,7 @@ -gpgpu_runtime_pending_launch_count_limit 2048 # Compute Capability --gpgpu_compute_capability_major 6 +-gpgpu_compute_capability_major 7 -gpgpu_compute_capability_minor 0 # SASS execution (only supported with CUDA >= 4.0) @@ -44,12 +44,13 @@ # shader core pipeline config -gpgpu_shader_registers 65536 --gpgpu_occupancy_sm_number 60 +-gpgpu_registers_per_block 65536 +-gpgpu_occupancy_sm_number 70 # This implies a maximum of 64 warps/SM -gpgpu_shader_core_pipeline 2048:32 -gpgpu_shader_cta 32 --gpgpu_simd_model 1 +-gpgpu_simd_model 1 # Pipeline widths and number of FUs # 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 @@ -91,6 +92,8 @@ -mem_unit_ports 4 -gpgpu_cache:dl1 S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 -gpgpu_shmem_size 98304 +-gpgpu_shmem_sizeDefault 98304 +-gpgpu_shmem_per_block 65536 -gmem_skip_L1D 0 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index b06f048..6c21dcb 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -44,6 +44,7 @@ # shader core pipeline config -gpgpu_shader_registers 65536 +-gpgpu_registers_per_block 65536 -gpgpu_occupancy_sm_number 70 # This implies a maximum of 64 warps/SM @@ -91,6 +92,8 @@ -mem_unit_ports 4 -gpgpu_cache:dl1 S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 -gpgpu_shmem_size 98304 +-gpgpu_shmem_sizeDefault 98304 +-gpgpu_shmem_per_block 65536 -gmem_skip_L1D 0 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index c70a570..45511d4 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -383,7 +383,7 @@ struct _cuda_device_id *GPGPUSim_Init() prop->sharedMemPerMultiprocessor = the_gpu->shared_mem_size(); #endif prop->sharedMemPerBlock = the_gpu->shared_mem_per_block(); - prop->regsPerBlock = the_gpu->num_registers_per_core(); + prop->regsPerBlock = the_gpu->num_registers_per_block(); prop->warpSize = the_gpu->wrp_size(); prop->clockRate = the_gpu->shader_clock(); #if (CUDART_VERSION >= 2010) @@ -1014,7 +1014,7 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetAttribute(int *value, enum cudaDevic prop = dev->get_prop(); switch (attr) { case 1: - *value= prop->maxThreadsDim[0] * prop->maxThreadsDim[1] * prop->maxThreadsDim[2] * prop->maxGridSize[0] * prop->maxGridSize[1] * prop->maxGridSize[2]; + *value= prop->maxThreadsPerBlock; break; case 2: *value= prop->maxThreadsDim[0]; @@ -1504,13 +1504,13 @@ __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun ) { dim3 gridDim = config.grid_dim(); dim3 blockDim = config.block_dim(); - if (gridDim.x * gridDim.y * gridDim.z == 0 || blockDim.x * blockDim.y * blockDim.z == 0) - { + //if (gridDim.x * gridDim.y * gridDim.z == 0 || blockDim.x * blockDim.y * blockDim.z == 0) + //{ //can't launch - printf("can't launch a empty kernel\n"); - g_cuda_launch_stack.pop_back(); - return g_last_cudaError = cudaErrorInvalidConfiguration; - } + // printf("can't launch a empty kernel\n"); + // g_cuda_launch_stack.pop_back(); + // return g_last_cudaError = cudaErrorInvalidConfiguration; + //} } struct CUstream_st *stream = config.get_stream(); if(g_stream_manager->is_blocking()) @@ -3151,9 +3151,12 @@ size_t getMaxThreadsPerBlock(struct cudaFuncAttributes *attr) { size_t max = prop.maxThreadsPerBlock; - if ((prop.regsPerBlock / attr->numRegs) < max) + if (attr->numRegs && (prop.regsPerBlock / attr->numRegs) < max) max = prop.regsPerBlock / attr->numRegs; + if (attr->sharedSizeBytes && (prop.sharedMemPerBlock / attr->sharedSizeBytes) < max) + max = prop.sharedMemPerBlock / attr->sharedSizeBytes; + return max; } diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 6de5845..72cb32b 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -320,7 +320,7 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-adaptive_volta_cache_config", OPT_BOOL, &adaptive_volta_cache_config, "adaptive_volta_cache_config", "0"); - option_parser_register(opp, "-gpgpu_shmem_size", OPT_UINT32, &gpgpu_shmem_sizeDefault, + option_parser_register(opp, "-gpgpu_shmem_sizeDefault", OPT_UINT32, &gpgpu_shmem_sizeDefault, "Size of shared memory per shader core (default 16kB)", "16384"); option_parser_register(opp, "-gpgpu_shmem_size_PrefL1", OPT_UINT32, &gpgpu_shmem_sizePrefL1, @@ -1065,7 +1065,7 @@ void gpgpu_sim::change_cache_config(FuncCache cache_config) if(cache_config != m_shader_config->m_L1D_config.get_cache_status()){ printf("FLUSH L1 Cache at configuration change between kernels\n"); for (unsigned i=0;in_simt_clusters;i++) { - m_cluster[i]->cache_flush(); + m_cluster[i]->cache_invalidate(); } } -- cgit v1.3