diff options
| -rw-r--r-- | CHANGES | 9 | ||||
| -rw-r--r-- | configs/GeForceGTX750Ti/config_fermi_islip.icnt | 70 | ||||
| -rw-r--r-- | configs/GeForceGTX750Ti/gpgpusim.config | 35 | ||||
| -rw-r--r-- | src/gpgpu-sim/shader.h | 4 |
4 files changed, 94 insertions, 24 deletions
@@ -8,15 +8,14 @@ Version 3.2.3+edits (development branch) versus 3.2.3 - Initial support for CUDA 5.0,5.5,6.0 and 7.5 to get basic sdk running (e.g., template, vectorAdd, ...). The issues required for CUDA 5.5 support were identified by the loneStarGPU group at The University of Texas at Austin and Texas State University. - Removed intersim2 svn repository files - Changed the makefile for cuobjdump_toptxplus,libcuda,intersim2 so that it outputs temporary files into the build directory -- Bug fixes: - - Fixed a bug where sm_version was hard coded to sm_20. Now, it extracts the highest sm version that is lower than - the forced_max_capability configuration in GPGPUSim. - +- Branching config file for GeForceGTX750Ti +- Branching correlation script for GeForceGTX750Ti. Modified config to fit Maxwell architecture. Modified shader.h to allow larger CTA per warp. - Bug fixes: - Fixed bug #81, fix ordering of pushing branch entries to the stack - Fixed a bug where for each icache miss we also count a hit - Fixed bug #88, Ejection buffer and Boundary buffer in Intersim2 initialization with a wrong node number -- Branching config file for GeForceGTX750Ti + - Fixed a bug where sm_version was hard coded to sm_20. Now, it extracts the highest sm version that is lower than + the forced_max_capability configuration in GPGPUSim. Version 3.2.3 versus 3.2.2 - Bug fixes: diff --git a/configs/GeForceGTX750Ti/config_fermi_islip.icnt b/configs/GeForceGTX750Ti/config_fermi_islip.icnt new file mode 100644 index 0000000..7820e4e --- /dev/null +++ b/configs/GeForceGTX750Ti/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 = 32; + +// 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 = 8; + +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 = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// 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/GeForceGTX750Ti/gpgpusim.config b/configs/GeForceGTX750Ti/gpgpusim.config index 436cb41..5b5ee90 100644 --- a/configs/GeForceGTX750Ti/gpgpusim.config +++ b/configs/GeForceGTX750Ti/gpgpusim.config @@ -9,22 +9,22 @@ -gpgpu_ptx_save_converted_ptxplus 0 # high level architecture configuration --gpgpu_n_clusters 15 +-gpgpu_n_clusters 5 -gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 6 --gpgpu_n_sub_partition_per_mchannel 2 +-gpgpu_n_mem 2 +-gpgpu_n_sub_partition_per_mchannel 1 # Fermi clock domains #-gpgpu_clock_domains <Core Clock>:<Interconnect Clock>:<L2 Clock>:<DRAM Clock> # 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 +-gpgpu_clock_domains 1080.0:1080.0:1080.0:1335.0 # shader core pipeline config --gpgpu_shader_registers 32768 +-gpgpu_shader_registers 65536 # This implies a maximum of 48 warps/SM --gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_core_pipeline 2048:32 -gpgpu_shader_cta 8 -gpgpu_simd_model 1 @@ -32,13 +32,13 @@ # ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB -gpgpu_pipeline_widths 2,1,1,2,1,1,2 -gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 +-gpgpu_num_sfu_units 8 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_latency_int 6,12,13,13,210 -ptx_opcode_initiation_int 1,2,2,1,8 --ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_latency_fp 6,12,6,6,374 -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 @@ -49,14 +49,15 @@ # ** 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 32:128:4,L:L:m:N:H,A:32:8,8 --gpgpu_shmem_size 49152 +-gpgpu_shmem_size 65536 # The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected #-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 #-gpgpu_shmem_size 16384 # 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 +# Ignore above. Maxwell has 2MB of L2 cache. Configuration is unknown, so making a guess. +-gpgpu_cache:dl2 256:128:16,L:B:m:W:L,A:32:4,4:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 @@ -82,8 +83,8 @@ -inter_config_file config_fermi_islip.icnt # memory partition latency config --rop_latency 120 --dram_latency 100 +-rop_latency 80 +-dram_latency 60 # dram model config -gpgpu_dram_scheduler 1 @@ -96,7 +97,7 @@ -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_n_mem_per_ctrlr 4 -gpgpu_dram_buswidth 4 -gpgpu_dram_burst_length 8 -dram_data_command_freq_ratio 4 # GDDR5 is QDR @@ -108,8 +109,8 @@ -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" -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 +# Maxwell has four schedulers per core +-gpgpu_num_sched_per_core 4 # Two Level Scheduler with active and pending pools #-gpgpu_scheduler two_level_active:6:0:1 # Loose round robbin scheduler @@ -124,7 +125,7 @@ -visualizer_enabled 0 # power model configs --power_simulation_enabled 1 +-power_simulation_enabled 0 -gpuwattch_xml_file gpuwattch_gtx480.xml # tracing functionality diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 1aa468b..38d09e9 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -267,7 +267,7 @@ private: inline unsigned hw_tid_from_wid(unsigned wid, unsigned warp_size, unsigned i){return wid * warp_size + i;}; inline unsigned wid_from_hw_tid(unsigned tid, unsigned warp_size){return tid/warp_size;}; -const unsigned WARP_PER_CTA_MAX = 48; +const unsigned WARP_PER_CTA_MAX = 64; typedef std::bitset<WARP_PER_CTA_MAX> warp_set_t; int register_bank(int regnum, int wid, unsigned num_banks, unsigned bank_warp_shift); @@ -1331,7 +1331,7 @@ struct shader_core_config : public core_config struct shader_core_stats_pod { - void* shader_core_stats_pod_start[0]; // DO NOT MOVE FROM THE TOP - spaceless pointer to the start of this structure + void* shader_core_stats_pod_start[]; // DO NOT MOVE FROM THE TOP - spaceless pointer to the start of this structure unsigned long long *shader_cycles; unsigned *m_num_sim_insn; // number of scalar thread instructions committed by this shader core unsigned *m_num_sim_winsn; // number of warp instructions committed by this shader core |