diff options
| author | Mahmoud Khairy A. Abdallah <[email protected]> | 2017-10-30 17:08:07 -0400 |
|---|---|---|
| committer | GitHub Enterprise <[email protected]> | 2017-10-30 17:08:07 -0400 |
| commit | ff6c395da7b70d9cdbb72a8cfc3721a7f27a2a25 (patch) | |
| tree | 5721f7d821a7318e4814ff163fd961ab92604f60 /configs | |
| parent | c56115c9df274708ada5964f4f46e2d08b1acd03 (diff) | |
Update gpgpusim.config
Diffstat (limited to 'configs')
| -rw-r--r-- | configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config | 23 |
1 files changed, 14 insertions, 9 deletions
diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config index 782edf6..2f09ab6 100644 --- a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config @@ -35,18 +35,23 @@ # ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB ## Pascal GP102 has 4 SP SIMD units and 4 SFU units ## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,1,4,4,1,8 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 +-gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 +-gpgpu_num_dp_units 1 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 -ptx_opcode_latency_int 4,13,4,5,145 -ptx_opcode_initiation_int 1,1,1,1,4 -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 4,8,4,4,130 +-ptx_opcode_initiation_dp 8,8,8,8,130 +-ptx_opcode_initiation_sfu 4 +-ptx_opcode_latency_sfu 8 # <nsets>:<bsize>:<assoc>,<rep>:<wr>:<alloc>:<wr_alloc>:<set_index_fn>,<mshr>:<N>:<merge>,<mq>:**<fifo_entry> # ** Optional parameter - Required when mshr_type==Texture Fifo @@ -54,21 +59,21 @@ # Pascal GP102 has 96KB Shared memory # Pascal GP102 has 64KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:128:8,16 -gpgpu_shmem_size 98304 -gmem_skip_L1D 1 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2 S:64:128:16,L:B:m:W:L,A:128:8,4:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 # 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +-gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 # 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 # 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 +-gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 # enable operand collector ## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units |