merging in the mainline

1 files changed, 24 insertions, 16 deletions

diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config
index 782edf6..4407870 100644
--- a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config
+++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config
@@ -21,7 +21,7 @@
 #-gpgpu_clock_domains <Core Clock>:<Interconnect Clock>:<L2 Clock>:<DRAM Clock>
 # Pascal NVIDIA TITAN X clock domains are adopted from 
 # https://en.wikipedia.org/wiki/GeForce_10_series
--gpgpu_clock_domains 1417.0:2834.0:1417.0:2500.0
+-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0
 
 # shader core pipeline config
 -gpgpu_shader_registers 65536
@@ -32,21 +32,26 @@
 -gpgpu_simd_model 1 
 
 # Pipeline widths and number of FUs
-# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB
+# 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
 ## 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_pipeline_widths 4,1,4,1,4,1,4,1,9
 -gpgpu_num_sp_units 4
 -gpgpu_num_sfu_units 4
+-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
@@ -54,33 +59,36 @@
 # 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 N: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,A: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,A:2:64,4
 
 # enable operand collector 
 ## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units
--gpgpu_operand_collector_num_units_sp 20
--gpgpu_operand_collector_num_units_sfu 4
+-gpgpu_operand_collector_num_units_sp 12
+-gpgpu_operand_collector_num_units_sfu 6
 -gpgpu_operand_collector_num_units_mem 8
+-gpgpu_operand_collector_num_units_dp 6
 -gpgpu_operand_collector_num_in_ports_sp 4
 -gpgpu_operand_collector_num_out_ports_sp 4
--gpgpu_operand_collector_num_in_ports_sfu 1
--gpgpu_operand_collector_num_out_ports_sfu 1
+-gpgpu_operand_collector_num_in_ports_sfu 4
+-gpgpu_operand_collector_num_out_ports_sfu 4
 -gpgpu_operand_collector_num_in_ports_mem 1
 -gpgpu_operand_collector_num_out_ports_mem 1
+-gpgpu_operand_collector_num_in_ports_dp 1
+-gpgpu_operand_collector_num_out_ports_dp 1
 # gpgpu_num_reg_banks should be increased to 32
 -gpgpu_num_reg_banks 32
 
@@ -97,7 +105,7 @@
 -inter_config_file config_fermi_islip.icnt
 
 # memory partition latency config 
--rop_latency 120
+-rop_latency 100
 -dram_latency 100
 
 # dram model config
@@ -122,8 +130,8 @@
 
 # Use the same GDDR5 timing from hynix H5GQ1H24AFR
 # disable bank groups for now, 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=1:CCDL=0:RTPL=0"
+-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52:
+                        CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3"
 
 # Pascal GP102 has four schedulers per core
 -gpgpu_num_sched_per_core 4