Add SM6_GTX1080Ti/gpgpusim.config back

Author: Yechen Liu

configs/tested-cfgs/SM6_GTX1080Ti/gpgpusim.config

@@ -0,0 +1,289 @@
+# This config models the Volta
+# For more info about volta architecture:
+# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf
+# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# 
+# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf
+# https://en.wikipedia.org/wiki/Volta_(microarchitecture)
+# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf
+# https://devblogs.nvidia.com/inside-volta/
+# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf
+
+# functional simulator specification
+-gpgpu_ptx_instruction_classification 0
+-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
+-gpgpu_kernel_launch_latency 5000
+-gpgpu_TB_launch_latency 0
+
+# Compute Capability
+-gpgpu_compute_capability_major 7
+-gpgpu_compute_capability_minor 0
+
+# PTX execution-driven
+-gpgpu_ptx_convert_to_ptxplus 0
+-gpgpu_ptx_save_converted_ptxplus 0
+
+# high level architecture configuration
+-gpgpu_n_clusters 80
+-gpgpu_n_cores_per_cluster 1
+-gpgpu_n_mem 32
+-gpgpu_n_sub_partition_per_mchannel 2 
+
+# volta clock domains
+#-gpgpu_clock_domains <Core Clock>:<Interconnect Clock>:<L2 Clock>:<DRAM Clock>
+-gpgpu_clock_domains 1132.0:1132.0:1132.0:850.0
+# boost mode
+# -gpgpu_clock_domains 1628.0:1628.0:1628.0:850.0
+
+# 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
+-gpgpu_shader_core_pipeline 2048:32 
+-gpgpu_shader_cta 32
+-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
+## Volta GV100 has 4 SP SIMD units, 4 SFU units, 4 DP units per core, 4 Tensor core units
+## we need to scale the number of pipeline registers to be equal to the number of SP units
+-gpgpu_pipeline_widths 4,4,4,4,4,4,4,4,4,4,8,4,4
+-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
+
+# Instruction latencies and initiation intervals
+# "ADD,MAX,MUL,MAD,DIV"
+# All Div operations are executed on SFU unit
+-ptx_opcode_latency_int 4,13,4,5,145,21
+-ptx_opcode_initiation_int 2,2,2,2,8,4
+-ptx_opcode_latency_fp 4,13,4,5,39
+-ptx_opcode_initiation_fp 2,2,2,2,4
+-ptx_opcode_latency_dp 8,19,8,8,330
+-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
+
+# Volta has sub core model, in which each scheduler has its own register file and EUs
+# i.e. schedulers are isolated
+-gpgpu_sub_core_model 1
+# disable specialized operand collectors and use generic operand collectors instead
+-gpgpu_enable_specialized_operand_collector 0
+-gpgpu_operand_collector_num_units_gen 8
+-gpgpu_operand_collector_num_in_ports_gen 8
+-gpgpu_operand_collector_num_out_ports_gen 8
+# volta has 8 banks, 4 schedulers, two banks per scheduler
+# we increase #banks to 16 to mitigate the effect of Regisrer File Cache (RFC) which we do not implement in the current version
+-gpgpu_num_reg_banks 16
+-gpgpu_reg_file_port_throughput 2
+
+# shared memory bankconflict detection 
+-gpgpu_shmem_num_banks 32
+-gpgpu_shmem_limited_broadcast 0
+-gpgpu_shmem_warp_parts 1
+-gpgpu_coalesce_arch 60
+
+# Volta has four schedulers per core
+-gpgpu_num_sched_per_core 4
+# Greedy then oldest scheduler
+-gpgpu_scheduler gto
+## In Volta, a warp scheduler can issue 1 inst per cycle
+-gpgpu_max_insn_issue_per_warp 1
+-gpgpu_dual_issue_diff_exec_units 1
+
+## L1/shared memory configuration
+# <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
+# Defualt config is 32KB DL1 and 96KB shared memory
+# In Volta, we assign the remaining shared memory to L1 cache 
+# if the assigned shd mem = 0, then L1 cache = 128KB
+# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x 
+# disable this mode in case of multi kernels/apps execution
+-gpgpu_adaptive_cache_config 1
+# Volta unified cache has four banks
+-gpgpu_l1_banks 4
+-gpgpu_cache:dl1  S:1:128:256,L:L:s:N:L,A:256:8,16:0,32
+-gpgpu_shmem_size 98304
+-gpgpu_shmem_sizeDefault 98304
+-gpgpu_shmem_per_block 65536
+-gpgpu_gmem_skip_L1D 0
+-gpgpu_n_cluster_ejection_buffer_size 32
+-gpgpu_l1_latency 20
+-gpgpu_smem_latency 20
+-gpgpu_flush_l1_cache 1
+
+# 32 sets, each 128 bytes 24-way for each memory sub partition (96 KB per memory sub partition). This gives us 6MB L2 cache
+-gpgpu_cache:dl2 S:32:128:24,L:B:m:L:P,A:192:4,32:0,32
+-gpgpu_cache:dl2_texture_only 0
+-gpgpu_dram_partition_queues 64:64:64:64
+-gpgpu_perf_sim_memcpy 1
+-gpgpu_memory_partition_indexing 2
+
+# 128 KB Inst.
+-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4
+-gpgpu_inst_fetch_throughput 4
+# 128 KB Tex
+# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod
+-gpgpu_tex_cache:l1 N:4:128:256,L:R:m:N:L,T:512:8,128:2
+# 64 KB Const
+-gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4
+-gpgpu_perfect_inst_const_cache 1
+
+# interconnection
+#-network_mode 1 
+#-inter_config_file config_volta_islip.icnt
+# use built-in local xbar
+-network_mode 2
+-icnt_in_buffer_limit 512
+-icnt_out_buffer_limit 512
+-icnt_subnets 2
+-icnt_flit_size 40
+-icnt_arbiter_algo 1
+
+# memory partition latency config 
+-gpgpu_l2_rop_latency 160
+-dram_latency 100
+
+# dram model config
+-gpgpu_dram_scheduler 1
+-gpgpu_frfcfs_dram_sched_queue_size 64
+-gpgpu_dram_return_queue_size 192
+
+# for HBM, three stacks, 24 channles, each (128 bits) 16 bytes width
+-gpgpu_n_mem_per_ctrlr 1
+-gpgpu_dram_buswidth 16
+-gpgpu_dram_burst_length 2
+-dram_data_command_freq_ratio 2  # HBM is DDR
+-gpgpu_mem_address_mask 1
+-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCB.CCCSSSSS
+
+# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf)
+# Timing for 1 GHZ
+# tRRDl and tWTR are missing, need to be added
+#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47:
+#                        CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4"
+
+# Timing for 850 MHZ, V100 HBM runs at 850 MHZ
+-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=12:RAS=28:RP=12:RC=40:
+                        CL=12:WL=2:CDLR=3:WR=10:nbkgrp=4:CCDL=2:RTPL=3"
+
+# HBM has dual bus interface, in which it can issue two col and row commands at a time
+-dram_dual_bus_interface 1
+# select lower bits for bnkgrp to increase bnkgrp parallelism
+-dram_bnk_indexing_policy 0
+-dram_bnkgrp_indexing_policy 1
+
+#-dram_seperate_write_queue_enable 1
+#-dram_write_queue_size 64:56:32
+
+# stat collection
+-gpgpu_memlatency_stat 14 
+-gpgpu_runtime_stat 500
+-enable_ptx_file_line_stats 1
+-visualizer_enabled 0
+
+# power model configs, disable it untill we create a real energy model for Volta
+-power_simulation_enabled 0
+
+# tracing functionality
+#-trace_enabled 1
+#-trace_components WARP_SCHEDULER,SCOREBOARD,LIVENESS
+#-trace_sampling_core 0
+
+### items for functional and timing simulation of UVM ###
+
+# gddr size should be less than or equal to 1GB, in the unit of MB/GB
+-gddr_size 1GB
+
+# size of gddr page, only 4KB and 2MB available
+-page_size 4KB
+
+# number of tlb entries per SM
+-tlb_size 4096
+
+# average page table walk latency (in core cycle)
+# for 4K page, set to 100 and for 2M page, set to 66
+-page_table_walk_latency 100
+
+# page eviction policy
+# 0 - lru 2MB (default)
+# 1 - lru tree-based neighborhood
+# 2 - lru sequential locality 64K
+# 3 - random 4KB
+# 4 - LFU 2MB
+# 5 - lru 4KB
+-eviction_policy 0
+
+# invalidate clean pages directly instead of writing back
+-invalidate_clean 0
+
+# reserve percentage (e.g. 10 or 20) of accesses pages from eviction in hope that they will be accessed in   next iteration
+-reserve_accessed_page_percent 0
+
+# percentage of free page buffer to trigger the page eviction (e.g. 5 or 10)
+-percentage_of_free_page_buffer 0
+
+# pcie bandwidth per direction
+-pcie_bandwidth 16.0GB/s
+
+# enable/disable GMMU statistics profiling for UVM
+-sim_prof_enable 1
+
+# disable deadlock check for UVM
+-gpgpu_deadlock_detect 0
+
+# latency to process kernel launch (150us or 222150 core cycles)
+#-gpgpu_kernel_launch_latency 222150
+
+# hardware prefetcher
+# 0 - disabled
+# 1 - tree-based neighborhood (default)
+# 2 - sequential locality 64K
+# 3 - random 4 K
+-hardware_prefetch 1
+
+# hardware prefetcher under over-subscription
+# 0 - disable upon eviction (default)
+# 1 - tree-based neighborhood
+# 2 - sequential locality 64K
+# 3 - random 4 K
+-hwprefetch_oversub 1
+
+# latency in core cycle to handle page fault (45us)
+# encompass the overhead of stalling threads, deciding memory address, page table walk, maintaining page     flags, transfer chunks and orders
+-page_fault_latency 66645
+
+# enabling accurate simulation for stalling warps and serializing accesses for page fault handling (default  0)
+-enable_accurate_simulation 0
+
+# Enable direct CPU-memory access from GPU
+# 0 - disable
+# 1 - adaptive
+# 2 - always
+# 3 - after oversubscription
+-enable_dma 0
+
+# Access counter threshold for migrating the page from cpu to gpu
+-migrate_threshold 8
+
+# Oversubscription Multiplicative Penalty Factor for Adaptive DMA
+-multiply_dma_penalty 2
+
+# enabling access pattern detection, policy engine, and adaptive memory management
+-enable_smart_runtime 0 
+
+# enabling skip cycles when all warps stall and wait for page fault come back
+-skip_cycles_enable 1