diff options
| author | Tim Rogers <[email protected]> | 2017-05-09 13:39:46 -0400 |
|---|---|---|
| committer | GitHub <[email protected]> | 2017-05-09 13:39:46 -0400 |
| commit | bcd1867f8f800f2f4eda9da4e8b2b2d6a567e622 (patch) | |
| tree | 81d2baa94e8f8f298aff5b44e164df43ae30a57d | |
| parent | 54529f8b3556427526fd416118c444ce0eb10a21 (diff) | |
| parent | 96d528311239d6ff82d7bec807a2509b344c9a60 (diff) | |
Merge branch 'dev' into dev
38 files changed, 2467 insertions, 116 deletions
diff --git a/.travis.yml b/.travis.yml new file mode 100644 index 0000000..18fcd7a --- /dev/null +++ b/.travis.yml @@ -0,0 +1,11 @@ +dist: trusty +before_install: + - wget http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1404/x86_64/cuda-repo-ubuntu1404_7.5-18_amd64.deb + - sudo dpkg -i cuda-repo-ubuntu1404_7.5-18_amd64.deb + - sudo apt-get -qq update + - sudo apt-get install -y cuda + - sudo apt-get install -y build-essential xutils-dev bison zlib1g-dev flex libglu1-mesa-dev +env: + - CUDA_INSTALL_PATH=/usr/local/cuda-7.5/ +language: cpp +script: source ./setup_environment && make @@ -8,14 +8,15 @@ 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. +- Additional changes to config file to reflect Maxwell uarch. - 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 + - 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: @@ -155,6 +155,7 @@ $(SIM_LIB_DIR)/libcudart.so: makedirs $(LIBS) cudalib if [ ! -f $(SIM_LIB_DIR)/libcudart.so.5.0 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.5.0; fi if [ ! -f $(SIM_LIB_DIR)/libcudart.so.5.5 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.5.5; fi if [ ! -f $(SIM_LIB_DIR)/libcudart.so.6.0 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.6.0; fi + if [ ! -f $(SIM_LIB_DIR)/libcudart.so.6.5 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.6.5; fi if [ ! -f $(SIM_LIB_DIR)/libcudart.so.7.5 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.7.5; fi if [ ! -f $(SIM_LIB_DIR)/libcudart.so.8.0 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.8.0; fi @@ -24,6 +24,12 @@ Tor M. Aamodt, Vijay Janapa Reddi, GPUWattch: Enabling Energy Optimizations in GPGPUs, In proceedings of the ACM/IEEE International Symposium on Computer Architecture (ISCA 2013), Tel-Aviv, Israel, June 23-27, 2013. +If you use the support for CUDA dynamic parallelism in your research, please cite: + +Jin Wang and Sudhakar Yalamanchili, Characterization and Analysis of Dynamic +Parallelism in Unstructured GPU Applications, 2014 IEEE International Symposium +on Workload Characterization (IISWC), November 2014. + If you use figures plotted using AerialVision in your publications, please cite: Aaron Ariel, Wilson W. L. Fung, Andrew Turner, Tor M. Aamodt, Visualizing diff --git a/configs/GeForceGTX750Ti/config_fermi_islip.icnt b/configs/GeForceGTX750Ti/config_fermi_islip.icnt new file mode 100644 index 0000000..069ca02 --- /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 = 7; +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 new file mode 100644 index 0000000..8b030b6 --- /dev/null +++ b/configs/GeForceGTX750Ti/gpgpusim.config @@ -0,0 +1,130 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 52 + + +# 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 5 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 2 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Maxwell clock domains +#-gpgpu_clock_domains <Core Clock>:<Interconnect Clock>:<L2 Clock>:<DRAM Clock> +# In Maxwell, each pipeline has 32 execution units, so the Core clock is just the reported number. +-gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 8 +-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 +-gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_num_sp_units 8 +-gpgpu_num_sfu_units 32 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 6,12,13,13,210 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 6,12,6,6,374 +-ptx_opcode_initiation_fp 1,1,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + +-gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 +-gmem_skip_L1D 1 +-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 + +# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache +-gpgpu_cache:dl2 1024:128:8,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 +-gpgpu_tex_cache:l1 16:128:32,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A: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_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 150 +-dram_latency 130 + +# 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 16 +-gpgpu_dram_return_queue_size 300 + +# for Maxwell, bus width is 128bits, this is 64 bytes (32 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 32 +-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" + +# 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 +#-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 modeling is currently disabled for GTX750Ti. The gpuwattch_gtx750ti.xml file present is only a placeholder and has NOT been tested for accuracy. +# proper modeling of Maxwell power is planned, but should be considered completely unsupported at present. +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx750Ti.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml b/configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml new file mode 100755 index 0000000..e2b2324 --- /dev/null +++ b/configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml @@ -0,0 +1,539 @@ +<!-- THIS CONFIG IS CURRENTLY PRESENT ONLY AS A PLACEHOLDER. IT HAS NOT BEEN TESTED FOR ACCURACY VERSUS GTX750TI HARDWARE. ANY USE OF THIS CONFIG WILL LIKELY GIVE HIGHLY INACCURATE RESULTS. --> +<?xml version="1.0" ?> +<component id="root" name="root"> + <component id="system" name="system"> + <!--McPAT will skip the components if number is set to 0 --> + <param name="GPU_Architecture" value="1"/><!-- 0-G80; 1-Fermi; others not supported --> + <param name="number_of_cores" value="16"/> + <param name="architecture" value="1"/> <!-- fermi:1 quadro:2 other: undefined--> + <param name="number_of_L1Directories" value="0"/> + <param name="number_of_L2Directories" value="0"/> + <param name="number_of_L2s" value="1"/> <!-- This number means how many L2 clusters in each cluster there can be multiple banks/ports --> + <param name="number_of_L3s" value="0"/> <!-- This number means how many L3 clusters --> + <param name="number_of_NoCs" value="1"/> + <param name="homogeneous_cores" value="1"/><!--1 means homo --> + <param name="homogeneous_L2s" value="1"/> + <param name="homogeneous_L1Directorys" value="1"/> + <param name="homogeneous_L2Directorys" value="1"/> + <param name="homogeneous_L3s" value="1"/> + <param name="homogeneous_ccs" value="1"/><!--cache coherece hardware --> + <param name="homogeneous_NoCs" value="1"/> + <param name="core_tech_node" value="40"/><!-- nm --> + <param name="target_core_clockrate" value="700"/><!--MHz --> + <param name="temperature" value="380"/> <!-- Kelvin --> + <param name="number_cache_levels" value="2"/> + <param name="interconnect_projection_type" value="0"/><!--0: agressive wire technology; 1: conservative wire technology --> + <param name="device_type" value="0"/><!--0: HP(High Performance Type); 1: LSTP(Low standby power) 2: LOP (Low Operating Power) --> + <param name="longer_channel_device" value="1"/><!-- 0 no use; 1 use when possible --> + <param name="machine_bits" value="32"/> + <param name="virtual_address_width" value="32"/> + <param name="physical_address_width" value="32"/> + <param name="virtual_memory_page_size" value="4096"/> + <param name="idle_core_power" value="1.59"/><!-- idle core power for GTX479 --> + <!--param name="scaling_coefficients" value="10,0.0884816,10,10,8,10,4.12782,10,2.48832,10,10,10,4.29982,0.387764,0.0714269,0.14302,0.01,0.546811,0.485351,0.806633,0.818073,1.9207,100,100,100,87.9303,100,10,4.3548,10"/--> + <param name="TOT_INST" value="2.00" /> + <param name="FP_INT" value="4.57" /> + <param name="IC_H" value="2.14" /> + <param name="IC_M" value="22.47" /> + <param name="DC_RH" value="22.14" /> + <param name="DC_RM" value="24.66" /> + <param name="DC_WH" value="1.53" /> + <param name="DC_WM" value="39.79" /> + <param name="TC_H" value="10.21" /> + <param name="TC_M" value="24.66" /> + <param name="CC_H" value="11.07" /> + <param name="CC_M" value="12.33" /> + <param name="SHRD_ACC" value="7.04" /> + <param name="REG_RD" value="0.14" /> + <param name="REG_WR" value="0.21" /> + <param name="NON_REG_OPs" value="2.11" /> + <param name="SP_ACC" value="2.38" /> + <param name="SFU_ACC" value="0.51" /> + <param name="FPU_ACC" value="0.64" /> + <param name="MEM_RD" value="0.33" /> + <param name="MEM_WR" value="0.40" /> + <param name="MEM_PRE" value="0.11" /> + <param name="L2_RH" value="13.79" /> + <param name="L2_RM" value="35.18" /> + <param name="L2_WH" value="43.07" /> + <param name="L2_WM" value="28.72" /> + <param name="NOC_A" value="305.48" /> + <param name="PIPE_A" value="2.57" /> + <param name="IDLE_CORE_N" value="1"/> + <param name="CONST_DYNAMICN" value="11" /> + <stat name="num_idle_cores" value="0"/><!-- Average Number of idle cores during this period --> + <stat name="total_cycles" value="total_cycles_match_mcpat"/> + <stat name="idle_cycles" value="idle_cycles_match_mcpat"/> + <stat name="busy_cycles" value="busy_cycles_match_mcpat"/> + <!--This page size(B) is complete different from the page size in Main memo secction. this page size is the size of + virtual memory from OS/Archi perspective; the page size in Main memo secction is the actuall physical line in a DRAM bank --> + <!-- *********************** cores ******************* --> + <component id="system.core0" name="core0"> + <!-- Core property --> + <param name="clock_rate" value="700"/> + <param name="instruction_length" value="32"/> + <param name="opcode_width" value="9"/> + <!-- address width determins the tag_width in Cache, LSQ and buffers in cache controller + default value is machine_bits, if not set --> + <param name="machine_type" value="1"/><!-- 1 inorder; 0 OOO--> + <!-- inorder/OoO --> + <param name="number_hardware_threads" value="32"/> + <!-- number_instruction_fetch_ports(icache ports) is always 1 in single-thread processor, + it only may be more than one in SMT processors. BTB ports always equals to fetch ports since + branch information in consective branch instructions in the same fetch group can be read out from BTB once.--> + <param name="fetch_width" value="1"/> + <!-- fetch_width determins the size of cachelines of L1 cache block --> + <param name="number_instruction_fetch_ports" value="1"/> + <param name="decode_width" value="1"/> + <!-- decode_width determins the number of ports of the + renaming table (both RAM and CAM) scheme --> + <param name="issue_width" value="2"/> + <!-- issue_width determins the number of ports of Issue window and other logic + as in the complexity effective proccessors paper; issue_width==dispatch_width --> + <param name="commit_width" value="2"/> + <!-- commit_width determins the number of ports of register files --> + <param name="fp_issue_width" value="1"/> + <param name="prediction_width" value="0"/> + <!-- number of branch instructions can be predicted simultannouesl--> + <!-- Current version of McPAT does not distinguish int and floating point pipelines + Theses parameters are reserved for future use.--> + <param name="pipelines_per_core" value="1,1"/> + <!--integer_pipeline and floating_pipelines, if the floating_pipelines is 0, then the pipeline is shared--> + <param name="pipeline_depth" value="8,8"/> + <!-- pipeline depth of int and fp, if pipeline is shared, the second number is the average cycles of fp ops --> + <!-- issue and exe unit--> + <param name="ALU_per_core" value="32"/> + <!-- contains an adder, a shifter, and a logical unit --> + <param name="MUL_per_core" value="4"/> + <!-- For MUL and Div --> + <param name="FPU_per_core" value="32"/> + <!-- buffer between IF and ID stage --> + <param name="instruction_buffer_size" value="1"/> + <!-- buffer between ID and sche/exe stage --> + <param name="decoded_stream_buffer_size" value="1"/> + <param name="instruction_window_scheme" value="0"/><!-- 0 PHYREG based, 1 RSBASED--> + <!-- McPAT support 2 types of OoO cores, RS based and physical reg based--> + <param name="instruction_window_size" value="1"/> + <param name="fp_instruction_window_size" value="1"/> + <!-- the instruction issue Q as in Alpha 21264; The RS as in Intel P6 --> + <param name="ROB_size" value="0"/> + <!-- each in-flight instruction has an entry in ROB --> + <!-- registers --> + <!-- SM parameters Added by Syed Gilani --> + <param name="rf_banks" value="32"/> + <param name="simd_width" value="32"/> + <param name="collector_units" value="32"/> + <param name="core_clock_ratio" value="2"/> + <param name="warp_size" value="32"/> + + <param name="archi_Regs_IRF_size" value="32768"/> + <param name="archi_Regs_FRF_size" value="32"/> + <!-- if OoO processor, phy_reg number is needed for renaming logic, + renaming logic is for both integer and floating point insts. --> + <param name="phy_Regs_IRF_size" value="32"/> + <param name="phy_Regs_FRF_size" value="32"/> + <!-- rename logic --> + <param name="rename_scheme" value="0"/> + <!-- can be RAM based(0) or CAM based(1) rename scheme + RAM-based scheme will have free list, status table; + CAM-based scheme have the valid bit in the data field of the CAM + both RAM and CAM need RAM-based checkpoint table, checkpoint_depth=# of in_flight instructions; + Detailed RAT Implementation see TR --> + <param name="register_windows_size" value="0"/> + <!-- how many windows in the windowed register file, sun processors; + no register windowing is used when this number is 0 --> + <!-- In OoO cores, loads and stores can be issued whether inorder(Pentium Pro) or (OoO)out-of-order(Alpha), + They will always try to exeute out-of-order though. --> + <param name="LSU_order" value="inorder"/> + <param name="store_buffer_size" value="32"/> + <!-- By default, in-order cores do not have load buffers --> + <param name="load_buffer_size" value="32"/> + <!-- number of ports refer to sustainable concurrent memory accesses --> + <param name="memory_ports" value="2"/> + <!-- max_allowed_in_flight_memo_instructions determins the # of ports of load and store buffer + as well as the ports of Dcache which is connected to LSU --> + <!-- dual-pumped Dcache can be used to save the extra read/write ports --> + <param name="RAS_size" value="1"/> + <!-- general stats, defines simulation periods;require total, idle, and busy cycles for senity check --> + <!-- please note: if target architecture is X86, then all the instrucions refer to (fused) micro-ops --> + <stat name="total_instructions" value="total_instructions_match_mcpat"/> + <stat name="int_instructions" value="int_instruction_match_mcpat"/> + <stat name="fp_instructions" value="flt_instruction_match_mcpat"/> + <stat name="branch_instructions" value="branch_instruction_match_mcpat"/> + <stat name="branch_mispredictions" value="0"/> + <stat name="load_instructions" value="load_instruction_match_mcpat"/> + <stat name="store_instructions" value="store_instruction_match_mcpat"/> + <stat name="committed_instructions" value="total_instructions_match_mcpat"/> + <stat name="committed_int_instructions" value="int_instruction_match_mcpat"/> + <stat name="committed_fp_instructions" value="flt_instruction_match_mcpat"/> + <stat name="pipeline_duty_cycle" value="0.6"/><!--<=1, runtime_ipc/peak_ipc; averaged for all cores if homogenous --> + <!-- the following cycle stats are used for heterogeneouse cores only, + please ignore them if homogeneouse cores --> + <stat name="total_cycles" value="total_cycles_match_mcpat"/> + <stat name="idle_cycles" value="idle_cycles_match_mcpat"/> + <stat name="busy_cycles" value="busy_cycles_match_mcpat"/> + <!-- instruction buffer stats --> + <!-- ROB stats, both RS and Phy based OoOs have ROB + performance simulator should capture the difference on accesses, + otherwise, McPAT has to guess based on number of commited instructions. --> + <stat name="ROB_reads" value="263886"/> + <stat name="ROB_writes" value="263886"/> + <!-- RAT accesses --> + <stat name="rename_accesses" value="263886"/> + <stat name="fp_rename_accesses" value="263886"/> + <!-- decode and rename stage use this, should be total ic - nop --> + <!-- Inst window stats --> + <stat name="inst_window_reads" value="263886"/> + <stat name="inst_window_writes" value="263886"/> + <stat name="inst_window_wakeup_accesses" value="263886"/> + <stat name="fp_inst_window_reads" value="263886"/> + <stat name="fp_inst_window_writes" value="263886"/> + <stat name="fp_inst_window_wakeup_accesses" value="263886"/> + <!-- RF accesses --> + <stat name="int_regfile_reads" value="int_register_read_access_match_mcpat"/> + <stat name="float_regfile_reads" value="int_register_write_access_match_mcpat"/> + <stat name="int_regfile_writes" value="float_register_read_access_match_mcpat"/> + <stat name="float_regfile_writes" value="float_register_write_access_match_mcpat"/> + + <!-- The following stat is for operand collector power - Added by Syed --> + <stat name="non_rf_operands" value="0"/> + + <!-- accesses to the working reg --> + <stat name="function_calls" value="0"/> + <stat name="context_switches" value="0"/> <!--not used in the McPAT --> + <!-- Number of Windowes switches (number of function calls and returns)--> + <!-- Alu stats by default, the processor has one FPU that includes the divider and + multiplier. The fpu accesses should include accesses to multiplier and divider --> + <stat name="ialu_accesses" value="ialu_accesses_match_mcpat"/> + <stat name="fpu_accesses" value="fpu_accesses_match_mcpat"/> + <stat name="mul_accesses" value="mul_accesses_match_mcpat"/> + <stat name="cdb_alu_accesses" value="0"/> + <stat name="cdb_mul_accesses" value="0"/> + <stat name="cdb_fpu_accesses" value="0"/> + <!-- multiple cycle accesses should be counted multiple times, + otherwise, McPAT can use internal counter for different floating point instructions + to get final accesses. But that needs detailed info for floating point inst mix --> + <!-- currently the performance simulator should + make sure all the numbers are final numbers, + including the explicit read/write accesses, + and the implicite accesses such as replacements and etc. + Future versions of McPAT may be able to reason the implicite access + based on param and stats of last level cache + The same rule applies to all cache access stats too! --> + <!-- following is AF for max power computation. + Do not change them, unless you understand them--> + <stat name="IFU_duty_cycle" value="0.25"/> + <stat name="LSU_duty_cycle" value="0.25"/> + <stat name="MemManU_I_duty_cycle" value="1"/> + <stat name="MemManU_D_duty_cycle" value="0.25"/> + <stat name="ALU_duty_cycle" value="0.9"/> + <stat name="MUL_duty_cycle" value="0.5"/> + <stat name="FPU_duty_cycle" value="1"/><!-- FPU numbers are already average --> + <stat name="ALU_cdb_duty_cycle" value="0.9"/> + <stat name="MUL_cdb_duty_cycle" value="0.5"/> + <stat name="FPU_cdb_duty_cycle" value="15"/> + <component id="system.core0.predictor" name="PBT"> + <!-- branch predictor; tournament predictor see Alpha implementation --> + <param name="local_predictor_size" value="10,3"/> + <param name="local_predictor_entries" value="1024"/> + <param name="global_predictor_entries" value="4096"/> + <param name="global_predictor_bits" value="2"/> + <param name="chooser_predictor_entries" value="4096"/> + <param name="chooser_predictor_bits" value="2"/> + <!-- These parameters can be combined like below in next version + <param name="load_predictor" value="10,3,1024"/> + <param name="global_predictor" value="4096,2"/> + <param name="predictor_chooser" value="4096,2"/> + --> + </component> + <component id="system.core0.itlb" name="itlb"> + <param name="number_entries" value="1"/> + <stat name="total_accesses" value="0"/> + <stat name="total_misses" value="0"/> + <stat name="conflicts" value="0"/> + <!-- there is no write requests to itlb although writes happen to itlb after miss, + which is actually a replacement --> + </component> + <component id="system.core0.icache" name="icache"> + <!-- there is no write requests to itlb although writes happen to it after miss, + which is actually a replacement --> + <param name="icache_config" value="16384,32,4,1,1,3,8,0"/> + <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy --> + <!-- cache_policy;//0 no write or write-though with non-write allocate;1 write-back with write-allocate --> + <param name="buffer_sizes" value="16, 16, 16,0"/> + <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size--> + <stat name="read_accesses" value="total_instructions_match_mcpat"/> + <stat name="read_misses" value="0"/> + <stat name="conflicts" value="0"/> + </component> + <component id="system.core0.dtlb" name="dtlb"> + <param name="number_entries" value="1"/> + <stat name="total_accesses" value="0"/> + <stat name="total_misses" value="0"/> + <stat name="conflicts" value="0"/> + </component> + <component id="system.core0.ccache" name="ccache"> + <!-- all the buffer related are optional --> + <param name="ccache_config" value="16384,32,4,1,1,3,8,0"/> + <param name="buffer_sizes" value="16, 16, 16, 0"/> + <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size--> + <stat name="read_accesses" value="ccache_read_accesses_match_mcpat"/> + <stat name="write_accesses" value="0"/> + <stat name="read_misses" value="ccache_read_misses_match_mcpat"/> + <stat name="write_misses" value="0"/> + <stat name="conflicts" value="0"/> + </component> + <component id="system.core0.tcache" name="tcache"> + <!-- all the buffer related are optional --> + <param name="tcache_config" value="16384,32,4,1,1,3,8,0"/> + <param name="buffer_sizes" value="16, 16, 16, 0"/> + <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size--> + <stat name="read_accesses" value="tcache_read_accesses_match_mcpat"/> + <stat name="write_accesses" value="0"/> + <stat name="read_misses" value="tcache_read_misses_match_mcpat"/> + <stat name="write_misses" value="0"/> + <stat name="conflicts" value="0"/> + </component> + <!--model the shared memory by mimicing dcache--> + <component id="system.core0.sharedmemory" name="sharedmemory"> + <!-- all the buffer related are optional --> + <param name="sharedmemory_config" value="49152,16,1,16,1,3,16,0"/> + <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy --> + <param name="buffer_sizes" value="16, 16, 16, 16"/> + <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size--> + <stat name="read_accesses" value="sharedmemory_read_access_match_mcpat"/> + <stat name="write_accesses" value="sharedmemory_write_access_match_mcpat"/> + <stat name="read_misses" value="0"/> + <stat name="write_misses" value="0"/> + <stat name="conflicts" value="0"/> + </component> + <component id="system.core0.dcache" name="dcache"> + <!-- all the buffer related are optional --> + <param name="dcache_config" value="16384,32,4,1,1,3,8,0"/> + <param name="buffer_sizes" value="16, 16, 16, 0"/> + <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size--> + <stat name="read_accesses" value="dcache_read_access_match_mcpat"/> + <stat name="write_accesses" value="dcache_write_access_match_mcpat"/> + <stat name="read_misses" value="dcache_read_miss_match_mcpat"/> + <stat name="write_misses" value="dcache_write_miss_match_mcpat"/> + <stat name="conflicts" value="0"/> + </component> + <component id="system.core0.BTB" name="BTB"> + <!-- all the buffer related are optional --> + <param name="BTB_config" value="8192,4,2,1, 1,3"/> + <!-- the parameters are capacity,block_width,associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,--> + </component> + </component> + <component id="system.L1Directory0" name="L1Directory0"> + <param name="Directory_type" value="0"/> + <!--0 cam based shadowed tag. 1 directory cache --> + <param name="Dir_config" value="2048,1,0,1, 4, 4,8"/> + <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,--> + <param name="buffer_sizes" value="8, 8, 8, 8"/> + <!-- all the buffer related are optional --> + <param name="clockrate" value="1400"/> + <param name="ports" value="1,1,1"/> + <!-- number of r, w, and rw search ports --> + <param name="device_type" value="0"/> + <!-- altough there are multiple access types, + Performance simulator needs to cast them into reads or writes + e.g. the invalidates can be considered as writes --> + <stat name="read_accesses" value="800000"/> + <stat name="write_accesses" value="27276"/> + <stat name="read_misses" value="1632"/> + <stat name="write_misses" value="183"/> + <stat name="conflicts" value="20"/> + <stat name="duty_cycle" value="0.45"/> + </component> + <component id="system.L2Directory0" name="L2Directory0"> + <param name="Directory_type" value="1"/> + <!--0 cam based shadowed tag. 1 directory cache --> + <param name="Dir_config" value="1048576,16,16,1,2, 100"/> + <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,--> + <param name="buffer_sizes" value="8, 8, 8, 8"/> + <!-- all the buffer related are optional --> + <param name="clockrate" value="1400"/> + <param name="ports" value="1,1,1"/> + <!-- number of r, w, and rw search ports --> + <param name="device_type" value="0"/> + <!-- altough there are multiple access types, + Performance simulator needs to cast them into reads or writes + e.g. the invalidates can be considered as writes --> + <stat name="read_accesses" value="0"/> + <stat name="write_accesses" value="0"/> + <stat name="read_misses" value="0"/> + <stat name="write_misses" value="0"/> + <stat name="conflicts" value="0"/> + <stat name="duty_cycle" value="0.45"/> + </component> + <component id="system.L20" name="L20"> + <!-- all the buffer related are optional --> + <param name="L2_config" value="131072,256,8,1, 4,23, 64, 1"/> + <!-- consider 4-way bank interleaving for Niagara 1 --> + <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy --> + <param name="buffer_sizes" value="16, 16, 16, 16"/> + <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size--> + <param name="clockrate" value="1400"/> + <param name="ports" value="1,1,1"/> + <!-- number of r, w, and rw ports --> + <param name="device_type" value="0"/> + <stat name="read_accesses" value="200000"/> + <stat name="write_accesses" value="0"/> + <stat name="read_misses" value="0"/> + <stat name="write_misses" value="0"/> + <stat name="conflicts" value="0"/> + <stat name="duty_cycle" value="0.5"/> + </component> + +<!--**********************************************************************--> +<component id="system.L30" name="L30"> + <param name="L3_config" value="1048576,64,16,1, 2,100, 64,1"/> + <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy --> + <param name="clockrate" value="3500"/> + <param name="ports" value="1,1,1"/> + <!-- number of r, w, and rw ports --> + <param name="device_type" value="0"/> + <param name="buffer_sizes" value="16, 16, 16, 16"/> + <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size--> + <stat name="read_accesses" value="58824"/> + <stat name="write_accesses" value="27276"/> + <stat name="read_misses" value="1632"/> + <stat name="write_misses" value="183"/> + <stat name="conflicts" value="0"/> + <stat name="duty_cycle" value="0.35"/> + </component> + + +<!--**********************************************************************--> + <component id="system.NoC0" name="noc0"> + <param name="clockrate" value="700"/> + <param name="type" value="1"/> + <!-- 1 NoC, O bus --> + <param name="horizontal_nodes" value="2"/> + <param name="vertical_nodes" value="1"/> + <param name="has_global_link" value="0"/> + <!-- 1 has global link, 0 does not have global link --> + <param name="link_throughput" value="1"/><!--w.r.t clock --> + <param name="link_latency" value="1"/><!--w.r.t clock --> + <!-- througput >= latency --> + <!-- Router architecture --> + <param name="input_ports" value="6"/> + <param name="output_ports" value="6"/> + <param name="virtual_channel_per_port" value="1"/> + <!-- input buffer; in classic routers only input ports need buffers --> + <param name="flit_bits" value="32"/> + <param name="input_buffer_entries_per_vc" value="1"/><!--VCs within the same ports share input buffers whose size is propotional to the number of VCs--> + <param name="chip_coverage" value="1"/> + <!-- When multiple NOC present, one NOC will cover part of the whole chip. chip_coverage <=1 --> + <stat name="total_accesses" value="0"/> + <!-- This is the number of total accesses within the whole network not for each router --> + <stat name="duty_cycle" value="0.6"/> + </component> +<!--**********************************************************************--> +<!--**********************************************************************--> + + <component id="system.mem" name="mem"> + <!-- Main memory property --> + <param name="mem_tech_node" value="40"/> + <param name="device_clock" value="200"/><!--MHz, this is clock rate of the actual memory device, not the FSB --> + <param name="peak_transfer_rate" value="3200"/><!--MB/S--> + <param name="internal_prefetch_of_DRAM_chip" value="4"/> + <!-- 2 for DDR, 4 for DDR2, 8 for DDR3...--> + <!-- the device clock, peak_transfer_rate, and the internal prefetch decide the DIMM property --> + <!-- above numbers can be easily found from Wikipedia --> + <param name="capacity_per_channel" value="4096"/> <!-- MB --> + <!-- capacity_per_Dram_chip=capacity_per_channel/number_of_dimms/number_ranks/Dram_chips_per_rank + Current McPAT assumes single DIMMs are used.--> + <param name="number_ranks" value="2"/> + <param name="num_banks_of_DRAM_chip" value="6"/> + <param name="Block_width_of_DRAM_chip" value="64"/> <!-- B --> + <param name="output_width_of_DRAM_chip" value="8"/> + <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip--> + <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip--> + <param name="page_size_of_DRAM_chip" value="8"/> <!-- 8 or 16 --> + <param name="burstlength_of_DRAM_chip" value="8"/> + <stat name="memory_accesses" value="1052"/> + <stat name="memory_reads" value="1052"/> + <stat name="memory_writes" value="1052"/> + </component> + <component id="system.mc" name="mc"> + <!-- Memeory controllers are for DDR(2,3...) DIMMs --> + <!-- current version of McPAT uses published values for base parameters of memory controller + improvments on MC will be added in later versions. --> + <param name="type" value="0"/> <!-- 1: low power; 0 high performance --> + <param name="mc_clock" value="1848"/><!--DIMM IO bus clock rate MHz DDR2-400 for Niagara 1--> + <param name="peak_transfer_rate" value="29568"/><!--MB/S Syed: GTX 470 has 177.4GB/s mem transfer rate with 6 MCs --> + <param name="block_size" value="64"/><!--B--> + <param name="number_mcs" value="6"/><!-- 6 GDDR5 memory controllers --> + <!-- current McPAT only supports homogeneous memory controllers --> + <param name="memory_channels_per_mc" value="2"/> + <param name="number_ranks" value="1"/> + <param name="withPHY" value="0"/> + <!-- # of ranks of each channel--> + <param name="req_window_size_per_channel" value="16"/> + <param name="IO_buffer_size_per_channel" value="16"/> + <param name="databus_width" value="32"/> + <param name="addressbus_width" value="32"/> + <param name="PRT_entries" value="32"/> + <!-- # of empirical DRAM model parameter --> + <param name="dram_cmd_coeff" value="0"/> + <param name="dram_act_coeff" value="0"/> + <param name="dram_nop_coeff" value="0"/> + <param name="dram_activity_coeff" value="0"/> + <param name="dram_pre_coeff" value="3.8475e-8f"/> + <param name="dram_rd_coeff" value="7.74707143e-8f"/> + <param name="dram_wr_coeff" value="3.54664286e-8f"/> + <param name="dram_req_coeff" value="0"/> + <param name="dram_const_coeff" value="0"/> + + <!-- McPAT will add the control bus width to the addressbus width automatically --> + <stat name="memory_accesses" value="memory_accesses_match_mcpat"/> + <stat name="memory_reads" value="memory_reads_match_mcpat"/> + <stat name="memory_writes" value="memory_writes_match_mcpat"/> + <!-- McPAT does not track individual mc, instead, it takes the total accesses and calculate + the average power per MC or per channel. This is sufficent for most application. + Further trackdown can be easily added in later versions. --> + </component> +<!--**********************************************************************--> + <component id="system.niu" name="niu"> + <!-- On chip 10Gb Ethernet NIC, including XAUI Phy and MAC controller --> + <!-- For a minimum IP packet size of 84B at 10Gb/s, a new packet arrives every 67.2ns. + the low bound of clock rate of a 10Gb MAC is 150Mhz --> + <param name="type" value="0"/> <!-- 1: low power; 0 high performance --> + <param name="clockrate" value="350"/> + <param name="number_units" value="0"/> <!-- unlike PCIe and memory controllers, each Ethernet controller only have one port --> + <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 --> + <stat name="total_load_perc" value="0.7"/> <!-- ratio of total achived load to total achivable bandwidth --> + <!-- McPAT does not track individual nic, instead, it takes the total accesses and calculate + the average power per nic or per channel. This is sufficent for most application. --> + </component> +<!--**********************************************************************--> + <component id="system.pcie" name="pcie"> + <!-- On chip PCIe controller, including Phy--> + <!-- For a minimum PCIe packet size of 84B at 8Gb/s per lane (PCIe 3.0), a new packet arrives every 84ns. + the low bound of clock rate of a PCIe per lane logic is 120Mhz --> + <param name="type" value="0"/> <!-- 1: low power; 0 high performance --> + <param name="withPHY" value="1"/> + <param name="clockrate" value="350"/> + <param name="number_units" value="0"/> + <param name="num_channels" value="8"/> <!-- 2 ,4 ,8 ,16 ,32 --> + <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 --> + <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth --> + <!-- McPAT does not track individual pcie controllers, instead, it takes the total accesses and calculate + the average power per pcie controller or per channel. This is sufficent for most application. --> + </component> +<!--**********************************************************************--> + <component id="system.flashc" name="flashc"> + <param name="number_flashcs" value="0"/> + <param name="type" value="1"/> <!-- 1: low power; 0 high performance --> + <param name="withPHY" value="1"/> + <param name="peak_transfer_rate" value="200"/><!--Per controller sustainable reak rate MB/S --> + <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 --> + <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth --> + <!-- McPAT does not track individual flash controller, instead, it takes the total accesses and calculate + the average power per fc or per channel. This is sufficent for most application --> + </component> +<!--**********************************************************************--> + + </component> +</component> diff --git a/cuobjdump_to_ptxplus/ptx_parser.h b/cuobjdump_to_ptxplus/ptx_parser.h index 1c96b46..418a733 100644 --- a/cuobjdump_to_ptxplus/ptx_parser.h +++ b/cuobjdump_to_ptxplus/ptx_parser.h @@ -110,6 +110,11 @@ void add_alignment_spec( int ) {PTX_PARSE_DPRINTF(" ");} void add_pragma( const char *a ) {PTX_PARSE_DPRINTF(" ");} void add_constptr(const char* identifier1, const char* identifier2, int offset) {PTX_PARSE_DPRINTF(" ");} +//Jin: handle instructino group for cdp +void start_inst_group(){PTX_PARSE_DPRINTF(" ");}; +void end_inst_group(){PTX_PARSE_DPRINTF(" ");}; + + /*non-dummy stuff below this point*/ extern cuobjdumpInstList *g_headerList; diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 75a2f31..cbe8a11 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -109,6 +109,7 @@ #include <stdarg.h> #include <iostream> #include <string> +#include <regex> #include <sstream> #include <fstream> #ifdef OPENGL_SUPPORT @@ -135,6 +136,7 @@ #include "../src/cuda-sim/ptx_parser.h" #include "../src/gpgpusim_entrypoint.h" #include "../src/stream_manager.h" +#include "../src/abstract_hardware_model.h" #include <pthread.h> #include <semaphore.h> @@ -262,10 +264,19 @@ struct CUctx_st { { if( m_code.find(fat_cubin_handle) != m_code.end() ) { symbol *s = m_code[fat_cubin_handle]->lookup(deviceFun); - assert( s != NULL ); - function_info *f = s->get_pc(); - assert( f != NULL ); - m_kernel_lookup[hostFun] = f; + if(s != NULL) { + function_info *f = s->get_pc(); + assert( f != NULL ); + m_kernel_lookup[hostFun] = f; + } + else { + printf("Warning: cannot find deviceFun %s\n", deviceFun); + m_kernel_lookup[hostFun] = NULL; + } + // assert( s != NULL ); + // function_info *f = s->get_pc(); + // assert( f != NULL ); + // m_kernel_lookup[hostFun] = f; } else { m_kernel_lookup[hostFun] = NULL; } @@ -321,9 +332,9 @@ class _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 = 2; - prop->minor = 0; - prop->totalGlobalMem = 0x40000000 /* 1 GB */; + prop->major = 5; + prop->minor = 2; + prop->totalGlobalMem = 0x80000000 /* 2 GB */; prop->memPitch = 0; prop->maxThreadsPerBlock = 512; prop->maxThreadsDim[0] = 512; @@ -440,6 +451,10 @@ extern "C" { * * * * *******************************************************************************/ +cudaError_t cudaPeekAtLastError(void) +{ + return g_last_cudaError; +} __host__ cudaError_t CUDARTAPI cudaMalloc(void **devPtr, size_t size) { @@ -534,6 +549,22 @@ __host__ cudaError_t CUDARTAPI cudaMemcpy(void *dst, const void *src, size_t cou g_stream_manager->push( stream_operation((size_t)src,dst,count,0) ); else if( kind == cudaMemcpyDeviceToDevice ) g_stream_manager->push( stream_operation((size_t)src,(size_t)dst,count,0) ); + else if ( kind == cudaMemcpyDefault ) { + if ((size_t)src >= GLOBAL_HEAP_START) { + if ((size_t)dst >= GLOBAL_HEAP_START) + g_stream_manager->push( stream_operation((size_t)src,(size_t)dst,count,0) ); // device to device + else + g_stream_manager->push( stream_operation((size_t)src,dst,count,0) ); // device to host + } + else { + if ((size_t)dst >= GLOBAL_HEAP_START) + g_stream_manager->push( stream_operation(src,(size_t)dst,count,0) ); + else { + printf("GPGPU-Sim PTX: cudaMemcpy - ERROR : unsupported transfer: host to host\n"); + abort(); + } + } + } else { printf("GPGPU-Sim PTX: cudaMemcpy - ERROR : unsupported cudaMemcpyKind\n"); abort(); @@ -968,6 +999,10 @@ __host__ cudaError_t CUDARTAPI cudaStreamCreate(cudaStream_t *stream) return g_last_cudaError = cudaSuccess; } +__host__ __device__ cudaError_t CUDARTAPI cudaStreamCreateWithFlags(cudaStream_t *stream, unsigned int flags) { + return cudaStreamCreate(stream); +} + __host__ cudaError_t CUDARTAPI cudaStreamDestroy(cudaStream_t stream) { #if (CUDART_VERSION >= 3000) @@ -980,7 +1015,8 @@ __host__ cudaError_t CUDARTAPI cudaStreamSynchronize(cudaStream_t stream) { #if (CUDART_VERSION >= 3000) if( stream == NULL ) - return g_last_cudaError = cudaErrorInvalidResourceHandle; + synchronize(); + return g_last_cudaError = cudaSuccess; stream->synchronize(); #else printf("GPGPU-Sim PTX: WARNING: Asynchronous kernel execution not supported (%s)\n", __my_func__); @@ -1328,7 +1364,12 @@ void extract_code_using_cuobjdump(){ printf("Running md5sum using \"%s\"\n", command); system(command); // Running cuobjdump using dynamic link to current process - snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass %s > %s", app_binary.c_str(), fname); + // Needs the option '-all' to extract PTX from CDP-enabled binary + extern bool g_cdp_enabled; + if(!g_cdp_enabled) + snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass %s > %s", app_binary.c_str(), fname); + else + snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass -all %s > %s", app_binary.c_str(), fname); bool parse_output = true; int result = system(command); if(result) { @@ -1497,6 +1538,78 @@ std::list<cuobjdumpSection*> pruneSectionList(std::list<cuobjdumpSection*> cuobj return prunedList; } +//! Merge all PTX sections that have a specific identifier into one file +std::list<cuobjdumpSection*> mergeMatchingSections(std::list<cuobjdumpSection*> cuobjdumpSectionList, std::string identifier){ + char *ptxcode = ""; + std::list<cuobjdumpSection*>::iterator old_iter; + cuobjdumpPTXSection* old_ptxsection = NULL; + cuobjdumpPTXSection* ptxsection; + std::list<cuobjdumpSection*> mergedList; + + for ( std::list<cuobjdumpSection*>::iterator iter = cuobjdumpSectionList.begin(); + iter != cuobjdumpSectionList.end(); + iter++){ + if((ptxsection=dynamic_cast<cuobjdumpPTXSection*>(*iter)) != NULL && + strcmp(ptxsection->getIdentifier().c_str(), identifier.c_str()) == 0){ + // Read and remove the last PTX section + if (old_ptxsection != NULL) { + ptxcode = readfile(old_ptxsection->getPTXfilename()); + // remove ptx file? + delete *old_iter; + } + + // Append all the PTX from the last PTX section into the current PTX section + // Add 50 to ptxcode to ignore the information regarding version/target/address_size + if (strlen(ptxcode) >= 50) { + FILE *ptxfile = fopen((ptxsection->getPTXfilename()).c_str(), "a"); + fprintf(ptxfile, "%s", ptxcode + 50); + fclose(ptxfile); + } + + old_iter = iter; + old_ptxsection = ptxsection; + } + // Store all non-PTX sections and PTX sections with non-matching identifiers + else { + mergedList.push_back(*iter); + } + } + + // Store the final PTX section + mergedList.push_back(*old_iter); + + return mergedList; +} + +//! Merge any PTX sections with matching identifiers +std::list<cuobjdumpSection*> mergeSections(std::list<cuobjdumpSection*> cuobjdumpSectionList){ + std::vector<std::string> identifier; + cuobjdumpPTXSection* ptxsection; + + // Add all identifiers present in PTX sections to a vector + for ( std::list<cuobjdumpSection*>::iterator iter = cuobjdumpSectionList.begin(); + iter != cuobjdumpSectionList.end(); + iter++){ + if((ptxsection=dynamic_cast<cuobjdumpPTXSection*>(*iter)) != NULL){ + std::string current_id = ptxsection->getIdentifier(); + + // If we haven't yet seen a given identifier, add it to the vector + if (std::find(identifier.begin(), identifier.end(), current_id) == identifier.end()) { + identifier.push_back(current_id); + } + } + } + + // Call mergeMatchingSections on all identifiers in the vector + for ( std::vector<std::string>::iterator iter = identifier.begin(); + iter != identifier.end(); + iter++) { + cuobjdumpSectionList = mergeMatchingSections(cuobjdumpSectionList, *iter); + } + + return cuobjdumpSectionList; +} + //! Within the section list, find the ELF section corresponding to a given identifier cuobjdumpELFSection* findELFSectionInList(std::list<cuobjdumpSection*> sectionlist, const std::string identifier){ @@ -1521,7 +1634,7 @@ cuobjdumpELFSection* findELFSection(const std::string identifier){ if (sec!=NULL)return sec; sec = findELFSectionInList(libSectionList, identifier); if (sec!=NULL)return sec; - std::cout << "Cound not find " << identifier << std::endl; + std::cout << "Could not find " << identifier << std::endl; assert(0 && "Could not find the required ELF section"); return NULL; } @@ -1537,6 +1650,14 @@ cuobjdumpPTXSection* findPTXSectionInList(std::list<cuobjdumpSection*> sectionli if((ptxsection=dynamic_cast<cuobjdumpPTXSection*>(*iter)) != NULL){ if(ptxsection->getIdentifier() == identifier) return ptxsection; + else { + extern bool g_cdp_enabled; + if(g_cdp_enabled) { + printf("Warning: __cudaRegisterFatBinary needs %s, but find PTX section with %s\n", + identifier.c_str(), ptxsection->getIdentifier().c_str()); + return ptxsection; + } + } } } return NULL; @@ -1548,7 +1669,7 @@ cuobjdumpPTXSection* findPTXSection(const std::string identifier){ if (sec!=NULL)return sec; sec = findPTXSectionInList(libSectionList, identifier); if (sec!=NULL)return sec; - std::cout << "Cound not find " << identifier << std::endl; + std::cout << "Could not find " << identifier << std::endl; assert(0 && "Could not find the required PTX section"); return NULL; } @@ -1560,10 +1681,12 @@ void cuobjdumpInit(){ CUctx_st *context = GPGPUSim_Context(); extract_code_using_cuobjdump(); //extract all the output of cuobjdump to _cuobjdump_*.* cuobjdumpSectionList = pruneSectionList(cuobjdumpSectionList, context); + cuobjdumpSectionList = mergeSections(cuobjdumpSectionList); } std::map<int, std::string> fatbinmap; std::map<int, bool>fatbin_registered; +std::map<std::string, symbol_table*> name_symtab; //! Keep track of the association between filename and cubin handle void cuobjdumpRegisterFatBinary(unsigned int handle, char* filename){ @@ -1576,6 +1699,13 @@ void cuobjdumpParseBinary(unsigned int handle){ if(fatbin_registered[handle]) return; fatbin_registered[handle] = true; CUctx_st *context = GPGPUSim_Context(); + std::string fname = fatbinmap[handle]; + + if (name_symtab.find(fname) != name_symtab.end()) { + symbol_table *symtab = name_symtab[fname]; + context->add_binary(symtab, handle); + return; + } unsigned max_capability = 0; for ( std::list<cuobjdumpSection*>::iterator iter = cuobjdumpSectionList.begin(); @@ -1584,16 +1714,13 @@ void cuobjdumpParseBinary(unsigned int handle){ unsigned capability = (*iter)->getArch(); if (capability > max_capability) max_capability = capability; } - if (max_capability > 20) printf("WARNING: No guarantee that PTX will be parsed for SM version %u\n", max_capability); - std::string fname = fatbinmap[handle]; cuobjdumpPTXSection* ptx = findPTXSection(fname); - symbol_table *symtab; char *ptxcode; const char *override_ptx_name = getenv("PTX_SIM_KERNELFILE"); - if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL) { + if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL) { ptxcode = readfile(ptx->getPTXfilename()); } else { printf("GPGPU-Sim PTX: overriding embedded ptx with '%s' (PTX_SIM_USE_PTX_FILE is set)\n", override_ptx_name); @@ -1619,6 +1746,7 @@ void cuobjdumpParseBinary(unsigned int handle){ } load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); + name_symtab[fname] = symtab; //TODO: Remove temporarily files as per configurations } @@ -1821,10 +1949,15 @@ void __cudaRegisterTexture( int ext ) //passes in a newly created textureReference { + std::string devStr (deviceName); + #if (CUDART_VERSION > 4020) + if (devStr.size() > 2 && devStr.data()[0] == ':' && devStr.data()[1] == ':') + devStr = devStr.replace(0, 2, ""); + #endif CUctx_st *context = GPGPUSim_Context(); gpgpu_t *gpu = context->get_device()->get_gpgpu(); printf("GPGPU-Sim PTX: in __cudaRegisterTexture:\n"); - gpu->gpgpu_ptx_sim_bindNameToTexture(deviceName, hostVar, dim, norm, ext); + gpu->gpgpu_ptx_sim_bindNameToTexture(devStr.data(), hostVar, dim, norm, ext); printf("GPGPU-Sim PTX: int dim = %d\n", dim); printf("GPGPU-Sim PTX: int norm = %d\n", norm); printf("GPGPU-Sim PTX: int ext = %d\n", ext); @@ -2013,6 +2146,11 @@ __host__ cudaError_t CUDARTAPI cudaFuncSetCacheConfig(const char *func, enum cud context->get_device()->get_gpgpu()->set_cache_config(context->get_kernel(func)->get_name(), (FuncCache)cacheConfig); return g_last_cudaError = cudaSuccess; } + +//Jin: hack for cdp +__host__ cudaError_t CUDARTAPI cudaDeviceSetLimit(enum cudaLimit limit, size_t value) { + return g_last_cudaError = cudaSuccess; +} #endif #endif diff --git a/libcuda/cuobjdump.y b/libcuda/cuobjdump.y index 9d61f25..31760f7 100644 --- a/libcuda/cuobjdump.y +++ b/libcuda/cuobjdump.y @@ -82,8 +82,7 @@ section : PTXHEADER { snprintf(filename, 1024, "_cuobjdump_%d.elf", elfserial); elffile = fopen(filename, "w"); setCuobjdumpelffilename(filename); - } headerinfo identifier{ - } elfcode { + } headerinfo compressedkeyword identifier elfcode { fclose(elffile); snprintf(filename, 1024, "_cuobjdump_%d.sass", elfserial++); sassfile = fopen(filename, "w"); diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index 84d165c..fe6f8ab 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -550,6 +550,9 @@ void warp_inst_t::completed( unsigned long long cycle ) const ptx_file_line_stats_add_latency(pc, latency * active_count()); } +//Jin: CDP support +bool g_cdp_enabled; +unsigned g_kernel_launch_latency; unsigned kernel_info_t::m_next_uid = 1; @@ -565,11 +568,18 @@ kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info * m_num_cores_running=0; m_uid = m_next_uid++; m_param_mem = new memory_space_impl<8192>("param",64*1024); + + //Jin: parent and child kernel management for CDP + m_parent_kernel = NULL; + + //Jin: launch latency management + m_launch_latency = g_kernel_launch_latency; } kernel_info_t::~kernel_info_t() { assert( m_active_threads.empty() ); + destroy_cta_streams(); delete m_param_mem; } @@ -578,6 +588,106 @@ std::string kernel_info_t::name() const return m_kernel_entry->get_name(); } +//Jin: parent and child kernel management for CDP +void kernel_info_t::set_parent(kernel_info_t * parent, + dim3 parent_ctaid, dim3 parent_tid) { + m_parent_kernel = parent; + m_parent_ctaid = parent_ctaid; + m_parent_tid = parent_tid; + parent->set_child(this); +} + +void kernel_info_t::set_child(kernel_info_t * child) { + m_child_kernels.push_back(child); +} + +void kernel_info_t::remove_child(kernel_info_t * child) { + assert(std::find(m_child_kernels.begin(), m_child_kernels.end(), child) + != m_child_kernels.end()); + m_child_kernels.remove(child); +} + +bool kernel_info_t::is_finished() { + if(done() && children_all_finished()) + return true; + else + return false; +} + +bool kernel_info_t::children_all_finished() { + if(!m_child_kernels.empty()) + return false; + + return true; +} + +void kernel_info_t::notify_parent_finished() { + if(m_parent_kernel) { + extern unsigned long long g_total_param_size; + g_total_param_size -= ((m_kernel_entry->get_args_aligned_size() + 255)/256*256); + m_parent_kernel->remove_child(this); + g_stream_manager->register_finished_kernel(m_parent_kernel->get_uid()); + } +} + +CUstream_st * kernel_info_t::create_stream_cta(dim3 ctaid) { + assert(get_default_stream_cta(ctaid)); + CUstream_st * stream = new CUstream_st(); + g_stream_manager->add_stream(stream); + assert(m_cta_streams.find(ctaid) != m_cta_streams.end()); + assert(m_cta_streams[ctaid].size() >= 1); //must have default stream + m_cta_streams[ctaid].push_back(stream); + + return stream; +} + +CUstream_st * kernel_info_t::get_default_stream_cta(dim3 ctaid) { + if(m_cta_streams.find(ctaid) != m_cta_streams.end()) { + assert(m_cta_streams[ctaid].size() >= 1); //already created, must have default stream + return *(m_cta_streams[ctaid].begin()); + } + else { + m_cta_streams[ctaid] = std::list<CUstream_st *>(); + CUstream_st * stream = new CUstream_st(); + g_stream_manager->add_stream(stream); + m_cta_streams[ctaid].push_back(stream); + return stream; + } +} + +bool kernel_info_t::cta_has_stream(dim3 ctaid, CUstream_st* stream) { + if(m_cta_streams.find(ctaid) == m_cta_streams.end()) + return false; + + std::list<CUstream_st *> &stream_list = m_cta_streams[ctaid]; + if(std::find(stream_list.begin(), stream_list.end(), stream) + == stream_list.end()) + return false; + else + return true; +} + +void kernel_info_t::print_parent_info() { + if(m_parent_kernel) { + printf("Parent %d: \'%s\', Block (%d, %d, %d), Thread (%d, %d, %d)\n", + m_parent_kernel->get_uid(), m_parent_kernel->name().c_str(), + m_parent_ctaid.x, m_parent_ctaid.y, m_parent_ctaid.z, + m_parent_tid.x, m_parent_tid.y, m_parent_tid.z); + } +} + +void kernel_info_t::destroy_cta_streams() { + printf("Destroy streams for kernel %d: ", get_uid()); size_t stream_size = 0; + for(auto s = m_cta_streams.begin(); s != m_cta_streams.end(); s++) { + stream_size += s->second.size(); + for(auto ss = s->second.begin(); ss != s->second.end(); ss++) + g_stream_manager->destroy_stream(*ss); + s->second.clear(); + } + printf("size %lu\n", stream_size); + m_cta_streams.clear(); +} + simt_stack::simt_stack( unsigned wid, unsigned warpSize) { m_warp_id=wid; diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index ba4ea29..aaa4b00 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -154,15 +154,35 @@ enum _memory_op_t { #include <stdlib.h> #include <map> #include <deque> +#include <algorithm> #if !defined(__VECTOR_TYPES_H__) struct dim3 { unsigned int x, y, z; }; #endif +struct dim3comp { + bool operator() (const dim3 & a, const dim3 & b) const + { + if(a.z < b.z) + return true; + else if(a.y < b.y) + return true; + else if (a.x < b.x) + return true; + else + return false; + } +}; void increment_x_then_y_then_z( dim3 &i, const dim3 &bound); +//Jin: child kernel information for CDP +#include "stream_manager.h" +class stream_manager; +struct CUstream_st; +extern stream_manager * g_stream_manager; + class kernel_info_t { public: // kernel_info_t() @@ -250,6 +270,35 @@ private: std::list<class ptx_thread_info *> m_active_threads; class memory_space *m_param_mem; + +public: + //Jin: parent and child kernel management for CDP + void set_parent(kernel_info_t * parent, dim3 parent_ctaid, dim3 parent_tid); + void set_child(kernel_info_t * child); + void remove_child(kernel_info_t * child); + bool is_finished(); + bool children_all_finished(); + void notify_parent_finished(); + CUstream_st * create_stream_cta(dim3 ctaid); + CUstream_st * get_default_stream_cta(dim3 ctaid); + bool cta_has_stream(dim3 ctaid, CUstream_st* stream); + void destroy_cta_streams(); + void print_parent_info(); + kernel_info_t * get_parent() { return m_parent_kernel; } + +private: + kernel_info_t * m_parent_kernel; + dim3 m_parent_ctaid; + dim3 m_parent_tid; + std::list<kernel_info_t *> m_child_kernels; //child kernel launched + std::map< dim3, std::list<CUstream_st *>, dim3comp > m_cta_streams; //streams created in each CTA + +//Jin: kernel timing +public: + unsigned long long launch_cycle; + unsigned long long start_cycle; + unsigned long long end_cycle; + unsigned m_launch_latency; }; struct core_config { @@ -334,7 +383,7 @@ protected: std::deque<simt_stack_entry> m_stack; }; -#define GLOBAL_HEAP_START 0x80000000 +#define GLOBAL_HEAP_START 0x703E20000 // start allocating from this address (lower values used for allocating globals in .ptx file) #define SHARED_MEM_SIZE_MAX (64*1024) #define LOCAL_MEM_SIZE_MAX (8*1024) @@ -825,6 +874,7 @@ public: m_mem_accesses_created=false; m_cache_hit=false; m_is_printf=false; + m_is_cdp = 0; } virtual ~warp_inst_t(){ } @@ -997,6 +1047,11 @@ protected: std::list<mem_access_t> m_accessq; static unsigned sm_next_uid; + + //Jin: cdp support +public: + int m_is_cdp; + }; void move_warp( warp_inst_t *&dst, warp_inst_t *&src ); @@ -1051,6 +1106,7 @@ class core_t { warp_inst_t getExecuteWarp(unsigned warpId); void get_pdom_stack_top_info( unsigned warpId, unsigned *pc, unsigned *rpc ) const; kernel_info_t * get_kernel_info(){ return m_kernel;} + class ptx_thread_info ** get_thread_info() { return m_thread; } unsigned get_warp_size() const { return m_warp_size; } void and_reduction(unsigned ctaid, unsigned barid, bool value) { reduction_storage[ctaid][barid] &= value; } void or_reduction(unsigned ctaid, unsigned barid, bool value) { reduction_storage[ctaid][barid] |= value; } diff --git a/src/cuda-sim/Makefile b/src/cuda-sim/Makefile index 166e256..f479294 100644 --- a/src/cuda-sim/Makefile +++ b/src/cuda-sim/Makefile @@ -62,7 +62,7 @@ ifeq ($(GNUC_CPP0X),1) endif endif -OBJS := $(OUTPUT_DIR)/ptx_parser.o $(OUTPUT_DIR)/ptx_loader.o $(OUTPUT_DIR)/cuda_device_printf.o $(OUTPUT_DIR)/instructions.o $(OUTPUT_DIR)/cuda-sim.o $(OUTPUT_DIR)/ptx_ir.o $(OUTPUT_DIR)/ptx_sim.o $(OUTPUT_DIR)/memory.o $(OUTPUT_DIR)/ptx-stats.o $(OUTPUT_DIR)/decuda_pred_table/decuda_pred_table.o $(OUTPUT_DIR)/ptx.tab.o $(OUTPUT_DIR)/lex.ptx_.o $(OUTPUT_DIR)/ptxinfo.tab.o $(OUTPUT_DIR)/lex.ptxinfo_.o +OBJS := $(OUTPUT_DIR)/ptx_parser.o $(OUTPUT_DIR)/ptx_loader.o $(OUTPUT_DIR)/cuda_device_printf.o $(OUTPUT_DIR)/instructions.o $(OUTPUT_DIR)/cuda-sim.o $(OUTPUT_DIR)/ptx_ir.o $(OUTPUT_DIR)/ptx_sim.o $(OUTPUT_DIR)/memory.o $(OUTPUT_DIR)/ptx-stats.o $(OUTPUT_DIR)/decuda_pred_table/decuda_pred_table.o $(OUTPUT_DIR)/ptx.tab.o $(OUTPUT_DIR)/lex.ptx_.o $(OUTPUT_DIR)/ptxinfo.tab.o $(OUTPUT_DIR)/lex.ptxinfo_.o $(OUTPUT_DIR)/cuda_device_runtime.o OPT += -DCUDART_VERSION=$(CUDART_VERSION) @@ -145,5 +145,6 @@ $(OUTPUT_DIR)/ptx_sim.o: $(OUTPUT_DIR)/ptx.tab.c $(OUTPUT_DIR)/cuda-sim.o: $(OUTPUT_DIR)/ptx.tab.c $(OUTPUT_DIR)/lex.ptxinfo_.o: $(OUTPUT_DIR)/ptx.tab.c $(OUTPUT_DIR)/lex.ptx_.o: $(OUTPUT_DIR)/ptx.tab.c +$(OUTPUT_DIR)/cuda_device_runtime.o: $(OUTPUT_DIR)/ptx.tab.c include $(OUTPUT_DIR)/Makefile.makedepend diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 715be98..7ec3ce9 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -48,6 +48,7 @@ #include "../gpgpusim_entrypoint.h" #include "decuda_pred_table/decuda_pred_table.h" #include "../stream_manager.h" +#include "cuda_device_runtime.h" int gpgpu_ptx_instruction_classification; void ** g_inst_classification_stat = NULL; @@ -63,6 +64,8 @@ unsigned gpgpu_param_num_shaders = 0; char *opcode_latency_int, *opcode_latency_fp, *opcode_latency_dp; char *opcode_initiation_int, *opcode_initiation_fp, *opcode_initiation_dp; +char *cdp_latency_str; +unsigned cdp_latency[5]; void ptx_opcocde_latency_options (option_parser_t opp) { option_parser_register(opp, "-ptx_opcode_latency_int", OPT_CSTR, &opcode_latency_int, @@ -89,6 +92,12 @@ void ptx_opcocde_latency_options (option_parser_t opp) { "Opcode initiation intervals for double precision floating points <ADD,MAX,MUL,MAD,DIV>" "Default 8,8,8,8,130", "8,8,8,8,130"); + option_parser_register(opp, "-cdp_latency", OPT_CSTR, &cdp_latency_str, + "CDP API latency <cudaStreamCreateWithFlags, \ +cudaGetParameterBufferV2_init_perWarp, cudaGetParameterBufferV2_perKernel, \ +cudaLaunchDeviceV2_init_perWarp, cudaLaunchDevicV2_perKernel>" + "Default 7200,8000,100,12000,1600", + "7200,8000,100,12000,1600"); } static address_type get_converge_point(address_type pc); @@ -608,6 +617,9 @@ void ptx_instruction::set_opcode_and_latency() sscanf(opcode_initiation_dp, "%u,%u,%u,%u,%u", &dp_init[0],&dp_init[1],&dp_init[2], &dp_init[3],&dp_init[4]); + sscanf(cdp_latency_str, "%u,%u,%u,%u,%u", + &cdp_latency[0],&cdp_latency[1],&cdp_latency[2], + &cdp_latency[3],&cdp_latency[4]); if(!m_operands.empty()){ std::vector<operand_info>::iterator it; @@ -638,19 +650,21 @@ void ptx_instruction::set_opcode_and_latency() case MEMBAR_OP: op = MEMORY_BARRIER_OP; break; case CALL_OP: { - if(m_is_printf) + if(m_is_printf || m_is_cdp) { op = ALU_OP; + } else op = CALL_OPS; break; } case CALLP_OP: { - if(m_is_printf) + if(m_is_printf || m_is_cdp) { op = ALU_OP; - else - op = CALL_OPS; - break; + } + else + op = CALL_OPS; + break; } case RET_OP: case RETP_OP: op = RET_OPS;break; case ADD_OP: case ADDP_OP: case ADDC_OP: case SUB_OP: case SUBC_OP: @@ -719,7 +733,7 @@ void ptx_instruction::set_opcode_and_latency() break; } break; - case MAD_OP: case MADP_OP: + case MAD_OP: case MADC_OP: case MADP_OP: //MAD latency switch(get_type()){ case F32_TYPE: @@ -769,6 +783,10 @@ void ptx_instruction::set_opcode_and_latency() initiation_interval = dp_init[2]; op = SFU_OP; break; + case SHFL_OP: + latency = 32; + initiation_interval = 15; + break; default: break; } @@ -845,8 +863,10 @@ void ptx_instruction::pre_decode() switch ( get_opcode() ) { #define OP_DEF(OP,FUNC,STR,DST,CLASSIFICATION) case OP: has_dst = (DST!=0); break; +#define OP_W_DEF(OP,FUNC,STR,DST,CLASSIFICATION) case OP: has_dst = (DST!=0); break; #include "opcodes.def" #undef OP_DEF +#undef OP_W_DEF default: printf( "Execution error: Invalid opcode (0x%x)\n", get_opcode() ); break; @@ -1065,6 +1085,32 @@ void function_info::add_param_data( unsigned argn, struct gpgpu_ptx_sim_arg *arg } } +unsigned function_info::get_args_aligned_size() { + + if(m_args_aligned_size >= 0) + return m_args_aligned_size; + + unsigned param_address = 0; + unsigned int total_size = 0; + for( std::map<unsigned,param_info>::iterator i=m_ptx_kernel_param_info.begin(); i!=m_ptx_kernel_param_info.end(); i++ ) { + param_info &p = i->second; + std::string name = p.get_name(); + symbol *param = m_symtab->lookup(name.c_str()); + + size_t arg_size = p.get_size() / 8; // size of param in bytes + total_size = (total_size + arg_size - 1) / arg_size * arg_size; //aligned + p.add_offset(total_size); + param->set_address(param_address + total_size); + total_size += arg_size; + } + + m_args_aligned_size = (total_size + 3) / 4 * 4; //final size aligned to word + + return m_args_aligned_size; + +} + + void function_info::finalize( memory_space *param_mem ) { unsigned param_address = 0; @@ -1087,13 +1133,17 @@ void function_info::finalize( memory_space *param_mem ) size = (size<(p.get_size()/8))?size:(p.get_size()/8); } // copy the parameter over word-by-word so that parameter that crosses a memory page can be copied over + //Jin: copy parameter using aligned rules const size_t word_size = 4; + param_address = (param_address + size - 1) / size * size; //aligned with size for (size_t idx = 0; idx < size; idx += word_size) { const char *pdata = reinterpret_cast<const char*>(param_value.pdata) + idx; // cast to char * for ptr arithmetic param_mem->write(param_address + idx, word_size, pdata,NULL,NULL); } + unsigned offset = p.get_offset(); + assert(offset == param_address); param->set_address(param_address); - param_address += size; + param_address += size; } } @@ -1240,8 +1290,10 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) } switch ( pI->get_opcode() ) { #define OP_DEF(OP,FUNC,STR,DST,CLASSIFICATION) case OP: FUNC(pI,this); op_classification = CLASSIFICATION; break; +#define OP_W_DEF(OP,FUNC,STR,DST,CLASSIFICATION) case OP: FUNC(pI,get_core(),inst); op_classification = CLASSIFICATION; break; #include "opcodes.def" #undef OP_DEF +#undef OP_W_DEF default: printf( "Execution error: Invalid opcode (0x%x)\n", pI->get_opcode() ); break; } delete pJ; @@ -1408,6 +1460,7 @@ unsigned ptx_sim_init_thread( kernel_info_t &kernel, static std::map<unsigned,memory_space*> shared_memory_lookup; static std::map<unsigned,ptx_cta_info*> ptx_cta_lookup; + static std::map<unsigned,ptx_warp_info*> ptx_warp_lookup; static std::map<unsigned,std::map<unsigned,memory_space*> > local_memory_lookup; if ( *thread_info != NULL ) { @@ -1455,7 +1508,8 @@ unsigned ptx_sim_init_thread( kernel_info_t &kernel, unsigned max_cta_per_sm = num_threads/cta_size; // e.g., 256 / 48 = 5 assert( max_cta_per_sm > 0 ); - unsigned sm_idx = (tid/cta_size)*gpgpu_param_num_shaders + sid; + //unsigned sm_idx = (tid/cta_size)*gpgpu_param_num_shaders + sid; + unsigned sm_idx = hw_cta_id*gpgpu_param_num_shaders + sid; if ( shared_memory_lookup.find(sm_idx) == shared_memory_lookup.end() ) { if ( g_debug_execution >= 1 ) { @@ -1486,7 +1540,16 @@ unsigned ptx_sim_init_thread( kernel_info_t &kernel, kernel.increment_thread_id(); new_tid += tid; ptx_thread_info *thd = new ptx_thread_info(kernel); - + + ptx_warp_info *warp_info = NULL; + if ( ptx_warp_lookup.find(hw_warp_id) == ptx_warp_lookup.end() ) { + warp_info = new ptx_warp_info(); + ptx_warp_lookup[hw_warp_id] = warp_info; + } else { + warp_info = ptx_warp_lookup[hw_warp_id]; + } + thd->m_warp_info = warp_info; + memory_space *local_mem = NULL; std::map<unsigned,memory_space*>::iterator l = local_mem_lookup.find(new_tid); if ( l != local_mem_lookup.end() ) { @@ -1746,14 +1809,19 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) g_the_gpu->getShaderCoreConfig()->warp_size ); cta.execute(); + +#if (CUDART_VERSION >= 5000) + launch_all_device_kernels(); +#endif } //registering this kernel as done - extern stream_manager *g_stream_manager; //openCL kernel simulation calls don't register the kernel so we don't register its exit - if(!openCL) - g_stream_manager->register_finished_kernel(kernel.get_uid()); + if(!openCL) { + extern stream_manager *g_stream_manager; + g_stream_manager->register_finished_kernel(kernel.get_uid()); + } //******PRINTING******* printf( "GPGPU-Sim: Done functional simulation (%u instructions simulated).\n", g_ptx_sim_num_insn ); @@ -1864,10 +1932,14 @@ unsigned translate_pc_to_ptxlineno(unsigned pc) // ptxinfo parser +extern std::map<unsigned,const char*> get_duplicate(); + int g_ptxinfo_error_detected; static char *g_ptxinfo_kname = NULL; static struct gpgpu_ptx_sim_info g_ptxinfo; +static std::map<unsigned,const char*> g_duplicate; +static const char *g_last_dup_type; const char *get_ptxinfo_kname() { @@ -1897,6 +1969,21 @@ struct gpgpu_ptx_sim_info get_ptxinfo() return g_ptxinfo; } +std::map<unsigned,const char*> get_duplicate() +{ + return g_duplicate; +} + +void ptxinfo_linenum( unsigned linenum ) +{ + g_duplicate[linenum] = g_last_dup_type; +} + +void ptxinfo_dup_type( const char *dup_type ) +{ + g_last_dup_type = dup_type; +} + void ptxinfo_function(const char *fname ) { clear_ptxinfo(); diff --git a/src/cuda-sim/cuda_device_runtime.cc b/src/cuda-sim/cuda_device_runtime.cc new file mode 100644 index 0000000..4a8ffe5 --- /dev/null +++ b/src/cuda-sim/cuda_device_runtime.cc @@ -0,0 +1,320 @@ +//Jin: cuda_device_runtime.cc +//Defines CUDA device runtime APIs for CDP support + + +#include <iostream> +#include <map> + +unsigned long long g_total_param_size = 0; +unsigned long long g_max_total_param_size = 0; + + +#if (CUDART_VERSION >= 5000) +#define __CUDA_RUNTIME_API_H__ + +#include <builtin_types.h> +#include <driver_types.h> +#include "../gpgpu-sim/gpu-sim.h" +#include "cuda-sim.h" +#include "ptx_ir.h" +#include "../stream_manager.h" +#include "cuda_device_runtime.h" + +#define DEV_RUNTIME_REPORT(a) \ + if( g_debug_execution ) { \ + std::cout << __FILE__ << ", " << __LINE__ << ": " << a << "\n"; \ + std::cout.flush(); \ + } + +class device_launch_config_t { + +public: + device_launch_config_t() {} + + device_launch_config_t(dim3 _grid_dim, + dim3 _block_dim, + unsigned int _shared_mem, + function_info * _entry): + grid_dim(_grid_dim), + block_dim(_block_dim), + shared_mem(_shared_mem), + entry(_entry) {} + + dim3 grid_dim; + dim3 block_dim; + unsigned int shared_mem; + function_info * entry; + +}; + +class device_launch_operation_t { + +public: + device_launch_operation_t() {} + device_launch_operation_t(kernel_info_t *_grid, + CUstream_st * _stream) : + grid(_grid), stream(_stream) {} + + kernel_info_t * grid; //a new child grid + + CUstream_st * stream; + +}; + + +std::map<void *, device_launch_config_t> g_cuda_device_launch_param_map; +std::list<device_launch_operation_t> g_cuda_device_launch_op; +extern stream_manager *g_stream_manager; + +//Handling device runtime api: +//void * cudaGetParameterBufferV2(void *func, dim3 gridDimension, dim3 blockDimension, unsigned int sharedMemSize) +void gpgpusim_cuda_getParameterBufferV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func) +{ + DEV_RUNTIME_REPORT("Calling cudaGetParameterBufferV2"); + + unsigned n_return = target_func->has_return(); + assert(n_return); + unsigned n_args = target_func->num_args(); + assert( n_args == 4 ); + + function_info * child_kernel_entry; + struct dim3 grid_dim, block_dim; + unsigned int shared_mem; + + for( unsigned arg=0; arg < n_args; arg ++ ) { + const operand_info &actual_param_op = pI->operand_lookup(n_return+1+arg); //param# + const symbol *formal_param = target_func->get_arg(arg); //cudaGetParameterBufferV2_param_# + unsigned size=formal_param->get_size_in_bytes(); + assert( formal_param->is_param_local() ); + assert( actual_param_op.is_param_local() ); + addr_t from_addr = actual_param_op.get_symbol()->get_address(); + + if(arg == 0) {//function_info* for the child kernel + unsigned long long buf; + assert(size == sizeof(function_info *)); + thread->m_local_mem->read(from_addr, size, &buf); + child_kernel_entry = (function_info *)buf; + assert(child_kernel_entry); + DEV_RUNTIME_REPORT("child kernel name " << child_kernel_entry->get_name()); + } + else if(arg == 1) { //dim3 grid_dim for the child kernel + assert(size == sizeof(struct dim3)); + thread->m_local_mem->read(from_addr, size, & grid_dim); + DEV_RUNTIME_REPORT("grid (" << grid_dim.x << ", " << grid_dim.y << ", " << grid_dim.z << ")"); + } + else if(arg == 2) { //dim3 block_dim for the child kernel + assert(size == sizeof(struct dim3)); + thread->m_local_mem->read(from_addr, size, & block_dim); + DEV_RUNTIME_REPORT("block (" << block_dim.x << ", " << block_dim.y << ", " << block_dim.z << ")"); + } + else if(arg == 3) { //unsigned int shared_mem + assert(size == sizeof(unsigned int)); + thread->m_local_mem->read(from_addr, size, & shared_mem); + DEV_RUNTIME_REPORT("shared memory " << shared_mem); + } + } + + //get total child kernel argument size and malloc buffer in global memory + unsigned child_kernel_arg_size = child_kernel_entry->get_args_aligned_size(); + void * param_buffer = thread->get_gpu()->gpu_malloc(child_kernel_arg_size); + g_total_param_size += ((child_kernel_arg_size + 255) / 256 * 256); + DEV_RUNTIME_REPORT("child kernel arg size total " << child_kernel_arg_size << ", parameter buffer allocated at " << param_buffer); + if(g_total_param_size > g_max_total_param_size) + g_max_total_param_size = g_total_param_size; + + //store param buffer address and launch config + device_launch_config_t device_launch_config(grid_dim, block_dim, shared_mem, child_kernel_entry); + assert(g_cuda_device_launch_param_map.find(param_buffer) == g_cuda_device_launch_param_map.end()); + g_cuda_device_launch_param_map[param_buffer] = device_launch_config; + + //copy the buffer address to retval0 + const operand_info &actual_return_op = pI->operand_lookup(0); //retval0 + const symbol *formal_return = target_func->get_return_var(); //void * + unsigned int return_size = formal_return->get_size_in_bytes(); + DEV_RUNTIME_REPORT("cudaGetParameterBufferV2 return value has size of " << return_size); + assert(actual_return_op.is_param_local()); + assert(actual_return_op.get_symbol()->get_size_in_bytes() == return_size && return_size == sizeof(void *)); + addr_t ret_param_addr = actual_return_op.get_symbol()->get_address(); + thread->m_local_mem->write(ret_param_addr, return_size, ¶m_buffer, NULL, NULL); + +} + +//Handling device runtime api: +//cudaError_t cudaLaunchDeviceV2(void *parameterBuffer, cudaStream_t stream) +void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func) { + DEV_RUNTIME_REPORT("Calling cudaLaunchDeviceV2"); + + unsigned n_return = target_func->has_return(); + assert(n_return); + unsigned n_args = target_func->num_args(); + assert( n_args == 2 ); + + kernel_info_t * device_grid = NULL; + function_info * device_kernel_entry = NULL; + void * parameter_buffer; + struct CUstream_st * child_stream; + device_launch_config_t config; + device_launch_operation_t device_launch_op; + + for( unsigned arg=0; arg < n_args; arg ++ ) { + const operand_info &actual_param_op = pI->operand_lookup(n_return+1+arg); //param# + const symbol *formal_param = target_func->get_arg(arg); //cudaLaunchDeviceV2_param_# + unsigned size=formal_param->get_size_in_bytes(); + assert( formal_param->is_param_local() ); + assert( actual_param_op.is_param_local() ); + addr_t from_addr = actual_param_op.get_symbol()->get_address(); + + if(arg == 0) {//paramter buffer for child kernel (in global memory) + //get parameter_buffer from the cudaLaunchDeviceV2_param0 + assert(size == sizeof(void *)); + thread->m_local_mem->read(from_addr, size, ¶meter_buffer); + assert((size_t)parameter_buffer >= GLOBAL_HEAP_START); + DEV_RUNTIME_REPORT("Parameter buffer locating at global memory " << parameter_buffer); + + //get child grid info through parameter_buffer address + assert(g_cuda_device_launch_param_map.find(parameter_buffer) != g_cuda_device_launch_param_map.end()); + config = g_cuda_device_launch_param_map[parameter_buffer]; + //device_grid = op.grid; + device_kernel_entry = config.entry; + DEV_RUNTIME_REPORT("find device kernel " << device_kernel_entry->get_name()); + + //copy data in parameter_buffer to device kernel param memory + unsigned device_kernel_arg_size = device_kernel_entry->get_args_aligned_size(); + DEV_RUNTIME_REPORT("device_kernel_arg_size " << device_kernel_arg_size); + memory_space *device_kernel_param_mem; + + //create child kernel_info_t and index it with parameter_buffer address + device_grid = new kernel_info_t(config.grid_dim, config.block_dim, device_kernel_entry); + device_grid->launch_cycle = gpu_sim_cycle + gpu_tot_sim_cycle; + kernel_info_t & parent_grid = thread->get_kernel(); + DEV_RUNTIME_REPORT("child kernel launched by " << parent_grid.name() << ", cta (" << + thread->get_ctaid().x << ", " << thread->get_ctaid().y << ", " << thread->get_ctaid().z << + "), thread (" << thread->get_tid().x << ", " << thread->get_tid().y << ", " << thread->get_tid().z << + ")"); + device_grid->set_parent(&parent_grid, thread->get_ctaid(), thread->get_tid()); + device_launch_op = device_launch_operation_t(device_grid, NULL); + device_kernel_param_mem = device_grid->get_param_memory(); //kernel param + size_t param_start_address = 0; + //copy in word + for(unsigned n = 0; n < device_kernel_arg_size; n += 4) { + unsigned int oneword; + thread->get_gpu()->get_global_memory()->read((size_t)parameter_buffer + n, 4, &oneword); + device_kernel_param_mem->write(param_start_address + n, 4, &oneword, NULL, NULL); + } + } + else if(arg == 1) { //cudaStream for the child kernel + + assert(size == sizeof(cudaStream_t)); + thread->m_local_mem->read(from_addr, size, &child_stream); + + kernel_info_t & parent_kernel = thread->get_kernel(); + if(child_stream == 0) { //default stream on device for current CTA + child_stream = parent_kernel.get_default_stream_cta(thread->get_ctaid()); + DEV_RUNTIME_REPORT("launching child kernel " << device_grid->get_uid() << + " to default stream of the cta " << child_stream->get_uid() << ": " << child_stream); + } + else { + assert(parent_kernel.cta_has_stream(thread->get_ctaid(), child_stream)); + DEV_RUNTIME_REPORT("launching child kernel " << device_grid->get_uid() << + " to stream " << child_stream->get_uid() << ": " << child_stream); + } + + device_launch_op.stream = child_stream; + } + + } + + + //launch child kernel + g_cuda_device_launch_op.push_back(device_launch_op); + g_cuda_device_launch_param_map.erase(parameter_buffer); + + //set retval0 + const operand_info &actual_return_op = pI->operand_lookup(0); //retval0 + const symbol *formal_return = target_func->get_return_var(); //cudaError_t + unsigned int return_size = formal_return->get_size_in_bytes(); + DEV_RUNTIME_REPORT("cudaLaunchDeviceV2 return value has size of " << return_size); + assert(actual_return_op.is_param_local()); + assert(actual_return_op.get_symbol()->get_size_in_bytes() == return_size + && return_size == sizeof(cudaError_t)); + cudaError_t error = cudaSuccess; + addr_t ret_param_addr = actual_return_op.get_symbol()->get_address(); + thread->m_local_mem->write(ret_param_addr, return_size, &error, NULL, NULL); + +} + + +//Handling device runtime api: +//cudaError_t cudaStreamCreateWithFlags ( cudaStream_t* pStream, unsigned int flags) +//flags can only be cudaStreamNonBlocking +void gpgpusim_cuda_streamCreateWithFlags(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func) { + DEV_RUNTIME_REPORT("Calling cudaStreamCreateWithFlags"); + + unsigned n_return = target_func->has_return(); + assert(n_return); + unsigned n_args = target_func->num_args(); + assert( n_args == 2 ); + + size_t generic_pStream_addr; + addr_t pStream_addr; + unsigned int flags; + for( unsigned arg=0; arg < n_args; arg ++ ) { + const operand_info &actual_param_op = pI->operand_lookup(n_return+1+arg); //param# + const symbol *formal_param = target_func->get_arg(arg); //cudaStreamCreateWithFlags_param_# + unsigned size=formal_param->get_size_in_bytes(); + assert( formal_param->is_param_local() ); + assert( actual_param_op.is_param_local() ); + addr_t from_addr = actual_param_op.get_symbol()->get_address(); + + if(arg == 0) {//cudaStream_t * pStream, address of cudaStream_t + assert(size == sizeof(cudaStream_t *)); + thread->m_local_mem->read(from_addr, size, &generic_pStream_addr); + + //pStream should be non-zero address in local memory + pStream_addr = generic_to_local(thread->get_hw_sid(), thread->get_hw_tid(), generic_pStream_addr); + + DEV_RUNTIME_REPORT("pStream locating at local memory " << pStream_addr); + } + else if(arg == 1) { //unsigned int flags, should be cudaStreamNonBlocking + assert(size == sizeof(unsigned int)); + thread->m_local_mem->read(from_addr, size, &flags); + assert(flags == cudaStreamNonBlocking); + } + } + + //create stream and write back to param0 + CUstream_st * stream = thread->get_kernel().create_stream_cta(thread->get_ctaid()); + DEV_RUNTIME_REPORT("Create stream " << stream->get_uid() << ": " << stream); + thread->m_local_mem->write(pStream_addr, sizeof(cudaStream_t), &stream, NULL, NULL); + + //set retval0 + const operand_info &actual_return_op = pI->operand_lookup(0); //retval0 + const symbol *formal_return = target_func->get_return_var(); //cudaError_t + unsigned int return_size = formal_return->get_size_in_bytes(); + DEV_RUNTIME_REPORT("cudaStreamCreateWithFlags return value has size of " << return_size); + assert(actual_return_op.is_param_local()); + assert(actual_return_op.get_symbol()->get_size_in_bytes() == return_size + && return_size == sizeof(cudaError_t)); + cudaError_t error = cudaSuccess; + addr_t ret_param_addr = actual_return_op.get_symbol()->get_address(); + thread->m_local_mem->write(ret_param_addr, return_size, &error, NULL, NULL); + +} + + +void launch_one_device_kernel() { + if(!g_cuda_device_launch_op.empty()) { + device_launch_operation_t &op = g_cuda_device_launch_op.front(); + + stream_operation stream_op = stream_operation(op.grid, g_ptx_sim_mode, op.stream); + g_stream_manager->push(stream_op); + g_cuda_device_launch_op.pop_front(); + } +} + +void launch_all_device_kernels() { + while(!g_cuda_device_launch_op.empty()) { + launch_one_device_kernel(); + } +} +#endif diff --git a/src/cuda-sim/cuda_device_runtime.h b/src/cuda-sim/cuda_device_runtime.h new file mode 100644 index 0000000..6dbcd71 --- /dev/null +++ b/src/cuda-sim/cuda_device_runtime.h @@ -0,0 +1,11 @@ +//Jin: cuda_device_runtime.h +//Defines CUDA device runtime APIs for CDP support +#if (CUDART_VERSION >= 5000) +#pragma once + +void gpgpusim_cuda_getParameterBufferV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); +void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); +void gpgpusim_cuda_streamCreateWithFlags(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); +void launch_all_device_kernels(); +void launch_one_device_kernel(); +#endif diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index cf7f04a..011c285 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -41,14 +41,19 @@ #include "../gpgpu-sim/gpu-sim.h" #include "../gpgpu-sim/shader.h" +//Jin: include device runtime for CDP +#include "cuda_device_runtime.h" + #include <stdarg.h> unsigned ptx_instruction::g_num_ptx_inst_uid=0; const char *g_opcode_string[NUM_OPCODES] = { #define OP_DEF(OP,FUNC,STR,DST,CLASSIFICATION) STR, +#define OP_W_DEF(OP,FUNC,STR,DST,CLASSIFICATION) STR, #include "opcodes.def" #undef OP_DEF +#undef OP_W_DEF }; void inst_not_implemented( const ptx_instruction * pI ) ; @@ -148,6 +153,8 @@ ptx_reg_t ptx_thread_info::get_operand_value( const operand_info &op, operand_in result.u64 = op.get_symbol()->get_address(); } else if ( op.is_local() ) { result.u64 = op.get_symbol()->get_address(); + } else if ( op.is_function_address() ) { + result.u64 = (size_t)op.get_symbol()->get_pc(); } else { const char *name = op.name().c_str(); printf("GPGPU-Sim PTX: ERROR ** get_operand_value : unknown operand type for %s\n", name ); @@ -1336,7 +1343,82 @@ void bar_impl( const ptx_instruction *pIin, ptx_thread_info *thread ) thread->m_last_dram_callback.instruction = pIin; } -void bfe_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { inst_not_implemented(pI); } +void bfe_impl( const ptx_instruction *pI, ptx_thread_info *thread ) +{ + unsigned i_type = pI->get_type(); + unsigned msb = (i_type == U32_TYPE || i_type == S32_TYPE) ? 31 : 63; + const operand_info &dst = pI->dst(); + const operand_info &src1 = pI->src1(); + const operand_info &src2 = pI->src2(); + const operand_info &src3 = pI->src3(); + ptx_reg_t a = thread->get_operand_value(src1, dst, i_type, thread, 1); + ptx_reg_t b = thread->get_operand_value(src2, dst, i_type, thread, 1); + ptx_reg_t c = thread->get_operand_value(src3, dst, i_type, thread, 1); + unsigned pos = b.u32 & 0xFF; + unsigned len = c.u32 & 0xFF; + unsigned d = 0; + switch (i_type) + { + case U32_TYPE: + { + unsigned mask; + d = a.u32 >> pos; + mask = 0xFFFFFFFF >> (32 - len); + d &= mask; + break; + } + case U64_TYPE: + { + unsigned long mask; + d = a.u64 >> pos; + mask = 0xFFFFFFFFFFFFFFFF >> (64 - len); + d &= mask; + break; + } + case S32_TYPE: + { + unsigned mask; + unsigned min = MY_MIN_I(pos + len - 1, msb); + unsigned sbit = len == 0 ? 0 : (a.s32 >> min) & 0x1; + d = a.s32 >> pos; + if (sbit > 0) + { + mask = 0xFFFFFFFF << len; + d |= mask; + } + else + { + mask = 0xFFFFFFFF >> (32 - len); + d &= mask; + } + break; + } + case S64_TYPE: + { + unsigned long mask; + unsigned min = MY_MIN_I(pos + len - 1, msb); + unsigned sbit = len == 0 ? 0 : (a.s64 >> min) & 0x1; + d = a.s64 >> pos; + if (sbit > 0) + { + mask = 0xFFFFFFFFFFFFFFFF << len; + d |= mask; + } + else + { + mask = 0xFFFFFFFFFFFFFFFF >> (64 - len); + d &= mask; + } + break; + } + default: + printf("Operand type not supported for BFE instruction.\n"); + abort(); + return; + } + thread->set_operand_value(dst,d, i_type, thread, pI); +} + void bfi_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { inst_not_implemented(pI); } void bfind_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { inst_not_implemented(pI); } @@ -1407,7 +1489,23 @@ void call_impl( const ptx_instruction *pI, ptx_thread_info *thread ) if( fname == "vprintf" ) { gpgpusim_cuda_vprintf(pI, thread, target_func); return; - } + } + +#if (CUDART_VERSION >= 5000) + //Jin: handle device runtime apis for CDP + else if(fname == "cudaGetParameterBufferV2") { + gpgpusim_cuda_getParameterBufferV2(pI, thread, target_func); + return; + } + else if(fname == "cudaLaunchDeviceV2") { + gpgpusim_cuda_launchDeviceV2(pI, thread, target_func); + return; + } + else if(fname == "cudaStreamCreateWithFlags") { + gpgpusim_cuda_streamCreateWithFlags(pI, thread, target_func); + return; + } +#endif // read source arguements into register specified in declaration of function arg_buffer_list_t arg_values; @@ -2426,6 +2524,11 @@ void madp_impl( const ptx_instruction *pI, ptx_thread_info *thread ) mad_def(pI, thread, true); } +void madc_impl( const ptx_instruction *pI, ptx_thread_info *thread ) +{ + mad_def(pI, thread, true); +} + void mad_def( const ptx_instruction *pI, ptx_thread_info *thread, bool use_carry ) { const operand_info &dst = pI->dst(); @@ -3436,6 +3539,81 @@ void set_impl( const ptx_instruction *pI, ptx_thread_info *thread ) } +void shfl_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) +{ + unsigned i_type = pI->get_type(); + int tid = inst.warp_id() * core->get_warp_size(); + ptx_thread_info *thread = core->get_thread_info()[tid]; + ptx_warp_info *warp_info = thread->m_warp_info; + int lane = warp_info->get_done_threads(); + thread = core->get_thread_info()[tid + lane]; + + const operand_info &dst = pI->dst(); + const operand_info &src1 = pI->src1(); + const operand_info &src2 = pI->src2(); + const operand_info &src3 = pI->src3(); + int bval = (thread->get_operand_value(src2, dst, i_type, thread, 1)).u32; + int cval = (thread->get_operand_value(src3, dst, i_type, thread, 1)).u32; + int mask = cval >> 8; + bval &= 0x1F; + cval &= 0x1F; + + int maxLane = (lane & mask) | (cval & ~mask); + int minLane = lane & mask; + + int src_idx; + unsigned p; + switch(pI->shfl_op()) { + case UP_OPTION: + src_idx = lane - bval; + p = (src_idx >= maxLane); + break; + case DOWN_OPTION: + src_idx = lane + bval; + p = (src_idx <= maxLane); + break; + case BFLY_OPTION: + src_idx = lane ^ bval; + p = (src_idx <= maxLane); + break; + case IDX_OPTION: + src_idx = minLane | (bval & ~mask); + p = (src_idx <= maxLane); + break; + default: + printf("GPGPU-Sim PTX: ERROR: Invalid shfl option\n"); + assert(0); + break; + } + // copy from own lane + if (!p) src_idx = lane; + + // copy input from lane src_idx + ptx_reg_t data; + if (inst.active(src_idx)) { + ptx_thread_info *source = core->get_thread_info()[tid + src_idx]; + data = source->get_operand_value(src1, dst, i_type, source, 1); + } else { + printf("GPGPU-Sim PTX: WARNING: shfl input value unpredictable for inactive threads in a warp\n"); + data.u32 = 0; + } + thread->set_operand_value(dst, data, i_type, thread, pI); + + /* + TODO: deal with predicates appropriately using the following pseudocode: + if (!isGuardPredicateTrue(src_idx)) { + printf("GPGPU-Sim PTX: WARNING: shfl input value unpredictable for predicated-off threads in a warp\n"); + } + if (dest predicate selected) data.pred = p; + */ + + // keep track of the number of threads that have executed in the warp + warp_info->inc_done_threads(); + if (warp_info->get_done_threads() == inst.active_count()) { + warp_info->reset_done_threads(); + } +} + void shl_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { ptx_reg_t a, b, d; diff --git a/src/cuda-sim/opcodes.def b/src/cuda-sim/opcodes.def index 7aaa14f..e1b1422 100644 --- a/src/cuda-sim/opcodes.def +++ b/src/cuda-sim/opcodes.def @@ -69,6 +69,7 @@ OP_DEF(LDU_OP,ldu_impl,"ldu",1,5) OP_DEF(LG2_OP,lg2_impl,"lg2",1,4) OP_DEF(MAD24_OP,mad24_impl,"mad24",1,2) OP_DEF(MAD_OP,mad_impl,"mad",1,2) +OP_DEF(MADC_OP,madc_impl,"madc",1,2) OP_DEF(MADP_OP,madp_impl,"madp",1,2) OP_DEF(MAX_OP,max_impl,"max",1,1) OP_DEF(MEMBAR_OP,membar_impl,"membar",1,3) @@ -97,6 +98,7 @@ OP_DEF(SAD_OP,sad_impl,"sad",1,1) OP_DEF(SELP_OP,selp_impl,"selp",1,1) OP_DEF(SETP_OP,setp_impl,"setp",1,1) OP_DEF(SET_OP,set_impl,"set",1,1) +OP_W_DEF(SHFL_OP,shfl_impl,"shfl",1,10) OP_DEF(SHL_OP,shl_impl,"shl",1,1) OP_DEF(SHR_OP,shr_impl,"shr",1,1) OP_DEF(SIN_OP,sin_impl,"sin",1,4) diff --git a/src/cuda-sim/opcodes.h b/src/cuda-sim/opcodes.h index 871091c..aa133da 100644 --- a/src/cuda-sim/opcodes.h +++ b/src/cuda-sim/opcodes.h @@ -30,9 +30,11 @@ enum opcode_t { #define OP_DEF(OP,FUNC,STR,DST,CLASSIFICATION) OP, +#define OP_W_DEF(OP,FUNC,STR,DST,CLASSIFICATION) OP, #include "opcodes.def" - NUM_OPCODES + NUM_OPCODES #undef OP_DEF +#undef OP_W_DEF }; enum special_regs { diff --git a/src/cuda-sim/ptx.l b/src/cuda-sim/ptx.l index dfed936..5471d6f 100644 --- a/src/cuda-sim/ptx.l +++ b/src/cuda-sim/ptx.l @@ -86,7 +86,8 @@ ldu TC; ptx_lval.int_value = LDU_OP; return OPCODE; lg2 TC; ptx_lval.int_value = LG2_OP; return OPCODE; mad24 TC; ptx_lval.int_value = MAD24_OP; return OPCODE; mad TC; ptx_lval.int_value = MAD_OP; return OPCODE; -madp TC; ptx_lval.int_value = MADP_OP; return OPCODE; +madc TC; ptx_lval.int_value = MADC_OP; return OPCODE; +madp TC; ptx_lval.int_value = MADP_OP; return OPCODE; max TC; ptx_lval.int_value = MAX_OP; return OPCODE; membar TC; ptx_lval.int_value = MEMBAR_OP; return OPCODE; min TC; ptx_lval.int_value = MIN_OP; return OPCODE; @@ -114,6 +115,7 @@ sad TC; ptx_lval.int_value = SAD_OP; return OPCODE; selp TC; ptx_lval.int_value = SELP_OP; return OPCODE; setp TC; ptx_lval.int_value = SETP_OP; return OPCODE; set TC; ptx_lval.int_value = SET_OP; return OPCODE; +shfl TC; ptx_lval.int_value = SHFL_OP; return OPCODE; shl TC; ptx_lval.int_value = SHL_OP; return OPCODE; shr TC; ptx_lval.int_value = SHR_OP; return OPCODE; sin TC; ptx_lval.int_value = SIN_OP; return OPCODE; @@ -180,7 +182,9 @@ breakaddr TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE; \.union TC; return UNION_DIRECTIVE; /* not in PTX 2.1 */ \.version TC; return VERSION_DIRECTIVE; \.visible TC; return VISIBLE_DIRECTIVE; +\.weak TC; return WEAK_DIRECTIVE; \.address_size TC; return ADDRESS_SIZE_DIRECTIVE; +\.weak TC; return WEAK_DIRECTIVE; \.constptr TC; return CONSTPTR_DIRECTIVE; /* Ptx plus directive for pointer to constant memory */ \.ptr TC; return PTR_DIRECTIVE; /* Added for new OpenCL genrated code */ @@ -249,6 +253,7 @@ breakaddr TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE; \.v4 TC; return V4_TYPE; \.half TC; return HALF_OPTION; /* ptxplus */ +\.cc TC; return EXTP_OPTION; /* extended precision option */ \.equ TC; return EQU_OPTION; \.neu TC; return NEU_OPTION; @@ -324,6 +329,13 @@ breakaddr TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE; \.wb TC; return WB_OPTION; \.wt TC; return WT_OPTION; +\.nc TC; return NC_OPTION; + +\.up TC; return UP_OPTION; +\.down TC; return DOWN_OPTION; +\.bfly TC; return BFLY_OPTION; +\.idx TC; return IDX_OPTION; + \.popc TC; return ATOMIC_POPC; \.and TC; return ATOMIC_AND; \.or TC; return ATOMIC_OR; @@ -386,7 +398,9 @@ breakaddr TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE; } <IN_COMMENT>{ "*/" BEGIN(INITIAL); -[^*\n]+ // eat comment in chunks +"CPTX_BEGIN" printf("BEGINNING CUSTOM PTX.\n"); BEGIN(INITIAL); +[^C*\n]+ // eat comment in chunks +"C" "*" // eat the lone star \n TC; } diff --git a/src/cuda-sim/ptx.y b/src/cuda-sim/ptx.y index 79faddf..e00aa4b 100644 --- a/src/cuda-sim/ptx.y +++ b/src/cuda-sim/ptx.y @@ -47,6 +47,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. %token PTR_DIRECTIVE %token ENTRY_DIRECTIVE %token EXTERN_DIRECTIVE +%token WEAK_DIRECTIVE %token FILE_DIRECTIVE %token FUNC_DIRECTIVE %token GLOBAL_DIRECTIVE @@ -71,6 +72,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. %token VERSION_DIRECTIVE %token ADDRESS_SIZE_DIRECTIVE %token VISIBLE_DIRECTIVE +%token WEAK_DIRECTIVE %token <string_value> IDENTIFIER %token <int_value> INT_OPERAND %token <float_value> FLOAT_OPERAND @@ -103,6 +105,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. %token COMMA %token PRED %token HALF_OPTION +%token EXTP_OPTION %token EQ_OPTION %token NE_OPTION %token LT_OPTION @@ -191,6 +194,11 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. %token CV_OPTION; %token WB_OPTION; %token WT_OPTION; +%token NC_OPTION; +%token UP_OPTION; +%token DOWN_OPTION; +%token BFLY_OPTION; +%token IDX_OPTION; %type <int_value> function_decl_header %type <ptr_value> function_decl @@ -240,9 +248,12 @@ function_ident_param: IDENTIFIER { add_function_name($1); } LEFT_PAREN {func_hea function_decl_header: ENTRY_DIRECTIVE { $$ = 1; g_func_decl=1; func_header(".entry"); } | VISIBLE_DIRECTIVE ENTRY_DIRECTIVE { $$ = 1; g_func_decl=1; func_header(".entry"); } + | WEAK_DIRECTIVE ENTRY_DIRECTIVE { $$ = 1; g_func_decl=1; func_header(".entry"); } | FUNC_DIRECTIVE { $$ = 0; g_func_decl=1; func_header(".func"); } | VISIBLE_DIRECTIVE FUNC_DIRECTIVE { $$ = 0; g_func_decl=1; func_header(".func"); } + | WEAK_DIRECTIVE FUNC_DIRECTIVE { $$ = 0; g_func_decl=1; func_header(".func"); } | EXTERN_DIRECTIVE FUNC_DIRECTIVE { $$ = 2; g_func_decl=1; func_header(".func"); } + | WEAK_DIRECTIVE FUNC_DIRECTIVE { $$ = 0; g_func_decl=1; func_header(".func"); } ; param_list: /*empty*/ @@ -268,8 +279,8 @@ statement_list: directive_statement { add_directive(); } | instruction_statement { add_instruction(); } | statement_list directive_statement { add_directive(); } | statement_list instruction_statement { add_instruction(); } - | statement_list statement_block - | statement_block + | statement_list {start_inst_group();} statement_block {end_inst_group();} + | {start_inst_group();} statement_block {end_inst_group();} ; directive_statement: variable_declaration SEMI_COLON @@ -321,6 +332,7 @@ var_spec: space_spec | type_spec | align_spec | EXTERN_DIRECTIVE { add_extern_spec(); } + | WEAK_DIRECTIVE ; align_spec: ALIGN_DIRECTIVE INT_OPERAND { add_alignment_spec($2); } @@ -438,6 +450,7 @@ option: type_spec | atomic_operation_spec ; | TO_OPTION { add_option(TO_OPTION); } | HALF_OPTION { add_option(HALF_OPTION); } + | EXTP_OPTION { add_option(EXTP_OPTION); } | CA_OPTION { add_option(CA_OPTION); } | CG_OPTION { add_option(CG_OPTION); } | CS_OPTION { add_option(CS_OPTION); } @@ -445,6 +458,11 @@ option: type_spec | CV_OPTION { add_option(CV_OPTION); } | WB_OPTION { add_option(WB_OPTION); } | WT_OPTION { add_option(WT_OPTION); } + | NC_OPTION { add_option(NC_OPTION); } + | UP_OPTION { add_option(UP_OPTION); } + | DOWN_OPTION { add_option(DOWN_OPTION); } + | BFLY_OPTION { add_option(BFLY_OPTION); } + | IDX_OPTION { add_option(IDX_OPTION); } ; atomic_operation_spec: ATOMIC_AND { add_option(ATOMIC_AND); } diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index 751b3f4..1805ce9 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -90,6 +90,11 @@ symbol_table::symbol_table( const char *scope_name, unsigned entry_point, symbol m_const_next = 0; m_global_next = 0x100; m_local_next = 0; + m_tex_next = 0; + + //Jin: handle instruction group for cdp + m_inst_group_id = 0; + m_parent = parent; if ( m_parent ) { m_shared_next = m_parent->m_shared_next; @@ -170,6 +175,41 @@ void symbol_table::add_function( function_info *func, const char *filename, unsi m_symbols[ func->get_name() ] = s; } +//Jin: handle instruction group for cdp +symbol_table* symbol_table::start_inst_group() { + char inst_group_name[1024]; + snprintf(inst_group_name, 1024, "%s_inst_group_%u", m_scope_name.c_str(), m_inst_group_id); + + //previous added + assert(m_inst_group_symtab.find(std::string(inst_group_name)) == m_inst_group_symtab.end()); + symbol_table *sym_table = new symbol_table(inst_group_name, 3/*inst group*/, this ); + + sym_table->m_global_next = m_global_next; + sym_table->m_shared_next = m_shared_next; + sym_table->m_local_next = m_local_next; + sym_table->m_reg_allocator = m_reg_allocator; + sym_table->m_tex_next = m_tex_next; + sym_table->m_const_next = m_const_next; + + m_inst_group_symtab[std::string(inst_group_name)] = sym_table; + + return sym_table; +} + +symbol_table * symbol_table::end_inst_group() { + symbol_table * sym_table = m_parent; + + sym_table->m_global_next = m_global_next; + sym_table->m_shared_next = m_shared_next; + sym_table->m_local_next = m_local_next; + sym_table->m_reg_allocator = m_reg_allocator; + sym_table->m_tex_next = m_tex_next; + sym_table->m_const_next = m_const_next; + sym_table->m_inst_group_id++; + + return sym_table; +} + void register_ptx_function( const char *name, function_info *impl ); // either libcuda or libopencl bool symbol_table::add_function_decl( const char *name, int entry_point, function_info **func_info, symbol_table **sym_table ) @@ -1167,6 +1207,16 @@ ptx_instruction::ptx_instruction( int opcode, case HALF_OPTION: m_inst_size = 4; // bytes break; + case EXTP_OPTION: + break; + case NC_OPTION: + break; + case UP_OPTION: + case DOWN_OPTION: + case BFLY_OPTION: + case IDX_OPTION: + m_shfl_op = last_ptx_inst_option; + break; default: assert(0); break; @@ -1201,6 +1251,12 @@ ptx_instruction::ptx_instruction( int opcode, if (fname =="vprintf"){ m_is_printf = true; } + if(fname == "cudaStreamCreateWithFlags") + m_is_cdp = 1; + if(fname == "cudaGetParameterBufferV2") + m_is_cdp = 2; + if(fname == "cudaLaunchDeviceV2") + m_is_cdp = 4; } } @@ -1248,6 +1304,7 @@ function_info::function_info(int entry_point ) m_kernel_info.regs = 0; m_kernel_info.smem = 0; m_local_mem_framesize = 0; + m_args_aligned_size = -1; } unsigned function_info::print_insn( unsigned pc, FILE * fp ) const diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 601a13d..9ad1571 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -291,6 +291,7 @@ private: std::list<operand_info> m_initializer; static unsigned sm_next_uid; + }; class symbol_table { @@ -330,6 +331,11 @@ public: iterator const_iterator_end() { return m_consts.end();} void dump(); + + //Jin: handle instruction group for cdp + symbol_table* start_inst_group(); + symbol_table* end_inst_group(); + private: unsigned m_reg_allocator; unsigned m_shared_next; @@ -347,6 +353,10 @@ private: std::list<symbol*> m_consts; std::map<std::string,function_info*> m_function_info_lookup; std::map<std::string,symbol_table*> m_function_symtab_lookup; + + //Jin: handle instruction group for cdp + unsigned m_inst_group_id; + std::map<std::string,symbol_table*> m_inst_group_symtab; }; class operand_info { @@ -690,7 +700,7 @@ public: } bool is_immediate_address() const { - return m_immediate_address; + return m_immediate_address; } bool is_literal() const { return m_type == int_t || @@ -993,6 +1003,7 @@ public: unsigned saturation_mode() const { return m_saturation_mode;} unsigned dimension() const { return m_geom_spec;} unsigned barrier_op() const {return m_barrier_op;} + unsigned shfl_op() const {return m_shfl_op;} enum vote_mode_t { vote_any, vote_all, vote_uni, vote_ballot }; enum vote_mode_t vote_mode() const { return m_vote_mode; } @@ -1058,6 +1069,7 @@ private: unsigned m_compare_op; unsigned m_saturation_mode; unsigned m_barrier_op; + unsigned m_shfl_op; std::list<int> m_scalar_type; memory_space_t m_space_spec; @@ -1200,6 +1212,8 @@ public: { return m_args.size(); } + unsigned get_args_aligned_size(); + const symbol* get_arg( unsigned n ) const { assert( n < m_args.size() ); @@ -1288,6 +1302,9 @@ private: static std::vector<ptx_instruction*> s_g_pc_to_insn; // a direct mapping from PC to instruction static unsigned sm_next_uid; + + //parameter size for device kernels + int m_args_aligned_size; }; class arg_buffer_t { diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc index f7bf70e..6c1b595 100644 --- a/src/cuda-sim/ptx_loader.cc +++ b/src/cuda-sim/ptx_loader.cc @@ -46,6 +46,8 @@ bool g_override_embedded_ptx = false; extern int ptx_parse(); extern int ptx__scan_string(const char*); +extern std::map<unsigned,const char*> get_duplicate(); + const char *g_ptxinfo_filename; extern int ptxinfo_parse(); extern int ptxinfo_debug; @@ -182,6 +184,109 @@ symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source return symtab; } +void fix_duplicate_errors(char fname2[1024]) { + char tempfile[1024] = "_temp_ptx"; + char commandline[1024]; + + // change the name of the ptx file to _temp_ptx + snprintf(commandline,1024,"mv %s %s",fname2,tempfile); + printf("Running: %s\n", commandline); + int result = system(commandline); + if (result != 0) { + printf("GPGPU-Sim PTX: ERROR ** while changing filename from %s to %s", fname2, tempfile); + exit(1); + } + + // store all of the ptx into a char array + FILE *ptxsource = fopen(tempfile,"r"); + fseek(ptxsource, 0, SEEK_END); + long filesize = ftell(ptxsource); + rewind(ptxsource); + char *ptxdata = (char*)malloc((filesize+1)*sizeof(char)); + fread(ptxdata, filesize, 1, ptxsource); + fclose(ptxsource); + + FILE *ptxdest = fopen(fname2,"w"); + std::map<unsigned,const char*> duplicate = get_duplicate(); + unsigned offset; + unsigned oldlinenum = 1; + unsigned linenum; + char *startptr = ptxdata; + char *funcptr; + char *tempptr = ptxdata - 1; + char *lineptr = ptxdata - 1; + + // recreate the ptx file without duplications + for ( std::map<unsigned,const char*>::iterator iter = duplicate.begin(); + iter != duplicate.end(); + iter++){ + // find the line of the next error + linenum = iter->first; + for (int i = oldlinenum; i < linenum; i++) { + lineptr = strchr(lineptr + 1, '\n'); + } + + // find the end of the current section to be copied over + // then find the start of the next section that will be copied + if (strcmp("function", iter->second) == 0) { + // get location of most recent .func + while (tempptr < lineptr && tempptr != NULL) { + funcptr = tempptr; + tempptr = strstr(funcptr + 1, ".func"); + } + + // get the start of the previous line + offset = 0; + while (*(funcptr - offset) != '\n') offset++; + + fwrite(startptr, sizeof(char), funcptr - offset + 1 - startptr, ptxdest); + + //find next location of startptr + if (*(lineptr + 3) == ';') { + // for function definitions + startptr = lineptr + 5; + } else if (*(lineptr + 3) == '{') { + // for functions enclosed with curly brackets + offset = 5; + unsigned bracket = 1; + while (bracket != 0) { + if (*(lineptr + offset) == '{') bracket++; + else if (*(lineptr + offset) == '}') bracket--; + offset++; + } + startptr = lineptr + offset + 1; + } else { + printf("GPGPU-Sim PTX: ERROR ** Unrecognized function format\n"); + abort(); + } + } else if (strcmp("variable", iter->second) == 0) { + fwrite(startptr, sizeof(char), (int)(lineptr + 1 - startptr), ptxdest); + + //find next location of startptr + offset = 1; + while (*(lineptr + offset) != '\n') offset++; + startptr = lineptr + offset + 1; + } else { + printf("GPGPU-Sim PTX: ERROR ** Unsupported duplicate type: %s\n", iter->second); + } + + oldlinenum = linenum; + } + // copy over the rest of the file + fwrite(startptr, sizeof(char), ptxdata + filesize - startptr, ptxdest); + + // cleanup + free(ptxdata); + fclose(ptxdest); + snprintf(commandline,1024,"rm -f %s",tempfile); + printf("Running: %s\n", commandline); + result = system(commandline); + if (result != 0) { + printf("GPGPU-Sim PTX: ERROR ** while deleting %s", tempfile); + exit(1); + } +} + void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version ) { char fname[1024]; @@ -217,7 +322,11 @@ void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num #if CUDART_VERSION >= 3000 if (sm_version == 0) sm_version = 20; - snprintf(extra_flags,1024,"--gpu-name=sm_%u",sm_version); + extern bool g_cdp_enabled; + if(!g_cdp_enabled) + snprintf(extra_flags,1024,"--gpu-name=sm_%u",sm_version); + else + snprintf(extra_flags,1024,"--compile-only --gpu-name=sm_%u",sm_version); #endif snprintf(commandline,1024,"$CUDA_INSTALL_PATH/bin/ptxas %s -v %s --output-file /dev/null 2> %s", @@ -225,14 +334,29 @@ void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num printf("GPGPU-Sim PTX: generating ptxinfo using \"%s\"\n", commandline); result = system(commandline); if( result != 0 ) { - printf("GPGPU-Sim PTX: ERROR ** while loading PTX (b) %d\n", result); - printf(" Ensure ptxas is in your path.\n"); - exit(1); + // 65280 = duplicate errors + if (result == 65280) { + ptxinfo_in = fopen(tempfile_ptxinfo,"r"); + g_ptxinfo_filename = tempfile_ptxinfo; + ptxinfo_parse(); + + fix_duplicate_errors(fname2); + snprintf(commandline,1024,"$CUDA_INSTALL_PATH/bin/ptxas %s -v %s --output-file /dev/null 2> %s", + extra_flags, fname2, tempfile_ptxinfo); + printf("GPGPU-Sim PTX: regenerating ptxinfo using \"%s\"\n", commandline); + result = system(commandline); + } + if (result != 0) { + printf("GPGPU-Sim PTX: ERROR ** while loading PTX (b) %d\n", result); + printf(" Ensure ptxas is in your path.\n"); + exit(1); + } } ptxinfo_in = fopen(tempfile_ptxinfo,"r"); g_ptxinfo_filename = tempfile_ptxinfo; ptxinfo_parse(); + if( ! g_save_embedded_ptx ) { snprintf(commandline,1024,"rm -f %s %s %s", fname, fname2, tempfile_ptxinfo); printf("GPGPU-Sim PTX: removing ptxinfo using \"%s\"\n", commandline); diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc index 824714a..baa3bcd 100644 --- a/src/cuda-sim/ptx_parser.cc +++ b/src/cuda-sim/ptx_parser.cc @@ -120,6 +120,20 @@ symbol_table *init_parser( const char *ptx_filename ) #define DEF(X,Y) g_ptx_token_decode[X] = Y; #include "ptx_parser_decode.def" #undef DEF + g_ptx_token_decode[undefined_space] = "undefined_space"; + g_ptx_token_decode[undefined_space] = "undefined_space=0"; + g_ptx_token_decode[reg_space] = "reg_space"; + g_ptx_token_decode[local_space] = "local_space"; + g_ptx_token_decode[shared_space] = "shared_space"; + g_ptx_token_decode[param_space_unclassified] = "param_space_unclassified"; + g_ptx_token_decode[param_space_kernel] = "param_space_kernel"; + g_ptx_token_decode[param_space_local] = "param_space_local"; + g_ptx_token_decode[const_space] = "const_space"; + g_ptx_token_decode[tex_space] = "tex_space"; + g_ptx_token_decode[surf_space] = "surf_space"; + g_ptx_token_decode[global_space] = "global_space"; + g_ptx_token_decode[generic_space] = "generic_space"; + g_ptx_token_decode[instruction_space] = "instruction_space"; return g_global_symbol_table; } @@ -187,6 +201,17 @@ void add_function_name( const char *name ) g_global_symbol_table->add_function( g_func_info, g_filename, ptx_lineno ); } +//Jin: handle instruction group for cdp +void start_inst_group() { + PTX_PARSE_DPRINTF("start_instruction_group"); + g_current_symbol_table = g_current_symbol_table->start_inst_group(); +} + +void end_inst_group() { + PTX_PARSE_DPRINTF("end_instruction_group"); + g_current_symbol_table = g_current_symbol_table->end_inst_group(); +} + void add_directive() { PTX_PARSE_DPRINTF("add_directive"); diff --git a/src/cuda-sim/ptx_parser.h b/src/cuda-sim/ptx_parser.h index fef7635..32f3903 100644 --- a/src/cuda-sim/ptx_parser.h +++ b/src/cuda-sim/ptx_parser.h @@ -94,6 +94,10 @@ void change_operand_neg( ); void set_immediate_operand_type( ); void version_header(double a); +//Jin: handle instructino group for cdp +void start_inst_group(); +void end_inst_group(); + #define NON_ARRAY_IDENTIFIER 1 #define ARRAY_IDENTIFIER_NO_DIM 2 #define ARRAY_IDENTIFIER 3 diff --git a/src/cuda-sim/ptx_sim.cc b/src/cuda-sim/ptx_sim.cc index 9b32e68..a3e43aa 100644 --- a/src/cuda-sim/ptx_sim.cc +++ b/src/cuda-sim/ptx_sim.cc @@ -128,6 +128,26 @@ unsigned ptx_cta_info::get_sm_idx() const return m_sm_idx; } +ptx_warp_info::ptx_warp_info() +{ + reset_done_threads(); +} + +unsigned ptx_warp_info::get_done_threads() const +{ + return m_done_threads; +} + +void ptx_warp_info::inc_done_threads() +{ + m_done_threads++; +} + +void ptx_warp_info::reset_done_threads() +{ + m_done_threads = 0; +} + unsigned g_ptx_thread_info_uid_next=1; unsigned g_ptx_thread_info_delete_count=0; @@ -153,6 +173,7 @@ ptx_thread_info::ptx_thread_info( kernel_info_t &kernel ) m_last_memory_space = undefined_space; m_branch_taken = 0; m_shared_mem = NULL; + m_warp_info = NULL; m_cta_info = NULL; m_local_mem = NULL; m_symbol_table = NULL; @@ -240,7 +261,7 @@ unsigned ptx_thread_info::get_builtin( int builtin_id, unsigned dim_mod ) } case GRIDID_REG: return m_gridid; - case LANEID_REG: feature_not_implemented( "%laneid" ); return 0; + case LANEID_REG: return get_hw_tid() % m_core->get_warp_size(); case LANEMASK_EQ_REG: feature_not_implemented( "%lanemask_eq" ); return 0; case LANEMASK_LE_REG: feature_not_implemented( "%lanemask_le" ); return 0; case LANEMASK_LT_REG: feature_not_implemented( "%lanemask_lt" ); return 0; diff --git a/src/cuda-sim/ptx_sim.h b/src/cuda-sim/ptx_sim.h index f926e6d..e6eb02e 100644 --- a/src/cuda-sim/ptx_sim.h +++ b/src/cuda-sim/ptx_sim.h @@ -167,6 +167,17 @@ private: std::set<ptx_thread_info*> m_dangling_pointers; }; +class ptx_warp_info { +public: + ptx_warp_info(); + unsigned get_done_threads() const; + void inc_done_threads(); + void reset_done_threads(); + +private: + unsigned m_done_threads; +}; + class symbol; struct stack_entry { @@ -418,6 +429,9 @@ public: void or_reduction(unsigned ctaid, unsigned barid, bool value) {m_core->or_reduction(ctaid,barid,value);} void popc_reduction(unsigned ctaid, unsigned barid, bool value) {m_core->popc_reduction(ctaid,barid,value);} + //Jin: get corresponding kernel grid for CDP purpose + kernel_info_t & get_kernel() { return m_kernel; } + public: addr_t m_last_effective_address; bool m_branch_taken; @@ -425,6 +439,7 @@ public: dram_callback_t m_last_dram_callback; memory_space *m_shared_mem; memory_space *m_local_mem; + ptx_warp_info *m_warp_info; ptx_cta_info *m_cta_info; ptx_reg_t m_last_set_operand_value; diff --git a/src/cuda-sim/ptxinfo.l b/src/cuda-sim/ptxinfo.l index 99ee1fc..33c2748 100644 --- a/src/cuda-sim/ptxinfo.l +++ b/src/cuda-sim/ptxinfo.l @@ -59,6 +59,11 @@ unsigned ptxinfo_col = 0; "gmem" TC; return GMEM; "line" TC; return LINE; "for" TC; return FOR; +"textures" TC; return TEXTURES; +"error : Duplicate definition of" TC; return DUPLICATE; +"function" TC; ptxinfo_lval.string_value = strdup(yytext); return FUNCTION; +"variable" TC; ptxinfo_lval.string_value = strdup(yytext); return VARIABLE; +"fatal : Ptx assembly aborted due to errors" TC; return FATAL; [_A-Za-z$%][_0-9A-Za-z$]* TC; ptxinfo_lval.string_value = strdup(yytext); return IDENTIFIER; [-]{0,1}[0-9]+ TC; ptxinfo_lval.int_value = atoi(yytext); return INT_OPERAND; diff --git a/src/cuda-sim/ptxinfo.y b/src/cuda-sim/ptxinfo.y index 294412d..d241d8c 100644 --- a/src/cuda-sim/ptxinfo.y +++ b/src/cuda-sim/ptxinfo.y @@ -54,11 +54,16 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. %token LINE %token <string_value> WARNING %token FOR +%token TEXTURES +%token DUPLICATE +%token <string_value> FUNCTION +%token <string_value> VARIABLE +%token FATAL %{ #include <stdlib.h> #include <string.h> - + static unsigned g_declared; static unsigned g_system; int ptxinfo_lex(void); @@ -70,6 +75,8 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. void ptxinfo_smem( unsigned declared, unsigned system ); void ptxinfo_cmem( unsigned nbytes, unsigned bank ); int ptxinfo_error(const char*); + void ptxinfo_linenum( unsigned ); + void ptxinfo_dup_type( const char* ); %} %% @@ -81,20 +88,26 @@ input: /* empty */ line: HEADER INFO COLON line_info | HEADER IDENTIFIER COMMA LINE INT_OPERAND SEMICOLON WARNING | HEADER WARNING { printf("GPGPU-Sim: ptxas %s\n", $2); } + | HEADER IDENTIFIER COMMA LINE INT_OPERAND SEMICOLON DUPLICATE duplicate { ptxinfo_linenum($5); } + | HEADER FATAL ; line_info: function_name | function_info { ptxinfo_addinfo(); } + | gmem_info ; function_name: FUNC QUOTE IDENTIFIER QUOTE { ptxinfo_function($3); } - | FUNC QUOTE IDENTIFIER QUOTE FOR QUOTE IDENTIFIER QUOTE {ptxinfo_function($3); } + | FUNC QUOTE IDENTIFIER QUOTE FOR QUOTE IDENTIFIER QUOTE { ptxinfo_function($3); } ; function_info: info | function_info COMMA info ; +gmem_info: INT_OPERAND BYTES GMEM + ; + info: USED INT_OPERAND REGS { ptxinfo_regs($2); } | tuple LMEM { ptxinfo_lmem(g_declared,g_system); } | tuple SMEM { ptxinfo_smem(g_declared,g_system); } @@ -104,10 +117,15 @@ info: USED INT_OPERAND REGS { ptxinfo_regs($2); } | INT_OPERAND BYTES SMEM { ptxinfo_smem($1,0); } | INT_OPERAND BYTES CMEM { ptxinfo_cmem($1,0); } | INT_OPERAND REGS { ptxinfo_regs($1); } + | INT_OPERAND TEXTURES {} ; tuple: INT_OPERAND PLUS INT_OPERAND BYTES { g_declared=$1; g_system=$3; } +duplicate: FUNCTION QUOTE IDENTIFIER QUOTE { ptxinfo_dup_type($1); } + | VARIABLE QUOTE IDENTIFIER QUOTE { ptxinfo_dup_type($1); } + ; + %% diff --git a/src/gpgpu-sim/Makefile b/src/gpgpu-sim/Makefile index bead38a..f10a8a4 100644 --- a/src/gpgpu-sim/Makefile +++ b/src/gpgpu-sim/Makefile @@ -59,6 +59,7 @@ ifneq ($(GPGPUSIM_POWER_MODEL),) endif OPTFLAGS += -g3 -fPIC +OPTFLAGS += -DCUDART_VERSION=$(CUDART_VERSION) CPP = g++ $(SNOW) OEXT = o diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index eafb909..58a5d16 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -61,6 +61,7 @@ #include "power_stat.h" #include "visualizer.h" #include "stats.h" +#include "../cuda-sim/cuda_device_runtime.h" #ifdef GPGPUSIM_POWER_MODEL #include "power_interface.h" @@ -366,6 +367,10 @@ void shader_core_config::reg_options(class OptionParser * opp) "For complete list of prioritization values see shader.h enum scheduler_prioritization_type" "Default: gto", "gto"); + + option_parser_register(opp, "-gpgpu_concurrent_kernel_sm", OPT_BOOL, &gpgpu_concurrent_kernel_sm, + "Support concurrent kernels on a SM (default = disabled)", + "0"); } void gpgpu_sim_config::reg_options(option_parser_t opp) @@ -438,6 +443,16 @@ void gpgpu_sim_config::reg_options(option_parser_t opp) &Trace::sampling_memory_partition, "The memory partition which is printed using MEMPART_DPRINTF. Default -1 (i.e. all)", "-1"); ptx_file_line_stats_options(opp); + + //Jin: kernel launch latency + extern unsigned g_kernel_launch_latency; + option_parser_register(opp, "-gpgpu_kernel_launch_latency", OPT_INT32, + &g_kernel_launch_latency, "Kernel launch latency in cycles. Default: 0", + "0"); + extern bool g_cdp_enabled; + option_parser_register(opp, "-gpgpu_cdp_enabled", OPT_BOOL, + &g_cdp_enabled, "Turn on CDP", + "0"); } ///////////////////////////////////////////////////////////////////////////// @@ -518,16 +533,27 @@ bool gpgpu_sim::get_more_cta_left() const kernel_info_t *gpgpu_sim::select_kernel() { + if(m_running_kernels[m_last_issued_kernel] && + !m_running_kernels[m_last_issued_kernel]->no_more_ctas_to_run()) { + unsigned launch_uid = m_running_kernels[m_last_issued_kernel]->get_uid(); + if(std::find(m_executed_kernel_uids.begin(), m_executed_kernel_uids.end(), launch_uid) == m_executed_kernel_uids.end()) { + m_running_kernels[m_last_issued_kernel]->start_cycle = gpu_sim_cycle + gpu_tot_sim_cycle; + m_executed_kernel_uids.push_back(launch_uid); + m_executed_kernel_names.push_back(m_running_kernels[m_last_issued_kernel]->name()); + } + return m_running_kernels[m_last_issued_kernel]; + } + for(unsigned n=0; n < m_running_kernels.size(); n++ ) { unsigned idx = (n+m_last_issued_kernel+1)%m_config.max_concurrent_kernel; if( kernel_more_cta_left(m_running_kernels[idx]) ){ m_last_issued_kernel=idx; + m_running_kernels[idx]->start_cycle = gpu_sim_cycle + gpu_tot_sim_cycle; // record this kernel for stat print if it is the first time this kernel is selected for execution unsigned launch_uid = m_running_kernels[idx]->get_uid(); - if (std::find(m_executed_kernel_uids.begin(), m_executed_kernel_uids.end(), launch_uid) == m_executed_kernel_uids.end()) { - m_executed_kernel_uids.push_back(launch_uid); - m_executed_kernel_names.push_back(m_running_kernels[idx]->name()); - } + assert(std::find(m_executed_kernel_uids.begin(), m_executed_kernel_uids.end(), launch_uid) == m_executed_kernel_uids.end()); + m_executed_kernel_uids.push_back(launch_uid); + m_executed_kernel_names.push_back(m_running_kernels[idx]->name()); return m_running_kernels[idx]; } @@ -551,6 +577,7 @@ void gpgpu_sim::set_kernel_done( kernel_info_t *kernel ) std::vector<kernel_info_t*>::iterator k; for( k=m_running_kernels.begin(); k!=m_running_kernels.end(); k++ ) { if( *k == kernel ) { + kernel->end_cycle = gpu_sim_cycle + gpu_tot_sim_cycle; *k = NULL; break; } @@ -622,6 +649,10 @@ gpgpu_sim::gpgpu_sim( const gpgpu_sim_config &config ) *active_sms=0; last_liveness_message_time = 0; + + //Jin: functional simulation for CDP + m_functional_sim = false; + m_functional_sim_kernel = NULL; } int gpgpu_sim::shared_mem_size() const @@ -927,7 +958,8 @@ void gpgpu_sim::gpu_print_stat() printf("gpu_tot_ipc = %12.4f\n", (float)(gpu_tot_sim_insn+gpu_sim_insn) / (gpu_tot_sim_cycle+gpu_sim_cycle)); printf("gpu_tot_issued_cta = %lld\n", gpu_tot_issued_cta + m_total_cta_launched); - + extern unsigned long long g_max_total_param_size; + fprintf(statfout, "max_total_param_size = %llu\n", g_max_total_param_size); // performance counter for stalls due to congestion. printf("gpu_stall_dramfull = %d\n", gpu_stall_dramfull); @@ -1070,7 +1102,119 @@ void shader_core_ctx::mem_instruction_stats(const warp_inst_t &inst) abort(); } } +bool shader_core_ctx::can_issue_1block(kernel_info_t & kernel) { + + //Jin: concurrent kernels on one SM + if(m_config->gpgpu_concurrent_kernel_sm) { + if(m_config->max_cta(kernel) < 1) + return false; + + return occupy_shader_resource_1block(kernel, false); + } + else { + return (get_n_active_cta() < m_config->max_cta(kernel)); + } +} + +int shader_core_ctx::find_available_hwtid(unsigned int cta_size, bool occupy) { + + unsigned int step; + for(step = 0; step < m_config->n_thread_per_shader; + step += cta_size) { + + unsigned int hw_tid; + for(hw_tid = step; hw_tid < step + cta_size; + hw_tid++) { + if(m_occupied_hwtid.test(hw_tid)) + break; + } + if(hw_tid == step + cta_size) //consecutive non-active + break; + } + if(step >= m_config->n_thread_per_shader) //didn't find + return -1; + else { + if(occupy) { + for(unsigned hw_tid = step; hw_tid < step + cta_size; + hw_tid++) + m_occupied_hwtid.set(hw_tid); + } + return step; + } +} + +bool shader_core_ctx::occupy_shader_resource_1block(kernel_info_t & k, bool occupy) { + unsigned threads_per_cta = k.threads_per_cta(); + const class function_info *kernel = k.entry(); + unsigned int padded_cta_size = threads_per_cta; + unsigned int warp_size = m_config->warp_size; + if (padded_cta_size%warp_size) + padded_cta_size = ((padded_cta_size/warp_size)+1)*(warp_size); + + if(m_occupied_n_threads + padded_cta_size > m_config->n_thread_per_shader) + return false; + if(find_available_hwtid(padded_cta_size, false) == -1) + return false; + + const struct gpgpu_ptx_sim_info *kernel_info = ptx_sim_kernel_info(kernel); + + if(m_occupied_shmem + kernel_info->smem > m_config->gpgpu_shmem_size) + return false; + + unsigned int used_regs = padded_cta_size * ((kernel_info->regs+3)&~3); + if(m_occupied_regs + used_regs > m_config->gpgpu_shader_registers) + return false; + + if(m_occupied_ctas +1 > m_config->max_cta_per_core) + return false; + + if(occupy) { + m_occupied_n_threads += padded_cta_size; + m_occupied_shmem += kernel_info->smem; + m_occupied_regs += (padded_cta_size * ((kernel_info->regs+3)&~3)); + m_occupied_ctas++; + + printf("GPGPU-Sim uArch: Shader %d occupied %d threads, %d shared mem, %d registers, %d ctas\n", + m_sid, m_occupied_n_threads, m_occupied_shmem, m_occupied_regs, m_occupied_ctas); + } + + return true; +} + +void shader_core_ctx::release_shader_resource_1block(unsigned hw_ctaid, kernel_info_t & k) { + + if(m_config->gpgpu_concurrent_kernel_sm) { + unsigned threads_per_cta = k.threads_per_cta(); + const class function_info *kernel = k.entry(); + unsigned int padded_cta_size = threads_per_cta; + unsigned int warp_size = m_config->warp_size; + if (padded_cta_size%warp_size) + padded_cta_size = ((padded_cta_size/warp_size)+1)*(warp_size); + + assert(m_occupied_n_threads >= padded_cta_size); + m_occupied_n_threads -= padded_cta_size; + + int start_thread = m_occupied_cta_to_hwtid[hw_ctaid]; + + for(unsigned hwtid = start_thread; hwtid < start_thread + padded_cta_size; + hwtid++) + m_occupied_hwtid.reset(hwtid); + m_occupied_cta_to_hwtid.erase(hw_ctaid); + + const struct gpgpu_ptx_sim_info *kernel_info = ptx_sim_kernel_info(kernel); + + assert(m_occupied_shmem >= (unsigned int)kernel_info->smem); + m_occupied_shmem -= kernel_info->smem; + + unsigned int used_regs = padded_cta_size * ((kernel_info->regs+3)&~3); + assert(m_occupied_regs >= used_regs); + m_occupied_regs -= used_regs; + + assert(m_occupied_ctas >= 1); + m_occupied_ctas--; + } +} //////////////////////////////////////////////////////////////////////////////////////////////// @@ -1083,11 +1227,23 @@ void shader_core_ctx::mem_instruction_stats(const warp_inst_t &inst) void shader_core_ctx::issue_block2core( kernel_info_t &kernel ) { - set_max_cta(kernel); + + if(!m_config->gpgpu_concurrent_kernel_sm) + set_max_cta(kernel); + else + assert(occupy_shader_resource_1block(kernel, true)); + + kernel.inc_running(); // find a free CTA context unsigned free_cta_hw_id=(unsigned)-1; - for (unsigned i=0;i<kernel_max_cta_per_shader;i++ ) { + + unsigned max_cta_per_core; + if(!m_config->gpgpu_concurrent_kernel_sm) + max_cta_per_core = kernel_max_cta_per_shader; + else + max_cta_per_core = m_config->max_cta_per_core; + for (unsigned i=0;i<max_cta_per_core;i++ ) { if( m_cta_status[i]==0 ) { free_cta_hw_id=i; break; @@ -1104,8 +1260,20 @@ void shader_core_ctx::issue_block2core( kernel_info_t &kernel ) int padded_cta_size = cta_size; if (cta_size%m_config->warp_size) padded_cta_size = ((cta_size/m_config->warp_size)+1)*(m_config->warp_size); - unsigned start_thread = free_cta_hw_id * padded_cta_size; - unsigned end_thread = start_thread + cta_size; + + unsigned int start_thread, end_thread; + + if(!m_config->gpgpu_concurrent_kernel_sm) { + start_thread = free_cta_hw_id * padded_cta_size; + end_thread = start_thread + cta_size; + } + else { + start_thread = find_available_hwtid(padded_cta_size, true); + assert((int)start_thread != -1); + end_thread = start_thread + cta_size; + assert(m_occupied_cta_to_hwtid.find(free_cta_hw_id) == m_occupied_cta_to_hwtid.end()); + m_occupied_cta_to_hwtid[free_cta_hw_id]= start_thread; + } // reset the microarchitecture state of the selected hardware thread and warp contexts reinit(start_thread, end_thread,false); @@ -1133,7 +1301,9 @@ void shader_core_ctx::issue_block2core( kernel_info_t &kernel ) m_n_active_cta++; shader_CTA_count_log(m_sid, 1); - printf("GPGPU-Sim uArch: core:%3d, cta:%2u initialized @(%lld,%lld)\n", m_sid, free_cta_hw_id, gpu_sim_cycle, gpu_tot_sim_cycle ); + printf("GPGPU-Sim uArch: core:%3d, cta:%2u, start_tid:%4u, end_tid:%4u, initialized @(%lld,%lld)\n", + m_sid, free_cta_hw_id, start_thread, end_thread, gpu_sim_cycle, gpu_tot_sim_cycle ); + } /////////////////////////////////////////////////////////////////////////////////////////// @@ -1366,6 +1536,11 @@ void gpgpu_sim::cycle() } try_snap_shot(gpu_sim_cycle); spill_log_to_file (stdout, 0, gpu_sim_cycle); + +#if (CUDART_VERSION >= 5000) + //launch device kernel + launch_one_device_kernel(); +#endif } } diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 33fffd3..a2d1b9b 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -492,6 +492,7 @@ private: std::string executed_kernel_info_string(); //< format the kernel information into a string for stat printout void clear_executed_kernel_info(); //< clear the kernel information after stat printout + public: unsigned long long gpu_sim_insn; unsigned long long gpu_tot_sim_insn; @@ -504,6 +505,25 @@ public: void change_cache_config(FuncCache cache_config); void set_cache_config(std::string kernel_name); + //Jin: functional simulation for CDP +private: + //set by stream operation every time a functoinal simulation is done + bool m_functional_sim; + kernel_info_t * m_functional_sim_kernel; + +public: + bool is_functional_sim() { return m_functional_sim; } + kernel_info_t * get_functional_kernel() { return m_functional_sim_kernel; } + void functional_launch(kernel_info_t * k) { + m_functional_sim = true; + m_functional_sim_kernel = k; + } + void finish_functional_sim(kernel_info_t * k) { + assert(m_functional_sim); + assert(m_functional_sim_kernel == k); + m_functional_sim = false; + m_functional_sim_kernel = NULL; + } }; #endif diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index ff2fac7..d17e51d 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -296,6 +296,14 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_last_inst_gpu_sim_cycle = 0; m_last_inst_gpu_tot_sim_cycle = 0; + + //Jin: for concurrent kernels on a SM + m_occupied_n_threads = 0; + m_occupied_shmem = 0; + m_occupied_regs = 0; + m_occupied_ctas = 0; + m_occupied_hwtid.reset(); + m_occupied_cta_to_hwtid.clear(); } void shader_core_ctx::reinit(unsigned start_thread, unsigned end_thread, bool reset_not_completed ) @@ -303,6 +311,15 @@ void shader_core_ctx::reinit(unsigned start_thread, unsigned end_thread, bool re if( reset_not_completed ) { m_not_completed = 0; m_active_threads.reset(); + + //Jin: for concurrent kernels on a SM + m_occupied_n_threads = 0; + m_occupied_shmem = 0; + m_occupied_regs = 0; + m_occupied_ctas = 0; + m_occupied_hwtid.reset(); + m_occupied_cta_to_hwtid.clear(); + } for (unsigned i = start_thread; i<end_thread; i++) { m_threadState[i].n_insn = 0; @@ -620,7 +637,7 @@ void shader_core_ctx::fetch() if( m_threadState[tid].m_active == true ) { m_threadState[tid].m_active = false; unsigned cta_id = m_warp[warp_id].get_cta_id(); - register_cta_thread_exit(cta_id); + register_cta_thread_exit(cta_id, &(m_thread[tid]->get_kernel())); m_not_completed -= 1; m_active_threads.reset(tid); assert( m_thread[tid]!= NULL ); @@ -828,6 +845,13 @@ void scheduler_unit::cycle() unsigned max_issue = m_shader->m_config->gpgpu_max_insn_issue_per_warp; 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(); + //Jin: handle cdp latency; + if(pI && pI->m_is_cdp && warp(warp_id).m_cdp_latency > 0) { + assert(warp(warp_id).m_cdp_dummy); + warp(warp_id).m_cdp_latency--; + break; + } + bool valid = warp(warp_id).ibuffer_next_valid(); bool warp_inst_issued = false; unsigned pc,rpc; @@ -862,6 +886,25 @@ void scheduler_unit::cycle() bool sp_pipe_avail = m_sp_out->has_free(); bool sfu_pipe_avail = m_sfu_out->has_free(); if( sp_pipe_avail && (pI->op != SFU_OP) ) { + + //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; + } + // 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); issued++; @@ -1917,7 +1960,7 @@ void ldst_unit::cycle() } } -void shader_core_ctx::register_cta_thread_exit( unsigned cta_num ) +void shader_core_ctx::register_cta_thread_exit( unsigned cta_num, kernel_info_t * kernel) { assert( m_cta_status[cta_num] > 0 ); m_cta_status[cta_num]--; @@ -1925,22 +1968,35 @@ void shader_core_ctx::register_cta_thread_exit( unsigned cta_num ) m_n_active_cta--; m_barriers.deallocate_barrier(cta_num); shader_CTA_count_unlog(m_sid, 1); + printf("GPGPU-Sim uArch: Shader %d finished CTA #%d (%lld,%lld), %u CTAs running\n", m_sid, cta_num, gpu_sim_cycle, gpu_tot_sim_cycle, m_n_active_cta ); + if( m_n_active_cta == 0 ) { - assert( m_kernel != NULL ); - m_kernel->dec_running(); - printf("GPGPU-Sim uArch: Shader %u empty (release kernel %u \'%s\').\n", m_sid, m_kernel->get_uid(), - m_kernel->name().c_str() ); - if( !m_gpu->kernel_more_cta_left(m_kernel) ) { - if( !m_kernel->running() ) { - printf("GPGPU-Sim uArch: GPU detected kernel \'%s\' finished on shader %u.\n", m_kernel->name().c_str(), m_sid ); - m_gpu->set_kernel_done( m_kernel ); - } - } - m_kernel=NULL; + printf("GPGPU-Sim uArch: Shader %u empty (last released kernel %u \'%s\').\n", m_sid, kernel->get_uid(), + kernel->name().c_str() ); fflush(stdout); + + //Shader can only be empty when no more cta are dispatched + if(kernel != m_kernel) { + assert(m_kernel == NULL || !m_gpu->kernel_more_cta_left(m_kernel)); + } + m_kernel = NULL; + } + + //Jin: for concurrent kernels on sm + release_shader_resource_1block(cta_num, *kernel); + kernel->dec_running(); + if( !m_gpu->kernel_more_cta_left(kernel) ) { + if( !kernel->running() ) { + printf("GPGPU-Sim uArch: GPU detected kernel %u \'%s\' finished on shader %u.\n", kernel->get_uid(), + kernel->name().c_str(), m_sid ); + if(m_kernel == kernel) + m_kernel = NULL; + m_gpu->set_kernel_done( kernel ); + } } + } } @@ -3238,15 +3294,33 @@ unsigned simt_core_cluster::issue_block2core() unsigned num_blocks_issued=0; for( unsigned i=0; i < m_config->n_simt_cores_per_cluster; i++ ) { unsigned core = (i+m_cta_issue_next_core+1)%m_config->n_simt_cores_per_cluster; - if( m_core[core]->get_not_completed() == 0 ) { - if( m_core[core]->get_kernel() == NULL ) { - kernel_info_t *k = m_gpu->select_kernel(); - if( k ) - m_core[core]->set_kernel(k); + + kernel_info_t * kernel; + //Jin: fetch kernel according to concurrent kernel setting + if(m_config->gpgpu_concurrent_kernel_sm) {//concurrent kernel on sm + //always select latest issued kernel + kernel_info_t *k = m_gpu->select_kernel(); + kernel = k; + } + else { + //first select core kernel, if no more cta, get a new kernel + //only when core completes + kernel = m_core[core]->get_kernel(); + if( !m_gpu->kernel_more_cta_left(kernel) ) { + //wait till current kernel finishes + if(m_core[core]->get_not_completed() == 0) + { + kernel_info_t *k = m_gpu->select_kernel(); + if( k ) + m_core[core]->set_kernel(k); + kernel = k; + } } } - kernel_info_t *kernel = m_core[core]->get_kernel(); - if( m_gpu->kernel_more_cta_left(kernel) && (m_core[core]->get_n_active_cta() < m_config->max_cta(*kernel)) ) { + + if( m_gpu->kernel_more_cta_left(kernel) && +// (m_core[core]->get_n_active_cta() < m_config->max_cta(*kernel)) ) { + m_core[core]->can_issue_1block(*kernel)) { m_core[core]->issue_block2core(*kernel); num_blocks_issued++; m_cta_issue_next_core=core; diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 1aa468b..bdd8dbe 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -108,6 +108,10 @@ public: m_last_fetch=0; m_next=0; m_inst_at_barrier=NULL; + + //Jin: cdp support + m_cdp_latency = 0; + m_cdp_dummy = false; } void init( address_type start_pc, unsigned cta_id, @@ -124,6 +128,10 @@ public: n_completed -= active.count(); // active threads are not yet completed m_active_threads = active; m_done_exit=false; + + //Jin: cdp support + m_cdp_latency = 0; + m_cdp_dummy = false; } bool functional_done() const; @@ -260,6 +268,11 @@ private: unsigned m_stores_outstanding; // number of store requests sent but not yet acknowledged unsigned m_inst_in_pipeline; + + //Jin: cdp support +public: + unsigned int m_cdp_latency; + bool m_cdp_dummy; }; @@ -267,7 +280,7 @@ 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); @@ -1327,11 +1340,14 @@ struct shader_core_config : public core_config int simt_core_sim_order; unsigned mem2device(unsigned memid) const { return memid + n_simt_clusters; } + + //Jin: concurrent kernel on sm + bool gpgpu_concurrent_kernel_sm; }; 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 @@ -1574,6 +1590,7 @@ public: void cycle(); void reinit(unsigned start_thread, unsigned end_thread, bool reset_not_completed ); void issue_block2core( class kernel_info_t &kernel ); + void cache_flush(); void accept_fetch_response( mem_fetch *mf ); void accept_ldst_unit_response( class mem_fetch * mf ); @@ -1582,7 +1599,7 @@ public: { assert(k); m_kernel=k; - k->inc_running(); +// k->inc_running(); printf("GPGPU-Sim uArch: Shader %d bind to kernel %u \'%s\'\n", m_sid, m_kernel->get_uid(), m_kernel->name().c_str() ); } @@ -1749,7 +1766,7 @@ public: virtual void checkExecutionStatusAndUpdate(warp_inst_t &inst, unsigned t, unsigned tid); address_type next_pc( int tid ) const; void fetch(); - void register_cta_thread_exit( unsigned cta_num ); + void register_cta_thread_exit(unsigned cta_num, kernel_info_t * kernel ); void decode(); @@ -1831,6 +1848,22 @@ public: // is that the dynamic_warp_id is a running number unique to every warp // run on this shader, where the warp_id is the static warp slot. unsigned m_dynamic_warp_id; + + //Jin: concurrent kernels on a sm +public: + bool can_issue_1block(kernel_info_t & kernel); + bool occupy_shader_resource_1block(kernel_info_t & kernel, bool occupy); + void release_shader_resource_1block(unsigned hw_ctaid, kernel_info_t & kernel); + int find_available_hwtid(unsigned int cta_size, bool occupy); +private: + unsigned int m_occupied_n_threads; + unsigned int m_occupied_shmem; + unsigned int m_occupied_regs; + unsigned int m_occupied_ctas; + std::bitset<MAX_THREAD_PER_SM> m_occupied_hwtid; + std::map<unsigned int, unsigned int> m_occupied_cta_to_hwtid; + + }; class simt_core_cluster { @@ -1851,6 +1884,7 @@ public: bool icnt_injection_buffer_full(unsigned size, bool write); void icnt_inject_request_packet(class mem_fetch *mf); + // for perfect memory interface bool response_queue_full() { return ( m_response_fifo.size() >= m_config->n_simt_ejection_buffer_size ); diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index 6ba38eb..04845e7 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -100,7 +100,7 @@ void *gpgpu_sim_thread_concurrent(void*) printf("GPGPU-Sim: *** simulation thread starting and spinning waiting for work ***\n"); fflush(stdout); } - while( g_stream_manager->empty_protected() && !g_sim_done ) + while( g_stream_manager->empty() && !g_sim_done ) ; if(g_debug_execution >= 3) { printf("GPGPU-Sim: ** START simulation thread (detected work) **\n"); @@ -127,6 +127,15 @@ void *gpgpu_sim_thread_concurrent(void*) if(g_stream_manager->operation(&sim_cycles) && !g_the_gpu->active()) break; + //functional simulation + if( g_the_gpu->is_functional_sim()) { + kernel_info_t * kernel = g_the_gpu->get_functional_kernel(); + assert(kernel); + gpgpu_cuda_ptx_sim_main_func(*kernel); + g_the_gpu->finish_functional_sim(kernel); + } + + //performance simulation if( g_the_gpu->active() ) { g_the_gpu->cycle(); sim_cycles = true; @@ -144,6 +153,7 @@ void *gpgpu_sim_thread_concurrent(void*) printf("GPGPU-Sim: ** STOP simulation thread (no work) **\n"); fflush(stdout); } + g_the_gpu->print_stats(); if(sim_cycles) { g_the_gpu->update_stats(); print_simulation_time(); diff --git a/src/stream_manager.cc b/src/stream_manager.cc index dd42f0a..3b6cbd5 100644 --- a/src/stream_manager.cc +++ b/src/stream_manager.cc @@ -95,6 +95,15 @@ stream_operation CUstream_st::next() return result; } +void CUstream_st::cancel_front() +{ + pthread_mutex_lock(&m_lock); + assert(m_pending); + m_pending = false; + pthread_mutex_unlock(&m_lock); + +} + void CUstream_st::print(FILE *fp) { pthread_mutex_lock(&m_lock); @@ -111,10 +120,10 @@ void CUstream_st::print(FILE *fp) } -void stream_operation::do_operation( gpgpu_sim *gpu ) +bool stream_operation::do_operation( gpgpu_sim *gpu ) { if( is_noop() ) - return; + return true; assert(!m_done && m_stream); if(g_debug_execution >= 3) @@ -151,17 +160,36 @@ void stream_operation::do_operation( gpgpu_sim *gpu ) m_stream->record_next_done(); break; case stream_kernel_launch: - if( gpu->can_start_kernel() ) { - gpu->set_cache_config(m_kernel->name()); - printf("kernel \'%s\' transfer to GPU hardware scheduler\n", m_kernel->name().c_str() ); - if( m_sim_mode ) - gpgpu_cuda_ptx_sim_main_func( *m_kernel ); - else + if( m_sim_mode ) { //Functional Sim + if(g_debug_execution >= 3) { + printf("kernel %d: \'%s\' transfer to GPU hardware scheduler\n", m_kernel->get_uid(), m_kernel->name().c_str() ); + m_kernel->print_parent_info(); + } + gpu->set_cache_config(m_kernel->name()); + gpu->functional_launch( m_kernel ); + } + else { //Performance Sim + if( gpu->can_start_kernel() && m_kernel->m_launch_latency == 0) { + if(g_debug_execution >= 3) { + printf("kernel %d: \'%s\' transfer to GPU hardware scheduler\n", m_kernel->get_uid(), m_kernel->name().c_str() ); + m_kernel->print_parent_info(); + } + gpu->set_cache_config(m_kernel->name()); gpu->launch( m_kernel ); + } + else { + if(m_kernel->m_launch_latency) + m_kernel->m_launch_latency--; + if(g_debug_execution >= 3) + printf("kernel %d: \'%s\', latency %u not ready to transfer to GPU hardware scheduler\n", + m_kernel->get_uid(), m_kernel->name().c_str(), m_kernel->m_launch_latency); + return false; + } } break; case stream_event: { - printf("event update\n"); + if(g_debug_execution >= 3) + printf("event update\n"); time_t wallclock = time((time_t *)NULL); m_event->update( gpu_tot_sim_cycle, wallclock ); m_stream->record_next_done(); @@ -172,6 +200,7 @@ void stream_operation::do_operation( gpgpu_sim *gpu ) } m_done=true; fflush(stdout); + return true; } void stream_operation::print( FILE *fp ) const @@ -199,11 +228,20 @@ stream_manager::stream_manager( gpgpu_sim *gpu, bool cuda_launch_blocking ) bool stream_manager::operation( bool * sim) { - pthread_mutex_lock(&m_lock); bool check=check_finished_kernel(); - if(check)m_gpu->print_stats(); + pthread_mutex_lock(&m_lock); +// if(check)m_gpu->print_stats(); stream_operation op =front(); - op.do_operation( m_gpu ); + if(!op.do_operation( m_gpu )) //not ready to execute + { + //cancel operation + if( op.is_kernel() ) { + unsigned grid_uid = op.get_kernel()->get_uid(); + m_grid_id_to_stream.erase(grid_uid); + } + op.get_stream()->cancel_front(); + + } pthread_mutex_unlock(&m_lock); //pthread_mutex_lock(&m_lock); // simulate a clock cycle on the GPU @@ -212,11 +250,9 @@ bool stream_manager::operation( bool * sim) bool stream_manager::check_finished_kernel() { - - unsigned grid_uid = m_gpu->finished_kernel(); - bool check=register_finished_kernel(grid_uid); - return check; - + unsigned grid_uid = m_gpu->finished_kernel(); + bool check=register_finished_kernel(grid_uid); + return check; } bool stream_manager::register_finished_kernel(unsigned grid_uid) @@ -226,13 +262,27 @@ bool stream_manager::register_finished_kernel(unsigned grid_uid) CUstream_st *stream = m_grid_id_to_stream[grid_uid]; kernel_info_t *kernel = stream->front().get_kernel(); assert( grid_uid == kernel->get_uid() ); - stream->record_next_done(); - m_grid_id_to_stream.erase(grid_uid); - delete kernel; - return true; - }else{ - return false; + + //Jin: should check children kernels for CDP + if(kernel->is_finished()) { +// std::ofstream kernel_stat("kernel_stat.txt", std::ofstream::out | std::ofstream::app); +// kernel_stat<< " kernel " << grid_uid << ": " << kernel->name(); +// if(kernel->get_parent()) +// kernel_stat << ", parent " << kernel->get_parent()->get_uid() << +// ", launch " << kernel->launch_cycle; +// kernel_stat<< ", start " << kernel->start_cycle << +// ", end " << kernel->end_cycle << ", retire " << gpu_sim_cycle + gpu_tot_sim_cycle << "\n"; +// printf("kernel %d finishes, retires from stream %d\n", grid_uid, stream->get_uid()); +// kernel_stat.flush(); +// kernel_stat.close(); + stream->record_next_done(); + m_grid_id_to_stream.erase(grid_uid); + kernel->notify_parent_finished(); + delete kernel; + return true; + } } + return false; } @@ -259,21 +309,22 @@ stream_operation stream_manager::front() { // called by gpu simulation thread stream_operation result; - if( concurrent_streams_empty() ) - m_service_stream_zero = true; +// if( concurrent_streams_empty() ) + m_service_stream_zero = true; if( m_service_stream_zero ) { - if( !m_stream_zero.empty() ) { - if( !m_stream_zero.busy() ) { + if( !m_stream_zero.empty() && !m_stream_zero.busy() ) { result = m_stream_zero.next(); if( result.is_kernel() ) { unsigned grid_id = result.get_kernel()->get_uid(); m_grid_id_to_stream[grid_id] = &m_stream_zero; } - } } else { m_service_stream_zero = false; } - } else { + } + + if(!m_service_stream_zero) + { std::list<struct CUstream_st*>::iterator s; for( s=m_streams.begin(); s != m_streams.end(); s++) { CUstream_st *stream = *s; diff --git a/src/stream_manager.h b/src/stream_manager.h index 701b33c..222a1b2 100644 --- a/src/stream_manager.h +++ b/src/stream_manager.h @@ -150,7 +150,7 @@ public: bool is_noop() const { return m_type == stream_no_op; } bool is_done() const { return m_done; } kernel_info_t *get_kernel() { return m_kernel; } - void do_operation( gpgpu_sim *gpu ); + bool do_operation( gpgpu_sim *gpu ); void print( FILE *fp ) const; struct CUstream_st *get_stream() { return m_stream; } void set_stream( CUstream_st *stream ) { m_stream = stream; } @@ -218,6 +218,7 @@ public: void push( const stream_operation &op ); void record_next_done(); stream_operation next(); + void cancel_front(); //front operation fails, cancle the pending status stream_operation &front() { return m_operations.front(); } void print( FILE *fp ); unsigned get_uid() const { return m_uid; } |