diff options
76 files changed, 6051 insertions, 566 deletions
@@ -32,4 +32,5 @@ cuobjdump_to_ptxplus/sass_parser.hh cuobjdump_to_ptxplus/sass_parser.output build/* +tags *.swp diff --git a/.travis.yml b/.travis.yml index 7a12a1c..b38f468 100644 --- a/.travis.yml +++ b/.travis.yml @@ -1,7 +1,7 @@ sudo: required before_install: - - docker pull tgrogers/gpgpu-sim_regress:latest + - docker pull tgrogers/gpgpu-sim_regress:volta_update language: cpp @@ -10,34 +10,13 @@ matrix: include: - services: docker env: - - CONFIG=configs.gtx480.yml + - CONFIG=GTX480 - CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ -# This config is just taking far too long... -# - services: docker -# env: -# - CONFIG=configs.gtx750ti.yml -# - CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ -# - PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - services: docker env: - - CONFIG=configs.quadro5600.yml - - CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - - PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - - services: docker - env: - - CONFIG=configs.quadro5800.yml - - CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - - PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - - services: docker - env: - - CONFIG=configs.teslac2050.yml - - CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - - PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - - services: docker - env: - - CONFIG=configs.gtx1080ti.yml - - CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ + - CONFIG=TITANV + - CUDA_INSTALL_PATH=/usr/local/cuda-9.1/ - PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-9.1/ -script: docker run -v `pwd`:/home/runner/gpgpu-sim_distribution:rw tgrogers/gpgpu-sim_regress:latest /bin/bash -c "./start_torque.sh; chown -R runner /home/runner/gpgpu-sim_distribution; su - runner -c 'export CUDA_INSTALL_PATH=$CUDA_INSTALL_PATH && export PTXAS_CUDA_INSTALL_PATH=$PTXAS_CUDA_INSTALL_PATH && source /home/runner/gpgpu-sim_distribution/setup_environment && make -j -C /home/runner/gpgpu-sim_distribution && cd /home/runner/gpgpu-sim_simulations/ && git pull && /home/runner/gpgpu-sim_simulations/util/job_launching/run_simulations.py -c /home/runner/gpgpu-sim_simulations/util/job_launching/regression_recipies/rodinia_2.0-ft/$CONFIG -N regress && /home/runner/gpgpu-sim_simulations/util/job_launching/monitor_func_test.py -v -N regress'" +script: docker run -v `pwd`:/home/runner/gpgpu-sim_distribution:rw tgrogers/gpgpu-sim_regress:volta_update /bin/bash -c "./start_torque.sh; chown -R runner /home/runner/gpgpu-sim_distribution; su - runner -c 'export CUDA_INSTALL_PATH=$CUDA_INSTALL_PATH && export PTXAS_CUDA_INSTALL_PATH=$PTXAS_CUDA_INSTALL_PATH && source /home/runner/gpgpu-sim_distribution/setup_environment && make -j -C /home/runner/gpgpu-sim_distribution && cd /home/runner/gpgpu-sim_simulations/ && git pull && /home/runner/gpgpu-sim_simulations/util/job_launching/run_simulations.py -C $CONFIG -B rodinia_2.0-ft -N regress && /home/runner/gpgpu-sim_simulations/util/job_launching/monitor_func_test.py -v -N regress'" @@ -1,4 +1,34 @@ LOG: +Version 4.0.0 (development branch) versus 3.2.3 +-Front-End: +1- Support .nc cache modifier and __ldg function to access the read-only L1D cache +2- Partially-support some SASS_60 in the PTXP_PLUS (not completed yet) +-GPU Core: +1- Fermi/Pascal coalescer: coalescing on 32-bytes sectors. +2- Adding separate dp unit pipeline. +3- diff dual issue: allow scheduler to issue diff insts at a time +4- Fair memory issue from multiple schedulers. +-Cache System: +1- Sector L1/L2 cache +2- Fetch-on-write and lazy-fetch-on-read write allocation policy. +3- Improving the L1 cache throughput (streaming L1 cache) +4- Performance model for CUDA memory copy. +5- Support memory partition indexing to reduce partition camping (POLY, XOR and PAE (ISCA’18) Indexing) +6- Adaptive cache configuration +-Memory: +1- Performance Model for HBM (mainly the dual-bus interface) +2- Separate Read/Write buffers. +3- Advanced bank indexing function. +-Statistics: +1- Adding more detailed cache statistics to define and analyze cache bottlenecks. +2- Adding more detailed memory statistics (BLP, RBL, etc) to define and analyze memory bottlenecks. +3- Addig new system stats: gpu occupancy, L2BW, etc +-Configs: +Adding the Pascal and Volta config files that has been correlated against real hardware. +See the correlation website here: +https://engineering.purdue.edu/tgrogers/group/correlator.html + + Version 3.2.3+edits (development branch) versus 3.2.3 - Support for running regression tests using Travis - Support added for CUDA dynamic parallelism (courtesy of Jin Wang from Georgia Tech) diff --git a/Jenkinsfile b/Jenkinsfile new file mode 100644 index 0000000..1969aea --- /dev/null +++ b/Jenkinsfile @@ -0,0 +1,126 @@ +pipeline { + agent { + label "purdue-cluster" + } + + options { + disableConcurrentBuilds() + } + + stages { + stage('simulator-build') { + steps { + parallel "4.2": { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/4.2_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + make -j' + }, "9.1" : { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + make -j' + } + } + } + stage('simulations-build'){ + steps{ + sh 'rm -rf gpgpu-sim_simulations' + sh 'git clone [email protected]:TimRogersGroup/gpgpu-sim_simulations.git && \ + cd gpgpu-sim_simulations && \ + git checkout purdue-cluster && \ + git pull && \ + ln -s /home/tgrogers-raid/a/common/data_dirs benchmarks/' + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/4.2_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + cd gpgpu-sim_simulations && \ + source ./benchmarks/src/setup_environment && \ + make -j -C ./benchmarks/src rodinia_2.0-ft sdk-4.2 && \ + make -C ./benchmarks/src data' + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + cd gpgpu-sim_simulations && \ + source ./benchmarks/src/setup_environment && \ + make -j -C ./benchmarks/src/ rodinia_2.0-ft sdk-4.2 && \ + make -C ./benchmarks/src data' + } + } + stage('regress'){ + steps { + parallel "4.2-rodinia": { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/4.2_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + ./gpgpu-sim_simulations/util/job_launching/run_simulations.py -B rodinia_2.0-ft -C GTX480,GTX480-PTXPLUS -N regress-$$ && \ + PLOTDIR="jenkins/${JOB_NAME}/${BUILD_NUMBER}/4.2-rodinia" && ssh [email protected] mkdir -p /home/dynamo/a/tgrogers/website/gpgpu-sim-plots/$PLOTDIR && \ + ./gpgpu-sim_simulations/util/job_launching/monitor_func_test.py -v -N regress-$$ -s stats-per-app-4.2.csv && \ + ./gpgpu-sim_simulations/util/plotting/plot-get-stats.py -c stats-per-app-4.2.csv -p [email protected]:~/website/gpgpu-sim-plots/$PLOTDIR -w https://engineering.purdue.edu/tgrogers/gpgpu-sim-plots/$PLOTDIR -n $PLOTDIR ' + }, "9.1-functest": { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + ./gpgpu-sim_simulations/util/job_launching/run_simulations.py -B rodinia_2.0-ft,sdk-4.2 -C TITANX,TITANX-L1ON -N regress-$$ && \ + PLOTDIR="jenkins/${JOB_NAME}/${BUILD_NUMBER}/9.1-rodinia" && ssh [email protected] mkdir -p /home/dynamo/a/tgrogers/website/gpgpu-sim-plots/$PLOTDIR && \ + ./gpgpu-sim_simulations/util/job_launching/monitor_func_test.py -v -s stats-per-app-9.1.csv -N regress-$$ && \ + ./gpgpu-sim_simulations/util/plotting/plot-get-stats.py -c stats-per-app-9.1.csv -p [email protected]:~/website/gpgpu-sim-plots/$PLOTDIR -w https://engineering.purdue.edu/tgrogers/gpgpu-sim-plots/$PLOTDIR -n $PLOTDIR' + } + } + } + stage('4.2-correlate'){ + steps { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/4.2_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + PLOTDIR="jenkins/${JOB_NAME}" &&\ + ./gpgpu-sim_simulations/util/job_launching/get_stats.py -R -K -k -B rodinia_2.0-ft -C GTX480,GTX480-PTXPLUS > stats-per-kernel-4.2.csv &&\ + ./gpgpu-sim_simulations/util/plotting/correlate_and_publish.sh stats-per-kernel-4.2.csv $PLOTDIR ${BUILD_NUMBER}' + } + } + stage('9.1-correlate'){ + steps { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + PLOTDIR="jenkins/${JOB_NAME}" &&\ + ./gpgpu-sim_simulations/util/job_launching/get_stats.py -R -K -k -B rodinia_2.0-ft,sdk-4.2 -C TITANX,TITANX-L1ON > stats-per-kernel-9.1.csv &&\ + ./gpgpu-sim_simulations/util/plotting/correlate_and_publish.sh stats-per-kernel-9.1.csv $PLOTDIR ${BUILD_NUMBER}' + } + } + stage('archive-and-delta') { + steps { + sh 'rm -rf gpgpu-sim-results-repo' + sh 'git clone [email protected]:purdue-aalp/gpgpu-sim-results-repo.git' + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/4.2_env_setup.sh &&\ + ./gpgpu-sim_simulations/util/job_launching/get_stats.py -R -K -k -B rodinia_2.0-ft -C GTX480-PTXPLUS > stats-per-kernel-4.2-ptxplus.csv &&\ + ./gpgpu-sim_simulations/util/job_launching/get_stats.py -R -K -k -B rodinia_2.0-ft -C GTX480 > stats-per-kernel-4.2-ptx.csv' + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + ./gpgpu-sim_simulations/util/job_launching/get_stats.py -R -K -k -B rodinia_2.0-ft,sdk-4.2 -C TITANX > stats-per-kernel-9.1-titanx.csv' + sh './gpgpu-sim_simulations/util/plotting/merge-stats.py -c ./gpgpu-sim-results-repo/jenkins/quick-regress/AALP/gpgpu-sim_distribution/dev-purdue-integration/stats-per-app-4.2.csv,./stats-per-app-4.2.csv -R > per-app-merge-4.2.csv' + sh './gpgpu-sim_simulations/util/plotting/merge-stats.py -c ./gpgpu-sim-results-repo/jenkins/quick-regress/AALP/gpgpu-sim_distribution/dev-purdue-integration/stats-per-app-9.1.csv,./stats-per-app-9.1.csv -R > per-app-merge-9.1.csv' + sh 'PLOTDIR="jenkins/${JOB_NAME}" &&\ + ./gpgpu-sim_simulations/util/plotting/plot-get-stats.py -c per-app-merge-4.2.csv -p [email protected]:~/website/gpgpu-sim-plots/$PLOTDIR/deltas -w https://engineering.purdue.edu/tgrogers/gpgpu-sim-plots/$PLOTDIR -n $PLOTDIR/deltas &&\ + ./gpgpu-sim_simulations/util/plotting/plot-get-stats.py -c per-app-merge-9.1.csv -p [email protected]:~/website/gpgpu-sim-plots/$PLOTDIR/deltas -w https://engineering.purdue.edu/tgrogers/gpgpu-sim-plots/$PLOTDIR/deltas -n $PLOTDIR/deltas &&\ + ./gpgpu-sim_simulations/util/plotting/merge-stats.py -c ./gpgpu-sim-results-repo/jenkins/quick-regress/AALP/gpgpu-sim_distribution/dev-purdue-integration/stats-per-kernel-4.2-ptx.csv,./stats-per-kernel-4.2-ptx.csv -R > per-kernel-merge-4.2-ptx.csv &&\ + ./gpgpu-sim_simulations/util/plotting/merge-stats.py -c ./gpgpu-sim-results-repo/jenkins/quick-regress/AALP/gpgpu-sim_distribution/dev-purdue-integration/stats-per-kernel-4.2-ptxplus.csv,./stats-per-kernel-4.2-ptxplus.csv -R > per-kernel-merge-4.2-ptxplus.csv &&\ + ./gpgpu-sim_simulations/util/plotting/merge-stats.py -c ./gpgpu-sim-results-repo/jenkins/quick-regress/AALP/gpgpu-sim_distribution/dev-purdue-integration/stats-per-kernel-9.1-titanx.csv,./stats-per-kernel-9.1-titanx.csv -R > per-kernel-merge-9.1-titanx.csv &&\ + ./gpgpu-sim_simulations/util/plotting/correlate_and_publish.sh per-kernel-merge-4.2-ptx.csv $PLOTDIR ${BUILD_NUMBER} &&\ + ./gpgpu-sim_simulations/util/plotting/correlate_and_publish.sh per-kernel-merge-4.2-ptxplus.csv $PLOTDIR ${BUILD_NUMBER} &&\ + ./gpgpu-sim_simulations/util/plotting/correlate_and_publish.sh per-kernel-merge-9.1-titanx.csv $PLOTDIR ${BUILD_NUMBER} &&\ + mkdir -p ./jenkins/quick-regress/${JOB_NAME}/ && cp stats-per-*.csv ./jenkins/quick-regress/${JOB_NAME}/ &&\ + cd ./gpgpu-sim-results-repo &&\ + git diff --quiet && git diff --staged --quiet || git commit -am "Jenkins automated checkin ${BUILD_NUMBER}" &&\ + git push' + } + } + } + post { + success { + emailext body: "See ${BUILD_URL}", + recipientProviders: [[$class: 'CulpritsRecipientProvider'], + [$class: 'RequesterRecipientProvider']], + subject: "[AALP Jenkins] Build #${BUILD_NUMBER} - Success!", + to: '[email protected]' + } + failure { + emailext body: "See ${BUILD_URL}", + recipientProviders: [[$class: 'CulpritsRecipientProvider'], + [$class: 'RequesterRecipientProvider']], + subject: "[AALP Jenkins] Build #${BUILD_NUMBER} - ${currentBuild.result}", + to: '[email protected]' + } + } +} @@ -5,8 +5,8 @@ AerialVision and a configurable and extensible energy model called GPUWattch. GPGPU-Sim and GPUWattch have been rigorously validated with performance and power measurements of real hardware GPUs. -This version of GPGPU-Sim has been tested with CUDA version 2.3, 3.1, 4.0, -5.0, 5.5, 6.0 and 7.5. +This version of GPGPU-Sim has been tested with CUDA version 4.2, +5.0, 5.5, 6.0 and 7.5, 8.0, 9.0, 9.1 Please see the copyright notice in the file COPYRIGHT distributed with this release in the same directory as this file. @@ -18,6 +18,12 @@ Analyzing CUDA Workloads Using a Detailed GPU Simulator, in IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), Boston, MA, April 19-21, 2009. +If you use the memory system in GPGPU-Sim, or the Volta/Pascal models, +please cite: +Mahmoud Khairy, Jain Akshay, Tor Aamodt, Timothy G Rogers, +Exploring Modern GPU Memory System Design Challenges through Accurate Modeling, arXiv:1810.07269, +https://arxiv.org/abs/1810.07269 + If you use the GPUWattch energy model in your research, please cite: Jingwen Leng, Tayler Hetherington, Ahmed ElTantawy, Syed Gilani, Nam Sung Kim, diff --git a/aerialvision/organizedata.py b/aerialvision/organizedata.py index 090b90f..ea947cd 100644 --- a/aerialvision/organizedata.py +++ b/aerialvision/organizedata.py @@ -97,7 +97,7 @@ def organizedata(fileVars): 'sparse':OrganizeSparse, # Vector data with 2D index (used by DRAM access stats) 'custom':0 } - data_type_char = {int:'I', float:'f'} + data_type_char = {int:'L', float:'d'} print "Organizing data into internal format..." diff --git a/configs/GTX480/config_fermi_islip.icnt b/configs/deprecated-cfgs/GTX480/config_fermi_islip.icnt index 7820e4e..7820e4e 100644 --- a/configs/GTX480/config_fermi_islip.icnt +++ b/configs/deprecated-cfgs/GTX480/config_fermi_islip.icnt diff --git a/configs/GTX480/gpgpusim.config b/configs/deprecated-cfgs/GTX480/gpgpusim.config index ee90c12..ee90c12 100644 --- a/configs/GTX480/gpgpusim.config +++ b/configs/deprecated-cfgs/GTX480/gpgpusim.config diff --git a/configs/GTX480/gpuwattch_gtx480.xml b/configs/deprecated-cfgs/GTX480/gpuwattch_gtx480.xml index 304e0fd..304e0fd 100755 --- a/configs/GTX480/gpuwattch_gtx480.xml +++ b/configs/deprecated-cfgs/GTX480/gpuwattch_gtx480.xml diff --git a/configs/GeForceGTX750Ti/config_fermi_islip.icnt b/configs/deprecated-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt index 069ca02..069ca02 100644 --- a/configs/GeForceGTX750Ti/config_fermi_islip.icnt +++ b/configs/deprecated-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt diff --git a/configs/GeForceGTX750Ti/gpgpusim.config b/configs/deprecated-cfgs/GeForceGTX750Ti/gpgpusim.config index c675aab..c675aab 100644 --- a/configs/GeForceGTX750Ti/gpgpusim.config +++ b/configs/deprecated-cfgs/GeForceGTX750Ti/gpgpusim.config diff --git a/configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml b/configs/deprecated-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml index e2b2324..e2b2324 100755 --- a/configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml +++ b/configs/deprecated-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml diff --git a/configs/QuadroFX5600/gpgpusim.config b/configs/deprecated-cfgs/QuadroFX5600/gpgpusim.config index 6f836ee..6f836ee 100644 --- a/configs/QuadroFX5600/gpgpusim.config +++ b/configs/deprecated-cfgs/QuadroFX5600/gpgpusim.config diff --git a/configs/QuadroFX5600/gpuwattch_quadrofx5600.xml b/configs/deprecated-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml index 2c5a6fc..2c5a6fc 100644 --- a/configs/QuadroFX5600/gpuwattch_quadrofx5600.xml +++ b/configs/deprecated-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml diff --git a/configs/QuadroFX5600/icnt_config_islip.icnt b/configs/deprecated-cfgs/QuadroFX5600/icnt_config_islip.icnt index de3bcc8..de3bcc8 100644 --- a/configs/QuadroFX5600/icnt_config_islip.icnt +++ b/configs/deprecated-cfgs/QuadroFX5600/icnt_config_islip.icnt diff --git a/configs/QuadroFX5800/config_quadro_islip.icnt b/configs/deprecated-cfgs/QuadroFX5800/config_quadro_islip.icnt index cfe9cac..cfe9cac 100644 --- a/configs/QuadroFX5800/config_quadro_islip.icnt +++ b/configs/deprecated-cfgs/QuadroFX5800/config_quadro_islip.icnt diff --git a/configs/QuadroFX5800/gpgpusim.config b/configs/deprecated-cfgs/QuadroFX5800/gpgpusim.config index fef1110..fef1110 100644 --- a/configs/QuadroFX5800/gpgpusim.config +++ b/configs/deprecated-cfgs/QuadroFX5800/gpgpusim.config diff --git a/configs/GeForceGTX1080Ti/config_fermi_islip.icnt b/configs/deprecated-cfgs/SM6_GTX1080/config_fermi_islip.icnt index 2a69ddd..2a69ddd 100644 --- a/configs/GeForceGTX1080Ti/config_fermi_islip.icnt +++ b/configs/deprecated-cfgs/SM6_GTX1080/config_fermi_islip.icnt diff --git a/configs/GeForceGTX1080Ti/gpgpusim.config b/configs/deprecated-cfgs/SM6_GTX1080/gpgpusim.config index fb044c6..fb044c6 100644 --- a/configs/GeForceGTX1080Ti/gpgpusim.config +++ b/configs/deprecated-cfgs/SM6_GTX1080/gpgpusim.config diff --git a/configs/GeForceGTX1080Ti/gpuwattch_gtx1080Ti.xml b/configs/deprecated-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml index 02619ff..02619ff 100755 --- a/configs/GeForceGTX1080Ti/gpuwattch_gtx1080Ti.xml +++ b/configs/deprecated-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml diff --git a/configs/deprecated-cfgs/SM6_P100/config_fermi_islip.icnt b/configs/deprecated-cfgs/SM6_P100/config_fermi_islip.icnt new file mode 100644 index 0000000..d26c8d9 --- /dev/null +++ b/configs/deprecated-cfgs/SM6_P100/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 = 60; +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/deprecated-cfgs/SM6_P100/gpgpusim.config b/configs/deprecated-cfgs/SM6_P100/gpgpusim.config new file mode 100644 index 0000000..a5e6736 --- /dev/null +++ b/configs/deprecated-cfgs/SM6_P100/gpgpusim.config @@ -0,0 +1,156 @@ +# This config models the Pascal GP100 +# For more info about this card, see Nvidia White paper +# https://images.nvidia.com/content/pdf/tesla/whitepaper/pascal-architecture-whitepaper.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 60 + +# 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 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 32 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Pscal clock domains +#-gpgpu_clock_domains <Core Clock>:<Interconnect Clock>:<L2 Clock>:<DRAM Clock> +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Nvidia_Tesla +-gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 + +# 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 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,2,1,2,2,1,4 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 2,2,2,2,130 + +# <nsets>:<bsize>:<assoc>,<rep>:<wr>:<alloc>:<wr_alloc>:<set_index_fn>,<mshr>:<N>:<merge>,<mq>:**<fifo_entry> +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +# Pascal GP100 has 48KB L1 cache +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 65536 +-gpgpu_shmem_size_PrefShared 65536 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 14 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_units_mem 10 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# 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 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# 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_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/deprecated-cfgs/SM6_P100/gpuwattch_gtx480.xml b/configs/deprecated-cfgs/SM6_P100/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/deprecated-cfgs/SM6_P100/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ +<?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/configs/Pascal_TITANX/config_fermi_islip.icnt b/configs/deprecated-cfgs/SM6_TITANX/config_fermi_islip.icnt index 602daee..602daee 100644 --- a/configs/Pascal_TITANX/config_fermi_islip.icnt +++ b/configs/deprecated-cfgs/SM6_TITANX/config_fermi_islip.icnt diff --git a/configs/Pascal_TITANX/gpgpusim.config b/configs/deprecated-cfgs/SM6_TITANX/gpgpusim.config index f78bd02..28912a3 100644 --- a/configs/Pascal_TITANX/gpgpusim.config +++ b/configs/deprecated-cfgs/SM6_TITANX/gpgpusim.config @@ -25,6 +25,7 @@ # shader core pipeline config -gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 61 # This implies a maximum of 64 warps/SM -gpgpu_shader_core_pipeline 2048:32 diff --git a/configs/deprecated-cfgs/SM6_TITANX/gpuwattch_gtx480.xml b/configs/deprecated-cfgs/SM6_TITANX/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/deprecated-cfgs/SM6_TITANX/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ +<?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/configs/deprecated-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/deprecated-cfgs/SM7_TITANV/config_fermi_islip.icnt new file mode 100644 index 0000000..fac792a --- /dev/null +++ b/configs/deprecated-cfgs/SM7_TITANV/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 = 64; +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/deprecated-cfgs/SM7_TITANV/gpgpusim.config b/configs/deprecated-cfgs/SM7_TITANV/gpgpusim.config new file mode 100644 index 0000000..aefb04a --- /dev/null +++ b/configs/deprecated-cfgs/SM7_TITANV/gpgpusim.config @@ -0,0 +1,165 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# 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 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains <Core Clock>:<Interconnect Clock>:<L2 Clock>:<DRAM Clock> +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 + +# 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 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,1,4,4,1,9 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 + +# <nsets>:<bsize>:<assoc>,<rep>:<wr>:<alloc>:<wr_alloc>:<set_index_fn>,<mshr>:<N>:<merge>,<mq>:**<fifo_entry> +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Volta GV100 has 64KB Shared memory +-gpgpu_cache:dl1 64:128:8,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefL1 64:128:16,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefShared 32:128:6,L:L:m:N:H,A:256:8,16:0 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 1 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 64:128:24,L:B:m:W:L,A:256:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 + +# 128 KB Inst. +-gpgpu_cache:il1 64:128:16,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 128:64:8,L:R:f:N:L,A:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 14 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_units_mem 10 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# 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 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# Fermi has two 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_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/TeslaC2050/config_fermi_islip.icnt b/configs/deprecated-cfgs/TeslaC2050/config_fermi_islip.icnt index a11bd8e..a11bd8e 100644 --- a/configs/TeslaC2050/config_fermi_islip.icnt +++ b/configs/deprecated-cfgs/TeslaC2050/config_fermi_islip.icnt diff --git a/configs/TeslaC2050/gpgpusim.config b/configs/deprecated-cfgs/TeslaC2050/gpgpusim.config index 6ac2c12..6ac2c12 100644 --- a/configs/TeslaC2050/gpgpusim.config +++ b/configs/deprecated-cfgs/TeslaC2050/gpgpusim.config diff --git a/configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt b/configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt new file mode 100644 index 0000000..c399db9 --- /dev/null +++ b/configs/tested-cfgs/SM2_GTX480/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 = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 27; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 64; + +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 = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// 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/tested-cfgs/SM2_GTX480/gpgpusim.config b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config new file mode 100644 index 0000000..4096b09 --- /dev/null +++ b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config @@ -0,0 +1,153 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + + +# 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 15 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains <Core Clock>:<Interconnect Clock>:<L2 Clock>:<DRAM Clock> +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 700.0:700.0:700.0:924.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 20 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +#For Fermi, DP unit =0, DP inst is executed on SFU +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# <nsets>:<bsize>:<assoc>,<rep>:<wr>:<alloc>:<wr_alloc>:<set_index_fn>,<mshr>:<N>:<merge>,<mq>:**<fifo_entry> +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,S:64:8,8 +-gpgpu_shmem_size 49152 +-icnt_flit_size 40 +-gmem_skip_L1D 0 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 35 +-smem_latency 26 +-gpgpu_flush_l1_cache 1 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,S:64:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 S:64:128:8,L:B:m:L:L,A:256:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 1 +-memory_partition_indexing 0 + +-gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,S:2:32,4 +-gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,T:128:4,128:2 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,S: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_coalesce_arch 20 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# 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 64 +-gpgpu_dram_return_queue_size 116 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-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" + +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# 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_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml b/configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ +<?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/configs/tested-cfgs/SM6_TITANX/config_fermi_islip.icnt b/configs/tested-cfgs/SM6_TITANX/config_fermi_islip.icnt new file mode 100644 index 0000000..dec4789 --- /dev/null +++ b/configs/tested-cfgs/SM6_TITANX/config_fermi_islip.icnt @@ -0,0 +1,73 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 64; +input_buffer_size = 256; +ejection_buffer_size = 64; +boundary_buffer_size = 64; + +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 = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// 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/tested-cfgs/SM6_TITANX/gpgpusim.config b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config new file mode 100644 index 0000000..3842508 --- /dev/null +++ b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config @@ -0,0 +1,185 @@ +# This config models the Pascal GP102 (NVIDIA TITAN X) +# For more info about this card, see Nvidia White paper +# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 61 +-gpgpu_ignore_resources_limitation 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +# P102 has two semi-indp scheds per core, and two cores per cluster +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains <Core Clock>:<Interconnect Clock>:<L2 Clock>:<DRAM Clock> +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 62 + +# This implies a maximum of 32 warps/SM +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 16 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 4 SFU units +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 +-gpgpu_num_dp_units 1 + + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,8,8,8,130 +-ptx_opcode_initiation_sfu 4 +-ptx_opcode_latency_sfu 8 + + +# latencies and cache configs are adopted from: +# https://arxiv.org/pdf/1804.06826.pdf +# <nsets>:<bsize>:<assoc>,<rep>:<wr>:<alloc>:<wr_alloc>:<set_index_fn>,<mshr>:<N>:<merge>,<mq>:**<fifo_entry> +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory divided over 2 cores, each has 48KB +# Pascal GP102 has 2 banks L1 cache, where each is 24KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres are used +-gpgpu_cache:dl1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 49152 +-gpgpu_shmem_size_PrefL1 49152 +-gpgpu_shmem_size_PrefShared 49152 +# By default, L1 cache is disabled in Pascal P102. +# requests with .nc modifier or __ldg mehtod will be cached in L1 cache even with gmem_skip_L1D=1 +-gmem_skip_L1D 1 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 82 +-smem_latency 24 +-gpgpu_flush_l1_cache 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:64,16:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 +-perf_sim_memcpy 1 + +# 4 KB Inst. +-gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_operand_collector_num_in_ports_sfu 2 +-gpgpu_operand_collector_num_out_ports_sfu 2 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +# Use Pascal Coalsce arhitetecture +-gpgpu_coalesce_arch 61 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 2 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# 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 64 +-gpgpu_dram_return_queue_size 240 + +# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) +# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: + CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" + +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal 102 has four schedulers per core +-gpgpu_num_sched_per_core 2 +# 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, disable it untill we create a real energy model for Pascal 102 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/tested-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/tested-cfgs/SM7_TITANV/config_fermi_islip.icnt new file mode 100644 index 0000000..2f25889 --- /dev/null +++ b/configs/tested-cfgs/SM7_TITANV/config_fermi_islip.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 256; +input_buffer_size = 256; +ejection_buffer_size = 256; +boundary_buffer_size = 256; + +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 = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// 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/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config new file mode 100644 index 0000000..c8351da --- /dev/null +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -0,0 +1,191 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# 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 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains <Core Clock>:<Interconnect Clock>:<L2 Clock>:<DRAM Clock> +# Volta NVIDIA GV100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:2000.0:1200.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 70 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Volta GV100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_dp_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 +-ptx_opcode_latency_sfu 100 +-ptx_opcode_initiation_sfu 8 + + +# <nsets>:<bsize>:<assoc>,<rep>:<wr>:<alloc>:<wr_alloc>:<set_index_fn>,<mshr>:<N>:<merge>,<mq>:**<fifo_entry> +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Defualt config is 32KB DL1 and 96KB shared memory +# In Volta, we assign the remaining shared memory to L1 cache +# if the assigned shd mem = 0, then L1 cache = 128KB +# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x +# disable this mode in case of multi kernels/apps execution +-adpative_volta_cache_config 1 +-gpgpu_cache:dl1 S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 98304 +-gmem_skip_L1D 0 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 28 +-smem_latency 19 +-gpgpu_flush_l1_cache 1 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 1 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 60 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# 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 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Volta V100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal has two 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, disable it untill we create a real energy model for Pascal 100 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/cuobjdump_to_ptxplus/ptx_parser.h b/cuobjdump_to_ptxplus/ptx_parser.h index a534e92..ee7a942 100644 --- a/cuobjdump_to_ptxplus/ptx_parser.h +++ b/cuobjdump_to_ptxplus/ptx_parser.h @@ -376,4 +376,8 @@ void func_header_info_int(const char* s, int i) g_headerList->getListEnd().addOperand(buff); } } + +void maxnt_id(int x, int y, int z) { + +} #endif //_PTX_PARSER_H_ diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index d67fd85..61af0ee 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -338,10 +338,19 @@ struct _cuda_device_id *GPGPUSim_Init() prop->minor = 2; prop->totalGlobalMem = 0x80000000 /* 2 GB */; prop->memPitch = 0; - prop->maxThreadsPerBlock = 512; - prop->maxThreadsDim[0] = 512; - prop->maxThreadsDim[1] = 512; - prop->maxThreadsDim[2] = 512; + if(prop->major >= 2) { + prop->maxThreadsPerBlock = 1024; + prop->maxThreadsDim[0] = 1024; + prop->maxThreadsDim[1] = 1024; + } + else + { + prop->maxThreadsPerBlock = 512; + prop->maxThreadsDim[0] = 512; + prop->maxThreadsDim[1] = 512; + } + + prop->maxThreadsDim[2] = 64; prop->maxGridSize[0] = 0x40000000; prop->maxGridSize[1] = 0x40000000; prop->maxGridSize[2] = 0x40000000; @@ -354,6 +363,9 @@ struct _cuda_device_id *GPGPUSim_Init() #if (CUDART_VERSION >= 2010) prop->multiProcessorCount = the_gpu->get_config().num_shader(); #endif +#if (CUDART_VERSION >= 4000) + prop->maxThreadsPerMultiProcessor = the_gpu->threads_per_core(); +#endif the_gpu->set_prop(prop); the_device = new _cuda_device_id(the_gpu); } @@ -2167,8 +2179,28 @@ cudaError_t CUDARTAPI cudaSetValidDevices(int *device_arr, int len) cudaError_t CUDARTAPI cudaSetDeviceFlags( int flags ) { - cuda_not_implemented(__my_func__,__LINE__); - return g_last_cudaError = cudaErrorUnknown; + // This flag is implicitly always on (unless you are using the driver API). It is safe for GPGPU-Sim to + // just ignore it. + if ( cudaDeviceMapHost == flags ) { + return g_last_cudaError = cudaSuccess; + } else { + cuda_not_implemented(__my_func__,__LINE__); + return g_last_cudaError = cudaErrorUnknown; + } +} + +size_t getMaxThreadsPerBlock(struct cudaFuncAttributes *attr) { + _cuda_device_id *dev = GPGPUSim_Init(); + struct cudaDeviceProp prop; + + prop = *dev->get_prop(); + + size_t max = prop.maxThreadsPerBlock; + + if ((prop.regsPerBlock / attr->numRegs) < max) + max = prop.regsPerBlock / attr->numRegs; + + return max; } cudaError_t CUDARTAPI cudaFuncGetAttributes(struct cudaFuncAttributes *attr, const char *hostFun ) @@ -2181,7 +2213,10 @@ cudaError_t CUDARTAPI cudaFuncGetAttributes(struct cudaFuncAttributes *attr, con attr->constSizeBytes = kinfo->cmem; attr->localSizeBytes = kinfo->lmem; attr->numRegs = kinfo->regs; - attr->maxThreadsPerBlock = 0; // from pragmas? + if(kinfo->maxthreads > 0) + attr->maxThreadsPerBlock = kinfo->maxthreads; + else + attr->maxThreadsPerBlock = getMaxThreadsPerBlock(attr); #if CUDART_VERSION >= 3000 attr->ptxVersion = kinfo->ptx_version; attr->binaryVersion = kinfo->sm_target; diff --git a/nightly.jenkinsfile b/nightly.jenkinsfile new file mode 100644 index 0000000..5221b3b --- /dev/null +++ b/nightly.jenkinsfile @@ -0,0 +1,75 @@ +pipeline { + agent { + label "purdue-cluster" + } + + options { + disableConcurrentBuilds() + overrideIndexTriggers(true) + } + + triggers { + pollSCM('0 1 * * *') + } + + stages { + stage('nightly-simulator-build') { + steps { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + make -j' + } + } + stage('nightly-simulations-build'){ + steps{ + sh 'rm -rf gpgpu-sim_simulations' + sh 'git clone [email protected]:TimRogersGroup/gpgpu-sim_simulations.git && \ + cd gpgpu-sim_simulations && \ + git checkout purdue-cluster && \ + git pull && \ + ln -s /home/tgrogers-raid/a/common/data_dirs benchmarks/' + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + cd gpgpu-sim_simulations && \ + source ./benchmarks/src/setup_environment && \ + make -i -j -C ./benchmarks/src/ all && \ + make -C ./benchmarks/src data' + } + } + stage('nightly-2B-insn-run'){ + steps { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + ./gpgpu-sim_simulations/util/job_launching/run_simulations.py -B `cat ./gpgpu-sim_simulations/util/job_launching/apps/all-apps.list` -C TITANX-2B,TITANX-L1ON-2B,P100-2B,TITANV-2B -N nightly-$$ && \ + PLOTDIR="jenkins/${JOB_NAME}/${BUILD_NUMBER}/getstats" && ssh [email protected] mkdir -p /home/dynamo/a/tgrogers/website/gpgpu-sim-plots/$PLOTDIR && \ + ./gpgpu-sim_simulations/util/job_launching/monitor_func_test.py -I -S 1800 -v -s stats-$$.csv -N nightly-$$ && \ + ./gpgpu-sim_simulations/util/plotting/plot-get-stats.py -c stats-$$.csv -p [email protected]:~/website/gpgpu-sim-plots/$PLOTDIR -w https://engineering.purdue.edu/tgrogers/gpgpu-sim-plots/$PLOTDIR -n $PLOTDIR' + } + } + stage('nightly-correlate'){ + steps { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + PLOTDIR="jenkins/${JOB_NAME}" &&\ + ./gpgpu-sim_simulations/util/job_launching/get_stats.py -R -K -k -B `cat ./gpgpu-sim_simulations/util/job_launching/apps/all-apps.list` -C TITANX-2B,TITANX-L1ON-2B,P100-2B,TITANV-2B > nightly-stats-per-kernel-9.1.csv &&\ + ./gpgpu-sim_simulations/util/plotting/correlate_and_publish.sh nightly-stats-per-kernel-9.1.csv $PLOTDIR ${BUILD_NUMBER}' + } + } + } + post { + success { + emailext body: "See ${BUILD_URL}", + recipientProviders: [[$class: 'CulpritsRecipientProvider'], + [$class: 'RequesterRecipientProvider']], + subject: "[AALP Jenkins] Build #${BUILD_NUMBER} - Success!", + to: '[email protected]' + } + failure { + emailext body: "See ${BUILD_URL}", + recipientProviders: [[$class: 'CulpritsRecipientProvider'], + [$class: 'RequesterRecipientProvider']], + subject: "[AALP Jenkins] Build #${BUILD_NUMBER} - ${currentBuild.result}", + to: '[email protected]' + } + } +} diff --git a/setup_environment b/setup_environment index f1af978..9578942 100644 --- a/setup_environment +++ b/setup_environment @@ -1,6 +1,6 @@ # see README before running this -ps -p $$ | awk '/bash/ || / sh/ || /zsh/ {exit 1;}' && echo "ERROR ** source setup_environment must be run in a bash, zsh or sh shell; see README" && exit +ps -p $$ | awk '/bash/ || / sh/ || /zsh/ {exit 1;}' && echo "WARNING ** source setup_environment must be run in a bash, zsh or sh shell; see README" export GPGPUSIM_SETUP_ENVIRONMENT_WAS_RUN= export GPGPUSIM_ROOT="$( cd "$( dirname "$BASH_SOURCE" )" && pwd )" diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index f7f1016..acb376a 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -316,12 +316,12 @@ void warp_inst_t::generate_mem_accesses() break; case global_space: case local_space: case param_space_local: - if( m_config->gpgpu_coalesce_arch == 13 ) { - if(isatomic()) - memory_coalescing_arch_13_atomic(is_write, access_type); - else - memory_coalescing_arch_13(is_write, access_type); - } else abort(); + if( m_config->gpgpu_coalesce_arch >= 13 && m_config->gpgpu_coalesce_arch <= 62) { + if(isatomic()) + memory_coalescing_arch_atomic(is_write, access_type); + else + memory_coalescing_arch(is_write, access_type); + } else abort(); break; @@ -345,7 +345,7 @@ void warp_inst_t::generate_mem_accesses() byte_mask.set(idx+i); } for( a=accesses.begin(); a != accesses.end(); ++a ) - m_accessq.push_back( mem_access_t(access_type,a->first,cache_block_size,is_write,a->second,byte_mask) ); + m_accessq.push_back( mem_access_t(access_type,a->first,cache_block_size,is_write,a->second, byte_mask, mem_access_sector_mask_t())); } if ( space.get_type() == global_space ) { @@ -354,15 +354,32 @@ void warp_inst_t::generate_mem_accesses() m_mem_accesses_created=true; } -void warp_inst_t::memory_coalescing_arch_13( bool is_write, mem_access_type access_type ) +void warp_inst_t::memory_coalescing_arch( bool is_write, mem_access_type access_type ) { // see the CUDA manual where it discusses coalescing rules before reading this unsigned segment_size = 0; unsigned warp_parts = m_config->mem_warp_parts; + bool sector_segment_size = false; + + if(m_config->gpgpu_coalesce_arch >= 20 && m_config->gpgpu_coalesce_arch < 39) + { + //Fermi and Kepler, L1 is normal and L2 is sector + if(m_config->gmem_skip_L1D || cache_op == CACHE_GLOBAL) + sector_segment_size = true; + else + sector_segment_size = false; + } + else if(m_config->gpgpu_coalesce_arch >= 40) + { + //Maxwell and Pascal, L1 and L2 are sectors + //all requests should be 32 bytes + sector_segment_size = true; + } + switch( data_size ) { case 1: segment_size = 32; break; - case 2: segment_size = 64; break; - case 4: case 8: case 16: segment_size = 128; break; + case 2: segment_size = sector_segment_size? 32 : 64; break; + case 4: case 8: case 16: segment_size = sector_segment_size? 32 : 128; break; } unsigned subwarp_size = m_config->warp_size / warp_parts; @@ -413,13 +430,13 @@ void warp_inst_t::memory_coalescing_arch_13( bool is_write, mem_access_type acce new_addr_type addr = t->first; const transaction_info &info = t->second; - memory_coalescing_arch_13_reduce_and_send(is_write, access_type, info, addr, segment_size); + memory_coalescing_arch_reduce_and_send(is_write, access_type, info, addr, segment_size); } } } -void warp_inst_t::memory_coalescing_arch_13_atomic( bool is_write, mem_access_type access_type ) +void warp_inst_t::memory_coalescing_arch_atomic( bool is_write, mem_access_type access_type ) { assert(space.get_type() == global_space); // Atomics allowed only for global memory @@ -488,13 +505,13 @@ void warp_inst_t::memory_coalescing_arch_13_atomic( bool is_write, mem_access_ty for(t=transaction_list.begin(); t!=transaction_list.end(); t++) { // For each transaction const transaction_info &info = *t; - memory_coalescing_arch_13_reduce_and_send(is_write, access_type, info, addr, segment_size); + memory_coalescing_arch_reduce_and_send(is_write, access_type, info, addr, segment_size); } } } } -void warp_inst_t::memory_coalescing_arch_13_reduce_and_send( bool is_write, mem_access_type access_type, const transaction_info &info, new_addr_type addr, unsigned segment_size ) +void warp_inst_t::memory_coalescing_arch_reduce_and_send( bool is_write, mem_access_type access_type, const transaction_info &info, new_addr_type addr, unsigned segment_size ) { assert( (addr & (segment_size-1)) == 0 ); @@ -543,7 +560,7 @@ void warp_inst_t::memory_coalescing_arch_13_reduce_and_send( bool is_write, mem_ assert(lower_half_used && upper_half_used); } } - m_accessq.push_back( mem_access_t(access_type,addr,size,is_write,info.active,info.bytes) ); + m_accessq.push_back( mem_access_t(access_type,addr,size,is_write,info.active,info.bytes, info.chunks) ); } void warp_inst_t::completed( unsigned long long cycle ) const @@ -577,6 +594,8 @@ kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info * //Jin: launch latency management m_launch_latency = g_kernel_launch_latency; + + volta_cache_config_set=false; } kernel_info_t::~kernel_info_t() diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 71d3d89..a612bac 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -82,6 +82,7 @@ enum uarch_op_t { ALU_OP=1, SFU_OP, TENSOR_CORE_OP, + DP_OP, ALU_SFU_OP, LOAD_OP, TENSOR_CORE_LOAD_OP, @@ -139,6 +140,7 @@ typedef enum special_operations_t special_ops; // Required to identify for the p enum operation_pipeline_t { UNKOWN_OP, SP__OP, + DP__OP, SFU__OP, TENSOR_CORE__OP, MEM__OP @@ -308,6 +310,8 @@ public: unsigned long long start_cycle; unsigned long long end_cycle; unsigned m_launch_latency; + + mutable bool volta_cache_config_set; }; struct core_config { @@ -347,6 +351,7 @@ struct core_config { unsigned gpgpu_cache_constl1_linesize; unsigned gpgpu_max_insn_issue_per_warp; + bool gmem_skip_L1D; // on = global memory access always skip the L1 cache }; // bounded stack that implements simt reconvergence using pdom mechanism from MICRO'07 paper @@ -398,6 +403,7 @@ protected: #define LOCAL_MEM_SIZE_MAX (8*1024) #define MAX_STREAMING_MULTIPROCESSORS 64 #define MAX_THREAD_PER_SM 2048 +#define MAX_WARP_PER_SM 64 #define TOTAL_LOCAL_MEM_PER_SM (MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX) #define TOTAL_SHARED_MEM (MAX_STREAMING_MULTIPROCESSORS*SHARED_MEM_SIZE_MAX) #define TOTAL_LOCAL_MEM (MAX_STREAMING_MULTIPROCESSORS*MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX) @@ -523,7 +529,14 @@ public: const struct textureReference* get_texref(const std::string &texname) const { std::map<std::string, const struct textureReference*>::const_iterator t=m_NameToTextureRef.find(texname); - assert( t != m_NameToTextureRef.end() ); + if( t == m_NameToTextureRef.end() ) { + // search for :: prefixed names + std::string temp("::" + texname); + t=m_NameToTextureRef.find(temp); + } + + assert(t != m_NameToTextureRef.end()); + return t->second; } const struct cudaArray* get_texarray( const struct textureReference *texref ) const @@ -574,6 +587,7 @@ struct gpgpu_ptx_sim_info int cmem; int gmem; int regs; + unsigned maxthreads; unsigned ptx_version; unsigned sm_target; }; @@ -625,6 +639,9 @@ private: const unsigned MAX_MEMORY_ACCESS_SIZE = 128; typedef std::bitset<MAX_MEMORY_ACCESS_SIZE> mem_access_byte_mask_t; +const unsigned SECTOR_CHUNCK_SIZE = 4; //four sectors +const unsigned SECTOR_SIZE = 32 ; //sector is 32 bytes width +typedef std::bitset<SECTOR_CHUNCK_SIZE> mem_access_sector_mask_t; #define NO_PARTIAL_WRITE (mem_access_byte_mask_t()) #define MEM_ACCESS_TYPE_TUP_DEF \ @@ -660,6 +677,7 @@ enum cache_operator_type { CACHE_ALL, // .ca CACHE_LAST_USE, // .lu CACHE_VOLATILE, // .cv + CACHE_L1, // .nc // loads and stores CACHE_STREAMING, // .cs @@ -689,8 +707,9 @@ public: unsigned size, bool wr, const active_mask_t &active_mask, - const mem_access_byte_mask_t &byte_mask ) - : m_warp_mask(active_mask), m_byte_mask(byte_mask) + const mem_access_byte_mask_t &byte_mask, + const mem_access_sector_mask_t §or_mask) + : m_warp_mask(active_mask), m_byte_mask(byte_mask), m_sector_mask(sector_mask) { init(); m_type = type; @@ -706,6 +725,7 @@ public: bool is_write() const { return m_write; } enum mem_access_type get_type() const { return m_type; } mem_access_byte_mask_t get_byte_mask() const { return m_byte_mask; } + mem_access_sector_mask_t get_sector_mask() const { return m_sector_mask; } void print(FILE *fp) const { @@ -739,6 +759,7 @@ private: mem_access_type m_type; active_mask_t m_warp_mask; mem_access_byte_mask_t m_byte_mask; + mem_access_sector_mask_t m_sector_mask; static unsigned sm_next_access_uid; }; @@ -961,9 +982,9 @@ public: }; void generate_mem_accesses(); - void memory_coalescing_arch_13( bool is_write, mem_access_type access_type ); - void memory_coalescing_arch_13_atomic( bool is_write, mem_access_type access_type ); - void memory_coalescing_arch_13_reduce_and_send( bool is_write, mem_access_type access_type, const transaction_info &info, new_addr_type addr, unsigned segment_size ); + void memory_coalescing_arch( bool is_write, mem_access_type access_type ); + void memory_coalescing_arch_atomic( bool is_write, mem_access_type access_type ); + void memory_coalescing_arch_reduce_and_send( bool is_write, mem_access_type access_type, const transaction_info &info, new_addr_type addr, unsigned segment_size ); void add_callback( unsigned lane_id, void (*function)(const class inst_t*, class ptx_thread_info*), diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index d5c5c3d..87a0b9c 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -62,8 +62,8 @@ addr_t g_debug_pc = 0xBEEF1518; unsigned g_ptx_sim_num_insn = 0; 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 *opcode_latency_int, *opcode_latency_fp, *opcode_latency_dp,*opcode_latency_sfu; +char *opcode_initiation_int, *opcode_initiation_fp, *opcode_initiation_dp,*opcode_initiation_sfu; char *cdp_latency_str; unsigned cdp_latency[5]; @@ -80,6 +80,10 @@ void ptx_opcocde_latency_options (option_parser_t opp) { "Opcode latencies for double precision floating points <ADD,MAX,MUL,MAD,DIV>" "Default 8,8,8,8,335", "8,8,8,8,335"); + option_parser_register(opp, "-ptx_opcode_latency_sfu", OPT_CSTR, &opcode_latency_sfu, + "Opcode latencies for SFU instructions" + "Default 8", + "8"); option_parser_register(opp, "-ptx_opcode_initiation_int", OPT_CSTR, &opcode_initiation_int, "Opcode initiation intervals for integers <ADD,MAX,MUL,MAD,DIV>" "Default 1,1,4,4,32", @@ -92,6 +96,10 @@ 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, "-ptx_opcode_initiation_sfu", OPT_CSTR, &opcode_initiation_sfu, + "Opcode initiation intervals for sfu instructions" + "Default 8", + "8"); option_parser_register(opp, "-cdp_latency", OPT_CSTR, &cdp_latency_str, "CDP API latency <cudaStreamCreateWithFlags, \ cudaGetParameterBufferV2_init_perWarp, cudaGetParameterBufferV2_perKernel, \ @@ -393,6 +401,10 @@ void gpgpu_t::memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t coun char *src_data = (char*)src; for (unsigned n=0; n < count; n ++ ) m_global_mem->write(dst_start_addr+n,1, src_data+n,NULL,NULL); + + // Copy into the performance model. + extern gpgpu_sim* g_the_gpu; + g_the_gpu->perf_memcpy_to_gpu(dst_start_addr, count); if(g_debug_execution >= 3) { printf( " done.\n"); fflush(stdout); @@ -408,6 +420,10 @@ void gpgpu_t::memcpy_from_gpu( void *dst, size_t src_start_addr, size_t count ) unsigned char *dst_data = (unsigned char*)dst; for (unsigned n=0; n < count; n ++ ) m_global_mem->read(src_start_addr+n,1,dst_data+n); + + // Copy into the performance model. + extern gpgpu_sim* g_the_gpu; + g_the_gpu->perf_memcpy_to_gpu(src_start_addr, count); if(g_debug_execution >= 3) { printf( " done.\n"); fflush(stdout); @@ -592,9 +608,11 @@ void ptx_instruction::set_opcode_and_latency() unsigned int_latency[5]; unsigned fp_latency[5]; unsigned dp_latency[5]; + unsigned sfu_latency; unsigned int_init[5]; unsigned fp_init[5]; unsigned dp_init[5]; + unsigned sfu_init; /* * [0] ADD,SUB * [1] MAX,Min @@ -611,7 +629,9 @@ void ptx_instruction::set_opcode_and_latency() sscanf(opcode_latency_dp, "%u,%u,%u,%u,%u", &dp_latency[0],&dp_latency[1],&dp_latency[2], &dp_latency[3],&dp_latency[4]); - sscanf(opcode_initiation_int, "%u,%u,%u,%u,%u,%u, %u", + sscanf(opcode_latency_sfu, "%u", + &sfu_latency); + sscanf(opcode_initiation_int, "%u,%u,%u,%u,%u", &int_init[0],&int_init[1],&int_init[2], &int_init[3],&int_init[4]); sscanf(opcode_initiation_fp, "%u,%u,%u,%u,%u", @@ -620,8 +640,10 @@ 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(opcode_initiation_sfu, "%u", + &sfu_init); sscanf(cdp_latency_str, "%u,%u,%u,%u,%u", - &cdp_latency[0],&cdp_latency[1],&cdp_latency[2], + &cdp_latency[0],&cdp_latency[1],&cdp_latency[2], &cdp_latency[3],&cdp_latency[4]); if(!m_operands.empty()){ @@ -684,6 +706,7 @@ void ptx_instruction::set_opcode_and_latency() case FF64_TYPE: latency = dp_latency[0]; initiation_interval = dp_init[0]; + op = DP_OP; break; case B32_TYPE: case U32_TYPE: @@ -705,6 +728,7 @@ void ptx_instruction::set_opcode_and_latency() case FF64_TYPE: latency = dp_latency[1]; initiation_interval = dp_init[1]; + op = DP_OP; break; case B32_TYPE: case U32_TYPE: @@ -721,13 +745,12 @@ void ptx_instruction::set_opcode_and_latency() case F32_TYPE: latency = fp_latency[2]; initiation_interval = fp_init[2]; - op = ALU_SFU_OP; break; case F64_TYPE: case FF64_TYPE: latency = dp_latency[2]; initiation_interval = dp_init[2]; - op = ALU_SFU_OP; + op = DP_OP; break; case B32_TYPE: case U32_TYPE: @@ -750,6 +773,7 @@ void ptx_instruction::set_opcode_and_latency() case FF64_TYPE: latency = dp_latency[3]; initiation_interval = dp_init[3]; + op = DP_OP; break; case B32_TYPE: case U32_TYPE: @@ -785,8 +809,8 @@ void ptx_instruction::set_opcode_and_latency() break; case SQRT_OP: case SIN_OP: case COS_OP: case EX2_OP: case LG2_OP: case RSQRT_OP: case RCP_OP: //Using double to approximate those - latency = dp_latency[2]; - initiation_interval = dp_init[2]; + latency = sfu_latency; + initiation_interval = sfu_init; op = SFU_OP; break; case MMA_OP: @@ -796,7 +820,7 @@ void ptx_instruction::set_opcode_and_latency() break; case SHFL_OP: latency = 32; - initiation_interval = 15; + initiation_interval = 4; break; default: break; @@ -889,6 +913,7 @@ void ptx_instruction::pre_decode() switch( m_cache_option ) { case CA_OPTION: cache_op = CACHE_ALL; break; + case NC_OPTION: cache_op = CACHE_L1; break; case CG_OPTION: cache_op = CACHE_GLOBAL; break; case CS_OPTION: cache_op = CACHE_STREAMING; break; case LU_OPTION: cache_op = CACHE_LAST_USE; break; @@ -1169,13 +1194,13 @@ void function_info::finalize( memory_space *param_mem ) // 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 + //param_address = (param_address + size - 1) / size * size; //aligned with size TODO: align not correct 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); + //assert(offset == param_address); param->set_address(param_address); param_address += size; } diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 4614f25..973eeab 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -33,6 +33,7 @@ #include "ptx.tab.h" #include <stdlib.h> #include <math.h> +#include <cmath> #include <fenv.h> #include "cuda-math.h" #include "../abstract_hardware_model.h" diff --git a/src/cuda-sim/ptx.y b/src/cuda-sim/ptx.y index 8b901a3..45392fb 100644 --- a/src/cuda-sim/ptx.y +++ b/src/cuda-sim/ptx.y @@ -237,7 +237,8 @@ function_defn: function_decl { set_symtab($1); func_header(".skip"); } statement block_spec: MAXNTID_DIRECTIVE INT_OPERAND COMMA INT_OPERAND COMMA INT_OPERAND {func_header_info_int(".maxntid", $2); func_header_info_int(",", $4); - func_header_info_int(",", $6); } + func_header_info_int(",", $6); + maxnt_id($2, $4, $6);} | MINNCTAPERSM_DIRECTIVE INT_OPERAND { func_header_info_int(".minnctapersm", $2); printf("GPGPU-Sim: Warning: .minnctapersm ignored. \n"); } | MAXNCTAPERSM_DIRECTIVE INT_OPERAND { func_header_info_int(".maxnctapersm", $2); printf("GPGPU-Sim: Warning: .maxnctapersm ignored. \n"); } ; diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index d12c741..1813f8c 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -222,6 +222,7 @@ bool symbol_table::add_function_decl( const char *name, int entry_point, functio } else { *func_info = new function_info(entry_point); (*func_info)->set_name(name); + (*func_info)->set_maxnt_id(0); m_function_info_lookup[key] = *func_info; } @@ -1234,16 +1235,16 @@ ptx_instruction::ptx_instruction( int opcode, m_inst_size = 4; // bytes break; case EXTP_OPTION: - break; + break; case NC_OPTION: - break; + m_cache_option = last_ptx_inst_option; + break; case UP_OPTION: case DOWN_OPTION: case BFLY_OPTION: case IDX_OPTION: m_shfl_op = last_ptx_inst_option; break; - case PRMT_F4E_MODE: case PRMT_B4E_MODE: case PRMT_RC8_MODE: diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 0767379..4c36cfc 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -1333,6 +1333,7 @@ public: const struct gpgpu_ptx_sim_info* get_kernel_info () const { + assert (m_kernel_info.maxthreads == maxnt_id); return &m_kernel_info; } @@ -1340,6 +1341,8 @@ public: m_kernel_info = info; m_kernel_info.ptx_version = 10*get_ptx_version().ver(); m_kernel_info.sm_target = get_ptx_version().target(); + // THIS DEPENDS ON ptxas being called after the PTX is parsed. + m_kernel_info.maxthreads = maxnt_id; } symbol_table *get_symtab() { @@ -1363,7 +1366,11 @@ public: } bool is_entry_point() const { return m_entry_point; } + void set_maxnt_id(unsigned maxthreads) { maxnt_id = maxthreads;} + unsigned get_maxnt_id() { return maxnt_id;} + private: + unsigned maxnt_id; unsigned m_uid; unsigned m_local_mem_framesize; bool m_entry_point; diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc index 9671ab7..cf40365 100644 --- a/src/cuda-sim/ptx_parser.cc +++ b/src/cuda-sim/ptx_parser.cc @@ -1012,6 +1012,10 @@ void target_header3(char* a, char* b, char* c) g_global_symbol_table->set_sm_target(a,b,c); } +void maxnt_id(int x, int y, int z) { + g_func_info->set_maxnt_id(x * y * z); +} + void func_header(const char* a) {} //intentional dummy function void func_header_info(const char* a) {} //intentional dummy function void func_header_info_int(const char* a, int b) {} //intentional dummy function diff --git a/src/cuda-sim/ptx_parser.h b/src/cuda-sim/ptx_parser.h index 8094b43..7b6e3a2 100644 --- a/src/cuda-sim/ptx_parser.h +++ b/src/cuda-sim/ptx_parser.h @@ -95,6 +95,7 @@ void change_double_operand_type( int addr_type ); void change_operand_neg( ); void set_immediate_operand_type( ); void version_header(double a); +void maxnt_id(int x, int y, int z); //Jin: handle instructino group for cdp void start_inst_group(); diff --git a/src/gpgpu-sim/addrdec.cc b/src/gpgpu-sim/addrdec.cc index 422576d..8651869 100644 --- a/src/gpgpu-sim/addrdec.cc +++ b/src/gpgpu-sim/addrdec.cc @@ -62,6 +62,9 @@ void linear_to_raw_address_translation::addrdec_setoption(option_parser_t opp) option_parser_register(opp, "-gpgpu_mem_address_mask", OPT_INT32, &gpgpu_mem_address_mask, "0 = old addressing mask, 1 = new addressing mask, 2 = new add. mask + flipped bank sel and chip sel bits", "0"); + option_parser_register(opp, "-memory_partition_indexing", OPT_UINT32, &memory_partition_indexing, + "0 = no indexing, 1 = bitwise xoring, 2 = IPoly, 3 = custom indexing", + "0"); } new_addr_type linear_to_raw_address_translation::partition_address( new_addr_type addr ) const @@ -103,6 +106,74 @@ void linear_to_raw_address_translation::addrdec_tlx(new_addr_type addr, addrdec_ tlx->burst= addrdec_packbits(addrdec_mask[BURST], rest_of_addr, addrdec_mkhigh[BURST], addrdec_mklow[BURST]); } + switch(memory_partition_indexing){ + case CONSECUTIVE: + //Do nothing + break; + case BITWISE_PERMUTATION: + { + assert(!gap); + tlx->chip = (tlx->chip) ^ (tlx->row & (m_n_channel-1)); + assert(tlx->chip < m_n_channel); + break; + } + case IPOLY: + { + /* + * Set Indexing function from "Pseudo-randomly interleaved memory." + * Rau, B. R et al. + * ISCA 1991 + * + * equations are adopted from: + * "Sacat: streaming-aware conflict-avoiding thrashing-resistant gpgpu cache management scheme." + * Khairy et al. + * IEEE TPDS 2017. + */ + if(m_n_channel == 32) { + std::bitset<64> a(tlx->row); + std::bitset<5> chip(tlx->chip); + chip[0] = a[13]^a[12]^a[11]^a[10]^a[9]^a[6]^a[5]^a[3]^a[0]^chip[0]; + chip[1] = a[14]^a[13]^a[12]^a[11]^a[10]^a[7]^a[6]^a[4]^a[1]^chip[1]; + chip[2] = a[14]^a[10]^a[9]^a[8]^a[7]^a[6]^a[3]^a[2]^a[0]^chip[2]; + chip[3] = a[11]^a[10]^a[9]^a[8]^a[7]^a[4]^a[3]^a[1]^chip[3]; + chip[4] = a[12]^a[11]^a[10]^a[9]^a[8]^a[5]^a[4]^a[2]^chip[4]; + tlx->chip = chip.to_ulong(); + + } + else{ /* Else incorrect number of channels for the hashing function */ + assert("\nGPGPU-Sim memory_partition_indexing error: The number of channels should be " + "32 for the hashing IPOLY index function.\n" && 0); + } + assert(tlx->chip < m_n_channel); + break; + } + case PAE: + { + //Page Address Entropy + //random selected bits from the page and bank bits + //similar to + //Liu, Yuxi, et al. "Get Out of the Valley: Power-Efficient Address Mapping for GPUs." ISCA 2018 + std::bitset<64> a(tlx->row); + std::bitset<5> chip(tlx->chip); + std::bitset<4> b(tlx->bk); + chip[0] = a[13]^a[10]^a[9]^a[5]^a[0]^b[3]^b[0]^chip[0]; + chip[1] = a[12]^a[11]^a[6]^a[1]^b[3]^b[2]^b[1]^chip[1]; + chip[2] = a[14]^a[9]^a[8]^a[7]^a[2]^b[1]^chip[2]; + chip[3] = a[11]^a[10]^a[8]^a[3]^b[2]^b[3]^chip[3]; + chip[4] = a[12]^a[9]^a[8]^a[5]^a[4]^b[1]^b[0]^chip[4]; + tlx->chip = chip.to_ulong(); + assert(tlx->chip < m_n_channel); + break; + } + case CUSTOM: + /* No custom set function implemented */ + //Do you custom index here + break; + default: + assert("\nUndefined set index function.\n" && 0); + break; + } + // combine the chip address and the lower bits of DRAM bank address to form the subpartition ID unsigned sub_partition_addr_mask = m_n_sub_partition_in_channel - 1; tlx->sub_partition = tlx->chip * m_n_sub_partition_in_channel diff --git a/src/gpgpu-sim/addrdec.h b/src/gpgpu-sim/addrdec.h index fd9af8d..bdc5fec 100644 --- a/src/gpgpu-sim/addrdec.h +++ b/src/gpgpu-sim/addrdec.h @@ -35,6 +35,14 @@ #include "../abstract_hardware_model.h" +enum partition_index_function{ + CONSECUTIVE = 0, + BITWISE_PERMUTATION, + IPOLY, + PAE, + CUSTOM +}; + struct addrdec_t { void print( FILE *fp ) const; @@ -72,6 +80,7 @@ private: const char *addrdec_option; int gpgpu_mem_address_mask; + partition_index_function memory_partition_indexing; bool run_test; int ADDR_CHIP_S; diff --git a/src/gpgpu-sim/delayqueue.h b/src/gpgpu-sim/delayqueue.h index b25f143..0caa5d4 100644 --- a/src/gpgpu-sim/delayqueue.h +++ b/src/gpgpu-sim/delayqueue.h @@ -161,6 +161,7 @@ public: } bool full() const { return (m_max_len && m_length >= m_max_len); } + bool is_avilable_size(unsigned size) const { return (m_max_len && m_length+size-1 >= m_max_len); } bool empty() const { return m_head == NULL; } unsigned get_n_element() const { return m_n_element; } unsigned get_length() const { return m_length; } diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index a0e024b..6c11b43 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -49,11 +49,45 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m m_stats = stats; m_config = config; + //rowblp + access_num=0; + hits_num=0; + read_num=0; + write_num=0; + hits_read_num=0; + hits_write_num=0; + banks_1time=0; + banks_acess_total=0; + banks_acess_total_after=0; + banks_time_ready=0; + banks_access_ready_total=0; + issued_two=0; + issued_total=0; + issued_total_row=0; + issued_total_col=0; + CCDc = 0; RRDc = 0; RTWc = 0; WTRc = 0; + wasted_bw_row=0; + wasted_bw_col=0; + util_bw=0; + idle_bw=0; + RCDc_limit=0; + CCDLc_limit=0; + CCDLc_limit_alone=0; + CCDc_limit=0; + WTRc_limit=0; + WTRc_limit_alone=0; + RCDWRc_limit=0; + RTWc_limit=0; + RTWc_limit_alone=0; + rwq_limit=0; + write_to_read_ratio_blp_rw_average=0; + bkgrp_parallsim_rw=0; + rw = READ; //read mode is default bkgrp = (bankgrp_t**) calloc(sizeof(bankgrp_t*), m_config->nbkgrp); @@ -74,12 +108,13 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m bk[i]->state = BANK_IDLE; bk[i]->bkgrpindex = i/(m_config->nbk/m_config->nbkgrp); } - prio = 0; + prio = 0; + rwq = new fifo_pipeline<dram_req_t>("rwq",m_config->CL,m_config->CL+1); mrqq = new fifo_pipeline<dram_req_t>("mrqq",0,2); returnq = new fifo_pipeline<mem_fetch>("dramreturnq",0,m_config->gpgpu_dram_return_queue_size==0?1024:m_config->gpgpu_dram_return_queue_size); m_frfcfs_scheduler = NULL; - if ( m_config->scheduler_type == DRAM_FRFCFS ) + if ( m_config->scheduler_type == DRAM_FRFCFS) m_frfcfs_scheduler = new frfcfs_scheduler(m_config,this,stats); n_cmd = 0; n_activity = 0; @@ -88,6 +123,8 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m n_pre = 0; n_rd = 0; n_wr = 0; + n_wr_WB=0; + n_rd_L2_A=0; n_req = 0; max_mrqs_temp = 0; bwutil = 0; @@ -113,13 +150,21 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m mrqq_Dist = StatCreate("mrqq_length",1, queue_limit()); else //queue length is unlimited; mrqq_Dist = StatCreate("mrqq_length",1,64); //track up to 64 entries + } -bool dram_t::full() const +bool dram_t::full(bool is_write) const { - if(m_config->scheduler_type == DRAM_FRFCFS ){ + if(m_config->scheduler_type == DRAM_FRFCFS){ if(m_config->gpgpu_frfcfs_dram_sched_queue_size == 0 ) return false; - return m_frfcfs_scheduler->num_pending() >= m_config->gpgpu_frfcfs_dram_sched_queue_size; + if(m_config->seperate_write_queue_enabled){ + if(is_write) + return m_frfcfs_scheduler->num_write_pending() >= m_config->gpgpu_frfcfs_dram_write_queue_size; + else + return m_frfcfs_scheduler->num_pending() >= m_config->gpgpu_frfcfs_dram_sched_queue_size; + } + else + return m_frfcfs_scheduler->num_pending() >= m_config->gpgpu_frfcfs_dram_sched_queue_size; } else return mrqq->full(); } @@ -127,7 +172,7 @@ bool dram_t::full() const unsigned dram_t::que_length() const { unsigned nreqs = 0; - if (m_config->scheduler_type == DRAM_FRFCFS ) { + if (m_config->scheduler_type == DRAM_FRFCFS) { nreqs = m_frfcfs_scheduler->num_pending(); } else { nreqs = mrqq->get_length(); @@ -146,7 +191,7 @@ unsigned int dram_t::queue_limit() const } -dram_req_t::dram_req_t( class mem_fetch *mf ) +dram_req_t::dram_req_t( class mem_fetch *mf, unsigned banks, unsigned dram_bnk_indexing_policy) { txbytes = 0; dqbytes = 0; @@ -154,7 +199,29 @@ dram_req_t::dram_req_t( class mem_fetch *mf ) const addrdec_t &tlx = mf->get_tlx_addr(); - bk = tlx.bk; + switch(dram_bnk_indexing_policy){ + case LINEAR_BK_INDEX: + { + bk = tlx.bk; + break; + } + case BITWISE_XORING_BK_INDEX: + { + //xoring bank bits with lower bits of the page + int lbank = log2(banks); + bk = tlx.bk ^ (tlx.row & ((1<<lbank)-1)); + break; + } + case CUSTOM_BK_INDEX: + /* No custom set function implemented */ + //Do you custom index here + break; + default: + assert("\nUndefined bank index function.\n" && 0); + break; + } + + row = tlx.row; col = tlx.col; nbytes = mf->get_data_size(); @@ -169,14 +236,15 @@ void dram_t::push( class mem_fetch *data ) { assert(id == data->get_tlx_addr().chip); // Ensure request is in correct memory partition - dram_req_t *mrq = new dram_req_t(data); + dram_req_t *mrq = new dram_req_t(data,m_config->nbk,m_config->dram_bnk_indexing_policy); + data->set_status(IN_PARTITION_MC_INTERFACE_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); - mrqq->push(mrq); + mrqq->push(mrq); // stats... n_req += 1; n_req_partial += 1; - if ( m_config->scheduler_type == DRAM_FRFCFS ) { + if ( m_config->scheduler_type == DRAM_FRFCFS) { unsigned nreqs = m_frfcfs_scheduler->num_pending(); if ( nreqs > max_mrqs_temp) max_mrqs_temp = nreqs; @@ -212,6 +280,7 @@ void dram_t::cycle() printf("\tDQ: BK%d Row:%03x Col:%03x", cmd->bk, cmd->row, cmd->col + cmd->dqbytes); #endif cmd->dqbytes += m_config->dram_atom_size; + if (cmd->dqbytes >= cmd->nbytes) { mem_fetch *data = cmd->data; data->set_status(IN_PARTITION_MC_RETURNQ,gpu_sim_cycle+gpu_tot_sim_cycle); @@ -240,7 +309,7 @@ void dram_t::cycle() printf("Error: Unknown DRAM scheduler type\n"); assert(0); } - if ( m_config->scheduler_type == DRAM_FRFCFS ) { + if ( m_config->scheduler_type == DRAM_FRFCFS) { unsigned nreqs = m_frfcfs_scheduler->num_pending(); if ( nreqs > max_mrqs) { max_mrqs = nreqs; @@ -258,130 +327,123 @@ void dram_t::cycle() unsigned k=m_config->nbk; bool issued = false; - // check if any bank is ready to issue a new read + //collect row buffer locality, BLP and other statistics + ///////////////////////////////////////////////////////////////////////// + unsigned int memory_pending=0; for (unsigned i=0;i<m_config->nbk;i++) { - unsigned j = (i + prio) % m_config->nbk; - unsigned grp = j>>m_config->bk_tag_length; - if (bk[j]->mrq) { //if currently servicing a memory request - bk[j]->mrq->data->set_status(IN_PARTITION_DRAM,gpu_sim_cycle+gpu_tot_sim_cycle); - // correct row activated for a READ - if ( !issued && !CCDc && !bk[j]->RCDc && - !(bkgrp[grp]->CCDLc) && - (bk[j]->curr_row == bk[j]->mrq->row) && - (bk[j]->mrq->rw == READ) && (WTRc == 0 ) && - (bk[j]->state == BANK_ACTIVE) && - !rwq->full() ) { - if (rw==WRITE) { - rw=READ; - rwq->set_min_length(m_config->CL); - } - rwq->push(bk[j]->mrq); - bk[j]->mrq->txbytes += m_config->dram_atom_size; - CCDc = m_config->tCCD; - bkgrp[grp]->CCDLc = m_config->tCCDL; - RTWc = m_config->tRTW; - bk[j]->RTPc = m_config->BL/m_config->data_command_freq_ratio; - bkgrp[grp]->RTPLc = m_config->tRTPL; - issued = true; - n_rd++; - bwutil += m_config->BL/m_config->data_command_freq_ratio; - bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; - bk[j]->n_access++; -#ifdef DRAM_VERIFY - PRINT_CYCLE=1; - printf("\tRD Bk:%d Row:%03x Col:%03x \n", - j, bk[j]->curr_row, - bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); -#endif - // transfer done - if ( !(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes) ) { - bk[j]->mrq = NULL; - } - } else - // correct row activated for a WRITE - if ( !issued && !CCDc && !bk[j]->RCDWRc && - !(bkgrp[grp]->CCDLc) && - (bk[j]->curr_row == bk[j]->mrq->row) && - (bk[j]->mrq->rw == WRITE) && (RTWc == 0 ) && - (bk[j]->state == BANK_ACTIVE) && - !rwq->full() ) { - if (rw==READ) { - rw=WRITE; - rwq->set_min_length(m_config->WL); - } - rwq->push(bk[j]->mrq); + if (bk[i]->mrq) + memory_pending++; + } + banks_1time += memory_pending; + if(memory_pending >0) + banks_acess_total++; - bk[j]->mrq->txbytes += m_config->dram_atom_size; - CCDc = m_config->tCCD; - bkgrp[grp]->CCDLc = m_config->tCCDL; - WTRc = m_config->tWTR; - bk[j]->WTPc = m_config->tWTP; - issued = true; - n_wr++; - bwutil += m_config->BL/m_config->data_command_freq_ratio; - bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; -#ifdef DRAM_VERIFY - PRINT_CYCLE=1; - printf("\tWR Bk:%d Row:%03x Col:%03x \n", - j, bk[j]->curr_row, - bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); -#endif - // transfer done - if ( !(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes) ) { - bk[j]->mrq = NULL; - } - } + unsigned int memory_pending_rw=0; + unsigned read_blp_rw=0; + unsigned write_blp_rw=0; + std::bitset<8> bnkgrp_rw_found; //assume max we have 8 bank groups - else - // bank is idle - if ( !issued && !RRDc && - (bk[j]->state == BANK_IDLE) && - !bk[j]->RPc && !bk[j]->RCc ) { -#ifdef DRAM_VERIFY - PRINT_CYCLE=1; - printf("\tACT BK:%d NewRow:%03x From:%03x \n", - j,bk[j]->mrq->row,bk[j]->curr_row); -#endif - // activate the row with current memory request - bk[j]->curr_row = bk[j]->mrq->row; - bk[j]->state = BANK_ACTIVE; - RRDc = m_config->tRRD; - bk[j]->RCDc = m_config->tRCD; - bk[j]->RCDWRc = m_config->tRCDWR; - bk[j]->RASc = m_config->tRAS; - bk[j]->RCc = m_config->tRC; - prio = (j + 1) % m_config->nbk; - issued = true; - n_act_partial++; - n_act++; - } + for (unsigned j=0;j<m_config->nbk;j++) { + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq && (((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && + (bk[j]->state == BANK_ACTIVE)))) + { + memory_pending_rw++; + read_blp_rw++; + bnkgrp_rw_found.set(grp); + } + else if + (bk[j]->mrq && (((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && + (bk[j]->state == BANK_ACTIVE)))) + { + memory_pending_rw++; + write_blp_rw++; + bnkgrp_rw_found.set(grp); + } + } + banks_time_rw += memory_pending_rw; + bkgrp_parallsim_rw += bnkgrp_rw_found.count(); + if(memory_pending_rw >0) + { + write_to_read_ratio_blp_rw_average += (double)write_blp_rw/(write_blp_rw+read_blp_rw); + banks_access_rw_total++; + } - else - // different row activated - if ( (!issued) && - (bk[j]->curr_row != bk[j]->mrq->row) && - (bk[j]->state == BANK_ACTIVE) && - (!bk[j]->RASc && !bk[j]->WTPc && - !bk[j]->RTPc && - !bkgrp[grp]->RTPLc) ) { - // make the bank idle again - bk[j]->state = BANK_IDLE; - bk[j]->RPc = m_config->tRP; - prio = (j + 1) % m_config->nbk; - issued = true; - n_pre++; - n_pre_partial++; -#ifdef DRAM_VERIFY - PRINT_CYCLE=1; - printf("\tPRE BK:%d Row:%03x \n", j,bk[j]->curr_row); -#endif - } - } else { - if (!CCDc && !RRDc && !RTWc && !WTRc && !bk[j]->RCDc && !bk[j]->RASc - && !bk[j]->RCc && !bk[j]->RPc && !bk[j]->RCDWRc) k--; - bk[j]->n_idle++; - } + unsigned int memory_Pending_ready=0; + for (unsigned j=0;j<m_config->nbk;j++) { + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq && ((!CCDc && !bk[j]->RCDc && + !(bkgrp[grp]->CCDLc) && + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && (WTRc == 0 ) && + (bk[j]->state == BANK_ACTIVE) && + !rwq->full()) + || + (!CCDc && !bk[j]->RCDWRc && + !(bkgrp[grp]->CCDLc) && + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && (RTWc == 0 ) && + (bk[j]->state == BANK_ACTIVE) && + !rwq->full()))) + { + memory_Pending_ready++; + } } + banks_time_ready += memory_Pending_ready; + if(memory_Pending_ready >0) + banks_access_ready_total++; + /////////////////////////////////////////////////////////////////////////////////// + + bool issued_col_cmd = false; + bool issued_row_cmd = false; + + if(m_config->dual_bus_interface) + { + //dual bus interface + //issue one row command and one column command + for (unsigned i=0;i<m_config->nbk;i++) { + unsigned j = (i + prio) % m_config->nbk; + issued_col_cmd = issue_col_command(j); + if(issued_col_cmd) break; + } + for (unsigned i=0;i<m_config->nbk;i++) { + unsigned j = (i + prio) % m_config->nbk; + issued_row_cmd = issue_row_command(j); + if(issued_row_cmd) break; + } + for (unsigned i=0;i<m_config->nbk;i++) { + unsigned j = (i + prio) % m_config->nbk; + if(!bk[j]->mrq) { + if (!CCDc && !RRDc && !RTWc && !WTRc && !bk[j]->RCDc && !bk[j]->RASc + && !bk[j]->RCc && !bk[j]->RPc && !bk[j]->RCDWRc) k--; + bk[j]->n_idle++; + } + } + } + else + { + //single bus interface + //issue only one row/column command + for (unsigned i=0;i<m_config->nbk;i++) { + unsigned j = (i + prio) % m_config->nbk; + if(!issued_col_cmd) + issued_col_cmd = issue_col_command(j); + + if(!issued_col_cmd && !issued_row_cmd) + issued_row_cmd = issue_row_command(j); + + if(!bk[j]->mrq) { + if (!CCDc && !RRDc && !RTWc && !WTRc && !bk[j]->RCDc && !bk[j]->RASc + && !bk[j]->RCc && !bk[j]->RPc && !bk[j]->RCDWRc) k--; + bk[j]->n_idle++; + } + + } + } + + issued = issued_row_cmd || issued_col_cmd; if (!issued) { n_nop++; n_nop_partial++; @@ -395,6 +457,85 @@ void dram_t::cycle() } n_cmd++; n_cmd_partial++; + if(issued) + { + issued_total++; + if(issued_col_cmd && issued_row_cmd) + issued_two++; + } + if(issued_col_cmd) issued_total_col++; + if(issued_row_cmd) issued_total_row++; + + + //Collect some statistics + //check the limitation, see where BW is wasted? + ///////////////////////////////////////////////////////// + unsigned int memory_pending_found=0; + for (unsigned i=0;i<m_config->nbk;i++) { + if (bk[i]->mrq) + memory_pending_found++; + } + if(memory_pending_found>0) + banks_acess_total_after++; + + bool memory_pending_rw_found=false; + for (unsigned j=0;j<m_config->nbk;j++) { + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq && (((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && + (bk[j]->state == BANK_ACTIVE)) + || + ( + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && + (bk[j]->state == BANK_ACTIVE)))) + memory_pending_rw_found=true; + } + + + if(issued_col_cmd || CCDc) + util_bw++; + else if (memory_pending_rw_found) + { + wasted_bw_col++; + for (unsigned j=0;j<m_config->nbk;j++) { + unsigned grp = get_bankgrp_number(j); + //read + if (bk[j]->mrq && (((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && + (bk[j]->state == BANK_ACTIVE)))) + { + if(bk[j]->RCDc) RCDc_limit++; + if(bkgrp[grp]->CCDLc) CCDLc_limit++; + if(WTRc) WTRc_limit++; + if(CCDc) CCDc_limit++; + if(rwq->full()) rwq_limit++; + if(bkgrp[grp]->CCDLc && !WTRc) CCDLc_limit_alone++; + if(!bkgrp[grp]->CCDLc && WTRc) WTRc_limit_alone++; + } + //write + else if (bk[j]->mrq && ((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && + (bk[j]->state == BANK_ACTIVE))) + { + if(bk[j]->RCDWRc) RCDWRc_limit++; + if(bkgrp[grp]->CCDLc) CCDLc_limit++; + if(RTWc) RTWc_limit++; + if(CCDc) CCDc_limit++; + if(rwq->full()) rwq_limit++; + if(bkgrp[grp]->CCDLc && !RTWc) CCDLc_limit_alone++; + if(!bkgrp[grp]->CCDLc && RTWc) RTWc_limit_alone++; + } + } + } + else if (memory_pending_found) + wasted_bw_row++; + else if (!memory_pending_found) + idle_bw++; + else + assert(1); + + ///////////////////////////////////////////////////////// // decrements counters once for each time dram_issueCMD is called DEC2ZERO(RRDc); @@ -420,26 +561,170 @@ void dram_t::cycle() #endif } +bool dram_t::issue_col_command(int j) +{ + bool issued = false; + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq) { //if currently servicing a memory request + bk[j]->mrq->data->set_status(IN_PARTITION_DRAM,gpu_sim_cycle+gpu_tot_sim_cycle); + // correct row activated for a READ + if ( !issued && !CCDc && !bk[j]->RCDc && + !(bkgrp[grp]->CCDLc) && + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && (WTRc == 0 ) && + (bk[j]->state == BANK_ACTIVE) && + !rwq->full() ) { + if (rw==WRITE) { + rw=READ; + rwq->set_min_length(m_config->CL); + } + rwq->push(bk[j]->mrq); + bk[j]->mrq->txbytes += m_config->dram_atom_size; + CCDc = m_config->tCCD; + bkgrp[grp]->CCDLc = m_config->tCCDL; + RTWc = m_config->tRTW; + bk[j]->RTPc = m_config->BL/m_config->data_command_freq_ratio; + bkgrp[grp]->RTPLc = m_config->tRTPL; + issued = true; + if(bk[j]->mrq->data->get_access_type() == L2_WR_ALLOC_R) + n_rd_L2_A++; + else + n_rd++; + + bwutil += m_config->BL/m_config->data_command_freq_ratio; + bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; + bk[j]->n_access++; + +#ifdef DRAM_VERIFY + PRINT_CYCLE=1; + printf("\tRD Bk:%d Row:%03x Col:%03x \n", + j, bk[j]->curr_row, + bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); +#endif + // transfer done + if ( !(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes) ) { + bk[j]->mrq = NULL; + } + } else + // correct row activated for a WRITE + if ( !issued && !CCDc && !bk[j]->RCDWRc && + !(bkgrp[grp]->CCDLc) && + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && (RTWc == 0 ) && + (bk[j]->state == BANK_ACTIVE) && + !rwq->full() ) { + if (rw==READ) { + rw=WRITE; + rwq->set_min_length(m_config->WL); + } + rwq->push(bk[j]->mrq); + + bk[j]->mrq->txbytes += m_config->dram_atom_size; + CCDc = m_config->tCCD; + bkgrp[grp]->CCDLc = m_config->tCCDL; + WTRc = m_config->tWTR; + bk[j]->WTPc = m_config->tWTP; + issued = true; + + if(bk[j]->mrq->data->get_access_type() == L2_WRBK_ACC) + n_wr_WB++; + else + n_wr++; + bwutil += m_config->BL/m_config->data_command_freq_ratio; + bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; +#ifdef DRAM_VERIFY + PRINT_CYCLE=1; + printf("\tWR Bk:%d Row:%03x Col:%03x \n", + j, bk[j]->curr_row, + bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); +#endif + // transfer done + if ( !(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes) ) { + bk[j]->mrq = NULL; + } + } + + } + + return issued; +} + +bool dram_t::issue_row_command(int j) +{ + bool issued = false; + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq) { //if currently servicing a memory request + bk[j]->mrq->data->set_status(IN_PARTITION_DRAM,gpu_sim_cycle+gpu_tot_sim_cycle); + // bank is idle + //else + if ( !issued && !RRDc && + (bk[j]->state == BANK_IDLE) && + !bk[j]->RPc && !bk[j]->RCc) { // +#ifdef DRAM_VERIFY + PRINT_CYCLE=1; + printf("\tACT BK:%d NewRow:%03x From:%03x \n", + j,bk[j]->mrq->row,bk[j]->curr_row); +#endif + // activate the row with current memory request + bk[j]->curr_row = bk[j]->mrq->row; + bk[j]->state = BANK_ACTIVE; + RRDc = m_config->tRRD; + bk[j]->RCDc = m_config->tRCD; + bk[j]->RCDWRc = m_config->tRCDWR; + bk[j]->RASc = m_config->tRAS; + bk[j]->RCc = m_config->tRC; + prio = (j + 1) % m_config->nbk; + issued = true; + n_act_partial++; + n_act++; + } + + else + // different row activated + if ( (!issued) && + (bk[j]->curr_row != bk[j]->mrq->row) && + (bk[j]->state == BANK_ACTIVE) && + (!bk[j]->RASc && !bk[j]->WTPc && + !bk[j]->RTPc && + !bkgrp[grp]->RTPLc) ) { + // make the bank idle again + bk[j]->state = BANK_IDLE; + bk[j]->RPc = m_config->tRP; + prio = (j + 1) % m_config->nbk; + issued = true; + n_pre++; + n_pre_partial++; +#ifdef DRAM_VERIFY + PRINT_CYCLE=1; + printf("\tPRE BK:%d Row:%03x \n", j,bk[j]->curr_row); +#endif + } + } + return issued; +} + + //if mrq is being serviced by dram, gets popped after CL latency fulfilled -class mem_fetch* dram_t::return_queue_pop() +class mem_fetch* dram_t::return_queue_pop() { return returnq->pop(); } -class mem_fetch* dram_t::return_queue_top() +class mem_fetch* dram_t::return_queue_top() { return returnq->top(); } + void dram_t::print( FILE* simFile) const { unsigned i; fprintf(simFile,"DRAM[%d]: %d bks, busW=%d BL=%d CL=%d, ", id, m_config->nbk, m_config->busW, m_config->BL, m_config->CL ); fprintf(simFile,"tRRD=%d tCCD=%d, tRCD=%d tRAS=%d tRP=%d tRC=%d\n", - m_config->tCCD, m_config->tRRD, m_config->tRCD, m_config->tRAS, m_config->tRP, m_config->tRC ); - fprintf(simFile,"n_cmd=%d n_nop=%d n_act=%d n_pre=%d n_req=%d n_rd=%d n_write=%d bw_util=%.4g\n", - n_cmd, n_nop, n_act, n_pre, n_req, n_rd, n_wr, + m_config->tRRD, m_config->tCCD, m_config->tRCD, m_config->tRAS, m_config->tRP, m_config->tRC ); + fprintf(simFile,"n_cmd=%d n_nop=%d n_act=%d n_pre=%d n_ref_event=%d n_req=%d n_rd=%d n_rd_L2_A=%d n_write=%d n_wr_bk=%d bw_util=%.4g\n", + n_cmd, n_nop, n_act, n_pre, n_ref, n_req, n_rd, n_rd_L2_A, n_wr, n_wr_WB, (float)bwutil/n_cmd); fprintf(simFile,"n_activity=%d dram_eff=%.4g\n", n_activity, (float)bwutil/n_activity); @@ -447,12 +732,66 @@ void dram_t::print( FILE* simFile) const fprintf(simFile, "bk%d: %da %di ",i,bk[i]->n_access,bk[i]->n_idle); } fprintf(simFile, "\n"); + fprintf(simFile, "\n------------------------------------------------------------------------\n"); + + printf("\nRow_Buffer_Locality = %.6f", (float)hits_num / access_num); + printf("\nRow_Buffer_Locality_read = %.6f", (float)hits_read_num / read_num); + printf("\nRow_Buffer_Locality_write = %.6f", (float)hits_write_num / write_num); + printf("\nBank_Level_Parallism = %.6f", (float)banks_1time / banks_acess_total); + printf("\nBank_Level_Parallism_Col = %.6f", (float)banks_time_rw / banks_access_rw_total); + printf("\nBank_Level_Parallism_Ready = %.6f", (float)banks_time_ready /banks_access_ready_total); + printf("\nwrite_to_read_ratio_blp_rw_average = %.6f", write_to_read_ratio_blp_rw_average /banks_access_rw_total); + printf("\nGrpLevelPara = %.6f \n", (float)bkgrp_parallsim_rw /banks_access_rw_total); + + printf("\nBW Util details:\n"); + printf("bwutil = %.6f \n", (float)bwutil/n_cmd); + printf("total_CMD = %d \n", n_cmd); + printf("util_bw = %d \n", util_bw); + printf("Wasted_Col = %d \n", wasted_bw_col); + printf("Wasted_Row = %d \n", wasted_bw_row); + printf("Idle = %d \n", idle_bw); + + printf("\nBW Util Bottlenecks: \n"); + printf("RCDc_limit = %d \n", RCDc_limit); + printf("RCDWRc_limit = %d \n", RCDWRc_limit); + printf("WTRc_limit = %d \n", WTRc_limit); + printf("RTWc_limit = %d \n", RTWc_limit); + printf("CCDLc_limit = %d \n", CCDLc_limit); + printf("rwq = %d \n", rwq_limit); + printf("CCDLc_limit_alone = %d \n", CCDLc_limit_alone); + printf("WTRc_limit_alone = %d \n", WTRc_limit_alone); + printf("RTWc_limit_alone = %d \n", RTWc_limit_alone); + + printf("\nCommands details: \n"); + printf("total_CMD = %d \n", n_cmd); + printf("n_nop = %d \n", n_nop); + printf("Read = %d \n", n_rd); + printf("Write = %d \n",n_wr); + printf("L2_Alloc = %d \n", n_rd_L2_A); + printf("L2_WB = %d \n", n_wr_WB); + printf("n_act = %d \n", n_act); + printf("n_pre = %d \n", n_pre); + printf("n_ref = %d \n", n_ref); + printf("n_req = %d \n", n_req ); + printf("total_req = %d \n", n_rd+n_wr+n_rd_L2_A+n_wr_WB); + + printf("\nDual Bus Interface Util: \n"); + printf("issued_total_row = %lu \n", issued_total_row); + printf("issued_total_col = %lu \n", issued_total_col); + printf("Row_Bus_Util = %.6f \n", (float)issued_total_row / n_cmd); + printf("CoL_Bus_Util = %.6f \n", (float)issued_total_col / n_cmd); + printf("Either_Row_CoL_Bus_Util = %.6f \n", (float)issued_total / n_cmd); + printf("Issued_on_Two_Bus_Simul_Util = %.6f \n", (float)issued_two /n_cmd); + printf("issued_two_Eff = %.6f \n", (float)issued_two /issued_total); + printf("queue_avg = %.6f \n\n", (float)ave_mrqs/n_cmd ); + + fprintf(simFile, "\n"); fprintf(simFile, "dram_util_bins:"); for (i=0;i<10;i++) fprintf(simFile, " %d", dram_util_bins[i]); fprintf(simFile, "\ndram_eff_bins:"); for (i=0;i<10;i++) fprintf(simFile, " %d", dram_eff_bins[i]); fprintf(simFile, "\n"); - if(m_config->scheduler_type== DRAM_FRFCFS) + if(m_config->scheduler_type== DRAM_FRFCFS) fprintf(simFile, "mrqq: max=%d avg=%g\n", max_mrqs, (float)ave_mrqs/n_cmd); } @@ -476,8 +815,8 @@ void dram_t::visualize() const void dram_t::print_stat( FILE* simFile ) { - fprintf(simFile,"DRAM (%d): n_cmd=%d n_nop=%d n_act=%d n_pre=%d n_req=%d n_rd=%d n_write=%d bw_util=%.4g ", - id, n_cmd, n_nop, n_act, n_pre, n_req, n_rd, n_wr, + fprintf(simFile,"DRAM (%d): n_cmd=%d n_nop=%d n_act=%d n_pre=%d n_ref=%d n_req=%d n_rd=%d n_write=%d bw_util=%.4g ", + id, n_cmd, n_nop, n_act, n_pre, n_ref, n_req, n_rd, n_wr, (float)bwutil/n_cmd); fprintf(simFile, "mrqq: %d %.4g mrqsmax=%d ", max_mrqs, (float)ave_mrqs/n_cmd, max_mrqs_temp); fprintf(simFile, "\n"); @@ -516,6 +855,7 @@ void dram_t::visualizer_print( gzFile visualizer_file ) n_pre_partial = 0; n_req_partial = 0; + // dram access type classification for (unsigned j = 0; j < m_config->nbk; j++) { gzprintf(visualizer_file,"dramglobal_acc_r: %u %u %u\n", id, j, @@ -553,3 +893,16 @@ void dram_t::set_dram_power_stats( unsigned &cmd, wr = n_wr; req = n_req; } + +unsigned dram_t::get_bankgrp_number(unsigned i) +{ + if(m_config->dram_bnkgrp_indexing_policy == HIGHER_BITS) { //higher bits + return i>>m_config->bk_tag_length; + } + else if (m_config->dram_bnkgrp_indexing_policy == LOWER_BITS) { //lower bits + return i&((m_config->nbkgrp-1)); + } + else { + assert(1); + } +} diff --git a/src/gpgpu-sim/dram.h b/src/gpgpu-sim/dram.h index 15c63e7..bee5b7b 100644 --- a/src/gpgpu-sim/dram.h +++ b/src/gpgpu-sim/dram.h @@ -31,9 +31,15 @@ #include "delayqueue.h" #include <set> +#include <vector> +#include <bitset> +#include <sstream> +#include <string> +#include <fstream> #include <zlib.h> #include <stdio.h> #include <stdlib.h> +#include<iomanip> #define READ 'R' //define read and write states #define WRITE 'W' @@ -42,7 +48,7 @@ class dram_req_t { public: - dram_req_t( class mem_fetch *data ); + dram_req_t( class mem_fetch *data , unsigned banks, unsigned dram_bnk_indexing_policy); unsigned int row; unsigned int col; @@ -87,6 +93,17 @@ struct bank_t unsigned int bkgrpindex; }; +enum bank_index_function{ + LINEAR_BK_INDEX = 0, + BITWISE_XORING_BK_INDEX, + CUSTOM_BK_INDEX +}; + +enum bank_grp_bits_position{ + HIGHER_BITS = 0, + LOWER_BITS +}; + class mem_fetch; class dram_t @@ -95,7 +112,7 @@ public: dram_t( unsigned int parition_id, const struct memory_config *config, class memory_stats_t *stats, class memory_partition_unit *mp ); - bool full() const; + bool full(bool is_write) const; void print( FILE* simFile ) const; void visualize() const; void print_stat( FILE* simFile ); @@ -106,6 +123,7 @@ public: class mem_fetch* return_queue_pop(); class mem_fetch* return_queue_top(); + void push( class mem_fetch *data ); void cycle(); void dram_log (int task); @@ -123,17 +141,24 @@ public: unsigned &wr, unsigned &req) const; -private: - void scheduler_fifo(); - void scheduler_frfcfs(); + const struct memory_config *m_config; +private: bankgrp_t **bkgrp; bank_t **bk; unsigned int prio; + unsigned get_bankgrp_number(unsigned i); + + void scheduler_fifo(); + void scheduler_frfcfs(); + + bool issue_col_command(int j); + bool issue_row_command(int j); + unsigned int RRDc; unsigned int CCDc; unsigned int RTWc; //read to write penalty applies across banks @@ -146,7 +171,7 @@ private: fifo_pipeline<dram_req_t> *rwq; fifo_pipeline<dram_req_t> *mrqq; //buffer to hold packets when DRAM processing is over - //should be filled with dram clock and popped with l2or icnt clock + //should be filled with dram clock and popped with l2or icnt clock fifo_pipeline<mem_fetch> *returnq; unsigned int dram_util_bins[10]; @@ -158,11 +183,51 @@ private: unsigned int n_nop; unsigned int n_act; unsigned int n_pre; + unsigned int n_ref; unsigned int n_rd; + unsigned int n_rd_L2_A; unsigned int n_wr; + unsigned int n_wr_WB; unsigned int n_req; unsigned int max_mrqs_temp; + //some statistics to collect to see where BW is wasted? + unsigned wasted_bw_row; + unsigned wasted_bw_col; + unsigned util_bw; + unsigned idle_bw; + unsigned RCDc_limit; + unsigned CCDLc_limit; + unsigned CCDLc_limit_alone; + unsigned CCDc_limit; + unsigned WTRc_limit; + unsigned WTRc_limit_alone; + unsigned RCDWRc_limit; + unsigned RTWc_limit; + unsigned RTWc_limit_alone; + unsigned rwq_limit; + + //row locality, BLP and other statistics + unsigned long access_num; + unsigned long read_num; + unsigned long write_num; + unsigned long long hits_num; + unsigned long long hits_read_num; + unsigned long long hits_write_num; + unsigned long long banks_1time; + unsigned long long banks_acess_total; + unsigned long long banks_acess_total_after; + unsigned long long banks_time_rw; + unsigned long long banks_access_rw_total; + unsigned long long banks_time_ready; + unsigned long long banks_access_ready_total; + unsigned long long issued_two; + unsigned long long issued_total; + unsigned long long issued_total_row; + unsigned long long issued_total_col; + double write_to_read_ratio_blp_rw_average; + unsigned long long bkgrp_parallsim_rw; + unsigned int bwutil; unsigned int max_mrqs; unsigned int ave_mrqs; diff --git a/src/gpgpu-sim/dram_sched.cc b/src/gpgpu-sim/dram_sched.cc index 8303e86..ff50050 100644 --- a/src/gpgpu-sim/dram_sched.cc +++ b/src/gpgpu-sim/dram_sched.cc @@ -36,6 +36,7 @@ frfcfs_scheduler::frfcfs_scheduler( const memory_config *config, dram_t *dm, mem m_config = config; m_stats = stats; m_num_pending = 0; + m_num_write_pending = 0; m_dram = dm; m_queue = new std::list<dram_req_t*>[m_config->nbk]; m_bins = new std::map<unsigned,std::list<std::list<dram_req_t*>::iterator> >[ m_config->nbk ]; @@ -49,15 +50,36 @@ frfcfs_scheduler::frfcfs_scheduler( const memory_config *config, dram_t *dm, mem curr_row_service_time[i] = 0; row_service_timestamp[i] = 0; } + if(m_config->seperate_write_queue_enabled) { + m_write_queue = new std::list<dram_req_t*>[m_config->nbk]; + m_write_bins = new std::map<unsigned,std::list<std::list<dram_req_t*>::iterator> >[ m_config->nbk ]; + m_last_write_row = new std::list<std::list<dram_req_t*>::iterator>*[ m_config->nbk ]; + + for ( unsigned i=0; i < m_config->nbk; i++ ) { + m_write_queue[i].clear(); + m_write_bins[i].clear(); + m_last_write_row[i] = NULL; + } + } + m_mode = READ_MODE; } void frfcfs_scheduler::add_req( dram_req_t *req ) { - m_num_pending++; - m_queue[req->bk].push_front(req); - std::list<dram_req_t*>::iterator ptr = m_queue[req->bk].begin(); - m_bins[req->bk][req->row].push_front( ptr ); //newest reqs to the front + if(m_config->seperate_write_queue_enabled && req->data->is_write()) { + assert(m_num_write_pending < m_config->gpgpu_frfcfs_dram_write_queue_size); + m_num_write_pending++; + m_write_queue[req->bk].push_front(req); + std::list<dram_req_t*>::iterator ptr = m_write_queue[req->bk].begin(); + m_write_bins[req->bk][req->row].push_front( ptr ); //newest reqs to the front + } else { + assert(m_num_pending < m_config->gpgpu_frfcfs_dram_sched_queue_size); + m_num_pending++; + m_queue[req->bk].push_front(req); + std::list<dram_req_t*>::iterator ptr = m_queue[req->bk].begin(); + m_bins[req->bk][req->row].push_front( ptr ); //newest reqs to the front + } } void frfcfs_scheduler::data_collection(unsigned int bank) @@ -78,41 +100,92 @@ void frfcfs_scheduler::data_collection(unsigned int bank) dram_req_t *frfcfs_scheduler::schedule( unsigned bank, unsigned curr_row ) { - if ( m_last_row[bank] == NULL ) { - if ( m_queue[bank].empty() ) + //row + bool rowhit = true; + std::list<dram_req_t*> *m_current_queue = m_queue; + std::map<unsigned,std::list<std::list<dram_req_t*>::iterator> > *m_current_bins = m_bins ; + std::list<std::list<dram_req_t*>::iterator> **m_current_last_row = m_last_row; + + if(m_config->seperate_write_queue_enabled) { + if(m_mode == READ_MODE && + ((m_num_write_pending >= m_config->write_high_watermark ) + // || (m_queue[bank].empty() && !m_write_queue[bank].empty()) + )) { + m_mode = WRITE_MODE; + } + else if(m_mode == WRITE_MODE && + (( m_num_write_pending < m_config->write_low_watermark ) + // || (!m_queue[bank].empty() && m_write_queue[bank].empty()) + )){ + m_mode = READ_MODE; + } + } + + if(m_mode == WRITE_MODE) { + m_current_queue = m_write_queue; + m_current_bins = m_write_bins ; + m_current_last_row = m_last_write_row; + } + + if ( m_current_last_row[bank] == NULL ) { + if ( m_current_queue[bank].empty() ) return NULL; - std::map<unsigned,std::list<std::list<dram_req_t*>::iterator> >::iterator bin_ptr = m_bins[bank].find( curr_row ); - if ( bin_ptr == m_bins[bank].end()) { - dram_req_t *req = m_queue[bank].back(); - bin_ptr = m_bins[bank].find( req->row ); - assert( bin_ptr != m_bins[bank].end() ); // where did the request go??? - m_last_row[bank] = &(bin_ptr->second); + std::map<unsigned,std::list<std::list<dram_req_t*>::iterator> >::iterator bin_ptr = m_current_bins[bank].find( curr_row ); + if ( bin_ptr == m_current_bins[bank].end()) { + dram_req_t *req = m_current_queue[bank].back(); + bin_ptr = m_current_bins[bank].find( req->row ); + assert( bin_ptr != m_current_bins[bank].end() ); // where did the request go??? + m_current_last_row[bank] = &(bin_ptr->second); data_collection(bank); + rowhit = false; } else { - m_last_row[bank] = &(bin_ptr->second); - + m_current_last_row[bank] = &(bin_ptr->second); + rowhit = true; } } - std::list<dram_req_t*>::iterator next = m_last_row[bank]->back(); + std::list<dram_req_t*>::iterator next = m_current_last_row[bank]->back(); dram_req_t *req = (*next); + //rowblp stats + m_dram->access_num++; + bool is_write = req->data->is_write(); + if(is_write) + m_dram->write_num++; + else + m_dram->read_num++; + + if(rowhit) { + m_dram->hits_num++; + if(is_write) + m_dram->hits_write_num++; + else + m_dram->hits_read_num++; + } + m_stats->concurrent_row_access[m_dram->id][bank]++; m_stats->row_access[m_dram->id][bank]++; - m_last_row[bank]->pop_back(); + m_current_last_row[bank]->pop_back(); - m_queue[bank].erase(next); - if ( m_last_row[bank]->empty() ) { - m_bins[bank].erase( req->row ); - m_last_row[bank] = NULL; + m_current_queue[bank].erase(next); + if ( m_current_last_row[bank]->empty() ) { + m_current_bins[bank].erase( req->row ); + m_current_last_row[bank] = NULL; } #ifdef DEBUG_FAST_IDEAL_SCHED if ( req ) printf("%08u : DRAM(%u) scheduling memory request to bank=%u, row=%u\n", (unsigned)gpu_sim_cycle, m_dram->id, req->bk, req->row ); #endif - assert( req != NULL && m_num_pending != 0 ); - m_num_pending--; + + if(m_config->seperate_write_queue_enabled && req->data->is_write()) { + assert( req != NULL && m_num_write_pending != 0 ); + m_num_write_pending--; + } + else { + assert( req != NULL && m_num_pending != 0 ); + m_num_pending--; + } return req; } @@ -129,7 +202,7 @@ void dram_t::scheduler_frfcfs() { unsigned mrq_latency; frfcfs_scheduler *sched = m_frfcfs_scheduler; - while ( !mrqq->empty() && (!m_config->gpgpu_frfcfs_dram_sched_queue_size || sched->num_pending() < m_config->gpgpu_frfcfs_dram_sched_queue_size)) { + while ( !mrqq->empty() ) { dram_req_t *req = mrqq->pop(); // Power stats @@ -160,6 +233,8 @@ void dram_t::scheduler_frfcfs() bk[b]->mrq = req; if (m_config->gpgpu_memlatency_stat) { mrq_latency = gpu_sim_cycle + gpu_tot_sim_cycle - bk[b]->mrq->timestamp; + m_stats->tot_mrq_latency += mrq_latency; + m_stats->tot_mrq_num++; bk[b]->mrq->timestamp = gpu_tot_sim_cycle + gpu_sim_cycle; m_stats->mrq_lat_table[LOGB2(mrq_latency)]++; if (mrq_latency > m_stats->max_mrq_latency) { diff --git a/src/gpgpu-sim/dram_sched.h b/src/gpgpu-sim/dram_sched.h index 3860f5b..63f5831 100644 --- a/src/gpgpu-sim/dram_sched.h +++ b/src/gpgpu-sim/dram_sched.h @@ -35,6 +35,11 @@ #include <list> #include <map> +enum memory_mode { + READ_MODE = 0, + WRITE_MODE +}; + class frfcfs_scheduler { public: frfcfs_scheduler( const memory_config *config, dram_t *dm, memory_stats_t *stats ); @@ -43,17 +48,24 @@ public: dram_req_t *schedule( unsigned bank, unsigned curr_row ); void print( FILE *fp ); unsigned num_pending() const { return m_num_pending;} + unsigned num_write_pending() const { return m_num_write_pending;} private: const memory_config *m_config; dram_t *m_dram; unsigned m_num_pending; + unsigned m_num_write_pending; std::list<dram_req_t*> *m_queue; std::map<unsigned,std::list<std::list<dram_req_t*>::iterator> > *m_bins; std::list<std::list<dram_req_t*>::iterator> **m_last_row; unsigned *curr_row_service_time; //one set of variables for each bank. unsigned *row_service_timestamp; //tracks when scheduler began servicing current row + std::list<dram_req_t*> *m_write_queue; + std::map<unsigned,std::list<std::list<dram_req_t*>::iterator> > *m_write_bins; + std::list<std::list<dram_req_t*>::iterator> **m_last_write_row; + + enum memory_mode m_mode; memory_stats_t *m_stats; }; diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index 8886398..ba81440 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -29,7 +29,6 @@ #include "stat-tool.h" #include <assert.h> -#define MAX_DEFAULT_CACHE_SIZE_MULTIBLIER 4 // used to allocate memory that is large enough to adapt the changes in cache size across kernels const char * cache_request_status_str(enum cache_request_status status) @@ -38,7 +37,8 @@ const char * cache_request_status_str(enum cache_request_status status) "HIT", "HIT_RESERVED", "MISS", - "RESERVATION_FAIL" + "RESERVATION_FAIL", + "SECTOR_MISS" }; assert(sizeof(static_cache_request_status_str) / sizeof(const char*) == NUM_CACHE_REQUEST_STATUS); @@ -47,6 +47,22 @@ const char * cache_request_status_str(enum cache_request_status status) return static_cache_request_status_str[status]; } +const char * cache_fail_status_str(enum cache_reservation_fail_reason status) +{ + static const char * static_cache_reservation_fail_reason_str[] = { + "LINE_ALLOC_FAIL", + "MISS_QUEUE_FULL", + "MSHR_ENRTY_FAIL", + "MSHR_MERGE_ENRTY_FAIL", + "MSHR_RW_PENDING" + }; + + assert(sizeof(static_cache_reservation_fail_reason_str) / sizeof(const char*) == NUM_CACHE_RESERVATION_FAIL_STATUS); + assert(status < NUM_CACHE_RESERVATION_FAIL_STATUS); + + return static_cache_reservation_fail_reason_str[status]; +} + unsigned l1d_cache_config::set_index(new_addr_type addr) const{ unsigned set_index = m_nset; // Default to linear set index function unsigned lower_xor = 0; @@ -54,10 +70,11 @@ unsigned l1d_cache_config::set_index(new_addr_type addr) const{ switch(m_set_index_function){ case FERMI_HASH_SET_FUNCTION: + case BITWISE_XORING_FUNCTION: /* * Set Indexing function from "A Detailed GPU Cache Model Based on Reuse Distance Theory" * Cedric Nugteren et al. - * ISCA 2014 + * HPCA 2014 */ if(m_nset == 32 || m_nset == 64){ // Lower xor value is bits 7-11 @@ -80,6 +97,36 @@ unsigned l1d_cache_config::set_index(new_addr_type addr) const{ } break; + case HASH_IPOLY_FUNCTION: + /* + * Set Indexing function from "Pseudo-randomly interleaved memory." + * Rau, B. R et al. + * ISCA 1991 + * + * "Sacat: streaming-aware conflict-avoiding thrashing-resistant gpgpu cache management scheme." + * Khairy et al. + * IEEE TPDS 2017. + */ + if(m_nset == 32 || m_nset == 64){ + std::bitset<64> a(addr); + std::bitset<6> index; + index[0] = a[25]^a[24]^a[23]^a[22]^a[21]^a[18]^a[17]^a[15]^a[12]^a[7]; //10 + index[1] = a[26]^a[25]^a[24]^a[23]^a[22]^a[19]^a[18]^a[16]^a[13]^a[8]; //10 + index[2] = a[26]^a[22]^a[21]^a[20]^a[19]^a[18]^a[15]^a[14]^a[12]^a[9]; //10 + index[3] = a[23]^a[22]^a[21]^a[20]^a[19]^a[16]^a[15]^a[13]^a[10]; //9 + index[4] = a[24]^a[23]^a[22]^a[21]^a[20]^a[17]^a[16]^a[14]^a[11]; //9 + + if(m_nset == 64) + index[5] = a[12]; + + set_index = index.to_ulong(); + + }else{ /* Else incorrect number of sets for the hashing function */ + assert("\nGPGPU-Sim cache configuration error: The number of sets should be " + "32 or 64 for the hashing set index function.\n" && 0); + } + break; + case CUSTOM_SET_FUNCTION: /* No custom set function implemented */ break; @@ -87,6 +134,10 @@ unsigned l1d_cache_config::set_index(new_addr_type addr) const{ case LINEAR_SET_FUNCTION: set_index = (addr >> m_line_sz_log2) & (m_nset-1); break; + + default: + assert("\nUndefined set index function.\n" && 0); + break; } // Linear function selected or custom set index function not implemented @@ -113,13 +164,16 @@ unsigned l2_cache_config::set_index(new_addr_type addr) const{ tag_array::~tag_array() { + unsigned cache_lines_num = m_config.get_max_num_lines(); + for(unsigned i=0; i<cache_lines_num; ++i) + delete m_lines[i]; delete[] m_lines; } tag_array::tag_array( cache_config &config, int core_id, int type_id, - cache_block_t* new_lines) + cache_block_t** new_lines) : m_config( config ), m_lines( new_lines ) { @@ -137,7 +191,21 @@ tag_array::tag_array( cache_config &config, : m_config( config ) { //assert( m_config.m_write_policy == READ_ONLY ); Old assert - m_lines = new cache_block_t[MAX_DEFAULT_CACHE_SIZE_MULTIBLIER*config.get_num_lines()]; + unsigned cache_lines_num = config.get_max_num_lines(); + m_lines = new cache_block_t*[cache_lines_num]; + if(config.m_cache_type == NORMAL) + { + for(unsigned i=0; i<cache_lines_num; ++i) + m_lines[i] = new line_cache_block(); + } + else if(config.m_cache_type == SECTOR) + { + for(unsigned i=0; i<cache_lines_num; ++i) + m_lines[i] = new sector_cache_block(); + } + else + assert(0); + init( core_id, type_id ); } @@ -147,15 +215,41 @@ void tag_array::init( int core_id, int type_id ) m_miss = 0; m_pending_hit = 0; m_res_fail = 0; + m_sector_miss = 0; // initialize snapshot counters for visualizer m_prev_snapshot_access = 0; m_prev_snapshot_miss = 0; m_prev_snapshot_pending_hit = 0; m_core_id = core_id; m_type_id = type_id; + is_used = false; } -enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx ) const { +void tag_array::add_pending_line(mem_fetch *mf){ + assert(mf); + new_addr_type addr = m_config.block_addr(mf->get_addr()); + line_table::const_iterator i = pending_lines.find(addr); + if ( i == pending_lines.end() ) { + pending_lines[addr] = mf->get_inst().get_uid(); + } +} + +void tag_array::remove_pending_line(mem_fetch *mf){ + assert(mf); + new_addr_type addr = m_config.block_addr(mf->get_addr()); + line_table::const_iterator i = pending_lines.find(addr); + if ( i != pending_lines.end() ) { + pending_lines.erase(addr); + } +} + +enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx, mem_fetch* mf, bool probe_mode) const { + mem_access_sector_mask_t mask = mf->get_access_sector_mask(); + return probe(addr, idx, mask, probe_mode, mf); +} + + +enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx, mem_access_sector_mask_t mask, bool probe_mode, mem_fetch* mf) const { //assert( m_config.m_write_policy == READ_ONLY ); unsigned set_index = m_config.set_index(addr); new_addr_type tag = m_config.tag(addr); @@ -169,35 +263,45 @@ enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx ) // check for hit or pending hit for (unsigned way=0; way<m_config.m_assoc; way++) { unsigned index = set_index*m_config.m_assoc+way; - cache_block_t *line = &m_lines[index]; + cache_block_t *line = m_lines[index]; if (line->m_tag == tag) { - if ( line->m_status == RESERVED ) { + if ( line->get_status(mask) == RESERVED ) { idx = index; return HIT_RESERVED; - } else if ( line->m_status == VALID ) { + } else if ( line->get_status(mask) == VALID ) { idx = index; return HIT; - } else if ( line->m_status == MODIFIED ) { + } else if ( line->get_status(mask) == MODIFIED) { + if(line->is_readable(mask)) { + idx = index; + return HIT; + } + else { + idx = index; + return SECTOR_MISS; + } + + } else if ( line->is_valid_line() && line->get_status(mask) == INVALID ) { idx = index; - return HIT; - } else { - assert( line->m_status == INVALID ); + return SECTOR_MISS; + }else { + assert( line->get_status(mask) == INVALID ); } } - if (line->m_status != RESERVED) { + if (!line->is_reserved_line()) { all_reserved = false; - if (line->m_status == INVALID) { + if (line->is_invalid_line()) { invalid_line = index; } else { // valid line : keep track of most appropriate replacement candidate if ( m_config.m_replacement_policy == LRU ) { - if ( line->m_last_access_time < valid_timestamp ) { - valid_timestamp = line->m_last_access_time; + if ( line->get_last_access_time() < valid_timestamp ) { + valid_timestamp = line->get_last_access_time(); valid_line = index; } } else if ( m_config.m_replacement_policy == FIFO ) { - if ( line->m_alloc_time < valid_timestamp ) { - valid_timestamp = line->m_alloc_time; + if ( line->get_alloc_time() < valid_timestamp ) { + valid_timestamp = line->get_alloc_time(); valid_line = index; } } @@ -215,40 +319,59 @@ enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx ) idx = valid_line; } else abort(); // if an unreserved block exists, it is either invalid or replaceable + + if(probe_mode && m_config.is_streaming()){ + line_table::const_iterator i = pending_lines.find(m_config.block_addr(addr)); + assert(mf); + if ( !mf->is_write() && i != pending_lines.end() ) { + if(i->second != mf->get_inst().get_uid()) + return SECTOR_MISS; + } + } + return MISS; } -enum cache_request_status tag_array::access( new_addr_type addr, unsigned time, unsigned &idx ) +enum cache_request_status tag_array::access( new_addr_type addr, unsigned time, unsigned &idx, mem_fetch* mf) { bool wb=false; - cache_block_t evicted; - enum cache_request_status result = access(addr,time,idx,wb,evicted); + evicted_block_info evicted; + enum cache_request_status result = access(addr,time,idx,wb,evicted,mf); assert(!wb); return result; } -enum cache_request_status tag_array::access( new_addr_type addr, unsigned time, unsigned &idx, bool &wb, cache_block_t &evicted ) +enum cache_request_status tag_array::access( new_addr_type addr, unsigned time, unsigned &idx, bool &wb, evicted_block_info &evicted, mem_fetch* mf ) { m_access++; + is_used = true; shader_cache_access_log(m_core_id, m_type_id, 0); // log accesses to cache - enum cache_request_status status = probe(addr,idx); + enum cache_request_status status = probe(addr,idx,mf); switch (status) { case HIT_RESERVED: m_pending_hit++; case HIT: - m_lines[idx].m_last_access_time=time; + m_lines[idx]->set_last_access_time(time, mf->get_access_sector_mask()); break; case MISS: m_miss++; shader_cache_access_log(m_core_id, m_type_id, 1); // log cache misses if ( m_config.m_alloc_policy == ON_MISS ) { - if( m_lines[idx].m_status == MODIFIED ) { + if( m_lines[idx]->is_modified_line()) { wb = true; - evicted = m_lines[idx]; + evicted.set_info(m_lines[idx]->m_block_addr, m_lines[idx]->get_modified_size()); } - m_lines[idx].allocate( m_config.tag(addr), m_config.block_addr(addr), time ); + m_lines[idx]->allocate( m_config.tag(addr), m_config.block_addr(addr), time, mf->get_access_sector_mask()); } break; + case SECTOR_MISS: + assert(m_config.m_cache_type == SECTOR); + m_sector_miss++; + shader_cache_access_log(m_core_id, m_type_id, 1); // log cache misses + if ( m_config.m_alloc_policy == ON_MISS ) { + ((sector_cache_block*)m_lines[idx])->allocate_sector( time, mf->get_access_sector_mask() ); + } + break; case RESERVATION_FAIL: m_res_fail++; shader_cache_access_log(m_core_id, m_type_id, 1); // log cache misses @@ -261,37 +384,70 @@ enum cache_request_status tag_array::access( new_addr_type addr, unsigned time, return status; } -void tag_array::fill( new_addr_type addr, unsigned time ) +void tag_array::fill( new_addr_type addr, unsigned time, mem_fetch* mf) { - assert( m_config.m_alloc_policy == ON_FILL ); + fill(addr, time, mf->get_access_sector_mask()); +} + +void tag_array::fill( new_addr_type addr, unsigned time, mem_access_sector_mask_t mask ) +{ + //assert( m_config.m_alloc_policy == ON_FILL ); unsigned idx; - enum cache_request_status status = probe(addr,idx); - assert(status==MISS); // MSHR should have prevented redundant memory request - m_lines[idx].allocate( m_config.tag(addr), m_config.block_addr(addr), time ); - m_lines[idx].fill(time); + enum cache_request_status status = probe(addr,idx,mask); + //assert(status==MISS||status==SECTOR_MISS); // MSHR should have prevented redundant memory request + if(status==MISS) + m_lines[idx]->allocate( m_config.tag(addr), m_config.block_addr(addr), time, mask ); + else if (status==SECTOR_MISS) { + assert(m_config.m_cache_type == SECTOR); + ((sector_cache_block*)m_lines[idx])->allocate_sector( time, mask ); + } + + m_lines[idx]->fill(time, mask); } -void tag_array::fill( unsigned index, unsigned time ) +void tag_array::fill( unsigned index, unsigned time, mem_fetch* mf) { assert( m_config.m_alloc_policy == ON_MISS ); - m_lines[index].fill(time); + m_lines[index]->fill(time, mf->get_access_sector_mask()); } + +//TODO: we need write back the flushed data to the upper level void tag_array::flush() { + if(!is_used) + return; + for (unsigned i=0; i < m_config.get_num_lines(); i++) - m_lines[i].m_status = INVALID; + if(m_lines[i]->is_modified_line()) { + for(unsigned j=0; j < SECTOR_CHUNCK_SIZE; j++) + m_lines[i]->set_status(INVALID, mem_access_sector_mask_t().set(j)) ; + } + + is_used = false; +} + +void tag_array::invalidate() +{ + if(!is_used) + return; + + for (unsigned i=0; i < m_config.get_num_lines(); i++) + for(unsigned j=0; j < SECTOR_CHUNCK_SIZE; j++) + m_lines[i]->set_status(INVALID, mem_access_sector_mask_t().set(j)) ; + + is_used = false; } float tag_array::windowed_miss_rate( ) const { unsigned n_access = m_access - m_prev_snapshot_access; - unsigned n_miss = m_miss - m_prev_snapshot_miss; + unsigned n_miss = (m_miss+m_sector_miss) - m_prev_snapshot_miss; // unsigned n_pending_hit = m_pending_hit - m_prev_snapshot_pending_hit; float missrate = 0.0f; if (n_access != 0) - missrate = (float) n_miss / n_access; + missrate = (float) (n_miss+m_sector_miss) / n_access; return missrate; } @@ -299,23 +455,24 @@ void tag_array::new_window() { m_prev_snapshot_access = m_access; m_prev_snapshot_miss = m_miss; + m_prev_snapshot_miss = m_miss + m_sector_miss; m_prev_snapshot_pending_hit = m_pending_hit; } void tag_array::print( FILE *stream, unsigned &total_access, unsigned &total_misses ) const { m_config.print(stream); - fprintf( stream, "\t\tAccess = %d, Miss = %d (%.3g), PendingHit = %d (%.3g)\n", - m_access, m_miss, (float) m_miss / m_access, + fprintf( stream, "\t\tAccess = %d, Miss = %d, Sector_Miss = %d, Total_Miss = %d (%.3g), PendingHit = %d (%.3g)\n", + m_access, m_miss, m_sector_miss, (m_miss+m_sector_miss), (float) (m_miss+m_sector_miss) / m_access, m_pending_hit, (float) m_pending_hit / m_access); - total_misses+=m_miss; + total_misses+=(m_miss+m_sector_miss); total_access+=m_access; } void tag_array::get_stats(unsigned &total_access, unsigned &total_misses, unsigned &total_hit_res, unsigned &total_res_fail) const{ // Update statistics from the tag array total_access = m_access; - total_misses = m_miss; + total_misses = (m_miss+m_sector_miss); total_hit_res = m_pending_hit; total_res_fail = m_res_fail; } @@ -324,16 +481,17 @@ void tag_array::get_stats(unsigned &total_access, unsigned &total_misses, unsign bool was_write_sent( const std::list<cache_event> &events ) { for( std::list<cache_event>::const_iterator e=events.begin(); e!=events.end(); e++ ) { - if( *e == WRITE_REQUEST_SENT ) + if( (*e).m_cache_event_type == WRITE_REQUEST_SENT ) return true; } return false; } -bool was_writeback_sent( const std::list<cache_event> &events ) +bool was_writeback_sent( const std::list<cache_event> &events, cache_event& wb_event) { for( std::list<cache_event>::const_iterator e=events.begin(); e!=events.end(); e++ ) { - if( *e == WRITE_BACK_REQUEST_SENT ) + if( (*e).m_cache_event_type == WRITE_BACK_REQUEST_SENT ) + wb_event = *e; return true; } return false; @@ -342,7 +500,16 @@ bool was_writeback_sent( const std::list<cache_event> &events ) bool was_read_sent( const std::list<cache_event> &events ) { for( std::list<cache_event>::const_iterator e=events.begin(); e!=events.end(); e++ ) { - if( *e == READ_REQUEST_SENT ) + if( (*e).m_cache_event_type == READ_REQUEST_SENT ) + return true; + } + return false; +} + +bool was_writeallocate_sent( const std::list<cache_event> &events ) +{ + for( std::list<cache_event>::const_iterator e=events.begin(); e!=events.end(); e++ ) { + if( (*e).m_cache_event_type == WRITE_ALLOCATE_SENT ) return true; } return false; @@ -375,11 +542,27 @@ void mshr_table::add( new_addr_type block_addr, mem_fetch *mf ){ } } +/// check is_read_after_write_pending +bool mshr_table::is_read_after_write_pending( new_addr_type block_addr){ + std::list<mem_fetch*> my_list = m_data[block_addr].m_list; + bool write_found = false; + for (std::list<mem_fetch*>::iterator it=my_list.begin(); it != my_list.end(); ++it) + { + if((*it)->is_write()) //Pending Write Request + write_found = true; + else if(write_found) //Pending Read Request and we found previous Write + return true; + } + + return false; + +} + /// Accept a new cache fill response: mark entry ready for processing void mshr_table::mark_ready( new_addr_type block_addr, bool &has_atomic ){ assert( !busy() ); table::iterator a = m_data.find(block_addr); - assert( a != m_data.end() ); // don't remove same request twice + assert( a != m_data.end() ); m_current_response.push_back( block_addr ); has_atomic = a->second.m_has_atomic; assert( m_current_response.size() <= m_data.size() ); @@ -417,9 +600,11 @@ void mshr_table::display( FILE *fp ) const{ /***************************************************************** Caches *****************************************************************/ cache_stats::cache_stats(){ m_stats.resize(NUM_MEM_ACCESS_TYPE); + m_fail_stats.resize(NUM_MEM_ACCESS_TYPE); for(unsigned i=0; i<NUM_MEM_ACCESS_TYPE; ++i){ m_stats[i].resize(NUM_CACHE_REQUEST_STATUS, 0); - } + m_fail_stats[i].resize(NUM_CACHE_RESERVATION_FAIL_STATUS, 0); + } m_cache_port_available_cycles = 0; m_cache_data_port_busy_cycles = 0; m_cache_fill_port_busy_cycles = 0; @@ -431,7 +616,8 @@ void cache_stats::clear(){ /// for(unsigned i=0; i<NUM_MEM_ACCESS_TYPE; ++i){ std::fill(m_stats[i].begin(), m_stats[i].end(), 0); - } + std::fill(m_fail_stats[i].begin(), m_fail_stats[i].end(), 0); + } m_cache_port_available_cycles = 0; m_cache_data_port_busy_cycles = 0; m_cache_fill_port_busy_cycles = 0; @@ -446,6 +632,14 @@ void cache_stats::inc_stats(int access_type, int access_outcome){ m_stats[access_type][access_outcome]++; } +void cache_stats::inc_fail_stats(int access_type, int fail_outcome){ + + if(!check_fail_valid(access_type, fail_outcome)) + assert(0 && "Unknown cache access type or access fail"); + + m_fail_stats[access_type][fail_outcome]++; +} + enum cache_request_status cache_stats::select_stats_status(enum cache_request_status probe, enum cache_request_status access) const { /// @@ -454,29 +648,47 @@ enum cache_request_status cache_stats::select_stats_status(enum cache_request_st /// if(probe == HIT_RESERVED && access != RESERVATION_FAIL) return probe; + else if(probe == SECTOR_MISS && access == MISS) + return probe; else return access; } -unsigned &cache_stats::operator()(int access_type, int access_outcome){ +unsigned &cache_stats::operator()(int access_type, int access_outcome, bool fail_outcome){ /// /// Simple method to read/modify the stat corresponding to (access_type, access_outcome) /// Used overloaded () to avoid the need for separate read/write member functions /// - if(!check_valid(access_type, access_outcome)) - assert(0 && "Unknown cache access type or access outcome"); + if(fail_outcome) { + if(!check_fail_valid(access_type, access_outcome)) + assert(0 && "Unknown cache access type or fail outcome"); - return m_stats[access_type][access_outcome]; + return m_fail_stats[access_type][access_outcome]; + } + else { + if(!check_valid(access_type, access_outcome)) + assert(0 && "Unknown cache access type or access outcome"); + + return m_stats[access_type][access_outcome]; + } } -unsigned cache_stats::operator()(int access_type, int access_outcome) const{ +unsigned cache_stats::operator()(int access_type, int access_outcome, bool fail_outcome) const{ /// /// Const accessor into m_stats. /// - if(!check_valid(access_type, access_outcome)) - assert(0 && "Unknown cache access type or access outcome"); + if(fail_outcome) { + if(!check_fail_valid(access_type, access_outcome)) + assert(0 && "Unknown cache access type or fail outcome"); - return m_stats[access_type][access_outcome]; + return m_fail_stats[access_type][access_outcome]; + } + else { + if(!check_valid(access_type, access_outcome)) + assert(0 && "Unknown cache access type or access outcome"); + + return m_stats[access_type][access_outcome]; + } } cache_stats cache_stats::operator+(const cache_stats &cs){ @@ -486,9 +698,12 @@ cache_stats cache_stats::operator+(const cache_stats &cs){ cache_stats ret; for(unsigned type=0; type<NUM_MEM_ACCESS_TYPE; ++type){ for(unsigned status=0; status<NUM_CACHE_REQUEST_STATUS; ++status){ - ret(type, status) = m_stats[type][status] + cs(type, status); + ret(type, status, false) = m_stats[type][status] + cs(type, status, false); + } + for(unsigned status=0; status<NUM_CACHE_RESERVATION_FAIL_STATUS; ++status){ + ret(type, status, true) = m_fail_stats[type][status] + cs(type, status, true); + } } - } ret.m_cache_port_available_cycles = m_cache_port_available_cycles + cs.m_cache_port_available_cycles; ret.m_cache_data_port_busy_cycles = m_cache_data_port_busy_cycles + cs.m_cache_data_port_busy_cycles; ret.m_cache_fill_port_busy_cycles = m_cache_fill_port_busy_cycles + cs.m_cache_fill_port_busy_cycles; @@ -501,8 +716,11 @@ cache_stats &cache_stats::operator+=(const cache_stats &cs){ /// for(unsigned type=0; type<NUM_MEM_ACCESS_TYPE; ++type){ for(unsigned status=0; status<NUM_CACHE_REQUEST_STATUS; ++status){ - m_stats[type][status] += cs(type, status); + m_stats[type][status] += cs(type, status, false); } + for(unsigned status=0; status<NUM_CACHE_RESERVATION_FAIL_STATUS; ++status){ + m_fail_stats[type][status] += cs(type, status, true); + } } m_cache_port_available_cycles += cs.m_cache_port_available_cycles; m_cache_data_port_busy_cycles += cs.m_cache_data_port_busy_cycles; @@ -528,19 +746,34 @@ void cache_stats::print_stats(FILE *fout, const char *cache_name) const{ cache_request_status_str((enum cache_request_status)status), m_stats[type][status]); if(status != RESERVATION_FAIL) - total_access[type]+= m_stats[type][status]; + total_access[type]+= m_stats[type][status]; } } for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { - if(total_access[type] > 0) - fprintf(fout, "\t%s[%s][%s] = %u\n", - m_cache_name.c_str(), - mem_access_type_str((enum mem_access_type)type), - "TOTAL_ACCESS", - total_access[type]); + if(total_access[type] > 0) + fprintf(fout, "\t%s[%s][%s] = %u\n", + m_cache_name.c_str(), + mem_access_type_str((enum mem_access_type)type), + "TOTAL_ACCESS", + total_access[type]); } } +void cache_stats::print_fail_stats(FILE *fout, const char *cache_name) const{ + std::string m_cache_name = cache_name; + for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { + for (unsigned fail = 0; fail < NUM_CACHE_RESERVATION_FAIL_STATUS; ++fail) { + if(m_fail_stats[type][fail] > 0){ + fprintf(fout, "\t%s[%s][%s] = %u\n", + m_cache_name.c_str(), + mem_access_type_str((enum mem_access_type)type), + cache_fail_status_str((enum cache_reservation_fail_reason)fail), + m_fail_stats[type][fail]); + } + } + } +} + void cache_sub_stats::print_port_stats(FILE *fout, const char *cache_name) const { float data_port_util = 0.0f; @@ -580,10 +813,10 @@ void cache_stats::get_sub_stats(struct cache_sub_stats &css) const{ for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { - if(status == HIT || status == MISS || status == HIT_RESERVED) + if(status == HIT || status == MISS || status == SECTOR_MISS || status == HIT_RESERVED) t_css.accesses += m_stats[type][status]; - if(status == MISS) + if(status == MISS || status == SECTOR_MISS) t_css.misses += m_stats[type][status]; if(status == HIT_RESERVED) @@ -611,6 +844,16 @@ bool cache_stats::check_valid(int type, int status) const{ return false; } +bool cache_stats::check_fail_valid(int type, int fail) const{ + /// + /// Verify a valid access_type/access_status + /// + if((type >= 0) && (type < NUM_MEM_ACCESS_TYPE) && (fail >= 0) && (fail < NUM_CACHE_RESERVATION_FAIL_STATUS)) + return true; + else + return false; +} + void cache_stats::sample_cache_port_utility(bool data_port_busy, bool fill_port_busy) { m_cache_port_available_cycles += 1; @@ -639,15 +882,18 @@ void baseline_cache::bandwidth_management::use_data_port(mem_fetch *mf, enum cac unsigned data_cycles = data_size / port_width + ((data_size % port_width > 0)? 1 : 0); m_data_port_occupied_cycles += data_cycles; } break; - case HIT_RESERVED: + case HIT_RESERVED: case MISS: { // the data array is accessed to read out the entire line for write-back - if (was_writeback_sent(events)) { - unsigned data_cycles = m_config.m_line_sz / port_width; + // in case of sector cache we need to write bank only the modified sectors + cache_event ev(WRITE_BACK_REQUEST_SENT); + if (was_writeback_sent(events, ev)) { + unsigned data_cycles = ev.m_evicted_block.m_modified_size / port_width; m_data_port_occupied_cycles += data_cycles; } } break; - case RESERVATION_FAIL: + case SECTOR_MISS: + case RESERVATION_FAIL: // Does not consume any port bandwidth break; default: @@ -660,7 +906,7 @@ void baseline_cache::bandwidth_management::use_data_port(mem_fetch *mf, enum cac void baseline_cache::bandwidth_management::use_fill_port(mem_fetch *mf) { // assume filling the entire line with the returned request - unsigned fill_cycles = m_config.m_line_sz / m_config.m_data_port_width; + unsigned fill_cycles = m_config.get_atom_sz() / m_config.m_data_port_width; m_fill_port_occupied_cycles += fill_cycles; } @@ -707,21 +953,43 @@ void baseline_cache::cycle(){ /// Interface for response from lower memory level (model bandwidth restictions in caller) void baseline_cache::fill(mem_fetch *mf, unsigned time){ + + if(m_config.m_mshr_type == SECTOR_ASSOC) { + assert(mf->get_original_mf()); + extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf->get_original_mf()); + assert( e != m_extra_mf_fields.end() ); + e->second.pending_read--; + + if(e->second.pending_read > 0) { + //wait for the other requests to come back + delete mf; + return; + } else { + mem_fetch *temp = mf; + mf = mf->get_original_mf(); + delete temp; + } + } + extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf); assert( e != m_extra_mf_fields.end() ); assert( e->second.m_valid ); mf->set_data_size( e->second.m_data_size ); + mf->set_addr( e->second.m_addr ); if ( m_config.m_alloc_policy == ON_MISS ) - m_tag_array->fill(e->second.m_cache_index,time); - else if ( m_config.m_alloc_policy == ON_FILL ) - m_tag_array->fill(e->second.m_block_addr,time); + m_tag_array->fill(e->second.m_cache_index,time,mf); + else if ( m_config.m_alloc_policy == ON_FILL ) { + m_tag_array->fill(e->second.m_block_addr,time,mf); + if(m_config.is_streaming()) + m_tag_array->remove_pending_line(mf); + } else abort(); bool has_atomic = false; m_mshrs.mark_ready(e->second.m_block_addr, has_atomic); if (has_atomic) { assert(m_config.m_alloc_policy == ON_MISS); - cache_block_t &block = m_tag_array->get_block(e->second.m_cache_index); - block.m_status = MODIFIED; // mark line as dirty for atomic operation + cache_block_t* block = m_tag_array->get_block(e->second.m_cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()); // mark line as dirty for atomic operation } m_extra_mf_fields.erase(mf); m_bandwidth_management.use_fill_port(mf); @@ -749,45 +1017,59 @@ void baseline_cache::send_read_request(new_addr_type addr, new_addr_type block_a unsigned time, bool &do_miss, std::list<cache_event> &events, bool read_only, bool wa){ bool wb=false; - cache_block_t e; + evicted_block_info e; send_read_request(addr, block_addr, cache_index, mf, time, do_miss, wb, e, events, read_only, wa); } /// Read miss handler. Check MSHR hit or MSHR available void baseline_cache::send_read_request(new_addr_type addr, new_addr_type block_addr, unsigned cache_index, mem_fetch *mf, - unsigned time, bool &do_miss, bool &wb, cache_block_t &evicted, std::list<cache_event> &events, bool read_only, bool wa){ + unsigned time, bool &do_miss, bool &wb, evicted_block_info &evicted, std::list<cache_event> &events, bool read_only, bool wa){ - bool mshr_hit = m_mshrs.probe(block_addr); - bool mshr_avail = !m_mshrs.full(block_addr); + new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); + bool mshr_hit = m_mshrs.probe(mshr_addr); + bool mshr_avail = !m_mshrs.full(mshr_addr); if ( mshr_hit && mshr_avail ) { if(read_only) - m_tag_array->access(block_addr,time,cache_index); + m_tag_array->access(block_addr,time,cache_index,mf); else - m_tag_array->access(block_addr,time,cache_index,wb,evicted); + m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf); - m_mshrs.add(block_addr,mf); + m_mshrs.add(mshr_addr,mf); do_miss = true; + } else if ( !mshr_hit && mshr_avail && (m_miss_queue.size() < m_config.m_miss_queue_size) ) { if(read_only) - m_tag_array->access(block_addr,time,cache_index); + m_tag_array->access(block_addr,time,cache_index,mf); else - m_tag_array->access(block_addr,time,cache_index,wb,evicted); + m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf); - m_mshrs.add(block_addr,mf); - m_extra_mf_fields[mf] = extra_mf_fields(block_addr,cache_index, mf->get_data_size()); - mf->set_data_size( m_config.get_line_sz() ); + m_mshrs.add(mshr_addr,mf); + if(m_config.is_streaming() && m_config.m_cache_type == SECTOR){ + m_tag_array->add_pending_line(mf); + } + m_extra_mf_fields[mf] = extra_mf_fields(mshr_addr,mf->get_addr(),cache_index, mf->get_data_size(), m_config); + mf->set_data_size( m_config.get_atom_sz() ); + mf->set_addr( mshr_addr ); m_miss_queue.push_back(mf); mf->set_status(m_miss_queue_status,time); if(!wa) - events.push_back(READ_REQUEST_SENT); + events.push_back(cache_event(READ_REQUEST_SENT)); + do_miss = true; } + else if(mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_MERGE_ENRTY_FAIL); + else if (!mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_ENRTY_FAIL); + else + assert(0); } /// Sends write request to lower level memory (write or writeback) void data_cache::send_write_request(mem_fetch *mf, cache_event request, unsigned time, std::list<cache_event> &events){ - events.push_back(request); + + events.push_back(request); m_miss_queue.push_back(mf); mf->set_status(m_miss_queue_status,time); } @@ -798,40 +1080,44 @@ void data_cache::send_write_request(mem_fetch *mf, cache_event request, unsigned /// Write-back hit: Mark block as modified cache_request_status data_cache::wr_hit_wb(new_addr_type addr, unsigned cache_index, mem_fetch *mf, unsigned time, std::list<cache_event> &events, enum cache_request_status status ){ new_addr_type block_addr = m_config.block_addr(addr); - m_tag_array->access(block_addr,time,cache_index); // update LRU state - cache_block_t &block = m_tag_array->get_block(cache_index); - block.m_status = MODIFIED; + m_tag_array->access(block_addr,time,cache_index,mf); // update LRU state + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()); return HIT; } /// Write-through hit: Directly send request to lower level memory cache_request_status data_cache::wr_hit_wt(new_addr_type addr, unsigned cache_index, mem_fetch *mf, unsigned time, std::list<cache_event> &events, enum cache_request_status status ){ - if(miss_queue_full(0)) + if(miss_queue_full(0)) { + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); return RESERVATION_FAIL; // cannot handle request this cycle + } new_addr_type block_addr = m_config.block_addr(addr); - m_tag_array->access(block_addr,time,cache_index); // update LRU state - cache_block_t &block = m_tag_array->get_block(cache_index); - block.m_status = MODIFIED; + m_tag_array->access(block_addr,time,cache_index,mf); // update LRU state + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()); // generate a write-through - send_write_request(mf, WRITE_REQUEST_SENT, time, events); + send_write_request(mf, cache_event(WRITE_REQUEST_SENT), time, events); return HIT; } /// Write-evict hit: Send request to lower level memory and invalidate corresponding block cache_request_status data_cache::wr_hit_we(new_addr_type addr, unsigned cache_index, mem_fetch *mf, unsigned time, std::list<cache_event> &events, enum cache_request_status status ){ - if(miss_queue_full(0)) + if(miss_queue_full(0)) { + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); return RESERVATION_FAIL; // cannot handle request this cycle + } // generate a write-through/evict - cache_block_t &block = m_tag_array->get_block(cache_index); - send_write_request(mf, WRITE_REQUEST_SENT, time, events); + cache_block_t* block = m_tag_array->get_block(cache_index); + send_write_request(mf, cache_event(WRITE_REQUEST_SENT), time, events); // Invalidate block - block.m_status = INVALID; + block->set_status(INVALID, mf->get_access_sector_mask()); return HIT; } @@ -850,34 +1136,46 @@ enum cache_request_status data_cache::wr_hit_global_we_local_wb(new_addr_type ad /// Write-allocate miss: Send write request to lower level memory // and send a read request for the same block enum cache_request_status -data_cache::wr_miss_wa( new_addr_type addr, +data_cache::wr_miss_wa_naive( new_addr_type addr, unsigned cache_index, mem_fetch *mf, unsigned time, std::list<cache_event> &events, enum cache_request_status status ) { new_addr_type block_addr = m_config.block_addr(addr); + new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); // Write allocate, maximum 3 requests (write miss, read request, write back request) // Conservatively ensure the worst-case request can be handled this cycle - bool mshr_hit = m_mshrs.probe(block_addr); - bool mshr_avail = !m_mshrs.full(block_addr); + bool mshr_hit = m_mshrs.probe(mshr_addr); + bool mshr_avail = !m_mshrs.full(mshr_addr); if(miss_queue_full(2) || (!(mshr_hit && mshr_avail) - && !(!mshr_hit && mshr_avail - && (m_miss_queue.size() < m_config.m_miss_queue_size)))) + && !(!mshr_hit && mshr_avail && (m_miss_queue.size() < m_config.m_miss_queue_size)))) { + //check what is the exactly the failure reason + if(miss_queue_full(2) ) + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + else if(mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_MERGE_ENRTY_FAIL); + else if (!mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_ENRTY_FAIL); + else + assert(0); + return RESERVATION_FAIL; + } - send_write_request(mf, WRITE_REQUEST_SENT, time, events); + send_write_request(mf, cache_event(WRITE_REQUEST_SENT), time, events); // Tries to send write allocate request, returns true on success and false on failure //if(!send_write_allocate(mf, addr, block_addr, cache_index, time, events)) // return RESERVATION_FAIL; const mem_access_t *ma = new mem_access_t( m_wr_alloc_type, mf->get_addr(), - mf->get_data_size(), + m_config.get_atom_sz(), false, // Now performing a read mf->get_access_warp_mask(), - mf->get_access_byte_mask() ); + mf->get_access_byte_mask(), + mf->get_access_sector_mask()); mem_fetch *n_mf = new mem_fetch( *ma, NULL, @@ -889,20 +1187,22 @@ data_cache::wr_miss_wa( new_addr_type addr, bool do_miss = false; bool wb = false; - cache_block_t evicted; + evicted_block_info evicted; // Send read request resulting from write miss send_read_request(addr, block_addr, cache_index, n_mf, time, do_miss, wb, evicted, events, false, true); + events.push_back(cache_event(WRITE_ALLOCATE_SENT)); + if( do_miss ){ // If evicted block is modified and not a write-through // (already modified lower level) if( wb && (m_config.m_write_policy != WRITE_THROUGH) ) { + assert(status == MISS); //SECTOR_MISS and HIT_RESERVED should not send write back mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, - m_wrbk_type,m_config.get_line_sz(),true); - m_miss_queue.push_back(wb); - wb->set_status(m_miss_queue_status,time); + m_wrbk_type,evicted.m_modified_size,true); + send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events); } return MISS; } @@ -910,6 +1210,177 @@ data_cache::wr_miss_wa( new_addr_type addr, return RESERVATION_FAIL; } + +enum cache_request_status +data_cache::wr_miss_wa_fetch_on_write( new_addr_type addr, + unsigned cache_index, mem_fetch *mf, + unsigned time, std::list<cache_event> &events, + enum cache_request_status status ) +{ + new_addr_type block_addr = m_config.block_addr(addr); + new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); + + if(mf->get_access_byte_mask().count() == m_config.get_atom_sz()) + { + //if the request writes to the whole cache line/sector, then, write and set cache line Modified. + //and no need to send read request to memory or reserve mshr + + if(miss_queue_full(0)) { + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + return RESERVATION_FAIL; // cannot handle request this cycle + } + + bool wb = false; + evicted_block_info evicted; + + cache_request_status status = m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf); + assert(status != HIT); + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()); + if(status == HIT_RESERVED) + block->set_ignore_on_fill(true, mf->get_access_sector_mask()); + + if( status != RESERVATION_FAIL ){ + // If evicted block is modified and not a write-through + // (already modified lower level) + if( wb && (m_config.m_write_policy != WRITE_THROUGH) ) { + mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, + m_wrbk_type,evicted.m_modified_size,true); + send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events); + } + return MISS; + } + return RESERVATION_FAIL; + } + else + { + bool mshr_hit = m_mshrs.probe(mshr_addr); + bool mshr_avail = !m_mshrs.full(mshr_addr); + if(miss_queue_full(1) + || (!(mshr_hit && mshr_avail) + && !(!mshr_hit && mshr_avail && (m_miss_queue.size() < m_config.m_miss_queue_size)))) { + //check what is the exactly the failure reason + if(miss_queue_full(1) ) + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + else if(mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_MERGE_ENRTY_FAIL); + else if (!mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_ENRTY_FAIL); + else + assert(0); + + return RESERVATION_FAIL; + } + + + //prevent Write - Read - Write in pending mshr + //allowing another write will override the value of the first write, and the pending read request will read incorrect result from the second write + if(m_mshrs.probe(mshr_addr) && m_mshrs.is_read_after_write_pending(mshr_addr) && mf->is_write()) + { + //assert(0); + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_RW_PENDING); + return RESERVATION_FAIL; + } + + const mem_access_t *ma = new mem_access_t( m_wr_alloc_type, + mf->get_addr(), + m_config.get_atom_sz(), + false, // Now performing a read + mf->get_access_warp_mask(), + mf->get_access_byte_mask(), + mf->get_access_sector_mask()); + + mem_fetch *n_mf = new mem_fetch( *ma, + NULL, + mf->get_ctrl_size(), + mf->get_wid(), + mf->get_sid(), + mf->get_tpc(), + mf->get_mem_config(), + NULL, + mf); + + + new_addr_type block_addr = m_config.block_addr(addr); + bool do_miss = false; + bool wb = false; + evicted_block_info evicted; + send_read_request( addr, + block_addr, + cache_index, + n_mf, time, do_miss, wb, evicted, events, false, true); + + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_modified_on_fill(true, mf->get_access_sector_mask()); + + events.push_back(cache_event(WRITE_ALLOCATE_SENT)); + + if( do_miss ){ + // If evicted block is modified and not a write-through + // (already modified lower level) + if(wb && (m_config.m_write_policy != WRITE_THROUGH) ){ + mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, + m_wrbk_type,evicted.m_modified_size,true); + send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events); + } + return MISS; + } + return RESERVATION_FAIL; + } +} + +enum cache_request_status +data_cache::wr_miss_wa_lazy_fetch_on_read( new_addr_type addr, + unsigned cache_index, mem_fetch *mf, + unsigned time, std::list<cache_event> &events, + enum cache_request_status status ) +{ + + new_addr_type block_addr = m_config.block_addr(addr); + new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); + + + //if the request writes to the whole cache line/sector, then, write and set cache line Modified. + //and no need to send read request to memory or reserve mshr + + if(miss_queue_full(0)) { + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + return RESERVATION_FAIL; // cannot handle request this cycle + } + + bool wb = false; + evicted_block_info evicted; + + cache_request_status m_status = m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf); + assert(m_status != HIT); + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()); + if(m_status == HIT_RESERVED) { + block->set_ignore_on_fill(true, mf->get_access_sector_mask()); + block->set_modified_on_fill(true, mf->get_access_sector_mask()); + } + + if(mf->get_access_byte_mask().count() == m_config.get_atom_sz()) + { + block->set_m_readable(true, mf->get_access_sector_mask()); + } else + { + block->set_m_readable(false, mf->get_access_sector_mask()); + } + + if( m_status != RESERVATION_FAIL ){ + // If evicted block is modified and not a write-through + // (already modified lower level) + if( wb && (m_config.m_write_policy != WRITE_THROUGH) ) { + mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, + m_wrbk_type,evicted.m_modified_size,true); + send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events); + } + return MISS; + } + return RESERVATION_FAIL; +} + /// No write-allocate miss: Simply send write request to lower level memory enum cache_request_status data_cache::wr_miss_no_wa( new_addr_type addr, @@ -919,11 +1390,14 @@ data_cache::wr_miss_no_wa( new_addr_type addr, std::list<cache_event> &events, enum cache_request_status status ) { - if(miss_queue_full(0)) - return RESERVATION_FAIL; // cannot handle request this cycle + if(miss_queue_full(0)) { + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + return RESERVATION_FAIL; // cannot handle request this cycle + } + // on miss, generate write through (no write buffering -- too many threads for that) - send_write_request(mf, WRITE_REQUEST_SENT, time, events); + send_write_request(mf, cache_event(WRITE_REQUEST_SENT), time, events); return MISS; } @@ -941,13 +1415,13 @@ data_cache::rd_hit_base( new_addr_type addr, enum cache_request_status status ) { new_addr_type block_addr = m_config.block_addr(addr); - m_tag_array->access(block_addr,time,cache_index); + m_tag_array->access(block_addr,time,cache_index,mf); // Atomics treated as global read/write requests - Perform read, mark line as // MODIFIED if(mf->isatomic()){ assert(mf->get_access_type() == GLOBAL_ACC_R); - cache_block_t &block = m_tag_array->get_block(cache_index); - block.m_status = MODIFIED; // mark line as dirty + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()) ; // mark line as dirty } return HIT; } @@ -963,15 +1437,17 @@ data_cache::rd_miss_base( new_addr_type addr, unsigned time, std::list<cache_event> &events, enum cache_request_status status ){ - if(miss_queue_full(1)) + if(miss_queue_full(1)) { // cannot handle request this cycle // (might need to generate two requests) + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); return RESERVATION_FAIL; + } new_addr_type block_addr = m_config.block_addr(addr); bool do_miss = false; bool wb = false; - cache_block_t evicted; + evicted_block_info evicted; send_read_request( addr, block_addr, cache_index, @@ -982,12 +1458,12 @@ data_cache::rd_miss_base( new_addr_type addr, // (already modified lower level) if(wb && (m_config.m_write_policy != WRITE_THROUGH) ){ mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, - m_wrbk_type,m_config.get_line_sz(),true); + m_wrbk_type,evicted.m_modified_size,true); send_write_request(wb, WRITE_BACK_REQUEST_SENT, time, events); } return MISS; } - return RESERVATION_FAIL; + return RESERVATION_FAIL; } /// Access cache for read_only_cache: returns RESERVATION_FAIL if @@ -998,16 +1474,16 @@ read_only_cache::access( new_addr_type addr, unsigned time, std::list<cache_event> &events ) { - assert( mf->get_data_size() <= m_config.get_line_sz()); + assert( mf->get_data_size() <= m_config.get_atom_sz()); assert(m_config.m_write_policy == READ_ONLY); assert(!mf->get_is_write()); new_addr_type block_addr = m_config.block_addr(addr); unsigned cache_index = (unsigned)-1; - enum cache_request_status status = m_tag_array->probe(block_addr,cache_index); + enum cache_request_status status = m_tag_array->probe(block_addr,cache_index,mf); enum cache_request_status cache_status = RESERVATION_FAIL; if ( status == HIT ) { - cache_status = m_tag_array->access(block_addr,time,cache_index); // update LRU state + cache_status = m_tag_array->access(block_addr,time,cache_index,mf); // update LRU state }else if ( status != RESERVATION_FAIL ) { if(!miss_queue_full(0)){ bool do_miss=false; @@ -1018,7 +1494,10 @@ read_only_cache::access( new_addr_type addr, cache_status = RESERVATION_FAIL; }else{ cache_status = RESERVATION_FAIL; + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); } + }else { + m_stats.inc_fail_stats(mf->get_access_type(), LINE_ALLOC_FAIL); } m_stats.inc_stats(mf->get_access_type(), m_stats.select_stats_status(status, cache_status)); @@ -1047,10 +1526,13 @@ data_cache::process_tag_probe( bool wr, access_status = (this->*m_wr_hit)( addr, cache_index, mf, time, events, probe_status ); - }else if ( probe_status != RESERVATION_FAIL ) { + }else if ( (probe_status != RESERVATION_FAIL) || (probe_status == RESERVATION_FAIL && m_config.m_write_alloc_policy == NO_WRITE_ALLOCATE) ) { access_status = (this->*m_wr_miss)( addr, cache_index, mf, time, events, probe_status ); + }else { + //the only reason for reservation fail here is LINE_ALLOC_FAIL (i.e all lines are reserved) + m_stats.inc_fail_stats(mf->get_access_type(), LINE_ALLOC_FAIL); } }else{ // Read if(probe_status == HIT){ @@ -1061,6 +1543,9 @@ data_cache::process_tag_probe( bool wr, access_status = (this->*m_rd_miss)( addr, cache_index, mf, time, events, probe_status ); + }else { + //the only reason for reservation fail here is LINE_ALLOC_FAIL (i.e all lines are reserved) + m_stats.inc_fail_stats(mf->get_access_type(), LINE_ALLOC_FAIL); } } @@ -1080,12 +1565,12 @@ data_cache::access( new_addr_type addr, std::list<cache_event> &events ) { - assert( mf->get_data_size() <= m_config.get_line_sz()); + assert( mf->get_data_size() <= m_config.get_atom_sz()); bool wr = mf->get_is_write(); new_addr_type block_addr = m_config.block_addr(addr); unsigned cache_index = (unsigned)-1; enum cache_request_status probe_status - = m_tag_array->probe( block_addr, cache_index ); + = m_tag_array->probe( block_addr, cache_index, mf, true); enum cache_request_status access_status = process_tag_probe( wr, probe_status, addr, cache_index, mf, time, events ); m_stats.inc_stats(mf->get_access_type(), @@ -1134,7 +1619,7 @@ enum cache_request_status tex_cache::access( new_addr_type addr, mem_fetch *mf, // at this point, we will accept the request : access tags and immediately allocate line new_addr_type block_addr = m_config.block_addr(addr); unsigned cache_index = (unsigned)-1; - enum cache_request_status status = m_tags.access(block_addr,time,cache_index); + enum cache_request_status status = m_tags.access(block_addr,time,cache_index,mf); enum cache_request_status cache_status = RESERVATION_FAIL; assert( status != RESERVATION_FAIL ); assert( status != HIT_RESERVED ); // as far as tags are concerned: HIT or MISS @@ -1142,12 +1627,12 @@ enum cache_request_status tex_cache::access( new_addr_type addr, mem_fetch *mf, if ( status == MISS ) { // we need to send a memory request... unsigned rob_index = m_rob.push( rob_entry(cache_index, mf, block_addr) ); - m_extra_mf_fields[mf] = extra_mf_fields(rob_index); + m_extra_mf_fields[mf] = extra_mf_fields(rob_index, m_config); mf->set_data_size(m_config.get_line_sz()); - m_tags.fill(cache_index,time); // mark block as valid + m_tags.fill(cache_index,time,mf); // mark block as valid m_request_fifo.push(mf); mf->set_status(m_request_queue_status,time); - events.push_back(READ_REQUEST_SENT); + events.push_back(cache_event(READ_REQUEST_SENT)); cache_status = MISS; } else { // the value *will* *be* in the cache already @@ -1174,7 +1659,7 @@ void tex_cache::cycle(){ unsigned rob_index = m_rob.next_pop_index(); const rob_entry &r = m_rob.peek(rob_index); assert( r.m_request == e.m_request ); - assert( r.m_block_addr == m_config.block_addr(e.m_request->get_addr()) ); + //assert( r.m_block_addr == m_config.block_addr(e.m_request->get_addr()) ); if ( r.m_ready ) { assert( r.m_index == e.m_cache_index ); m_cache[r.m_index].m_valid = true; @@ -1197,6 +1682,23 @@ void tex_cache::cycle(){ /// Place returning cache block into reorder buffer void tex_cache::fill( mem_fetch *mf, unsigned time ) { + if(m_config.m_mshr_type == SECTOR_TEX_FIFO) { + assert(mf->get_original_mf()); + extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf->get_original_mf()); + assert( e != m_extra_mf_fields.end() ); + e->second.pending_read--; + + if(e->second.pending_read > 0) { + //wait for the other requests to come back + delete mf; + return; + } else { + mem_fetch *temp = mf; + mf = mf->get_original_mf(); + delete temp; + } + } + extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf); assert( e != m_extra_mf_fields.end() ); assert( e->second.m_valid ); diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 7535a1d..e663cf6 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -36,9 +36,12 @@ #include "../tr1_hash_map.h" #include "addrdec.h" +#include <iostream> + +#define MAX_DEFAULT_CACHE_SIZE_MULTIBLIER 4 enum cache_block_state { - INVALID, + INVALID=0, RESERVED, VALID, MODIFIED @@ -49,13 +52,51 @@ enum cache_request_status { HIT_RESERVED, MISS, RESERVATION_FAIL, + SECTOR_MISS, NUM_CACHE_REQUEST_STATUS }; -enum cache_event { +enum cache_reservation_fail_reason { + LINE_ALLOC_FAIL= 0,// all line are reserved + MISS_QUEUE_FULL, // MISS queue (i.e. interconnect or DRAM) is full + MSHR_ENRTY_FAIL, + MSHR_MERGE_ENRTY_FAIL, + MSHR_RW_PENDING, + NUM_CACHE_RESERVATION_FAIL_STATUS +}; + +enum cache_event_type { WRITE_BACK_REQUEST_SENT, READ_REQUEST_SENT, - WRITE_REQUEST_SENT + WRITE_REQUEST_SENT, + WRITE_ALLOCATE_SENT +}; + +struct evicted_block_info { + new_addr_type m_block_addr; + unsigned m_modified_size; + evicted_block_info() { + m_block_addr = 0; + m_modified_size = 0; + } + void set_info(new_addr_type block_addr, unsigned modified_size){ + m_block_addr = block_addr; + m_modified_size = modified_size; + } +}; + +struct cache_event { + enum cache_event_type m_cache_event_type; + evicted_block_info m_evicted_block; //if it was write_back event, fill the the evicted block info + + cache_event(enum cache_event_type m_cache_event){ + m_cache_event_type = m_cache_event; + } + + cache_event(enum cache_event_type cache_event, evicted_block_info evicted_block){ + m_cache_event_type = cache_event; + m_evicted_block = evicted_block; + } }; const char * cache_request_status_str(enum cache_request_status status); @@ -65,33 +106,340 @@ struct cache_block_t { { m_tag=0; m_block_addr=0; - m_alloc_time=0; - m_fill_time=0; - m_last_access_time=0; - m_status=INVALID; } - void allocate( new_addr_type tag, new_addr_type block_addr, unsigned time ) + + virtual void allocate( new_addr_type tag, new_addr_type block_addr, unsigned time, mem_access_sector_mask_t sector_mask) = 0; + virtual void fill( unsigned time, mem_access_sector_mask_t sector_mask) = 0; + + virtual bool is_invalid_line() = 0; + virtual bool is_valid_line() = 0; + virtual bool is_reserved_line() = 0; + virtual bool is_modified_line() = 0; + + virtual enum cache_block_state get_status( mem_access_sector_mask_t sector_mask) = 0; + virtual void set_status(enum cache_block_state m_status, mem_access_sector_mask_t sector_mask) = 0; + + virtual unsigned get_last_access_time() = 0; + virtual void set_last_access_time(unsigned time, mem_access_sector_mask_t sector_mask) = 0; + virtual unsigned get_alloc_time() = 0; + virtual void set_ignore_on_fill(bool m_ignore, mem_access_sector_mask_t sector_mask) = 0; + virtual void set_modified_on_fill(bool m_modified, mem_access_sector_mask_t sector_mask) = 0; + virtual unsigned get_modified_size() = 0; + virtual void set_m_readable(bool readable, mem_access_sector_mask_t sector_mask)=0; + virtual bool is_readable(mem_access_sector_mask_t sector_mask)=0; + virtual void print_status()=0; + virtual ~cache_block_t() {} + + + new_addr_type m_tag; + new_addr_type m_block_addr; + +}; + +struct line_cache_block: public cache_block_t { + line_cache_block() + { + m_alloc_time=0; + m_fill_time=0; + m_last_access_time=0; + m_status=INVALID; + m_ignore_on_fill_status = false; + m_set_modified_on_fill = false; + m_readable = true; + } + void allocate( new_addr_type tag, new_addr_type block_addr, unsigned time, mem_access_sector_mask_t sector_mask) + { + m_tag=tag; + m_block_addr=block_addr; + m_alloc_time=time; + m_last_access_time=time; + m_fill_time=0; + m_status=RESERVED; + m_ignore_on_fill_status = false; + m_set_modified_on_fill = false; + } + void fill( unsigned time, mem_access_sector_mask_t sector_mask ) + { + //if(!m_ignore_on_fill_status) + // assert( m_status == RESERVED ); + + m_status = m_set_modified_on_fill? MODIFIED : VALID; + + m_fill_time=time; + } + virtual bool is_invalid_line() + { + return m_status == INVALID; + } + virtual bool is_valid_line() + { + return m_status == VALID; + } + virtual bool is_reserved_line() + { + return m_status == RESERVED; + } + virtual bool is_modified_line() + { + return m_status == MODIFIED; + } + + virtual enum cache_block_state get_status(mem_access_sector_mask_t sector_mask) + { + return m_status; + } + virtual void set_status(enum cache_block_state status, mem_access_sector_mask_t sector_mask) + { + m_status = status; + } + virtual unsigned get_last_access_time() + { + return m_last_access_time; + } + virtual void set_last_access_time(unsigned time, mem_access_sector_mask_t sector_mask) + { + m_last_access_time = time; + } + virtual unsigned get_alloc_time() + { + return m_alloc_time; + } + virtual void set_ignore_on_fill(bool m_ignore, mem_access_sector_mask_t sector_mask) + { + m_ignore_on_fill_status = m_ignore; + } + virtual void set_modified_on_fill(bool m_modified, mem_access_sector_mask_t sector_mask) + { + m_set_modified_on_fill = m_modified; + } + virtual unsigned get_modified_size() + { + return SECTOR_CHUNCK_SIZE * SECTOR_SIZE; //i.e. cache line size + } + virtual void set_m_readable(bool readable, mem_access_sector_mask_t sector_mask) + { + m_readable = readable; + } + virtual bool is_readable(mem_access_sector_mask_t sector_mask) { + return m_readable; + } + virtual void print_status() { + printf("m_block_addr is %llu, status = %u\n", m_block_addr, m_status); + } + + +private: + unsigned m_alloc_time; + unsigned m_last_access_time; + unsigned m_fill_time; + cache_block_state m_status; + bool m_ignore_on_fill_status; + bool m_set_modified_on_fill; + bool m_readable; +}; + +struct sector_cache_block : public cache_block_t { + sector_cache_block() + { + init(); + } + + void init() { + for(unsigned i =0; i< SECTOR_CHUNCK_SIZE; ++i) { + m_sector_alloc_time[i]= 0; + m_sector_fill_time[i]= 0; + m_last_sector_access_time[i]= 0; + m_status[i]= INVALID; + m_ignore_on_fill_status[i] = false; + m_set_modified_on_fill[i] = false; + m_readable[i] = true; + } + m_line_alloc_time=0; + m_line_last_access_time=0; + m_line_fill_time=0; + } + + virtual void allocate( new_addr_type tag, new_addr_type block_addr, unsigned time, mem_access_sector_mask_t sector_mask ) { - m_tag=tag; - m_block_addr=block_addr; - m_alloc_time=time; - m_last_access_time=time; - m_fill_time=0; - m_status=RESERVED; + allocate_line( tag, block_addr, time, sector_mask ); } - void fill( unsigned time ) + + void allocate_line( new_addr_type tag, new_addr_type block_addr, unsigned time, mem_access_sector_mask_t sector_mask ) + { + //allocate a new line + //assert(m_block_addr != 0 && m_block_addr != block_addr); + init(); + m_tag=tag; + m_block_addr=block_addr; + + unsigned sidx = get_sector_index(sector_mask); + + //set sector stats + m_sector_alloc_time[sidx]=time; + m_last_sector_access_time[sidx]=time; + m_sector_fill_time[sidx]=0; + m_status[sidx]=RESERVED; + m_ignore_on_fill_status[sidx] = false; + m_set_modified_on_fill[sidx] = false; + + //set line stats + m_line_alloc_time=time; //only set this for the first allocated sector + m_line_last_access_time=time; + m_line_fill_time=0; + } + + void allocate_sector(unsigned time, mem_access_sector_mask_t sector_mask ) + { + //allocate invalid sector of this allocated valid line + assert(is_valid_line()); + unsigned sidx = get_sector_index(sector_mask); + + //set sector stats + m_sector_alloc_time[sidx]=time; + m_last_sector_access_time[sidx]=time; + m_sector_fill_time[sidx]=0; + if(m_status[sidx]==MODIFIED) //this should be the case only for fetch-on-write policy //TO DO + m_set_modified_on_fill[sidx] = true; + else + m_set_modified_on_fill[sidx] = false; + + m_status[sidx]=RESERVED; + m_ignore_on_fill_status[sidx] = false; + //m_set_modified_on_fill[sidx] = false; + m_readable[sidx] = true; + + //set line stats + m_line_last_access_time=time; + m_line_fill_time=0; + } + + virtual void fill( unsigned time, mem_access_sector_mask_t sector_mask) { - assert( m_status == RESERVED ); - m_status=VALID; - m_fill_time=time; + unsigned sidx = get_sector_index(sector_mask); + + // if(!m_ignore_on_fill_status[sidx]) + // assert( m_status[sidx] == RESERVED ); + + m_status[sidx] = m_set_modified_on_fill[sidx]? MODIFIED : VALID; + + m_sector_fill_time[sidx]=time; + m_line_fill_time=time; + } + virtual bool is_invalid_line() { + //all the sectors should be invalid + for(unsigned i =0; i< SECTOR_CHUNCK_SIZE; ++i) { + if (m_status[i] != INVALID) + return false; + } + return true; + } + virtual bool is_valid_line() { return !(is_invalid_line()); } + virtual bool is_reserved_line() { + //if any of the sector is reserved, then the line is reserved + for(unsigned i =0; i< SECTOR_CHUNCK_SIZE; ++i) { + if (m_status[i] == RESERVED) + return true; + } + return false; + } + virtual bool is_modified_line() { + //if any of the sector is modified, then the line is modified + for(unsigned i =0; i< SECTOR_CHUNCK_SIZE; ++i) { + if (m_status[i] == MODIFIED) + return true; + } + return false; } - new_addr_type m_tag; - new_addr_type m_block_addr; - unsigned m_alloc_time; - unsigned m_last_access_time; - unsigned m_fill_time; - cache_block_state m_status; + virtual enum cache_block_state get_status(mem_access_sector_mask_t sector_mask) + { + unsigned sidx = get_sector_index(sector_mask); + + return m_status[sidx]; + } + + virtual void set_status(enum cache_block_state status, mem_access_sector_mask_t sector_mask) + { + unsigned sidx = get_sector_index(sector_mask); + m_status[sidx] = status; + } + + virtual unsigned get_last_access_time() + { + return m_line_last_access_time; + } + + virtual void set_last_access_time(unsigned time, mem_access_sector_mask_t sector_mask) + { + unsigned sidx = get_sector_index(sector_mask); + + m_last_sector_access_time[sidx] = time; + m_line_last_access_time = time; + } + + virtual unsigned get_alloc_time() + { + return m_line_alloc_time; + } + + virtual void set_ignore_on_fill(bool m_ignore, mem_access_sector_mask_t sector_mask) + { + unsigned sidx = get_sector_index(sector_mask); + m_ignore_on_fill_status[sidx] = m_ignore; + } + + virtual void set_modified_on_fill(bool m_modified, mem_access_sector_mask_t sector_mask) + { + unsigned sidx = get_sector_index(sector_mask); + m_set_modified_on_fill[sidx] = m_modified; + } + + virtual void set_m_readable(bool readable, mem_access_sector_mask_t sector_mask) + { + unsigned sidx = get_sector_index(sector_mask); + m_readable[sidx] = readable; + } + + virtual bool is_readable(mem_access_sector_mask_t sector_mask) { + unsigned sidx = get_sector_index(sector_mask); + return m_readable[sidx]; + } + + virtual unsigned get_modified_size() + { + unsigned modified=0; + for(unsigned i =0; i< SECTOR_CHUNCK_SIZE; ++i) { + if (m_status[i] == MODIFIED) + modified++; + } + return modified * SECTOR_SIZE; + } + + virtual void print_status() { + printf("m_block_addr is %llu, status = %u %u %u %u\n", m_block_addr, m_status[0], m_status[1], m_status[2], m_status[3]); + } + + +private: + unsigned m_sector_alloc_time[SECTOR_CHUNCK_SIZE]; + unsigned m_last_sector_access_time[SECTOR_CHUNCK_SIZE]; + unsigned m_sector_fill_time[SECTOR_CHUNCK_SIZE]; + unsigned m_line_alloc_time; + unsigned m_line_last_access_time; + unsigned m_line_fill_time; + cache_block_state m_status[SECTOR_CHUNCK_SIZE]; + bool m_ignore_on_fill_status[SECTOR_CHUNCK_SIZE]; + bool m_set_modified_on_fill[SECTOR_CHUNCK_SIZE]; + bool m_readable[SECTOR_CHUNCK_SIZE]; + + unsigned get_sector_index(mem_access_sector_mask_t sector_mask) + { + assert(sector_mask.count() == 1); + for(unsigned i =0; i< SECTOR_CHUNCK_SIZE; ++i) { + if(sector_mask.to_ulong() & (1<<i)) + return i; + } + } }; enum replacement_policy_t { @@ -109,26 +457,44 @@ enum write_policy_t { enum allocation_policy_t { ON_MISS, - ON_FILL + ON_FILL, + STREAMING }; enum write_allocate_policy_t { NO_WRITE_ALLOCATE, - WRITE_ALLOCATE + WRITE_ALLOCATE, + FETCH_ON_WRITE, + LAZY_FETCH_ON_READ }; enum mshr_config_t { - TEX_FIFO, - ASSOC // normal cache + TEX_FIFO, // Tex cache + ASSOC, // normal cache + SECTOR_TEX_FIFO, //Tex cache sends requests to high-level sector cache + SECTOR_ASSOC // normal cache sends requests to high-level sector cache }; enum set_index_function{ - FERMI_HASH_SET_FUNCTION = 0, - LINEAR_SET_FUNCTION, + LINEAR_SET_FUNCTION = 0, + BITWISE_XORING_FUNCTION, + HASH_IPOLY_FUNCTION, + FERMI_HASH_SET_FUNCTION, CUSTOM_SET_FUNCTION }; +enum cache_type{ + NORMAL = 0, + SECTOR +}; + +#define MAX_WARP_PER_SHADER 64 +#define INCT_TOTAL_BUFFER 64 +#define L2_TOTAL 64 +#define MAX_WARP_PER_SHADER 64 +#define MAX_WARP_PER_SHADER 64 + class cache_config { public: cache_config() @@ -140,27 +506,39 @@ public: m_config_stringPrefShared = NULL; m_data_port_width = 0; m_set_index_function = LINEAR_SET_FUNCTION; + m_is_streaming = false; } void init(char * config, FuncCache status) { cache_status= status; assert( config ); - char rp, wp, ap, mshr_type, wap, sif; + char ct, rp, wp, ap, mshr_type, wap, sif; - int ntok = sscanf(config,"%u:%u:%u,%c:%c:%c:%c:%c,%c:%u:%u,%u:%u,%u", - &m_nset, &m_line_sz, &m_assoc, &rp, &wp, &ap, &wap, + int ntok = sscanf(config,"%c:%u:%u:%u,%c:%c:%c:%c:%c,%c:%u:%u,%u:%u,%u", + &ct, &m_nset, &m_line_sz, &m_assoc, &rp, &wp, &ap, &wap, &sif,&mshr_type,&m_mshr_entries,&m_mshr_max_merge, &m_miss_queue_size, &m_result_fifo_entries, &m_data_port_width); - if ( ntok < 11 ) { + if ( ntok < 12 ) { if ( !strcmp(config,"none") ) { m_disabled = true; return; } exit_parse_error(); } + + switch (ct) { + case 'N': m_cache_type = NORMAL; break; + case 'S': m_cache_type = SECTOR; break; + default: exit_parse_error(); + } + switch (rp) { + case 'L': m_replacement_policy = LRU; break; + case 'F': m_replacement_policy = FIFO; break; + default: exit_parse_error(); + } switch (rp) { case 'L': m_replacement_policy = LRU; break; case 'F': m_replacement_policy = FIFO; break; @@ -177,21 +555,46 @@ public: switch (ap) { case 'm': m_alloc_policy = ON_MISS; break; case 'f': m_alloc_policy = ON_FILL; break; + case 's': m_alloc_policy = STREAMING; break; default: exit_parse_error(); } + if(m_alloc_policy == STREAMING) { + //For streaming cache, we set the alloc policy to be on-fill to remove all line_alloc_fail stalls + //we set the MSHRs to be equal to max allocated cache lines. This is possible by moving TAG to be shared between cache line and MSHR enrty (i.e. for each cache line, there is an MSHR rntey associated with it) + //This is the easiest think we can think about to model (mimic) L1 streaming cache in Pascal and Volta + //Based on our microbenchmakrs, MSHRs entries have been increasing substantially in Pascal and Volta + //For more information about streaming cache, see: + // http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + // https://ieeexplore.ieee.org/document/8344474/ + m_is_streaming = true; + m_alloc_policy = ON_FILL; + m_mshr_entries = m_nset*m_assoc*MAX_DEFAULT_CACHE_SIZE_MULTIBLIER; + if(m_cache_type == SECTOR) + m_mshr_entries *= SECTOR_CHUNCK_SIZE; + m_mshr_max_merge = MAX_WARP_PER_SM; + } switch (mshr_type) { - case 'F': m_mshr_type = TEX_FIFO; assert(ntok==13); break; + case 'F': m_mshr_type = TEX_FIFO; assert(ntok==14); break; + case 'T': m_mshr_type = SECTOR_TEX_FIFO; assert(ntok==14); break; case 'A': m_mshr_type = ASSOC; break; + case 'S' : m_mshr_type = SECTOR_ASSOC; break; default: exit_parse_error(); } m_line_sz_log2 = LOGB2(m_line_sz); m_nset_log2 = LOGB2(m_nset); m_valid = true; + m_atom_sz = (m_cache_type == SECTOR)? SECTOR_SIZE : m_line_sz; + original_m_assoc = m_assoc; + //For more details about difference between FETCH_ON_WRITE and WRITE VALIDAE policies + //Read: Jouppi, Norman P. "Cache write policies and performance". ISCA 93. + //WRITE_ALLOCATE is the old write policy in GPGPU-sim 3.x, that send WRITE and READ for every write request switch(wap){ - case 'W': m_write_alloc_policy = WRITE_ALLOCATE; break; case 'N': m_write_alloc_policy = NO_WRITE_ALLOCATE; break; - default: exit_parse_error(); + case 'W': m_write_alloc_policy = WRITE_ALLOCATE; break; + case 'F': m_write_alloc_policy = FETCH_ON_WRITE; break; + case 'L': m_write_alloc_policy = LAZY_FETCH_ON_READ; break; + default: exit_parse_error(); } // detect invalid configuration @@ -208,6 +611,15 @@ public: assert(0 && "Invalid cache configuration: Writeback cache cannot allocate new line on fill. "); } + if((m_write_alloc_policy == FETCH_ON_WRITE || m_write_alloc_policy == LAZY_FETCH_ON_READ )&& m_alloc_policy == ON_FILL) + { + assert(0 && "Invalid cache configuration: FETCH_ON_WRITE and LAZY_FETCH_ON_READ cannot work properly with ON_FILL policy. Cache must be ON_MISS. "); + } + if(m_cache_type == SECTOR) + { + assert(m_line_sz / SECTOR_SIZE == SECTOR_CHUNCK_SIZE && m_line_sz % SECTOR_SIZE == 0); + } + // default: port to data array width and granularity = line size if (m_data_port_width == 0) { m_data_port_width = m_line_sz; @@ -216,6 +628,7 @@ public: switch(sif){ case 'H': m_set_index_function = FERMI_HASH_SET_FUNCTION; break; + case 'P': m_set_index_function = HASH_IPOLY_FUNCTION; break; case 'C': m_set_index_function = CUSTOM_SET_FUNCTION; break; case 'L': m_set_index_function = LINEAR_SET_FUNCTION; break; default: exit_parse_error(); @@ -227,12 +640,21 @@ public: assert( m_valid ); return m_line_sz; } + unsigned get_atom_sz() const + { + assert( m_valid ); + return m_atom_sz; + } unsigned get_num_lines() const { assert( m_valid ); return m_nset * m_assoc; } - + unsigned get_max_num_lines() const + { + assert( m_valid ); + return MAX_DEFAULT_CACHE_SIZE_MULTIBLIER * m_nset * original_m_assoc; + } void print( FILE *fp ) const { fprintf( fp, "Size = %d B (%d Set x %d-way x %d byte line)\n", @@ -265,6 +687,32 @@ public: { return addr & ~(m_line_sz-1); } + new_addr_type mshr_addr( new_addr_type addr ) const + { + return addr & ~(m_atom_sz-1); + } + enum mshr_config_t get_mshr_type() const + { + return m_mshr_type; + } + void set_assoc(unsigned n) + { + //set new assoc. L1 cache dynamically resized in Volta + m_assoc = n; + } + unsigned get_nset() const + { + assert( m_valid ); + return m_nset; + } + unsigned get_total_size_inKB() const + { + assert( m_valid ); + return (m_assoc*m_nset*m_line_sz)/1024; + } + bool is_streaming() { + return m_is_streaming; + } FuncCache get_cache_status() {return cache_status;} char *m_config_string; char *m_config_stringPrefL1; @@ -285,11 +733,15 @@ protected: unsigned m_nset; unsigned m_nset_log2; unsigned m_assoc; + unsigned m_atom_sz; + unsigned original_m_assoc; + bool m_is_streaming; enum replacement_policy_t m_replacement_policy; // 'L' = LRU, 'F' = FIFO enum write_policy_t m_write_policy; // 'T' = write through, 'B' = write back, 'R' = read only enum allocation_policy_t m_alloc_policy; // 'm' = allocate on miss, 'f' = allocate on fill enum mshr_config_t m_mshr_type; + enum cache_type m_cache_type; write_allocate_policy_t m_write_alloc_policy; // 'W' = Write allocate, 'N' = No write allocate @@ -316,12 +768,14 @@ protected: friend class data_cache; friend class l1_cache; friend class l2_cache; + friend class memory_sub_partition; }; class l1d_cache_config : public cache_config{ public: l1d_cache_config() : cache_config(){} virtual unsigned set_index(new_addr_type addr) const; + unsigned l1_latency; }; class l2_cache_config : public cache_config { @@ -340,17 +794,20 @@ public: tag_array(cache_config &config, int core_id, int type_id ); ~tag_array(); - enum cache_request_status probe( new_addr_type addr, unsigned &idx ) const; - enum cache_request_status access( new_addr_type addr, unsigned time, unsigned &idx ); - enum cache_request_status access( new_addr_type addr, unsigned time, unsigned &idx, bool &wb, cache_block_t &evicted ); + enum cache_request_status probe( new_addr_type addr, unsigned &idx, mem_fetch* mf, bool probe_mode=false ) const; + enum cache_request_status probe( new_addr_type addr, unsigned &idx, mem_access_sector_mask_t mask, bool probe_mode=false, mem_fetch* mf = NULL ) const; + enum cache_request_status access( new_addr_type addr, unsigned time, unsigned &idx, mem_fetch* mf ); + enum cache_request_status access( new_addr_type addr, unsigned time, unsigned &idx, bool &wb, evicted_block_info &evicted, mem_fetch* mf ); - void fill( new_addr_type addr, unsigned time ); - void fill( unsigned idx, unsigned time ); + void fill( new_addr_type addr, unsigned time, mem_fetch* mf ); + void fill( unsigned idx, unsigned time, mem_fetch* mf ); + void fill( new_addr_type addr, unsigned time, mem_access_sector_mask_t mask ); unsigned size() const { return m_config.get_num_lines();} - cache_block_t &get_block(unsigned idx) { return m_lines[idx];} + cache_block_t* get_block(unsigned idx) { return m_lines[idx];} - void flush(); // flash invalidate all entries + void flush(); // flush all written entries + void invalidate(); // invalidate all entries void new_window(); void print( FILE *stream, unsigned &total_access, unsigned &total_misses ) const; @@ -358,6 +815,8 @@ public: void get_stats(unsigned &total_access, unsigned &total_misses, unsigned &total_hit_res, unsigned &total_res_fail) const; void update_cache_parameters(cache_config &config); + void add_pending_line(mem_fetch *mf); + void remove_pending_line(mem_fetch *mf); protected: // This constructor is intended for use only from derived classes that wish to // avoid unnecessary memory allocation that takes place in the @@ -365,19 +824,20 @@ protected: tag_array( cache_config &config, int core_id, int type_id, - cache_block_t* new_lines ); + cache_block_t** new_lines ); void init( int core_id, int type_id ); protected: cache_config &m_config; - cache_block_t *m_lines; /* nbanks x nset x assoc lines in total */ + cache_block_t **m_lines; /* nbanks x nset x assoc lines in total */ unsigned m_access; unsigned m_miss; unsigned m_pending_hit; // number of cache miss that hit a line that is allocated but not filled unsigned m_res_fail; + unsigned m_sector_miss; // performance counters for calculating the amount of misses within a time window unsigned m_prev_snapshot_access; @@ -386,11 +846,16 @@ protected: int m_core_id; // which shader core is using this int m_type_id; // what kind of cache is this (normal, texture, constant) + + bool is_used; //a flag if the whole cache has ever been accessed before + + typedef tr1_hash_map<new_addr_type,unsigned> line_table; + line_table pending_lines; }; class mshr_table { public: - mshr_table( unsigned num_entries, unsigned max_merged ) + mshr_table( unsigned num_entries, unsigned max_merged) : m_num_entries(num_entries), m_max_merged(max_merged) #if (tr1_hash_map_ismap == 0) @@ -414,6 +879,8 @@ public: /// Returns next ready access mem_fetch *next_access(); void display( FILE *fp ) const; + // Returns true if there is a pending read after write + bool is_read_after_write_pending(new_addr_type block_addr); void check_mshr_parameters( unsigned num_entries, unsigned max_merged ) { @@ -433,7 +900,9 @@ private: mshr_entry() : m_has_atomic(false) { } }; typedef tr1_hash_map<new_addr_type,mshr_entry> table; + typedef tr1_hash_map<new_addr_type,mshr_entry> line_table; table m_data; + line_table pending_lines; // it may take several cycles to process the merged requests bool m_current_response_ready; @@ -510,12 +979,14 @@ public: cache_stats(); void clear(); void inc_stats(int access_type, int access_outcome); + void inc_fail_stats(int access_type, int fail_outcome); enum cache_request_status select_stats_status(enum cache_request_status probe, enum cache_request_status access) const; - unsigned &operator()(int access_type, int access_outcome); - unsigned operator()(int access_type, int access_outcome) const; + unsigned &operator()(int access_type, int access_outcome, bool fail_outcome); + unsigned operator()(int access_type, int access_outcome, bool fail_outcome) const; cache_stats operator+(const cache_stats &cs); cache_stats &operator+=(const cache_stats &cs); void print_stats(FILE *fout, const char *cache_name = "Cache_stats") const; + void print_fail_stats(FILE *fout, const char *cache_name = "Cache_fail_stats") const; unsigned get_stats(enum mem_access_type *access_type, unsigned num_access_type, enum cache_request_status *access_status, unsigned num_access_status) const; void get_sub_stats(struct cache_sub_stats &css) const; @@ -523,8 +994,10 @@ public: void sample_cache_port_utility(bool data_port_busy, bool fill_port_busy); private: bool check_valid(int type, int status) const; + bool check_fail_valid(int type, int fail) const; std::vector< std::vector<unsigned> > m_stats; + std::vector< std::vector<unsigned> > m_fail_stats; unsigned long long m_cache_port_available_cycles; unsigned long long m_cache_data_port_busy_cycles; @@ -543,6 +1016,7 @@ public: bool was_write_sent( const std::list<cache_event> &events ); bool was_read_sent( const std::list<cache_event> &events ); +bool was_writeallocate_sent( const std::list<cache_event> &events ); /// Baseline cache /// Implements common functions for read_only_cache and data_cache @@ -552,7 +1026,7 @@ public: baseline_cache( const char *name, cache_config &config, int core_id, int type_id, mem_fetch_interface *memport, enum mem_fetch_status status ) : m_config(config), m_tag_array(new tag_array(config,core_id,type_id)), - m_mshrs(config.m_mshr_entries,config.m_mshr_max_merge), + m_mshrs(config.m_mshr_entries,config.m_mshr_max_merge), m_bandwidth_management(config) { init( name, config, memport, status ); @@ -564,7 +1038,7 @@ public: enum mem_fetch_status status ) { m_name = name; - assert(config.m_mshr_type == ASSOC); + assert(config.m_mshr_type == ASSOC || config.m_mshr_type == SECTOR_ASSOC); m_memport=memport; m_miss_queue_status = status; } @@ -594,6 +1068,7 @@ public: mem_fetch *next_access(){return m_mshrs.next_access();} // flash invalidate all entries in cache void flush(){m_tag_array->flush();} + void invalidate(){m_tag_array->invalidate();} void print(FILE *fp, unsigned &accesses, unsigned &misses) const; void display_state( FILE *fp ) const; @@ -612,6 +1087,15 @@ public: bool data_port_free() const { return m_bandwidth_management.data_port_free(); } bool fill_port_free() const { return m_bandwidth_management.fill_port_free(); } + // This is a gapping hole we are poking in the system to quickly handle + // filling the cache on cudamemcopies. We don't care about anything other than + // L2 state after the memcopy - so just force the tag array to act as though + // something is read or written without doing anything else. + void force_tag_access( new_addr_type addr, unsigned time, mem_access_sector_mask_t mask ) + { + m_tag_array->fill( addr, time, mask ); + } + protected: // Constructor that can be used by derived classes with custom tag arrays baseline_cache( const char *name, @@ -633,24 +1117,31 @@ protected: std::string m_name; cache_config &m_config; tag_array* m_tag_array; - mshr_table m_mshrs; + mshr_table m_mshrs; std::list<mem_fetch*> m_miss_queue; enum mem_fetch_status m_miss_queue_status; mem_fetch_interface *m_memport; struct extra_mf_fields { extra_mf_fields() { m_valid = false;} - extra_mf_fields( new_addr_type a, unsigned i, unsigned d ) + extra_mf_fields( new_addr_type a, new_addr_type ad, unsigned i, unsigned d, const cache_config& m_config) { m_valid = true; m_block_addr = a; + m_addr = ad; m_cache_index = i; m_data_size = d; + pending_read = m_config.m_mshr_type == SECTOR_ASSOC? m_config.m_line_sz/SECTOR_SIZE : 0; + } bool m_valid; new_addr_type m_block_addr; + new_addr_type m_addr; unsigned m_cache_index; unsigned m_data_size; + //this variable is used when a load request generates multiple load transactions + //For example, a read request from non-sector L1 request sends a request to sector L2 + unsigned pending_read; }; typedef std::map<mem_fetch*,extra_mf_fields> extra_mf_fields_lookup; @@ -668,7 +1159,7 @@ protected: unsigned time, bool &do_miss, std::list<cache_event> &events, bool read_only, bool wa); /// Read miss handler. Check MSHR hit or MSHR available void send_read_request(new_addr_type addr, new_addr_type block_addr, unsigned cache_index, mem_fetch *mf, - unsigned time, bool &do_miss, bool &wb, cache_block_t &evicted, std::list<cache_event> &events, bool read_only, bool wa); + unsigned time, bool &do_miss, bool &wb, evicted_block_info &evicted, std::list<cache_event> &events, bool read_only, bool wa); /// Sub-class containing all metadata for port bandwidth management class bandwidth_management @@ -760,8 +1251,10 @@ public: // Set write miss function switch(m_config.m_write_alloc_policy){ - case WRITE_ALLOCATE: m_wr_miss = &data_cache::wr_miss_wa; break; case NO_WRITE_ALLOCATE: m_wr_miss = &data_cache::wr_miss_no_wa; break; + case WRITE_ALLOCATE: m_wr_miss = &data_cache::wr_miss_wa_naive; break; + case FETCH_ON_WRITE: m_wr_miss = &data_cache::wr_miss_wa_fetch_on_write; break; + case LAZY_FETCH_ON_READ: m_wr_miss = &data_cache::wr_miss_wa_lazy_fetch_on_read; break; default: assert(0 && "Error: Must set valid cache write miss policy\n"); break; // Need to set a write miss function @@ -870,12 +1363,33 @@ protected: /// Sends read request, and possible write-back request, // to lower level memory for a write miss with write-allocate enum cache_request_status - wr_miss_wa( new_addr_type addr, - unsigned cache_index, - mem_fetch *mf, - unsigned time, - std::list<cache_event> &events, - enum cache_request_status status ); // write-allocate + wr_miss_wa_naive( new_addr_type addr, + unsigned cache_index, + mem_fetch *mf, + unsigned time, + std::list<cache_event> &events, + enum cache_request_status status ); // write-allocate-send-write-and-read-request + enum cache_request_status + wr_miss_wa_fetch_on_write( new_addr_type addr, + unsigned cache_index, + mem_fetch *mf, + unsigned time, + std::list<cache_event> &events, + enum cache_request_status status ); // write-allocate with fetch-on-every-write + enum cache_request_status + wr_miss_wa_lazy_fetch_on_read( new_addr_type addr, + unsigned cache_index, + mem_fetch *mf, + unsigned time, + std::list<cache_event> &events, + enum cache_request_status status ); // write-allocate with read-fetch-only + enum cache_request_status + wr_miss_wa_write_validate( new_addr_type addr, + unsigned cache_index, + mem_fetch *mf, + unsigned time, + std::list<cache_event> &events, + enum cache_request_status status ); // write-allocate that writes with no read fetch enum cache_request_status wr_miss_no_wa( new_addr_type addr, unsigned cache_index, @@ -991,7 +1505,7 @@ public: m_result_fifo(config.m_result_fifo_entries) { m_name = name; - assert(config.m_mshr_type == TEX_FIFO); + assert(config.m_mshr_type == TEX_FIFO || config.m_mshr_type == SECTOR_TEX_FIFO ); assert(config.m_write_policy == READ_ONLY); assert(config.m_alloc_policy == ON_MISS); m_memport=memport; @@ -1144,13 +1658,15 @@ private: struct extra_mf_fields { extra_mf_fields() { m_valid = false;} - extra_mf_fields( unsigned i ) + extra_mf_fields( unsigned i, const cache_config &m_config ) { m_valid = true; m_rob_index = i; + pending_read = m_config.m_mshr_type == SECTOR_TEX_FIFO? m_config.m_line_sz/SECTOR_SIZE : 0; } bool m_valid; unsigned m_rob_index; + unsigned pending_read; }; cache_stats m_stats; diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 63ba759..a8be4d2 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -150,6 +150,8 @@ void power_config::reg_options(class OptionParser * opp) void memory_config::reg_options(class OptionParser * opp) { + option_parser_register(opp, "-perf_sim_memcpy", OPT_BOOL, &m_perf_sim_memcpy, + "Fill the L2 cache on memcpy", "1"); option_parser_register(opp, "-gpgpu_dram_scheduler", OPT_INT32, &scheduler_type, "0 = fifo, 1 = FR-FCFS (defaul)", "1"); option_parser_register(opp, "-gpgpu_dram_partition_queues", OPT_CSTR, &gpgpu_L2_queue_config, @@ -202,7 +204,27 @@ void memory_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-dram_latency", OPT_UINT32, &dram_latency, "DRAM latency (default 30)", "30"); - + option_parser_register(opp, "-dual_bus_interface", OPT_UINT32, &dual_bus_interface, + "dual_bus_interface (default = 0) ", + "0"); + option_parser_register(opp, "-dram_bnk_indexing_policy", OPT_UINT32, &dram_bnk_indexing_policy, + "dram_bnk_indexing_policy (0 = normal indexing, 1 = Xoring with the higher bits) (Default = 0)", + "0"); + option_parser_register(opp, "-dram_bnkgrp_indexing_policy", OPT_UINT32, &dram_bnkgrp_indexing_policy, + "dram_bnkgrp_indexing_policy (0 = take higher bits, 1 = take lower bits) (Default = 0)", + "0"); + option_parser_register(opp, "-Seperate_Write_Queue_Enable", OPT_BOOL, &seperate_write_queue_enabled, + "Seperate_Write_Queue_Enable", + "0"); + option_parser_register(opp, "-Write_Queue_Size", OPT_CSTR, &write_queue_size_opt, + "Write_Queue_Size", + "32:28:16"); + option_parser_register(opp, "-Elimnate_rw_turnaround", OPT_BOOL, &elimnate_rw_turnaround, + "elimnate_rw_turnaround i.e set tWTR and tRTW = 0", + "0"); + option_parser_register(opp, "-icnt_flit_size", OPT_UINT32, &icnt_flit_size, + "icnt_flit_size", + "32"); m_address_mapping.addrdec_setoption(opp); } @@ -229,6 +251,12 @@ void shader_core_config::reg_options(class OptionParser * opp) "per-shader L1 data cache config " " {<nsets>:<bsize>:<assoc>,<rep>:<wr>:<alloc>:<wr_alloc>,<mshr>:<N>:<merge>,<mq> | none}", "none" ); + option_parser_register(opp, "-l1_latency", OPT_UINT32, &m_L1D_config.l1_latency, + "L1 Hit Latency", + "0"); + option_parser_register(opp, "-smem_latency", OPT_UINT32, &smem_latency, + "smem Latency", + "3"); option_parser_register(opp, "-gpgpu_cache:dl1PrefL1", OPT_CSTR, &m_L1D_config.m_config_stringPrefL1, "per-shader L1 data cache config " " {<nsets>:<bsize>:<assoc>,<rep>:<wr>:<alloc>:<wr_alloc>,<mshr>:<N>:<merge>,<mq> | none}", @@ -256,6 +284,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shader_registers", OPT_UINT32, &gpgpu_shader_registers, "Number of registers per shader core. Limits number of concurrent CTAs. (default 8192)", "8192"); + option_parser_register(opp, "-gpgpu_ignore_resources_limitation", OPT_BOOL, &gpgpu_ignore_resources_limitation, + "gpgpu_ignore_resources_limitation (default 0)", + "0"); option_parser_register(opp, "-gpgpu_shader_cta", OPT_UINT32, &max_cta_per_core, "Maximum number of concurrent CTAs in shader (default 8)", "8"); @@ -277,6 +308,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shmem_size", OPT_UINT32, &gpgpu_shmem_size, "Size of shared memory per shader core (default 16kB)", "16384"); + option_parser_register(opp, "-adpative_volta_cache_config", OPT_BOOL, &adpative_volta_cache_config, + "adpative_volta_cache_config", + "0"); option_parser_register(opp, "-gpgpu_shmem_size", OPT_UINT32, &gpgpu_shmem_sizeDefault, "Size of shared memory per shader core (default 16kB)", "16384"); @@ -295,6 +329,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shmem_warp_parts", OPT_INT32, &mem_warp_parts, "Number of portions a warp is divided into for shared memory bank conflict check ", "2"); + option_parser_register(opp, "-gpgpu_shmem_warp_parts", OPT_INT32, &mem_warp_parts, + "Number of portions a warp is divided into for shared memory bank conflict check ", + "2"); option_parser_register(opp, "-gpgpu_warpdistro_shader", OPT_INT32, &gpgpu_warpdistro_shader, "Specify which shader core to collect the warp size distribution from", "-1"); @@ -303,7 +340,7 @@ void shader_core_config::reg_options(class OptionParser * opp) "0"); option_parser_register(opp, "-gpgpu_local_mem_map", OPT_BOOL, &gpgpu_local_mem_map, "Mapping from local memory space address to simulated GPU physical address space (default = enabled)", - "1"); + "1"); option_parser_register(opp, "-gpgpu_num_reg_banks", OPT_INT32, &gpgpu_num_reg_banks, "Number of register banks (default = 8)", "8"); @@ -313,6 +350,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_operand_collector_num_units_sp", OPT_INT32, &gpgpu_operand_collector_num_units_sp, "number of collector units (default = 4)", "4"); + option_parser_register(opp, "-gpgpu_operand_collector_num_units_dp", OPT_INT32, &gpgpu_operand_collector_num_units_dp, + "number of collector units (default = 0)", + "0"); option_parser_register(opp, "-gpgpu_operand_collector_num_units_sfu", OPT_INT32, &gpgpu_operand_collector_num_units_sfu, "number of collector units (default = 4)", "4"); @@ -328,6 +368,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_operand_collector_num_in_ports_sp", OPT_INT32, &gpgpu_operand_collector_num_in_ports_sp, "number of collector unit in ports (default = 1)", "1"); + option_parser_register(opp, "-gpgpu_operand_collector_num_in_ports_dp", OPT_INT32, &gpgpu_operand_collector_num_in_ports_dp, + "number of collector unit in ports (default = 0)", + "0"); option_parser_register(opp, "-gpgpu_operand_collector_num_in_ports_sfu", OPT_INT32, &gpgpu_operand_collector_num_in_ports_sfu, "number of collector unit in ports (default = 1)", "1"); @@ -343,6 +386,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_operand_collector_num_out_ports_sp", OPT_INT32, &gpgpu_operand_collector_num_out_ports_sp, "number of collector unit in ports (default = 1)", "1"); + option_parser_register(opp, "-gpgpu_operand_collector_num_out_ports_dp", OPT_INT32, &gpgpu_operand_collector_num_out_ports_dp, + "number of collector unit in ports (default = 0)", + "0"); option_parser_register(opp, "-gpgpu_operand_collector_num_out_ports_sfu", OPT_INT32, &gpgpu_operand_collector_num_out_ports_sfu, "number of collector unit in ports (default = 1)", "1"); @@ -356,27 +402,33 @@ void shader_core_config::reg_options(class OptionParser * opp) "number of collector unit in ports (default = 0)", "0"); option_parser_register(opp, "-gpgpu_coalesce_arch", OPT_INT32, &gpgpu_coalesce_arch, - "Coalescing arch (default = 13, anything else is off for now)", + "Coalescing arch (GT200 = 13, Fermi = 20)", "13"); option_parser_register(opp, "-gpgpu_num_sched_per_core", OPT_INT32, &gpgpu_num_sched_per_core, "Number of warp schedulers per core", "1"); option_parser_register(opp, "-gpgpu_max_insn_issue_per_warp", OPT_INT32, &gpgpu_max_insn_issue_per_warp, - "Max number of instructions that can be issued per warp in one cycle by scheduler", - "2"); + "Max number of instructions that can be issued per warp in one cycle by scheduler (either 1 or 2)", + "2"); + option_parser_register(opp, "-gpgpu_dual_issue_diff_exec_units", OPT_BOOL, &gpgpu_dual_issue_diff_exec_units, + "should dual issue use two different execution unit resources (Default = 1)", + "1"); option_parser_register(opp, "-gpgpu_simt_core_sim_order", OPT_INT32, &simt_core_sim_order, "Select the simulation order of cores in a cluster (0=Fix, 1=Round-Robin)", "1"); option_parser_register(opp, "-gpgpu_pipeline_widths", OPT_CSTR, &pipeline_widths_string, "Pipeline widths " - "ID_OC_SP,ID_OC_SFU,ID_OC_TENSOR_CORE,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_TENSOR_CORE,OC_EX_MEM,EX_WB", - "1,1,1,1,1,1,1,1,1,1" ); + "ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB,ID_OC_TENSOR_CORE,OC_EX_TENSOR_CORE", + "1,1,1,1,1,1,1,1,1,1,1" ); option_parser_register(opp, "-gpgpu_tensor_core_avail", OPT_INT32, &gpgpu_tensor_core_avail, "Tensor Core Available (default=0)", "0"); option_parser_register(opp, "-gpgpu_num_sp_units", OPT_INT32, &gpgpu_num_sp_units, "Number of SP units (default=1)", "1"); + option_parser_register(opp, "-gpgpu_num_dp_units", OPT_INT32, &gpgpu_num_dp_units, + "Number of DP units (default=0)", + "0"); option_parser_register(opp, "-gpgpu_num_sfu_units", OPT_INT32, &gpgpu_num_sfu_units, "Number of SF units (default=1)", "1"); @@ -425,7 +477,6 @@ void gpgpu_sim_config::reg_options(option_parser_t opp) option_parser_register(opp, "-gpgpu_flush_l2_cache", OPT_BOOL, &gpgpu_flush_l2_cache, "Flush L2 cache at the end of each kernel call", "0"); - option_parser_register(opp, "-gpgpu_deadlock_detect", OPT_BOOL, &gpu_deadlock_detect, "Stop the simulation at deadlock (1=on (default), 0=off)", "1"); @@ -669,7 +720,7 @@ gpgpu_sim::gpgpu_sim( const gpgpu_sim_config &config ) m_running_kernels.resize( config.max_concurrent_kernel, NULL ); m_last_issued_kernel = 0; - m_last_cluster_issue = 0; + m_last_cluster_issue = m_shader_config->n_simt_clusters-1; // this causes first launch to use simt cluster 0 *average_pipeline_duty_cycle=0; *active_sms=0; @@ -813,6 +864,7 @@ void gpgpu_sim::update_stats() { partiton_replys_in_parallel_total += partiton_replys_in_parallel; partiton_reqs_in_parallel_util_total += partiton_reqs_in_parallel_util; gpu_tot_sim_cycle_parition_util += gpu_sim_cycle_parition_util ; + gpu_tot_occupancy += gpu_occupancy; gpu_sim_cycle = 0; partiton_reqs_in_parallel = 0; @@ -821,6 +873,7 @@ void gpgpu_sim::update_stats() { gpu_sim_cycle_parition_util = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; + gpu_occupancy = occupancy_stats(); } void gpgpu_sim::print_stats() @@ -994,6 +1047,9 @@ void gpgpu_sim::gpu_print_stat() printf("gpu_tot_sim_insn = %lld\n", gpu_tot_sim_insn+gpu_sim_insn); 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); + printf("gpu_occupancy = %.4f\% \n", gpu_occupancy.get_occ_fraction() * 100); + printf("gpu_tot_occupancy = %.4f\% \n", (gpu_occupancy + gpu_tot_occupancy).get_occ_fraction() * 100); + extern unsigned long long g_max_total_param_size; fprintf(statfout, "max_total_param_size = %llu\n", g_max_total_param_size); @@ -1032,6 +1088,8 @@ void gpgpu_sim::gpu_print_stat() } printf("\nTotal_core_cache_stats:\n"); core_cache_stats.print_stats(stdout, "Total_core_cache_stats_breakdown"); + printf("\nTotal_core_cache_fail_stats:\n"); + core_cache_stats.print_fail_stats(stdout, "Total_core_cache_fail_stats_breakdown"); shader_print_scheduler_stat( stdout, false ); m_shader_stats->print(stdout); @@ -1076,6 +1134,8 @@ void gpgpu_sim::gpu_print_stat() printf("L2_total_cache_reservation_fails = %u\n", total_l2_css.res_fails); printf("L2_total_cache_breakdown:\n"); l2_stats.print_stats(stdout, "L2_cache_stats_breakdown"); + printf("L2_total_cache_reservation_fail_breakdown:\n"); + l2_stats.print_fail_stats(stdout, "L2_cache_stats_fail_breakdown"); total_l2_css.print_port_stats(stdout, "L2_cache"); } } @@ -1429,7 +1489,7 @@ void gpgpu_sim::cycle() if (mf) { unsigned response_size = mf->get_is_write()?mf->get_ctrl_size():mf->size(); if ( ::icnt_has_buffer( m_shader_config->mem2device(i), response_size ) ) { - if (!mf->get_is_write()) + //if (!mf->get_is_write()) mf->set_return_timestamp(gpu_sim_cycle+gpu_tot_sim_cycle); mf->set_status(IN_ICNT_TO_SHADER,gpu_sim_cycle+gpu_tot_sim_cycle); ::icnt_push( m_shader_config->mem2device(i), mf->get_tpc(), mf, response_size ); @@ -1462,12 +1522,14 @@ void gpgpu_sim::cycle() for (unsigned i=0;i<m_memory_config->m_n_mem_sub_partition;i++) { //move memory request from interconnect into memory partition (if not backed up) //Note:This needs to be called in DRAM clock domain if there is no L2 cache in the system - if ( m_memory_sub_partition[i]->full() ) { + //In the worst case, we may need to push SECTOR_CHUNCK_SIZE requests, so ensure you have enough buffer for them + if ( m_memory_sub_partition[i]->full(SECTOR_CHUNCK_SIZE) ) { gpu_stall_dramfull++; } else { mem_fetch* mf = (mem_fetch*) icnt_pop( m_shader_config->mem2device(i) ); m_memory_sub_partition[i]->push( mf, gpu_sim_cycle + gpu_tot_sim_cycle ); - partiton_reqs_in_parallel_per_cycle++; + if(mf) + partiton_reqs_in_parallel_per_cycle++; } m_memory_sub_partition[i]->cache_cycle(gpu_sim_cycle+gpu_tot_sim_cycle); m_memory_sub_partition[i]->accumulate_L2cache_stats(m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX]); @@ -1494,6 +1556,8 @@ void gpgpu_sim::cycle() // Update core icnt/cache stats for GPUWattch m_cluster[i]->get_icnt_stats(m_power_stats->pwr_mem_stat->n_simt_to_mem[CURRENT_STAT_IDX][i], m_power_stats->pwr_mem_stat->n_mem_to_simt[CURRENT_STAT_IDX][i]); m_cluster[i]->get_cache_stats(m_power_stats->pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX]); + m_cluster[i]->get_current_occupancy(gpu_occupancy.aggregate_warp_slot_filled, gpu_occupancy.aggregate_theoretical_warp_slots); + } float temp=0; for (unsigned i=0;i<m_shader_config->num_shader();i++){ @@ -1521,12 +1585,12 @@ void gpgpu_sim::cycle() issue_block2core(); - // Depending on configuration, flush the caches once all of threads are completed. + // Depending on configuration, invalidate the caches once all of threads are completed. int all_threads_complete = 1; if (m_config.gpgpu_flush_l1_cache) { for (unsigned i=0;i<m_shader_config->n_simt_clusters;i++) { if (m_cluster[i]->get_not_completed() == 0) - m_cluster[i]->cache_flush(); + m_cluster[i]->cache_invalidate(); else all_threads_complete = 0 ; } @@ -1548,7 +1612,7 @@ void gpgpu_sim::cycle() int dlc = 0; for (unsigned i=0;i<m_memory_config->m_n_mem;i++) { dlc = m_memory_sub_partition[i]->flushL2(); - assert (dlc == 0); // need to model actual writes to DRAM here + assert (dlc == 0); // TODO: need to model actual writes to DRAM here printf("Dirty lines flushed from L2 %d is %d\n", i, dlc ); } } @@ -1560,15 +1624,20 @@ void gpgpu_sim::cycle() time_t curr_time; time(&curr_time); unsigned long long elapsed_time = MAX(curr_time - g_simulation_starttime, 1); - if ( (elapsed_time - last_liveness_message_time) >= m_config.liveness_message_freq ) { + if ( (elapsed_time - last_liveness_message_time) >= m_config.liveness_message_freq && DTRACE(LIVENESS) ) { days = elapsed_time/(3600*24); hrs = elapsed_time/3600 - 24*days; minutes = elapsed_time/60 - 60*(hrs + 24*days); sec = elapsed_time - 60*(minutes + 60*(hrs + 24*days)); - - DPRINTF(LIVENESS, "GPGPU-Sim uArch: cycles simulated: %lld inst.: %lld (ipc=%4.1f) sim_rate=%u (inst/sec) elapsed = %u:%u:%02u:%02u / %s", - gpu_tot_sim_cycle + gpu_sim_cycle, gpu_tot_sim_insn + gpu_sim_insn, + + unsigned long long active = 0, total = 0; + for (unsigned i=0;i<m_shader_config->n_simt_clusters;i++) { + m_cluster[i]->get_current_occupancy(active, total); + } + DPRINTF(LIVENESS, "uArch: inst.: %lld (ipc=%4.1f, occ=%0.4f\% [%llu / %llu]) sim_rate=%u (inst/sec) elapsed = %u:%u:%02u:%02u / %s", + gpu_tot_sim_insn + gpu_sim_insn, (double)gpu_sim_insn/(double)gpu_sim_cycle, + float(active)/float(total) * 100, active, total, (unsigned)((gpu_tot_sim_insn+gpu_sim_insn) / elapsed_time), (unsigned)days,(unsigned)hrs,(unsigned)minutes,(unsigned)sec, ctime(&curr_time)); @@ -1620,6 +1689,24 @@ void shader_core_ctx::dump_warp_state( FILE *fout ) const m_warp[w].print(fout); } + +void gpgpu_sim::perf_memcpy_to_gpu( size_t dst_start_addr, size_t count ) +{ + if (m_memory_config->m_perf_sim_memcpy) { + assert (dst_start_addr % 32 == 0); + + for ( unsigned counter = 0; counter < count; counter += 32 ) { + const unsigned wr_addr = dst_start_addr + counter; + addrdec_t raw_addr; + mem_access_sector_mask_t mask; + mask.set(wr_addr % 128 / 32); + m_memory_config->m_address_mapping.addrdec_tlx( wr_addr, &raw_addr ); + const unsigned partition_id = raw_addr.sub_partition / m_memory_config->m_n_sub_partition_per_memory_channel; + m_memory_partition_unit[ partition_id ]->handle_memcpy_to_gpu( wr_addr, raw_addr.sub_partition, mask ); + } + } +} + void gpgpu_sim::dump_pipeline( int mask, int s, int m ) const { /* diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 7d92c66..1bae1fa 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -198,8 +198,14 @@ struct memory_config { bk_tag_length = i-1; assert(nbkgrp>0 && "Number of bank groups cannot be zero"); tRCDWR = tRCD-(WL+1); + if(elimnate_rw_turnaround) + { + tRTW = 0; + tWTR = 0; + } else { tRTW = (CL+(BL/data_command_freq_ratio)+2-WL); - tWTR = (WL+(BL/data_command_freq_ratio)+tCDLR); + tWTR = (WL+(BL/data_command_freq_ratio)+tCDLR); + } tWTP = (WL+(BL/data_command_freq_ratio)+tWR); dram_atom_size = BL * busW * gpu_n_mem_per_ctrlr; // burst length x bus width x # chips per partition @@ -213,7 +219,9 @@ struct memory_config { m_L2_config.init(&m_address_mapping); m_valid = true; - icnt_flit_size = 32; // Default 32 + + sscanf(write_queue_size_opt,"%d:%d:%d", + &gpgpu_frfcfs_dram_write_queue_size,&write_high_watermark,&write_low_watermark); } void reg_options(class OptionParser * opp); @@ -264,12 +272,25 @@ struct memory_config { unsigned nbk; + bool elimnate_rw_turnaround; + unsigned data_command_freq_ratio; // frequency ratio between DRAM data bus and command bus (2 for GDDR3, 4 for GDDR5) unsigned dram_atom_size; // number of bytes transferred per read or write command linear_to_raw_address_translation m_address_mapping; unsigned icnt_flit_size; + + unsigned dram_bnk_indexing_policy; + unsigned dram_bnkgrp_indexing_policy; + bool dual_bus_interface; + + bool seperate_write_queue_enabled; + char *write_queue_size_opt; + unsigned gpgpu_frfcfs_dram_write_queue_size; + unsigned write_high_watermark; + unsigned write_low_watermark; + bool m_perf_sim_memcpy; }; // global counters and flags (please try not to add to this list!!!) @@ -362,6 +383,32 @@ private: friend class gpgpu_sim; }; +struct occupancy_stats { + occupancy_stats() : aggregate_warp_slot_filled(0), aggregate_theoretical_warp_slots(0){} + occupancy_stats( unsigned long long wsf, unsigned long long tws ) + : aggregate_warp_slot_filled(wsf), aggregate_theoretical_warp_slots(tws){} + + unsigned long long aggregate_warp_slot_filled; + unsigned long long aggregate_theoretical_warp_slots; + + float get_occ_fraction() const { + return float(aggregate_warp_slot_filled) / float(aggregate_theoretical_warp_slots); + } + + occupancy_stats& operator+=(const occupancy_stats& rhs) { + aggregate_warp_slot_filled += rhs.aggregate_warp_slot_filled; + aggregate_theoretical_warp_slots += rhs.aggregate_theoretical_warp_slots; + return *this; + } + + occupancy_stats operator+(const occupancy_stats& rhs) const{ + return occupancy_stats( aggregate_warp_slot_filled + rhs.aggregate_warp_slot_filled, + aggregate_theoretical_warp_slots + rhs.aggregate_theoretical_warp_slots + ); + } +}; + + class gpgpu_sim : public gpgpu_t { public: gpgpu_sim( const gpgpu_sim_config &config ); @@ -405,6 +452,8 @@ public: void gpu_print_stat(); void dump_pipeline( int mask, int s, int m ) const; + void perf_memcpy_to_gpu( size_t dst_start_addr, size_t count ); + //The next three functions added to be used by the functional simulation function //! Get shader core configuration @@ -498,6 +547,9 @@ public: unsigned long long gpu_tot_sim_insn; unsigned long long gpu_sim_insn_last_update; unsigned gpu_sim_insn_last_update_sid; + occupancy_stats gpu_occupancy; + occupancy_stats gpu_tot_occupancy; + FuncCache get_cache_config(std::string kernel_name); void set_cache_config(std::string kernel_name, FuncCache cacheConfig ); @@ -526,4 +578,5 @@ public: } }; + #endif diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index c5fc44e..25da107 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -74,6 +74,15 @@ memory_partition_unit::memory_partition_unit( unsigned partition_id, } } +void memory_partition_unit::handle_memcpy_to_gpu( size_t addr, unsigned global_subpart_id, mem_access_sector_mask_t mask ) +{ + unsigned p = global_sub_partition_id_to_local_id(global_subpart_id); + std::string mystring = + mask.to_string<char,std::string::traits_type,std::string::allocator_type>(); + MEMPART_DPRINTF("Copy Engine Request Received For Address=%llx, local_subpart=%u, global_subpart=%u, sector_mask=%s \n", addr, p, global_subpart_id, mystring.c_str()); + m_sub_partition[p]->force_l2_tag_update(addr,gpu_sim_cycle+gpu_tot_sim_cycle, mask); +} + memory_partition_unit::~memory_partition_unit() { delete m_dram; @@ -93,7 +102,9 @@ memory_partition_unit::arbitration_metadata::arbitration_metadata(const struct m m_private_credit_limit = 1; m_shared_credit_limit = config->gpgpu_frfcfs_dram_sched_queue_size + config->gpgpu_dram_return_queue_size - - (config->m_n_sub_partition_per_memory_channel - 1); + - (config->m_n_sub_partition_per_memory_channel - 1); + if(config->seperate_write_queue_enabled ) + m_shared_credit_limit += config->gpgpu_frfcfs_dram_write_queue_size; if (config->gpgpu_frfcfs_dram_sched_queue_size == 0 or config->gpgpu_dram_return_queue_size == 0) { @@ -220,7 +231,8 @@ void memory_partition_unit::dram_cycle() m_dram->cycle(); m_dram->dram_log(SAMPLELOG); - if( !m_dram->full() ) { + // mem_fetch *mf = m_sub_partition[spid]->L2_dram_queue_top(); + //if( !m_dram->full(mf->is_write()) ) { // L2->DRAM queue to DRAM latency queue // Arbitrate among multiple L2 subpartitions int last_issued_partition = m_arbitration_metadata.last_borrower(); @@ -228,6 +240,9 @@ void memory_partition_unit::dram_cycle() int spid = (p + last_issued_partition + 1) % m_config->m_n_sub_partition_per_memory_channel; if (!m_sub_partition[spid]->L2_dram_queue_empty() && can_issue_to_dram(spid)) { mem_fetch *mf = m_sub_partition[spid]->L2_dram_queue_top(); + if(m_dram->full(mf->is_write()) ) + break; + m_sub_partition[spid]->L2_dram_queue_pop(); MEMPART_DPRINTF("Issue mem_fetch request %p from sub partition %d to dram\n", mf, spid); dram_delay_t d; @@ -239,12 +254,13 @@ void memory_partition_unit::dram_cycle() break; // the DRAM should only accept one request per cycle } } - } + //} // DRAM latency queue - if( !m_dram_latency_queue.empty() && ( (gpu_sim_cycle+gpu_tot_sim_cycle) >= m_dram_latency_queue.front().ready_cycle ) && !m_dram->full() ) { - mem_fetch* mf = m_dram_latency_queue.front().req; - m_dram_latency_queue.pop_front(); + + if( !m_dram_latency_queue.empty() && ( (gpu_sim_cycle+gpu_tot_sim_cycle) >= m_dram_latency_queue.front().ready_cycle ) && !m_dram->full(m_dram_latency_queue.front().req->is_write()) ) { + mem_fetch* mf = m_dram_latency_queue.front().req; + m_dram_latency_queue.pop_front(); m_dram->push(mf); } } @@ -299,6 +315,7 @@ memory_sub_partition::memory_sub_partition( unsigned sub_partition_id, m_id = sub_partition_id; m_config=config; m_stats=stats; + m_memcpy_cycle_offset = 0; assert(m_id < m_config->m_n_mem_sub_partition); @@ -343,6 +360,14 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); m_L2_icnt_queue->push(mf); }else{ + if(m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE) + { + mem_fetch* original_wr_mf = mf->get_original_wr_mf(); + assert(original_wr_mf); + original_wr_mf->set_reply(); + original_wr_mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); + m_L2_icnt_queue->push(original_wr_mf); + } m_request_tracker.erase(mf); delete mf; } @@ -355,10 +380,11 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) if ( !m_config->m_L2_config.disabled() && m_L2cache->waiting_for_fill(mf) ) { if (m_L2cache->fill_port_free()) { mf->set_status(IN_PARTITION_L2_FILL_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); - m_L2cache->fill(mf,gpu_sim_cycle+gpu_tot_sim_cycle); + m_L2cache->fill(mf,gpu_sim_cycle+gpu_tot_sim_cycle+m_memcpy_cycle_offset); m_dram_L2_queue->pop(); } } else if ( !m_L2_icnt_queue->full() ) { + if(mf->is_write() && mf->get_type() == WRITE_ACK) mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); m_L2_icnt_queue->push(mf); m_dram_L2_queue->pop(); @@ -380,9 +406,10 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) bool port_free = m_L2cache->data_port_free(); if ( !output_full && port_free ) { std::list<cache_event> events; - enum cache_request_status status = m_L2cache->access(mf->get_addr(),mf,gpu_sim_cycle+gpu_tot_sim_cycle,events); + enum cache_request_status status = m_L2cache->access(mf->get_addr(),mf,gpu_sim_cycle+gpu_tot_sim_cycle+m_memcpy_cycle_offset,events); bool write_sent = was_write_sent(events); bool read_sent = was_read_sent(events); + MEM_SUBPART_DPRINTF("Probing L2 cache Address=%llx, status=%u\n", mf->get_addr(), status); if ( status == HIT ) { if( !write_sent ) { @@ -402,6 +429,11 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) m_icnt_L2_queue->pop(); } } else if ( status != RESERVATION_FAIL ) { + if(mf->is_write() && (m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE || m_config->m_L2_config.m_write_alloc_policy == LAZY_FETCH_ON_READ) && !was_writeallocate_sent(events)) { + mf->set_reply(); + mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); + m_L2_icnt_queue->push(mf); + } // L2 cache accepted request m_icnt_L2_queue->pop(); } else { @@ -432,6 +464,11 @@ bool memory_sub_partition::full() const return m_icnt_L2_queue->full(); } +bool memory_sub_partition::full(unsigned size) const +{ + return m_icnt_L2_queue->is_avilable_size(size); +} + bool memory_sub_partition::L2_dram_queue_empty() const { return m_L2_dram_queue->empty(); @@ -532,7 +569,15 @@ unsigned memory_sub_partition::flushL2() if (!m_config->m_L2_config.disabled()) { m_L2cache->flush(); } - return 0; // L2 is read only in this version + return 0; //TODO: write the flushed data to the main memory +} + +unsigned memory_sub_partition::invalidateL2() +{ + if (!m_config->m_L2_config.disabled()) { + m_L2cache->invalidate(); + } + return 0; } bool memory_sub_partition::busy() const @@ -540,21 +585,94 @@ bool memory_sub_partition::busy() const return !m_request_tracker.empty(); } -void memory_sub_partition::push( mem_fetch* req, unsigned long long cycle ) +std::vector<mem_fetch*> memory_sub_partition::breakdown_request_to_sector_requests(mem_fetch* mf) { - if (req) { - m_request_tracker.insert(req); - m_stats->memlatstat_icnt2mem_pop(req); - if( req->istexture() ) { - m_icnt_L2_queue->push(req); - req->set_status(IN_PARTITION_ICNT_TO_L2_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); - } else { - rop_delay_t r; - r.req = req; - r.ready_cycle = cycle + m_config->rop_latency; - m_rop.push(r); - req->set_status(IN_PARTITION_ROP_DELAY,gpu_sim_cycle+gpu_tot_sim_cycle); - } + std::vector<mem_fetch*> result; + + if(mf->get_data_size() == SECTOR_SIZE && mf->get_access_sector_mask().count() == 1) { + result.push_back(mf); + } else if (mf->get_data_size() == 128 || mf->get_data_size() == 64) { + //We only accept 32, 64 and 128 bytes reqs + unsigned start=0, end=0; + if(mf->get_data_size() == 128) { + start=0; end=3; + } else if (mf->get_data_size() == 64 && mf->get_access_sector_mask().to_string() == "1100") { + start=2; end=3; + } else if (mf->get_data_size() == 64 && mf->get_access_sector_mask().to_string() == "0011") { + start=0; end=1; + } else if (mf->get_data_size() == 64 && (mf->get_access_sector_mask().to_string() == "1111" || mf->get_access_sector_mask().to_string() == "0000")) { + if(mf->get_addr() % 128 == 0) { + start=0; end=1; + } else { + start=2; end=3; + } + } else + { + printf("Invalid sector received, address = 0x%06x, sector mask = %s, data size = %d", + mf->get_addr(), mf->get_access_sector_mask(), mf->get_data_size()); + assert(0 && "Undefined sector mask is received"); + } + + std::bitset<SECTOR_SIZE*SECTOR_CHUNCK_SIZE> byte_sector_mask; + byte_sector_mask.reset(); + for(unsigned k=start*SECTOR_SIZE; k< SECTOR_SIZE; ++k) + byte_sector_mask.set(k); + + for(unsigned j=start, i=0; j<= end ; ++j, ++i){ + + const mem_access_t *ma = new mem_access_t( mf->get_access_type(), + mf->get_addr() + SECTOR_SIZE*i, + SECTOR_SIZE, + mf->is_write(), + mf->get_access_warp_mask(), + mf->get_access_byte_mask() & byte_sector_mask, + std::bitset<SECTOR_CHUNCK_SIZE>().set(j)); + + mem_fetch *n_mf = new mem_fetch( *ma, + NULL, + mf->get_ctrl_size(), + mf->get_wid(), + mf->get_sid(), + mf->get_tpc(), + mf->get_mem_config(), + mf); + + result.push_back(n_mf); + byte_sector_mask <<= SECTOR_SIZE; + } + } else { + printf("Invalid sector received, address = 0x%06x, sector mask = %d, byte mask = , data size = %d", + mf->get_addr(), mf->get_access_sector_mask().count(), mf->get_data_size()); + assert(0 && "Undefined data size is received"); + } + + return result; +} + +void memory_sub_partition::push( mem_fetch* m_req, unsigned long long cycle ) +{ + if (m_req) { + m_stats->memlatstat_icnt2mem_pop(m_req); + std::vector<mem_fetch*> reqs; + if(m_config->m_L2_config.m_cache_type == SECTOR) + reqs = breakdown_request_to_sector_requests(m_req); + else + reqs.push_back(m_req); + + for(unsigned i=0; i<reqs.size(); ++i) { + mem_fetch* req = reqs[i]; + m_request_tracker.insert(req); + if( req->istexture() ) { + m_icnt_L2_queue->push(req); + req->set_status(IN_PARTITION_ICNT_TO_L2_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); + } else { + rop_delay_t r; + r.req = req; + r.ready_cycle = cycle + m_config->rop_latency; + m_rop.push(r); + req->set_status(IN_PARTITION_ROP_DELAY,gpu_sim_cycle+gpu_tot_sim_cycle); + } + } } } diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index 3df54b1..18c0a8b 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -72,6 +72,7 @@ public: void print_stat( FILE *fp ) { m_dram->print_stat(fp); } void visualize() const { m_dram->visualize(); } void print( FILE *fp ) const; + void handle_memcpy_to_gpu( size_t dst_start_addr, unsigned subpart_id, mem_access_sector_mask_t mask ); class memory_sub_partition * get_sub_partition(int sub_partition_id) { @@ -154,12 +155,14 @@ public: void cache_cycle( unsigned cycle ); bool full() const; + bool full(unsigned size) const; void push( class mem_fetch* mf, unsigned long long clock_cycle ); class mem_fetch* pop(); class mem_fetch* top(); void set_done( mem_fetch *mf ); unsigned flushL2(); + unsigned invalidateL2(); // interface to L2_dram_queue bool L2_dram_queue_empty() const; @@ -177,6 +180,12 @@ public: void accumulate_L2cache_stats(class cache_stats &l2_stats) const; void get_L2cache_sub_stats(struct cache_sub_stats &css) const; + void force_l2_tag_update(new_addr_type addr, unsigned time, mem_access_sector_mask_t mask) + { + m_L2cache->force_tag_access( addr, m_memcpy_cycle_offset + time, mask ); + m_memcpy_cycle_offset += 1; + } + private: // data unsigned m_id; //< the global sub partition ID @@ -207,6 +216,15 @@ private: std::set<mem_fetch*> m_request_tracker; friend class L2interface; + + std::vector<mem_fetch*> breakdown_request_to_sector_requests(mem_fetch* mf); + + // This is a cycle offset that has to be applied to the l2 accesses to account for + // the cudamemcpy read/writes. We want GPGPU-Sim to only count cycles for kernel execution + // but we want cudamemcpy to go through the L2. Everytime an access is made from cudamemcpy + // this counter is incremented, and when the l2 is accessed (in both cudamemcpyies and otherwise) + // this value is added to the gpgpu-sim cycle counters. + unsigned m_memcpy_cycle_offset; }; class L2interface : public mem_fetch_interface { diff --git a/src/gpgpu-sim/l2cache_trace.h b/src/gpgpu-sim/l2cache_trace.h index 3dac87d..2235cdc 100644 --- a/src/gpgpu-sim/l2cache_trace.h +++ b/src/gpgpu-sim/l2cache_trace.h @@ -34,6 +34,9 @@ #define MEMPART_PRINT_STR SIM_PRINT_STR " %d - " #define MEMPART_DTRACE(x) ( DTRACE(x) && (Trace::sampling_memory_partition == -1 || Trace::sampling_memory_partition == (int)get_mpid()) ) +#define MEM_SUBPART_PRINT_STR SIM_PRINT_STR " %d - " +#define MEM_SUBPART_DTRACE(x) ( DTRACE(x) && (Trace::sampling_memory_partition == -1 || Trace::sampling_memory_partition == (int)m_id) ) + // Intended to be called from inside components of a memory partition // Depends on a get_mpid() function #define MEMPART_DPRINTF(...) do {\ @@ -46,10 +49,23 @@ }\ } while (0) +#define MEM_SUBPART_DPRINTF(...) do {\ + if (MEM_SUBPART_DTRACE(MEMORY_PARTITION_UNIT)) {\ + printf( MEM_SUBPART_PRINT_STR,\ + gpu_sim_cycle + gpu_tot_sim_cycle,\ + Trace::trace_streams_str[Trace::MEMORY_SUBPARTITION_UNIT],\ + m_id );\ + printf(__VA_ARGS__);\ + }\ +} while (0) + #else #define MEMPART_DTRACE(x) (false) #define MEMPART_DPRINTF(x, ...) do {} while (0) +#define MEM_SUBPART_DTRACE(x) (false) +#define MEM_SUBPART_DPRINTF(x, ...) do {} while (0) + #endif diff --git a/src/gpgpu-sim/mem_fetch.cc b/src/gpgpu-sim/mem_fetch.cc index 729636d..a260a35 100644 --- a/src/gpgpu-sim/mem_fetch.cc +++ b/src/gpgpu-sim/mem_fetch.cc @@ -39,7 +39,10 @@ mem_fetch::mem_fetch( const mem_access_t &access, unsigned wid, unsigned sid, unsigned tpc, - const struct memory_config *config ) + const struct memory_config *config, + mem_fetch *m_original_mf, + mem_fetch *m_original_wr_mf) + { m_request_uid = sm_next_mf_request_uid++; m_access = access; @@ -61,6 +64,8 @@ mem_fetch::mem_fetch( const mem_access_t &access, m_status_change = gpu_sim_cycle + gpu_tot_sim_cycle; m_mem_config = config; icnt_flit_size = config->icnt_flit_size; + original_mf = m_original_mf; + original_wr_mf = m_original_wr_mf; } mem_fetch::~mem_fetch() diff --git a/src/gpgpu-sim/mem_fetch.h b/src/gpgpu-sim/mem_fetch.h index de98748..e5efffd 100644 --- a/src/gpgpu-sim/mem_fetch.h +++ b/src/gpgpu-sim/mem_fetch.h @@ -55,7 +55,9 @@ public: unsigned wid, unsigned sid, unsigned tpc, - const struct memory_config *config ); + const struct memory_config *config, + mem_fetch *original_mf = NULL, + mem_fetch *original_wr_mf = NULL); ~mem_fetch(); void set_status( enum mem_fetch_status status, unsigned long long cycle ); @@ -104,6 +106,7 @@ public: enum mem_access_type get_access_type() const { return m_access.get_type(); } const active_mask_t& get_access_warp_mask() const { return m_access.get_warp_mask(); } mem_access_byte_mask_t get_access_byte_mask() const { return m_access.get_byte_mask(); } + mem_access_sector_mask_t get_access_sector_mask() const { return m_access.get_sector_mask(); } address_type get_pc() const { return m_inst.empty()?-1:m_inst.pc; } const warp_inst_t &get_inst() { return m_inst; } @@ -112,6 +115,10 @@ public: const memory_config *get_mem_config(){return m_mem_config;} unsigned get_num_flits(bool simt_to_mem); + + mem_fetch* get_original_mf() { return original_mf; } + mem_fetch* get_original_wr_mf() { return original_wr_mf; } + private: // request source information unsigned m_request_uid; @@ -143,6 +150,10 @@ private: const struct memory_config *m_mem_config; unsigned icnt_flit_size; + + mem_fetch* original_mf; //this pointer is set up when a request is divided into sector requests at L2 cache (if the req size > L2 sector size), so the pointer refers to the original request + mem_fetch* original_wr_mf; //this pointer refers to the original write req, when fetch-on-write policy is used + }; #endif diff --git a/src/gpgpu-sim/mem_latency_stat.cc b/src/gpgpu-sim/mem_latency_stat.cc index fde0eff..c5452b9 100644 --- a/src/gpgpu-sim/mem_latency_stat.cc +++ b/src/gpgpu-sim/mem_latency_stat.cc @@ -75,6 +75,10 @@ memory_stats_t::memory_stats_t( unsigned n_shader, const struct shader_core_conf max_mf_latency = 0; max_icnt2mem_latency = 0; max_icnt2sh_latency = 0; + tot_icnt2mem_latency = 0; + tot_icnt2sh_latency = 0; + tot_mrq_num = 0; + tot_mrq_latency = 0; memset(mrq_lat_table, 0, sizeof(unsigned)*32); memset(dq_lat_table, 0, sizeof(unsigned)*32); memset(mf_lat_table, 0, sizeof(unsigned)*32); @@ -158,6 +162,7 @@ void memory_stats_t::memlatstat_read_done(mem_fetch *mf) mf_max_lat_table[mf->get_tlx_addr().chip][mf->get_tlx_addr().bk] = mf_latency; unsigned icnt2sh_latency; icnt2sh_latency = (gpu_tot_sim_cycle+gpu_sim_cycle) - mf->get_return_timestamp(); + tot_icnt2sh_latency += icnt2sh_latency; icnt2sh_lat_table[LOGB2(icnt2sh_latency)]++; if (icnt2sh_latency > max_icnt2sh_latency) max_icnt2sh_latency = icnt2sh_latency; @@ -191,6 +196,7 @@ void memory_stats_t::memlatstat_icnt2mem_pop(mem_fetch *mf) if (m_memory_config->gpgpu_memlatency_stat) { unsigned icnt2mem_latency; icnt2mem_latency = (gpu_tot_sim_cycle+gpu_sim_cycle) - mf->get_timestamp(); + tot_icnt2mem_latency += icnt2mem_latency; icnt2mem_lat_table[LOGB2(icnt2mem_latency)]++; if (icnt2mem_latency > max_icnt2mem_latency) max_icnt2mem_latency = icnt2mem_latency; @@ -216,14 +222,19 @@ void memory_stats_t::memlatstat_print( unsigned n_mem, unsigned gpu_mem_n_bk ) unsigned max_bank_accesses, min_bank_accesses, max_chip_accesses, min_chip_accesses; if (m_memory_config->gpgpu_memlatency_stat) { + printf("maxmflatency = %d \n", max_mf_latency); + printf("max_icnt2mem_latency = %d \n", max_icnt2mem_latency); printf("maxmrqlatency = %d \n", max_mrq_latency); - printf("maxdqlatency = %d \n", max_dq_latency); - printf("maxmflatency = %d \n", max_mf_latency); + //printf("maxdqlatency = %d \n", max_dq_latency); + printf("max_icnt2sh_latency = %d \n", max_icnt2sh_latency); if (num_mfs) { printf("averagemflatency = %lld \n", mf_total_lat/num_mfs); + printf("avg_icnt2mem_latency = %lld \n", tot_icnt2mem_latency/num_mfs); + if(tot_mrq_num) + printf("avg_mrq_latency = %lld \n", tot_mrq_latency/tot_mrq_num); + + printf("avg_icnt2sh_latency = %lld \n", tot_icnt2sh_latency/num_mfs); } - printf("max_icnt2mem_latency = %d \n", max_icnt2mem_latency); - printf("max_icnt2sh_latency = %d \n", max_icnt2sh_latency); printf("mrq_lat_table:"); for (i=0; i< 32; i++) { printf("%d \t", mrq_lat_table[i]); diff --git a/src/gpgpu-sim/mem_latency_stat.h b/src/gpgpu-sim/mem_latency_stat.h index 4968a3b..5b89202 100644 --- a/src/gpgpu-sim/mem_latency_stat.h +++ b/src/gpgpu-sim/mem_latency_stat.h @@ -56,6 +56,10 @@ public: unsigned max_dq_latency; unsigned max_mf_latency; unsigned max_icnt2mem_latency; + unsigned long long int tot_icnt2mem_latency; + unsigned long long int tot_icnt2sh_latency; + unsigned long long int tot_mrq_latency; + unsigned long long int tot_mrq_num; unsigned max_icnt2sh_latency; unsigned mrq_lat_table[32]; unsigned dq_lat_table[32]; diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index f9cfa58..82f9181 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -74,13 +74,14 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, shader_core_stats *stats ) : core_t( gpu, NULL, config->warp_size, config->n_thread_per_shader ), m_barriers( this, config->max_warps_per_shader, config->max_cta_per_core, config->max_barriers_per_cta, config->warp_size ), - m_dynamic_warp_id(0) + m_dynamic_warp_id(0), m_active_warps(0) { m_cluster = cluster; m_config = config; m_memory_config = mem_config; m_stats = stats; unsigned warp_size=config->warp_size; + Issue_Prio = 0; m_sid = shader_id; m_tpc = tpc_id; @@ -131,6 +132,8 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, CONCRETE_SCHEDULER_TWO_LEVEL_ACTIVE : sched_config.find("gto") != std::string::npos ? CONCRETE_SCHEDULER_GTO : + sched_config.find("old") != std::string::npos ? + CONCRETE_SCHEDULER_OLDEST_FIRST : sched_config.find("warp_limiting") != std::string::npos ? CONCRETE_SCHEDULER_WARP_LIMITING: NUM_CONCRETE_SCHEDULERS; @@ -147,6 +150,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_simt_stack, &m_warp, &m_pipeline_reg[ID_OC_SP], + &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], &m_pipeline_reg[ID_OC_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], @@ -162,6 +166,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_simt_stack, &m_warp, &m_pipeline_reg[ID_OC_SP], + &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], &m_pipeline_reg[ID_OC_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], @@ -178,6 +183,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_simt_stack, &m_warp, &m_pipeline_reg[ID_OC_SP], + &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], &m_pipeline_reg[ID_OC_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], @@ -185,6 +191,22 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, ) ); break; + case CONCRETE_SCHEDULER_OLDEST_FIRST: + schedulers.push_back( + new oldest_scheduler( m_stats, + this, + m_scoreboard, + m_simt_stack, + &m_warp, + &m_pipeline_reg[ID_OC_SP], + &m_pipeline_reg[ID_OC_DP], + &m_pipeline_reg[ID_OC_SFU], + &m_pipeline_reg[ID_OC_TENSOR_CORE], + &m_pipeline_reg[ID_OC_MEM], + i + ) + ); + break; case CONCRETE_SCHEDULER_WARP_LIMITING: schedulers.push_back( new swl_scheduler( m_stats, @@ -193,6 +215,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_simt_stack, &m_warp, &m_pipeline_reg[ID_OC_SP], + &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], &m_pipeline_reg[ID_OC_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], @@ -215,8 +238,9 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, } //op collector configuration - enum { SP_CUS, SFU_CUS, TENSOR_CORE_CUS, MEM_CUS, GEN_CUS }; + enum { SP_CUS, DP_CUS, SFU_CUS, MEM_CUS, TENSOR_CORE_CUS, GEN_CUS }; m_operand_collector.add_cu_set(SP_CUS, m_config->gpgpu_operand_collector_num_units_sp, m_config->gpgpu_operand_collector_num_out_ports_sp); + m_operand_collector.add_cu_set(DP_CUS, m_config->gpgpu_operand_collector_num_units_dp, m_config->gpgpu_operand_collector_num_out_ports_dp); m_operand_collector.add_cu_set(SFU_CUS, m_config->gpgpu_operand_collector_num_units_sfu, m_config->gpgpu_operand_collector_num_out_ports_sfu); m_operand_collector.add_cu_set(TENSOR_CORE_CUS, config->gpgpu_operand_collector_num_units_tensor_core, config->gpgpu_operand_collector_num_out_ports_tensor_core); m_operand_collector.add_cu_set(MEM_CUS, m_config->gpgpu_operand_collector_num_units_mem, m_config->gpgpu_operand_collector_num_out_ports_mem); @@ -234,6 +258,15 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, in_ports.clear(),out_ports.clear(),cu_sets.clear(); } + for (unsigned i = 0; i < m_config->gpgpu_operand_collector_num_in_ports_dp; i++) { + in_ports.push_back(&m_pipeline_reg[ID_OC_DP]); + out_ports.push_back(&m_pipeline_reg[OC_EX_DP]); + cu_sets.push_back((unsigned)DP_CUS); + cu_sets.push_back((unsigned)GEN_CUS); + m_operand_collector.add_port(in_ports,out_ports,cu_sets); + in_ports.clear(),out_ports.clear(),cu_sets.clear(); + } + for (unsigned i = 0; i < m_config->gpgpu_operand_collector_num_in_ports_sfu; i++) { in_ports.push_back(&m_pipeline_reg[ID_OC_SFU]); out_ports.push_back(&m_pipeline_reg[OC_EX_SFU]); @@ -279,7 +312,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_operand_collector.init( m_config->gpgpu_num_reg_banks, this ); // execute - m_num_function_units = m_config->gpgpu_num_sp_units + m_config->gpgpu_num_sfu_units + config->gpgpu_num_tensor_core_units + 1; // sp_unit, sfu, ldst_unit + m_num_function_units = m_config->gpgpu_num_sp_units + m_config->gpgpu_num_dp_units + m_config->gpgpu_num_sfu_units + m_config->gpgpu_num_tensor_core_units + 1; // sp_unit, sfu, ldst_unit //m_dispatch_port = new enum pipeline_stage_name_t[ m_num_function_units ]; //m_issue_port = new enum pipeline_stage_name_t[ m_num_function_units ]; @@ -291,6 +324,12 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_issue_port.push_back(OC_EX_SP); } + for (int k = 0; k < m_config->gpgpu_num_dp_units; k++) { + m_fu.push_back(new dp_unit( &m_pipeline_reg[EX_WB], m_config, this )); + m_dispatch_port.push_back(ID_OC_DP); + m_issue_port.push_back(OC_EX_DP); + } + for (int k = 0; k < m_config->gpgpu_num_sfu_units; k++) { m_fu.push_back(new sfu( &m_pipeline_reg[EX_WB], m_config, this )); m_dispatch_port.push_back(ID_OC_SFU); @@ -307,7 +346,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_fu.push_back(m_ldst_unit); m_dispatch_port.push_back(ID_OC_MEM); m_issue_port.push_back(OC_EX_MEM); - + assert(m_num_function_units == m_fu.size() and m_fu.size() == m_dispatch_port.size() and m_fu.size() == m_issue_port.size()); //there are as many result buses as the width of the EX_WB stage @@ -341,6 +380,7 @@ void shader_core_ctx::reinit(unsigned start_thread, unsigned end_thread, bool re m_occupied_ctas = 0; m_occupied_hwtid.reset(); m_occupied_cta_to_hwtid.clear(); + m_active_warps = 0; } for (unsigned i = start_thread; i<end_thread; i++) { @@ -375,6 +415,7 @@ void shader_core_ctx::init_warps( unsigned cta_id, unsigned start_thread, unsign m_warp[i].init(start_pc,cta_id,i,active_threads, m_dynamic_warp_id); ++m_dynamic_warp_id; m_not_completed += n_active; + ++m_active_warps; } } } @@ -402,6 +443,18 @@ void shader_core_ctx::get_pdom_stack_top_info( unsigned tid, unsigned *pc, unsig m_simt_stack[warp_id]->get_pdom_stack_top_info(pc,rpc); } +float shader_core_ctx::get_current_occupancy( unsigned long long & active, unsigned long long & total ) const +{ + // To match the achieved_occupancy in nvprof, only SMs that are active are counted toward the occupancy. + if ( m_active_warps > 0 ) { + total += m_warp.size(); + active += m_active_warps; + return float(active) / float(total); + } else { + return 0; + } +} + void shader_core_stats::print( FILE* fout ) const { unsigned long long thread_icount_uarch=0; @@ -436,15 +489,15 @@ void shader_core_stats::print( FILE* fout ) const fprintf(fout, "gpgpu_n_intrawarp_mshr_merge = %d\n", gpgpu_n_intrawarp_mshr_merge); fprintf(fout, "gpgpu_n_cmem_portconflict = %d\n", gpgpu_n_cmem_portconflict); - fprintf(fout, "gpgpu_stall_shd_mem[c_mem][bk_conf] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][BK_CONF]); - fprintf(fout, "gpgpu_stall_shd_mem[c_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][MSHR_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[c_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[c_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][DATA_PORT_STALL]); - fprintf(fout, "gpgpu_stall_shd_mem[t_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][MSHR_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[t_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[t_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][DATA_PORT_STALL]); + fprintf(fout, "gpgpu_stall_shd_mem[c_mem][resource_stall] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][BK_CONF]); + //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][MSHR_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][DATA_PORT_STALL]); + //fprintf(fout, "gpgpu_stall_shd_mem[t_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][MSHR_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[t_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[t_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][DATA_PORT_STALL]); fprintf(fout, "gpgpu_stall_shd_mem[s_mem][bk_conf] = %d\n", gpu_stall_shd_mem_breakdown[S_MEM][BK_CONF]); - fprintf(fout, "gpgpu_stall_shd_mem[gl_mem][bk_conf] = %d\n", + fprintf(fout, "gpgpu_stall_shd_mem[gl_mem][resource_stall] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][BK_CONF] + gpu_stall_shd_mem_breakdown[G_MEM_ST][BK_CONF] + gpu_stall_shd_mem_breakdown[L_MEM_LD][BK_CONF] + @@ -462,22 +515,22 @@ void shader_core_stats::print( FILE* fout ) const gpu_stall_shd_mem_breakdown[L_MEM_LD][DATA_PORT_STALL] + gpu_stall_shd_mem_breakdown[L_MEM_ST][DATA_PORT_STALL] ); // data port stall at data cache - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][MSHR_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][WB_ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][WB_CACHE_RSRV_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][MSHR_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][WB_ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][WB_CACHE_RSRV_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][MSHR_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][WB_ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][WB_CACHE_RSRV_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_st][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][MSHR_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_st][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][WB_ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][WB_CACHE_RSRV_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][MSHR_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][WB_ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][WB_CACHE_RSRV_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][MSHR_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][WB_ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][WB_CACHE_RSRV_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][MSHR_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][WB_ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][WB_CACHE_RSRV_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_st][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][MSHR_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_st][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][WB_ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][WB_CACHE_RSRV_FAIL]); fprintf(fout, "gpu_reg_bank_conflict_stalls = %d\n", gpu_reg_bank_conflict_stalls); @@ -488,6 +541,14 @@ void shader_core_stats::print( FILE* fout ) const for (unsigned i = 3; i < m_config->warp_size + 3; i++) fprintf(fout, "\tW%d:%d", i-2, shader_cycle_distro[i]); fprintf(fout, "\n"); + fprintf(fout, "single_issue_nums: "); + for (unsigned i = 0; i < m_config->gpgpu_num_sched_per_core; i++) + fprintf(fout, "WS%d:%d\t", i, single_issue_nums[i]); + fprintf(fout, "\n"); + fprintf(fout, "dual_issue_nums: "); + for (unsigned i = 0; i < m_config->gpgpu_num_sched_per_core; i++) + fprintf(fout, "WS%d:%d\t", i, dual_issue_nums[i]); + fprintf(fout, "\n"); m_outgoing_traffic_stats->print(fout); m_incoming_traffic_stats->print(fout); @@ -669,13 +730,15 @@ void shader_core_ctx::fetch() } if( did_exit ) m_warp[warp_id].set_done_exit(); + --m_active_warps; + assert(m_active_warps >= 0); } // this code fetches instructions from the i-cache or generates memory requests if( !m_warp[warp_id].functional_done() && !m_warp[warp_id].imiss_pending() && m_warp[warp_id].ibuffer_empty() ) { address_type pc = m_warp[warp_id].get_pc(); address_type ppc = pc + PROGRAM_MEM_START; - unsigned nbytes=16; + unsigned nbytes=16; unsigned offset_in_block = pc & (m_config->m_L1I_config.get_line_sz()-1); if( (offset_in_block+nbytes) > m_config->m_L1I_config.get_line_sz() ) nbytes = (m_config->m_L1I_config.get_line_sz()-offset_in_block); @@ -750,10 +813,19 @@ void shader_core_ctx::issue_warp( register_set& pipe_reg_set, const warp_inst_t* } void shader_core_ctx::issue(){ + + //Ensure fair round robin issu between schedulers + unsigned j; + for (unsigned i = 0; i < schedulers.size(); i++) { + j = (Issue_Prio + i) % schedulers.size(); + schedulers[j]->cycle(); + } + Issue_Prio = (Issue_Prio+1)% schedulers.size(); + //really is issue; - for (unsigned i = 0; i < schedulers.size(); i++) { - schedulers[i]->cycle(); - } + //for (unsigned i = 0; i < schedulers.size(); i++) { + // schedulers[i]->cycle(); + //} } shd_warp_t& scheduler_unit::warp(int i){ @@ -868,7 +940,10 @@ void scheduler_unit::cycle() unsigned warp_id = (*iter)->get_warp_id(); unsigned checked=0; unsigned issued=0; - unsigned max_issue = m_shader->m_config->gpgpu_max_insn_issue_per_warp; + exec_unit_type_t previous_issued_inst_exec_type = exec_unit_type_t::NONE; + unsigned max_issue = m_shader->m_config->gpgpu_max_insn_issue_per_warp; + bool diff_exec_units = m_shader->m_config->gpgpu_dual_issue_diff_exec_units; + 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; @@ -901,18 +976,21 @@ void scheduler_unit::cycle() ready_inst = true; const active_mask_t &active_mask = m_simt_stack[warp_id]->get_active_mask(); assert( warp(warp_id).inst_in_pipeline() ); - if ( (pI->op == LOAD_OP)||(pI->op ==TENSOR_CORE_LOAD_OP)|| (pI->op == STORE_OP)|| (pI->op==TENSOR_CORE_STORE_OP) || (pI->op == MEMORY_BARRIER_OP) ) { - if( m_mem_out->has_free() ) { - m_shader->issue_warp(*m_mem_out,pI,active_mask,warp_id); - issued++; - issued_inst=true; - warp_inst_issued = true; - } + if ( (pI->op == LOAD_OP) || (pI->op == STORE_OP) || (pI->op == MEMORY_BARRIER_OP)||(pI->op==TENSOR_CORE_LOAD_OP)||(pI->op==TENSOR_CORE_STORE_OP) ) { + if( m_mem_out->has_free() && (!diff_exec_units || previous_issued_inst_exec_type != exec_unit_type_t::MEM)) { + m_shader->issue_warp(*m_mem_out,pI,active_mask,warp_id); + issued++; + issued_inst=true; + warp_inst_issued = true; + previous_issued_inst_exec_type = exec_unit_type_t::MEM; + } } else { + bool sp_pipe_avail = m_sp_out->has_free(); bool sfu_pipe_avail = m_sfu_out->has_free(); bool tensor_core_pipe_avail = m_tensor_core_out->has_free(); - if( sp_pipe_avail && (pI->op != SFU_OP) && (pI->op != TENSOR_CORE_OP) ) { + bool dp_pipe_avail = m_dp_out->has_free(); + if( sp_pipe_avail && (pI->op != TENSOR_CORE_OP) && (pI->op != SFU_OP && pI->op != DP_OP) && (!diff_exec_units || previous_issued_inst_exec_type != exec_unit_type_t::SP)) { //Jin: special for CDP api if(pI->m_is_cdp && !warp(warp_id).m_cdp_dummy) { @@ -937,12 +1015,23 @@ void scheduler_unit::cycle() issued++; issued_inst=true; warp_inst_issued = true; - } else if ( (pI->op == SFU_OP) || (pI->op == ALU_SFU_OP) ) { + previous_issued_inst_exec_type = exec_unit_type_t::SP; + } else if ( (m_shader->m_config->gpgpu_num_dp_units != 0) && (pI->op == DP_OP) && (!diff_exec_units || previous_issued_inst_exec_type != exec_unit_type_t::DP)) { + if( dp_pipe_avail ) { + m_shader->issue_warp(*m_dp_out,pI,active_mask,warp_id); + issued++; + issued_inst=true; + warp_inst_issued = true; + previous_issued_inst_exec_type = exec_unit_type_t::DP; + } + } //If the DP units = 0 (like in Fermi archi), then change DP inst to SFU inst + else if ( ((m_shader->m_config->gpgpu_num_dp_units == 0 && pI->op == DP_OP) || (pI->op == SFU_OP) || (pI->op == ALU_SFU_OP)) && (!diff_exec_units || previous_issued_inst_exec_type != exec_unit_type_t::SFU)) { if( sfu_pipe_avail ) { m_shader->issue_warp(*m_sfu_out,pI,active_mask,warp_id); issued++; issued_inst=true; warp_inst_issued = true; + previous_issued_inst_exec_type = exec_unit_type_t::SFU; } } else if ( (pI->op == TENSOR_CORE_OP) ) { @@ -952,12 +1041,12 @@ void scheduler_unit::cycle() issued_inst=true; warp_inst_issued = true; } - } - } - } else { + } + }//end of else + } else { SCHED_DPRINTF( "Warp (warp_id %u, dynamic_warp_id %u) fails scoreboard\n", (*iter)->get_warp_id(), (*iter)->get_dynamic_warp_id() ); - } + } } } else if( valid ) { // this case can happen after a return instruction in diverged warp @@ -988,6 +1077,14 @@ void scheduler_unit::cycle() m_last_supervised_issued = supervised_iter; } } + + if(issued == 1) + m_stats->single_issue_nums[m_id]++; + else if(issued > 1) + m_stats->dual_issue_nums[m_id]++; + else + abort(); //issued should be > 0 + break; } } @@ -1045,6 +1142,16 @@ void gto_scheduler::order_warps() scheduler_unit::sort_warps_by_oldest_dynamic_id ); } +void oldest_scheduler::order_warps() +{ + order_by_priority( m_next_cycle_prioritized_warps, + m_supervised_warps, + m_last_supervised_issued, + m_supervised_warps.size(), + ORDERED_PRIORITY_FUNC_ONLY, + scheduler_unit::sort_warps_by_oldest_dynamic_id ); +} + void two_level_active_scheduler::do_on_warp_issued( unsigned warp_id, unsigned num_issued, @@ -1117,12 +1224,13 @@ swl_scheduler::swl_scheduler ( shader_core_stats* stats, shader_core_ctx* shader Scoreboard* scoreboard, simt_stack** simt, std::vector<shd_warp_t>* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* tensor_core_out, register_set* mem_out, int id, char* config_string ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, tensor_core_out, mem_out, id ) + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ) { unsigned m_prioritization_readin; int ret = sscanf( config_string, @@ -1386,8 +1494,14 @@ ldst_unit::process_cache_access( cache_t* cache, mem_stage_stall_type result = NO_RC_FAIL; bool write_sent = was_write_sent(events); bool read_sent = was_read_sent(events); - if( write_sent ) - m_core->inc_store_req( inst.warp_id() ); + if( write_sent ) { + unsigned inc_ack = (m_config->m_L1D_config.get_mshr_type() == SECTOR_ASSOC)? + (mf->get_data_size()/SECTOR_SIZE) : 1; + + for(unsigned i=0; i< inc_ack; ++i) + m_core->inc_store_req( inst.warp_id() ); + + } if ( status == HIT ) { assert( !read_sent ); inst.accessq_pop_back(); @@ -1399,7 +1513,7 @@ ldst_unit::process_cache_access( cache_t* cache, if( !write_sent ) delete mf; } else if ( status == RESERVATION_FAIL ) { - result = COAL_STALL; + result = BK_CONF; assert( !read_sent ); assert( !write_sent ); delete mf; @@ -1408,8 +1522,8 @@ ldst_unit::process_cache_access( cache_t* cache, //inst.clear_active( access.get_warp_mask() ); // threads in mf writeback when mf returns inst.accessq_pop_back(); } - if( !inst.accessq_empty() ) - result = BK_CONF; + if( !inst.accessq_empty() && result == NO_RC_FAIL) + result = COAL_STALL; return result; } @@ -1429,6 +1543,111 @@ mem_stage_stall_type ldst_unit::process_memory_access_queue( cache_t *cache, war return process_cache_access( cache, mf->get_addr(), inst, events, mf, status ); } +mem_stage_stall_type ldst_unit::process_memory_access_queue_l1cache( l1_cache *cache, warp_inst_t &inst ) +{ + mem_stage_stall_type result = NO_RC_FAIL; + if( inst.accessq_empty() ) + return result; + + mem_fetch *mf = m_mf_allocator->alloc(inst,inst.accessq_back()); + + if(m_config->m_L1D_config.l1_latency > 0) + { + if((l1_latency_queue[m_config->m_L1D_config.l1_latency-1]) == NULL) + { + l1_latency_queue[m_config->m_L1D_config.l1_latency-1] = mf; + + if( mf->get_inst().is_store() ) { + unsigned inc_ack = (m_config->m_L1D_config.get_mshr_type() == SECTOR_ASSOC)? + (mf->get_data_size()/SECTOR_SIZE) : 1; + + for(unsigned i=0; i< inc_ack; ++i) + m_core->inc_store_req( inst.warp_id() ); + } + + inst.accessq_pop_back(); + } + else + { + result = BK_CONF; + delete mf; + } + if( !inst.accessq_empty() && result !=BK_CONF) + result = COAL_STALL; + return result; + } + else + { + std::list<cache_event> events; + enum cache_request_status status = cache->access(mf->get_addr(),mf,gpu_sim_cycle+gpu_tot_sim_cycle,events); + return process_cache_access( cache, mf->get_addr(), inst, events, mf, status ); + } +} + +void ldst_unit::L1_latency_queue_cycle() +{ + //std::deque< std::pair<mem_fetch*,bool> >::iterator it = m_latency_queue.begin(); + if((l1_latency_queue[0]) != NULL) + { + mem_fetch* mf_next = l1_latency_queue[0]; + std::list<cache_event> events; + enum cache_request_status status = m_L1D->access(mf_next->get_addr(),mf_next,gpu_sim_cycle+gpu_tot_sim_cycle,events); + + bool write_sent = was_write_sent(events); + bool read_sent = was_read_sent(events); + + if ( status == HIT ) { + assert( !read_sent ); + l1_latency_queue[0] = NULL; + if ( mf_next->get_inst().is_load() ) { + for ( unsigned r=0; r < 4; r++) + if (mf_next->get_inst().out[r] > 0) + { + assert(m_pending_writes[mf_next->get_inst().warp_id()][mf_next->get_inst().out[r]]>0); + unsigned still_pending = --m_pending_writes[mf_next->get_inst().warp_id()][mf_next->get_inst().out[r]]; + if(!still_pending) + { + m_pending_writes[mf_next->get_inst().warp_id()].erase(mf_next->get_inst().out[r]); + m_scoreboard->releaseRegister(mf_next->get_inst().warp_id(),mf_next->get_inst().out[r]); + m_core->warp_inst_complete(mf_next->get_inst()); + } + } + } + + //For write hit in WB policy + if(mf_next->get_inst().is_store() && !write_sent) + { + unsigned dec_ack = (m_config->m_L1D_config.get_mshr_type() == SECTOR_ASSOC)? + (mf_next->get_data_size()/SECTOR_SIZE) : 1; + + mf_next->set_reply(); + + for(unsigned i=0; i< dec_ack; ++i) + m_core->store_ack(mf_next); + } + + if( !write_sent ) + delete mf_next; + + } else if ( status == RESERVATION_FAIL ) { + assert( !read_sent ); + assert( !write_sent ); + } else { + assert( status == MISS || status == HIT_RESERVED ); + l1_latency_queue[0] = NULL; + } + } + + for( unsigned stage = 0; stage<m_config->m_L1D_config.l1_latency-1; ++stage) + if( l1_latency_queue[stage] == NULL) { + l1_latency_queue[stage] = l1_latency_queue[stage+1] ; + l1_latency_queue[stage+1] = NULL; + } + +} + + + bool ldst_unit::constant_cycle( warp_inst_t &inst, mem_stage_stall_type &rc_fail, mem_stage_access_type &fail_type) { if( inst.empty() || ((inst.space.get_type() != const_space) && (inst.space.get_type() != param_space_kernel)) ) @@ -1478,7 +1697,7 @@ bool ldst_unit::memory_cycle( warp_inst_t &inst, mem_stage_stall_type &stall_rea bypassL1D = true; } else if (inst.space.is_global()) { // global memory access // skip L1 cache if the option is enabled - if (m_core->get_config()->gmem_skip_L1D) + if (m_core->get_config()->gmem_skip_L1D && (CACHE_L1 != inst.cache_op)) bypassL1D = true; } if( bypassL1D ) { @@ -1502,9 +1721,9 @@ bool ldst_unit::memory_cycle( warp_inst_t &inst, mem_stage_stall_type &stall_rea } } else { assert( CACHE_UNDEFINED != inst.cache_op ); - stall_cond = process_memory_access_queue(m_L1D,inst); + stall_cond = process_memory_access_queue_l1cache(m_L1D,inst); } - if( !inst.accessq_empty() ) + if( !inst.accessq_empty() && stall_cond == NO_RC_FAIL) stall_cond = COAL_STALL; if (stall_cond != NO_RC_FAIL) { stall_reason = stall_cond; @@ -1534,6 +1753,11 @@ void ldst_unit::flush(){ m_L1D->flush(); } +void ldst_unit::invalidate(){ + // Flush L1D cache + m_L1D->invalidate(); +} + simd_function_unit::simd_function_unit( const shader_core_config *config ) { m_config=config; @@ -1586,6 +1810,13 @@ void sp_unit::active_lanes_in_pipeline(){ m_core->incfuactivelanes_stat(active_count); m_core->incfumemactivelanes_stat(active_count); } +void dp_unit::active_lanes_in_pipeline(){ + unsigned active_count=pipelined_simd_unit::get_active_lanes_in_pipeline(); + assert(active_count<=m_core->get_config()->warp_size); + m_core->incspactivelanes_stat(active_count); + m_core->incfuactivelanes_stat(active_count); + m_core->incfumemactivelanes_stat(active_count); +} void sfu::active_lanes_in_pipeline(){ unsigned active_count=pipelined_simd_unit::get_active_lanes_in_pipeline(); @@ -1610,6 +1841,12 @@ sp_unit::sp_unit( register_set* result_port, const shader_core_config *config,sh m_name = "SP "; } +dp_unit::dp_unit( register_set* result_port, const shader_core_config *config,shader_core_ctx *core) + : pipelined_simd_unit(result_port,config,config->max_sfu_latency,core) +{ + m_name = "DP "; +} + void sp_unit :: issue(register_set& source_reg) { warp_inst_t** ready_reg = source_reg.get_ready(); @@ -1619,6 +1856,14 @@ void sp_unit :: issue(register_set& source_reg) pipelined_simd_unit::issue(source_reg); } +void dp_unit :: issue(register_set& source_reg) +{ + warp_inst_t** ready_reg = source_reg.get_ready(); + //m_core->incexecstat((*ready_reg)); + (*ready_reg)->op_pipe=DP__OP; + m_core->incsp_stat(m_core->get_config()->warp_size,(*ready_reg)->latency); + pipelined_simd_unit::issue(source_reg); +} pipelined_simd_unit::pipelined_simd_unit( register_set* result_port, const shader_core_config *config, unsigned max_latency,shader_core_ctx *core ) : simd_function_unit(config) @@ -1712,8 +1957,9 @@ ldst_unit::ldst_unit( mem_fetch_interface *icnt, const memory_config *mem_config, shader_core_stats *stats, unsigned sid, - unsigned tpc ) : pipelined_simd_unit(NULL,config,3,core), m_next_wb(config) + unsigned tpc ) : pipelined_simd_unit(NULL,config,config->smem_latency,core), m_next_wb(config) { + assert(config->smem_latency > 1); init( icnt, mf_allocator, core, @@ -1734,6 +1980,12 @@ ldst_unit::ldst_unit( mem_fetch_interface *icnt, m_icnt, m_mf_allocator, IN_L1D_MISS_QUEUE ); + + if(m_config->m_L1D_config.l1_latency > 0) + { + for(int i=0; i<m_config->m_L1D_config.l1_latency; i++ ) + l1_latency_queue.push_back((mem_fetch*)NULL); + } } } @@ -1916,12 +2168,12 @@ void ldst_unit::cycle() if( !m_response_fifo.empty() ) { mem_fetch *mf = m_response_fifo.front(); - if (mf->istexture()) { + if (mf->get_access_type() == TEXTURE_ACC_R) { if (m_L1T->fill_port_free()) { m_L1T->fill(mf,gpu_sim_cycle+gpu_tot_sim_cycle); m_response_fifo.pop_front(); } - } else if (mf->isconst()) { + } else if (mf->get_access_type() == CONST_ACC_R) { if (m_L1C->fill_port_free()) { mf->set_status(IN_SHADER_FETCHED,gpu_sim_cycle+gpu_tot_sim_cycle); m_L1C->fill(mf,gpu_sim_cycle+gpu_tot_sim_cycle); @@ -1960,7 +2212,11 @@ void ldst_unit::cycle() m_L1T->cycle(); m_L1C->cycle(); - if( m_L1D ) m_L1D->cycle(); + if( m_L1D ) { + m_L1D->cycle(); + if(m_config->m_L1D_config.l1_latency > 0) + L1_latency_queue_cycle(); + } warp_inst_t &pipe_reg = *m_dispatch_reg; enum mem_stage_stall_type rc_fail = NO_RC_FAIL; @@ -1983,9 +2239,9 @@ void ldst_unit::cycle() unsigned warp_id = pipe_reg.warp_id(); if( pipe_reg.is_load() ) { if( pipe_reg.space.get_type() == shared_space ) { - if( m_pipeline_reg[2]->empty() ) { + if( m_pipeline_reg[m_config->smem_latency-1]->empty() ) { // new shared memory request - move_warp(m_pipeline_reg[2],m_dispatch_reg); + move_warp(m_pipeline_reg[m_config->smem_latency-1],m_dispatch_reg); m_dispatch_reg->clear(); } } else { @@ -2562,9 +2818,42 @@ unsigned int shader_core_config::max_cta( const kernel_info_t &k ) const assert( result <= MAX_CTA_PER_SHADER ); if (result < 1) { printf ("GPGPU-Sim uArch: ERROR ** Kernel requires more resources than shader has.\n"); + if(gpgpu_ignore_resources_limitation) { + printf ("GPGPU-Sim uArch: gpgpu_ignore_resources_limitation is set, ignore the ERROR!\n"); + return 1; + } abort(); } + if(adpative_volta_cache_config && !k.volta_cache_config_set) { + //For Volta, we assign the remaining shared memory to L1 cache + //For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x + unsigned total_shmed = kernel_info->smem * result; + assert(total_shmed >=0 && total_shmed <= gpgpu_shmem_size); + assert(gpgpu_shmem_size == 98304); //Volta has 96 KB shared + assert(m_L1D_config.get_nset() == 4); //Volta L1 has four sets + if(total_shmed < gpgpu_shmem_size){ + if(total_shmed == 0) + m_L1D_config.set_assoc(256); //L1 is 128KB ans shd=0 + else if(total_shmed > 0 && total_shmed <= 8192) + m_L1D_config.set_assoc(240); //L1 is 120KB ans shd=8KB + else if(total_shmed > 8192 && total_shmed <= 16384) + m_L1D_config.set_assoc(224); //L1 is 112KB ans shd=16KB + else if(total_shmed > 16384 && total_shmed <= 32768) + m_L1D_config.set_assoc(192); //L1 is 96KB ans shd=32KB + else if(total_shmed > 32768 && total_shmed <= 65536) + m_L1D_config.set_assoc(128); //L1 is 64KB ans shd=64KB + else if(total_shmed > 65536 && total_shmed <= gpgpu_shmem_size) + m_L1D_config.set_assoc(64); //L1 is 32KB and shd=96KB + else + assert(0); + + printf ("GPGPU-Sim: Reconfigure L1 cache in Volta Archi to %uKB\n", m_L1D_config.get_total_size_inKB()); + } + + k.volta_cache_config_set = true; + } + return result; } @@ -2586,6 +2875,11 @@ void shader_core_ctx::cache_flush() m_ldst_unit->flush(); } +void shader_core_ctx::cache_invalidate() +{ + m_ldst_unit->invalidate(); +} + // modifiers std::list<opndcoll_rfu_t::op_t> opndcoll_rfu_t::arbiter_t::allocate_reads() { @@ -3340,6 +3634,15 @@ void simt_core_cluster::print_not_completed( FILE *fp ) const } } + +float simt_core_cluster::get_current_occupancy( unsigned long long& active, unsigned long long& total ) const { + float aggregate = 0.f; + for( unsigned i=0; i < m_config->n_simt_cores_per_cluster; i++ ) { + aggregate+=m_core[i]->get_current_occupancy( active, total ); + } + return aggregate / m_config->n_simt_cores_per_cluster; +} + unsigned simt_core_cluster::get_n_active_cta() const { unsigned n=0; @@ -3403,6 +3706,12 @@ void simt_core_cluster::cache_flush() m_core[i]->cache_flush(); } +void simt_core_cluster::cache_invalidate() +{ + for( unsigned i=0; i < m_config->n_simt_cores_per_cluster; i++ ) + m_core[i]->cache_invalidate(); +} + bool simt_core_cluster::icnt_injection_buffer_full(unsigned size, bool write) { unsigned request_size = size; diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 97e438f..437506c 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -70,6 +70,14 @@ #define WRITE_MASK_SIZE 8 +enum exec_unit_type_t +{ + NONE = 0, + SP = 1, + SFU = 2, + MEM = 3, + DP = 4 +}; class thread_ctx_t { public: @@ -308,6 +316,7 @@ enum concrete_scheduler CONCRETE_SCHEDULER_GTO, CONCRETE_SCHEDULER_TWO_LEVEL_ACTIVE, CONCRETE_SCHEDULER_WARP_LIMITING, + CONCRETE_SCHEDULER_OLDEST_FIRST, NUM_CONCRETE_SCHEDULERS }; @@ -317,13 +326,14 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector<shd_warp_t>* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* tensor_core_out, register_set* mem_out, int id) : m_supervised_warps(), m_stats(stats), m_shader(shader), m_scoreboard(scoreboard), m_simt_stack(simt), /*m_pipeline_reg(pipe_regs),*/ m_warp(warp), - m_sp_out(sp_out),m_sfu_out(sfu_out),m_tensor_core_out(tensor_core_out),m_mem_out(mem_out), m_id(id){} + m_sp_out(sp_out),m_dp_out(dp_out),m_sfu_out(sfu_out),m_tensor_core_out(tensor_core_out),m_mem_out(mem_out), m_id(id){} virtual ~scheduler_unit(){} virtual void add_supervised_warp_id(int i) { m_supervised_warps.push_back(&warp(i)); @@ -395,6 +405,7 @@ protected: //warp_inst_t** m_pipeline_reg; std::vector<shd_warp_t>* m_warp; register_set* m_sp_out; + register_set* m_dp_out; register_set* m_sfu_out; register_set* m_tensor_core_out; register_set* m_mem_out; @@ -408,11 +419,12 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector<shd_warp_t>* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* tensor_core_out, register_set* mem_out, int id ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, tensor_core_out, mem_out, id ){} + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ){} virtual ~lrr_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -426,11 +438,12 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector<shd_warp_t>* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* tensor_core_out, register_set* mem_out, int id ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, tensor_core_out, mem_out, id ){} + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ){} virtual ~gto_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -439,6 +452,25 @@ public: }; +class oldest_scheduler : public scheduler_unit { +public: + oldest_scheduler ( shader_core_stats* stats, shader_core_ctx* shader, + Scoreboard* scoreboard, simt_stack** simt, + std::vector<shd_warp_t>* warp, + register_set* sp_out, + register_set* dp_out, + register_set* sfu_out, + register_set* tensor_core_out, + register_set* mem_out, + int id ) + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ){} + virtual ~oldest_scheduler () {} + virtual void order_warps (); + virtual void done_adding_supervised_warps() { + m_last_supervised_issued = m_supervised_warps.begin(); + } + +}; class two_level_active_scheduler : public scheduler_unit { public: @@ -446,12 +478,13 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector<shd_warp_t>* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* tensor_core_out, register_set* mem_out, int id, char* config_str ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, tensor_core_out, mem_out, id ), + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ), m_pending_warps() { unsigned inner_level_readin; @@ -497,6 +530,7 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector<shd_warp_t>* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* tensor_core_out, register_set* mem_out, @@ -1060,6 +1094,23 @@ public: switch(inst.op) { case SFU_OP: break; case ALU_SFU_OP: break; + case DP_OP: break; //for compute <= 29 (i..e Fermi and GT200) + default: return false; + } + return pipelined_simd_unit::can_issue(inst); + } + virtual void active_lanes_in_pipeline(); + virtual void issue( register_set& source_reg ); +}; + +class dp_unit : public pipelined_simd_unit +{ +public: + dp_unit( register_set* result_port, const shader_core_config *config, shader_core_ctx *core ); + virtual bool can_issue( const warp_inst_t &inst ) const + { + switch(inst.op) { + case DP_OP: break; default: return false; } return pipelined_simd_unit::can_issue(inst); @@ -1098,6 +1149,7 @@ public: case STORE_OP: return false; case TENSOR_CORE_STORE_OP: return false; case MEMORY_BARRIER_OP: return false; + case DP_OP: return false; default: break; } return pipelined_simd_unit::can_issue(inst); @@ -1129,6 +1181,7 @@ public: void fill( mem_fetch *mf ); void flush(); + void invalidate(); void writeback(); // accessors @@ -1195,6 +1248,7 @@ protected: mem_fetch *mf, enum cache_request_status status ); mem_stage_stall_type process_memory_access_queue( cache_t *cache, warp_inst_t &inst ); + mem_stage_stall_type process_memory_access_queue_l1cache( l1_cache *cache, warp_inst_t &inst ); const memory_config *m_memory_config; class mem_fetch_interface *m_icnt; @@ -1223,13 +1277,18 @@ protected: // for debugging unsigned long long m_last_inst_gpu_sim_cycle; unsigned long long m_last_inst_gpu_tot_sim_cycle; + + std::deque<mem_fetch* > l1_latency_queue; + void L1_latency_queue_cycle(); }; enum pipeline_stage_name_t { ID_OC_SP=0, + ID_OC_DP, ID_OC_SFU, ID_OC_MEM, OC_EX_SP, + OC_EX_DP, OC_EX_SFU, OC_EX_MEM, EX_WB, @@ -1240,9 +1299,11 @@ enum pipeline_stage_name_t { const char* const pipeline_stage_name_decode[] = { "ID_OC_SP", + "ID_OC_DP", "ID_OC_SFU", "ID_OC_MEM", "OC_EX_SP", + "OC_EX_DP", "OC_EX_SFU", "OC_EX_MEM", "EX_WB", @@ -1340,27 +1401,29 @@ struct shader_core_config : public core_config mutable cache_config m_L1C_config; mutable l1d_cache_config m_L1D_config; - bool gmem_skip_L1D; // on = global memory access always skip the L1 cache - bool gpgpu_dwf_reg_bankconflict; int gpgpu_num_sched_per_core; int gpgpu_max_insn_issue_per_warp; + bool gpgpu_dual_issue_diff_exec_units; //op collector int gpgpu_operand_collector_num_units_sp; + int gpgpu_operand_collector_num_units_dp; int gpgpu_operand_collector_num_units_sfu; int gpgpu_operand_collector_num_units_tensor_core; int gpgpu_operand_collector_num_units_mem; int gpgpu_operand_collector_num_units_gen; unsigned int gpgpu_operand_collector_num_in_ports_sp; + unsigned int gpgpu_operand_collector_num_in_ports_dp; unsigned int gpgpu_operand_collector_num_in_ports_sfu; unsigned int gpgpu_operand_collector_num_in_ports_tensor_core; unsigned int gpgpu_operand_collector_num_in_ports_mem; unsigned int gpgpu_operand_collector_num_in_ports_gen; unsigned int gpgpu_operand_collector_num_out_ports_sp; + unsigned int gpgpu_operand_collector_num_out_ports_dp; unsigned int gpgpu_operand_collector_num_out_ports_sfu; unsigned int gpgpu_operand_collector_num_out_ports_tensor_core; unsigned int gpgpu_operand_collector_num_out_ports_mem; @@ -1368,6 +1431,7 @@ struct shader_core_config : public core_config int gpgpu_num_sp_units; int gpgpu_tensor_core_avail; + int gpgpu_num_dp_units; int gpgpu_num_sfu_units; int gpgpu_num_tensor_core_units; int gpgpu_num_mem_units; @@ -1379,6 +1443,7 @@ struct shader_core_config : public core_config unsigned gpgpu_num_reg_banks; bool gpgpu_reg_bank_use_warp_id; bool gpgpu_local_mem_map; + bool gpgpu_ignore_resources_limitation; unsigned max_sp_latency; unsigned max_sfu_latency; @@ -1391,10 +1456,14 @@ struct shader_core_config : public core_config int simt_core_sim_order; + unsigned smem_latency; + unsigned mem2device(unsigned memid) const { return memid + n_simt_clusters; } //Jin: concurrent kernel on sm bool gpgpu_concurrent_kernel_sm; + + bool adpative_volta_cache_config; }; struct shader_core_stats_pod { @@ -1457,6 +1526,8 @@ struct shader_core_stats_pod { unsigned *last_shader_cycle_distro; unsigned *num_warps_issuable; unsigned gpgpu_n_stall_shd_mem; + unsigned* single_issue_nums; + unsigned* dual_issue_nums; //memory access classification int gpgpu_n_mem_read_local; @@ -1528,6 +1599,8 @@ public: m_n_diverge = (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); shader_cycle_distro = (unsigned*) calloc(config->warp_size+3, sizeof(unsigned)); last_shader_cycle_distro = (unsigned*) calloc(m_config->warp_size+3, sizeof(unsigned)); + single_issue_nums = (unsigned*) calloc(config->gpgpu_num_sched_per_core,sizeof(unsigned)); + dual_issue_nums = (unsigned*) calloc(config->gpgpu_num_sched_per_core, sizeof(unsigned)); n_simt_to_mem = (long *)calloc(config->num_shader(), sizeof(long)); n_mem_to_simt = (long *)calloc(config->num_shader(), sizeof(long)); @@ -1651,6 +1724,7 @@ public: void issue_block2core( class kernel_info_t &kernel ); void cache_flush(); + void cache_invalidate(); void accept_fetch_response( mem_fetch *mf ); void accept_ldst_unit_response( class mem_fetch * mf ); void broadcast_barrier_reduction(unsigned cta_id, unsigned bar_id,warp_set_t warps); @@ -1679,6 +1753,7 @@ public: // accessors virtual bool warp_waiting_at_barrier( unsigned warp_id ) const; void get_pdom_stack_top_info( unsigned tid, unsigned *pc, unsigned *rpc ) const; + float get_current_occupancy( unsigned long long & active, unsigned long long & total ) const; // used by pipeline timing model components: // modifiers @@ -1885,10 +1960,14 @@ public: std::vector<register_set> m_pipeline_reg; Scoreboard *m_scoreboard; opndcoll_rfu_t m_operand_collector; + int m_active_warps; //schedule std::vector<scheduler_unit*> schedulers; + //issue + unsigned int Issue_Prio; + // execute unsigned m_num_function_units; std::vector<pipeline_stage_name_t> m_dispatch_port; @@ -1940,6 +2019,7 @@ public: void reinit(); unsigned issue_block2core(); void cache_flush(); + void cache_invalidate(); bool icnt_injection_buffer_full(unsigned size, bool write); void icnt_inject_request_packet(class mem_fetch *mf); @@ -1970,6 +2050,7 @@ public: void get_L1T_sub_stats(struct cache_sub_stats &css) const; void get_icnt_stats(long &n_simt_to_mem, long &n_mem_to_simt) const; + float get_current_occupancy( unsigned long long& active, unsigned long long & total ) const; private: unsigned m_cluster_id; diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index ede9f20..52e2f5e 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -172,6 +172,7 @@ void *gpgpu_sim_thread_concurrent(void*) g_sim_active = false; pthread_mutex_unlock(&g_sim_lock); } while( !g_sim_done ); + printf("GPGPU-Sim: *** simulation thread exiting ***\n"); fflush(stdout); diff --git a/src/gpuwattch/makefile b/src/gpuwattch/makefile index ab718cc..354c9ec 100644 --- a/src/gpuwattch/makefile +++ b/src/gpuwattch/makefile @@ -11,10 +11,10 @@ opt: $(TAR).mk obj_opt @$(MAKE) TAG=opt -C . -f $(TAR).mk obj_dbg: - mkdir $@ + mkdir -p $@ obj_opt: - mkdir $@ + mkdir -p $@ depend: @$(MAKE) TAG=opt -C . -f $(TAR).mk depend diff --git a/src/intersim2/Makefile b/src/intersim2/Makefile index bd42000..7d10b3f 100644 --- a/src/intersim2/Makefile +++ b/src/intersim2/Makefile @@ -36,7 +36,7 @@ DEBUG ?= 0 LEX = flex YACC = bison -y DEFINE = #-DTRACK_STALLS -DTRACK_BUFFERS -DTRACK_FLOWS -DTRACK_CREDITS -INCPATH = -I. -Iarbiters -Iallocators -Irouters -Inetworks -Ipower +INCPATH = -I. -Iarbiters -Iallocators -Irouters -Inetworks -Ipower -I$(GPGPUSIM_ROOT)/src ifeq ($(CREATE_LIBRARY),1) INCPATH += -I$(GPGPUSIM_ROOT)/src/gpgpu-sim/ diff --git a/src/intersim2/interconnect_interface.cpp b/src/intersim2/interconnect_interface.cpp index 4386821..1e1a2d7 100644 --- a/src/intersim2/interconnect_interface.cpp +++ b/src/intersim2/interconnect_interface.cpp @@ -44,6 +44,7 @@ #include "booksim.hpp" #include "intersim_config.hpp" #include "network.hpp" +#include "trace.h" InterconnectInterface* InterconnectInterface::New(const char* const config_file) { @@ -147,6 +148,8 @@ void InterconnectInterface::Push(unsigned input_deviceID, unsigned output_device { // it should have free buffer assert(HasBuffer(input_deviceID, size)); + + DPRINTF(INTERCONNECT, "Sent %d bytes from %d to %d", size, input_deviceID, output_deviceID); int output_icntID = _node_map[output_deviceID]; int input_icntID = _node_map[input_deviceID]; @@ -178,7 +181,11 @@ void InterconnectInterface::Push(unsigned input_deviceID, unsigned output_device case WRITE_REQUEST: packet_type = Flit::WRITE_REQUEST ;break; case READ_REPLY: packet_type = Flit::READ_REPLY ;break; case WRITE_ACK: packet_type = Flit::WRITE_REPLY ;break; - default: assert (0); + default: + { + cout<<"Type "<<mf->get_type()<<" is undefined!"<<endl; + assert (0 && "Type is undefined"); + } } //TODO: _include_queuing ? diff --git a/src/trace_streams.tup b/src/trace_streams.tup index fd68a16..074c7c8 100644 --- a/src/trace_streams.tup +++ b/src/trace_streams.tup @@ -29,6 +29,8 @@ TS_TUP_BEGIN( trace_streams_type ) TS_TUP( WARP_SCHEDULER ), TS_TUP( SCOREBOARD ), TS_TUP( MEMORY_PARTITION_UNIT ), + TS_TUP( MEMORY_SUBPARTITION_UNIT ), + TS_TUP( INTERCONNECT ), TS_TUP( LIVENESS ), TS_TUP( NUM_TRACE_STREAMS ) TS_TUP_END( trace_streams_type ) @@ -1 +1 @@ -const char *g_gpgpusim_version_string = "GPGPU-Sim Simulator Version 3.2.2 "; +const char *g_gpgpusim_version_string = "GPGPU-Sim Simulator Version 4.0.0 "; |