Add support for different SPR/ICX SKUs

closed/Intel/code/3d-unet-99.9/pytorch-cpu-kits19/README.md

@@ -86,4 +86,3 @@ bash make_preprocess.sh
 bash run_SPR56C_2S.sh acc
 bash run_SPR56C_2S.sh perf
 ```
-

closed/Intel/code/3d-unet-99.9/pytorch-cpu-kits19/env_acc.list

@@ -0,0 +1 @@
+RUN_TYPE=acc

closed/Intel/code/3d-unet-99.9/pytorch-cpu-kits19/env_perf.list

@@ -0,0 +1 @@
+RUN_TYPE=perf

closed/Intel/code/3d-unet-99.9/pytorch-cpu-kits19/run_offline.sh

@@ -0,0 +1,19 @@
+number_threads=`nproc --all`
+echo $number_threads
+number_instance=$((number_threads/8))
+echo $number_instance
+
+
+if [ -z "${RUN_TYPE}" ]; then
+    echo "NO RUN_TYPE ( perf/acc)found export RUN_TYPE=perf"
+    export RUN_TYPE=perf
+fi
+echo $RUN_TYPE
+bash run_mlperf.sh --type=${RUN_TYPE} \
+	           --precision=int8 \
+		   --user-conf=/workspace/user.conf \
+		   --num-instance=$number_instance \
+		   --cpus-per-instance=4 \
+                   --scenario=Offline
+
+

closed/Intel/code/3d-unet-99.9/pytorch-cpu-kits19/user.conf

@@ -1,5 +1,5 @@
 # Please set these fields depending on the performance of your system to
 # override default LoadGen settings.
 *.Offline.target_qps = 1.5
-#*.Offline.min_query_count = 1200
-#*.Offline.min_duration = 6000
+*.Offline.min_query_count = 1200
+*.Offline.min_duration = 6000

closed/Intel/code/README.md

@@ -0,0 +1,59 @@
+# MLPerf Inference Benchmarks on AWS instances
+
+## Setup Instructions
+
+
+
+## How to Run on AWS
+
+### Docker images list
+
+| IMAGE ID     | IMAGE REPO and TAG  |                                         
+| --------- | ------------------------------------------------ |
+| ce03546c4d75 | intel/intel-optimized-pytorch:mlperf-inference-2.1-3dunet_v2              | 
+| a17e4ea194cc | intel/intel-optimized-pytorch:mlperf-infernece-2.1-resnet50_v2 | 
+| 038cf950d509 | intel/intel-optimized-pytorch:mlperf-inference-2.1-dlrm_v2                   | 
+| 5dfdb372208a | intel/intel-optimized-pytorch:mlperf-inference-2.1-datacenter-retinanet_v2                   | 
+
+### 3dunet
+
+```
+cd 3d-unet-99.9/pytorch-cpu-kits19/
+```
+
+You could run the workload with prepared AWS AMI image.  
+please replace the IMAGE ID, YOUR SCRIPT and PATH accordingly.
+```
+../../run_docker.sh <IMAGE ID>  /workspace/<YOUR SCRIPT in current host folder> <PATH of 3dunet codes in docker image> aws_3dunet
+```
+Here is the example, and it should also work with the prepared AMI image.
+```
+../../run_docker.sh ce03546c4d75  /workspace/run_offline.sh /opt/workdir/code/3d-unet-99.9/pytorch-cpu-kits19 aws_3dunet
+```
+
+### Resnet50
+
+```
+cd resnet50/pytorch-cpu/
+```
+```
+../../run_docker.sh a17e4ea194cc  /workspace/run_offline.sh  /opt/workdir/code/resnet50/pytorch-cpu aws_rn50
+```
+
+### DLRM
+
+```
+cd dlrm-99.9/pytorch-cpu/
+```
+```
+../../run_docker.sh 038cf950d509  /workspace/run_offline.sh  /opt/workdir/code/dlrm-99.9/pytorch-cpu aws_dlrm
+```
+> Please change 'run_offline.sh' to 'run_offline_accuracy.sh', 'run_server.sh' or 'run_server_accuracy.sh' accordingly, based on your desired benchmark type and the mode.
+### Retinanet
+
+```
+cd retinanet/pytorch-cpu/
+```
+```
+../../run_docker.sh 5dfdb372208a  /workspace/run_offline.sh  /opt/workdir/code/retinanet/pytorch-cpu aws_ret
+```

closed/Intel/code/dlrm-99.9/pytorch-cpu/run_local.sh

@@ -1,9 +1,9 @@
 #!/bin/bash
 
-source ./run_common.sh
+source /workspace/run_common.sh
 
-common_opt="--config ./mlperf.conf"
-OUTPUT_DIR=$PWD/output/$name/$mode/$test_type
+common_opt="--config /workspace/mlperf.conf"
+OUTPUT_DIR=/workspace/output/$name/$mode/$test_type
 if [[ $test_type == "performance" ]]; then
     OUTPUT_DIR=$OUTPUT_DIR/run_1
 fi

closed/Intel/code/dlrm-99.9/pytorch-cpu/run_main.sh

@@ -31,4 +31,4 @@ if [ $1 == "server" ]; then
 fi
 
 echo "Running $mode bs=$batch_size $dtype $test_type $DNNL_MAX_CPU_ISA"
-./run_local.sh pytorch dlrm terabyte cpu $dtype $test_type --scenario $mode --max-ind-range=40000000 --samples-to-aggregate-quantile-file=../tools/dist_quantile.txt --max-batchsize=$batch_size $extra_option
+/workspace/run_local.sh pytorch dlrm terabyte cpu $dtype $test_type --scenario $mode --max-ind-range=40000000 --samples-to-aggregate-quantile-file=../tools/dist_quantile.txt --max-batchsize=$batch_size $extra_option

closed/Intel/code/dlrm-99.9/pytorch-cpu/run_mlperf.sh

@@ -26,17 +26,18 @@ do
 done
 
 WORK_DIR=$PWD
-
-source ./setup_dataset.sh
+pwd
+ls
+source /workspace/setup_dataset.sh
 
 if [ ${mode} = "offline" ]; then
     scenario="offline"
-    source ./setup_env_offline.sh
+    source /workspace/setup_env_offline_ICX.sh
 fi
 
 if [ ${mode} = "server" ]; then
     scenario="server"
-    source ./setup_env_server.sh
+    source /workspace/setup_env_server_ICX.sh
 fi
 
 if [ ${run_type} = "perf" ];then
@@ -52,5 +53,5 @@ if [ ${dtype} = "fp32" ];then
 fi
 
 echo ${mode} $BATCH_SIZE ${run_type} ${dtype} "mode"
-sudo ./run_clean.sh
-./run_main.sh ${scenario} ${accuracy} ${dtype}
+sudo /workspace/run_clean.sh
+/workspace/run_main.sh ${scenario} ${accuracy} ${dtype}

closed/Intel/code/dlrm-99.9/pytorch-cpu/run_offline.sh

@@ -0,0 +1 @@
+/workspace/run_mlperf.sh --mode=offline --type=perf --dtype=int8

closed/Intel/code/dlrm-99.9/pytorch-cpu/run_offline_accuracy.sh

@@ -0,0 +1 @@
+/workspace/run_mlperf.sh --mode=offline --type=acc --dtype=int8

closed/Intel/code/dlrm-99.9/pytorch-cpu/run_server.sh

@@ -0,0 +1 @@
+/workspace/run_mlperf.sh --mode=server --type=perf --dtype=int8

closed/Intel/code/dlrm-99.9/pytorch-cpu/run_server_accuracy.sh

@@ -0,0 +1 @@
+/workspace/run_mlperf.sh --mode=server --type=acc --dtype=int8

closed/Intel/code/dlrm-99.9/pytorch-cpu/setup_dataset.sh

@@ -1,2 +1,2 @@
 export DATA_DIR=/data/mlperf_data/dlrm/
-export MODEL_DIR=/data/mlperf_data/dlrm/
+export MODEL_DIR=/media

closed/Intel/code/dlrm-99.9/pytorch-cpu/setup_env_offline_ICX.sh

@@ -0,0 +1,12 @@
+number_threads=`nproc --all`
+number_cores=$((number_threads/2))
+number_sockets=`grep physical.id /proc/cpuinfo | sort -u | wc -l`
+cpu_per_socket=$((number_cores/number_sockets))
+
+export NUM_SOCKETS=$number_sockets        # i.e. 8
+export CPUS_PER_SOCKET=$cpu_per_socket   # i.e. 28
+export CPUS_PER_PROCESS=$cpu_per_socket  # which determine how much processes will be used
+                            # process-per-socket = CPUS_PER_SOCKET/CPUS_PER_PROCESS
+export CPUS_PER_INSTANCE=1  # instance-per-process number=CPUS_PER_PROCESS/CPUS_PER_INSTANCE
+                  # total-instance = instance-per-process * process-per-socket
+export BATCH_SIZE=16000

closed/Intel/code/dlrm-99.9/pytorch-cpu/setup_env_server_ICX.sh

@@ -0,0 +1,12 @@
+number_threads=`nproc --all`
+number_cores=$((number_threads/2))
+number_sockets=`grep physical.id /proc/cpuinfo | sort -u | wc -l`
+cpu_per_socket=$((number_cores/number_sockets))
+
+export NUM_SOCKETS=$number_sockets        # i.e. 8
+export CPUS_PER_SOCKET=$cpu_per_socket   # i.e. 28
+export CPUS_PER_PROCESS=$cpu_per_socket  # which determine how much processes will be used
+                                         # process-per-socket = CPUS_PER_SOCKET/CPUS_PER_PROCESS
+export CPUS_PER_INSTANCE=8  # instance-per-process number=CPUS_PER_PROCESS/CPUS_PER_INSTANCE
+                             # total-instance = instance-per-process * process-per-socket
+export BATCH_SIZE=8000

closed/Intel/code/dlrm-99.9/pytorch-cpu/user.conf.ICX40C_2S

@@ -0,0 +1,2 @@
+dlrm.Server.target_qps = 14000.0
+dlrm.Offline.target_qps = 100430.0

closed/Intel/code/offline_test.sh

@@ -0,0 +1,26 @@
+cd resnet50/pytorch-cpu/
+chmod +x run_offline_accuracy.sh
+chmod +x run_offline.sh
+../../run_docker.sh a17e4ea194cc  /workspace/run_offline_accuracy.sh  /opt/workdir/code/resnet50/pytorch-cpu aws_rn50  resnet50_acc.txt
+../../run_docker.sh a17e4ea194cc  /workspace/run_offline.sh  /opt/workdir/code/resnet50/pytorch-cpu aws_rn50  resnet50_perf.txt
+cd ../../
+
+cd dlrm-99.9/pytorch-cpu/
+chmod +x run_offline_accuracy.sh
+chmod +x run_offline.sh
+../../run_docker.sh 038cf950d509  /workspace/run_offline_accuracy.sh  /opt/workdir/code/dlrm-99.9/pytorch-cpu aws_dlrm  dlrm_acc.txt
+../../run_docker.sh 038cf950d509  /workspace/run_offline.sh  /opt/workdir/code/dlrm-99.9/pytorch-cpu aws_dlrm  dlrm_perf.txt
+cd ../../
+
+cd retinanet/pytorch-cpu/
+chmod +x run_offline_accuracy.sh
+chmod +x run_offline.sh
+../../run_docker.sh 5dfdb372208a  /workspace/run_offline_accuracy.sh  /opt/workdir/code/retinanet/pytorch-cpu aws_ret  retinanet_acc.txt
+../../run_docker.sh 5dfdb372208a  /workspace/run_offline.sh  /opt/workdir/code/retinanet/pytorch-cpu aws_ret  retinanet_perf.txt
+cd ../../
+
+cd 3d-unet-99.9/pytorch-cpu-kits19/
+chmod +x run_offline.sh
+../../run_docker.sh ce03546c4d75  /workspace/run_offline.sh /opt/workdir/code/3d-unet-99.9/pytorch-cpu-kits19 aws_3dunet  3dunet_acc.txt env_acc.list
+../../run_docker.sh ce03546c4d75  /workspace/run_offline.sh /opt/workdir/code/3d-unet-99.9/pytorch-cpu-kits19 aws_3dunet  3dunet_perf.txt env_perf.list 
+cd ../../

closed/Intel/code/resnet50/pytorch-cpu/README.md

@@ -33,6 +33,8 @@ git clone -b v2.6 https://github.com/oneapi-src/oneDNN.git onednn
   wget https://repo.anaconda.com/archive/Anaconda3-2022.05-Linux-x86_64.sh
   bash Anaconda3-2022.05-Linux-x86_64.sh
   ```
+  
+
 + Setup conda environment to install requirements, and build packages
   ```
   bash prepare_env.sh

closed/Intel/code/resnet50/pytorch-cpu/run_offline.sh

@@ -1,4 +1,8 @@
 #!/bin/bash
+number_threads=`nproc --all`
+number_cores=$((number_threads/2))
+number_sockets=`grep physical.id /proc/cpuinfo | sort -u | wc -l`
+cpu_per_socket=$((number_cores/number_sockets))
 
 export DATA_DIR=${PWD}/ILSVRC2012_img_val
 export RN50_START=${PWD}/models/resnet50-start-int8-model.pth
@@ -30,30 +34,30 @@ if [ -z "${RN50_FULL}" ]; then
     exit 1
 fi
 
-CONDA_ENV_NAME=rn50-mlperf
-source ~/anaconda3/etc/profile.d/conda.sh
-conda activate ${CONDA_ENV_NAME}
+#CONDA_ENV_NAME=rn50-mlperf
+#source ~/anaconda3/etc/profile.d/conda.sh
+#conda activate ${CONDA_ENV_NAME}
 
 export MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:9000000000,muzzy_decay_ms:9000000000"
 
-export LD_PRELOAD=${CONDA_PREFIX}/lib/libjemalloc.so
+#export LD_PRELOAD=${CONDA_PREFIX}/lib/libjemalloc.so
 
-export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libiomp5.so
+#export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libiomp5.so
 
 KMP_SETTING="KMP_AFFINITY=granularity=fine,compact,1,0"
 export KMP_BLOCKTIME=1
 export $KMP_SETTING
 
 CUR_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
-APP=${CUR_DIR}/build/bin/mlperf_runner
+APP=${PWD}/build/bin/mlperf_runner
 
 export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:${CONDA_PREFIX}/lib
 
 if [ -e "mlperf_log_summary.txt" ]; then
     rm mlperf_log_summary.txt
 fi
 
-numactl -C 0-55,56-111 -m 0,1 ${APP} --scenario Offline \
+numactl -m 0 ${APP} --scenario Offline \
 	--mode Performance \
 	--mlperf_conf ${CUR_DIR}/src/mlperf.conf \
 	--user_conf ${CUR_DIR}/src/user.conf \
@@ -62,7 +66,7 @@ numactl -C 0-55,56-111 -m 0,1 ${APP} --scenario Offline \
     --rn50-part3 ${RN50_END} \
     --rn50-full-model ${RN50_FULL} \
 	--data_path ${DATA_DIR} \
-	--num_instance 224 \
+	--num_instance $number_threads \
 	--warmup_iters 20 \
 	--cpus_per_instance 1\
 	--total_sample_count 50000 \

closed/Intel/code/resnet50/pytorch-cpu/run_offline_accuracy.sh

@@ -30,30 +30,30 @@ if [ -z "${RN50_FULL}" ]; then
     exit 1
 fi
 
-CONDA_ENV_NAME=rn50-mlperf
-source ~/anaconda3/etc/profile.d/conda.sh
-conda activate ${CONDA_ENV_NAME}
+#CONDA_ENV_NAME=rn50-mlperf
+#source ~/anaconda3/etc/profile.d/conda.sh
+#conda activate ${CONDA_ENV_NAME}
 
 export MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:9000000000,muzzy_decay_ms:9000000000"
 
-export LD_PRELOAD=${CONDA_PREFIX}/lib/libjemalloc.so
+#export LD_PRELOAD=${CONDA_PREFIX}/lib/libjemalloc.so
 
-export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libiomp5.so
+#export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libiomp5.so
 
 KMP_SETTING="KMP_AFFINITY=granularity=fine,compact,1,0"
 export KMP_BLOCKTIME=1
 export $KMP_SETTING
 
 CUR_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
-APP=${CUR_DIR}/build/bin/mlperf_runner
+APP=${PWD}/build/bin/mlperf_runner
 
 export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:${CONDA_PREFIX}/lib
 
 if [ -e mlperf_log_accuracy.json ]; then
     rm mlperf_log_accuracy.json
 fi
 
-numactl -C 0-55,56-111 -m 0,1 ${APP} --scenario Offline \
+numactl -C 0-47 -m 0 ${APP} --scenario Offline \
 	--mode Accuracy \
 	--mlperf_conf ${CUR_DIR}/src/mlperf.conf \
 	--user_conf ${CUR_DIR}/src/user.conf \

closed/Intel/code/resnet50/pytorch-cpu/run_server.sh

@@ -29,30 +29,33 @@ if [ -z "${RN50_FULL}" ]; then
     exit 1
 fi
 
-CONDA_ENV_NAME=rn50-mlperf
-source ~/anaconda3/etc/profile.d/conda.sh
-conda activate ${CONDA_ENV_NAME}
+# comment out the follwoing lines for AWS
+# CONDA_ENV_NAME=rn50-mlperf
+# source ~/anaconda3/etc/profile.d/conda.sh
+# conda activate ${CONDA_ENV_NAME}
 
 export MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:9000000000,muzzy_decay_ms:9000000000"
 
-export LD_PRELOAD=${CONDA_PREFIX}/lib/libjemalloc.so
-
-export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libiomp5.so
+# comment out the follwoing lines for AWS
+# export LD_PRELOAD=${CONDA_PREFIX}/lib/libjemalloc.so
+# export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libiomp5.so
 
 KMP_SETTING="KMP_AFFINITY=granularity=fine,compact,1,0"
 export KMP_BLOCKTIME=1
 export $KMP_SETTING
 
 CUR_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
-APP=${CUR_DIR}/build/bin/mlperf_runner
+
+# change {CUR_DIR} to {PWD} for AWS
+APP=${PWD}/build/bin/mlperf_runner
 
 export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:${CONDA_PREFIX}/lib
 
 if [ -e "mlperf_log_summary.txt" ]; then
     rm mlperf_log_summary.txt
 fi
 
-numactl -C 0-55,56-111 -m 0,1 ${APP} --scenario Server  \
+numactl -C 0-47 -m 0,1 ${APP} --scenario Server  \
 	--mode Performance  \
 	--mlperf_conf ${CUR_DIR}/src/mlperf.conf \
 	--user_conf ${CUR_DIR}/src/user.conf \