p47068915
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[MSLITE] tensorrt distribution test

This commit is contained in:
ling 2021-11-25 15:39:30 +08:00
parent 1a7ac7bcdc
commit ce33a63a7d
2 changed files with 40 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
mindspore/lite/test/config/models_tensorrt.cfg
View File

@ -1,3 +1,9 @@
gender_resnet34_lzl.onnx;1:input.1
ml_video_edit_person_divison_pic;1:blob1
ml_video_edit_person_divison_video;2:blob1,blob2
# model_info accuracy_limit run_mode
# model_name;input_info;input_shapes;mode accuracy_limit (DIS/ );RankSize
# Run in one cuda server
gender_resnet34_lzl.onnx
ml_video_edit_person_divison_pic
ml_video_edit_person_divison_video;2
# Run in distribution server
Gpu_Distribution_Matmul.mindir 0.01 DIS;2

37
mindspore/lite/test/st/scripts/tensorrt/run_benchmark_tensorrt.sh
View File

@ -13,9 +13,13 @@ function Run_TensorRT() {
if [[ $line_info == \#* || $line_info == "" ]]; then
continue
fi
#model_name.onnx;2:input1,input2;1,16,16,4;fp16 0.5
# model_info accuracy_limit run_mode
model_info=`echo ${line_info} | awk -F ' ' '{print $1}'`
spec_acc_limit=`echo ${line_info} | awk -F ' ' '{print $2}'`
run_mode=`echo ${line_info} | awk -F ' ' '{print $3}'`
# model_info detail
model_name=`echo ${model_info} | awk -F ';' '{print $1}'`
input_info=`echo ${model_info} | awk -F ';' '{print $2}'`
input_shapes=`echo ${model_info} | awk -F ';' '{print $3}'`
@ -24,6 +28,11 @@ function Run_TensorRT() {
if [[ ${model_name##*.} == "caffemodel" ]]; then
model_name=${model_name%.*}
fi
# run_mode detail
run_mode_name=`echo ${run_mode} | awk -F ';' '{print $1}'`
run_mode_size=`echo ${run_mode} | awk -F ';' '{print $2}'`
echo "Benchmarking ${model_name} ......"
model_file=${basepath}'/'${model_name}'.ms'
input_files=""
@ -51,13 +60,23 @@ function Run_TensorRT() {
enableFp16="true"
fi
echo 'CUDA_VISILE_DEVICE='${cuda_device_id}' ./benchmark --modelFile='${model_file}' --inputShapes='${input_shapes}' --inDataFile='${input_files}' --benchmarkDataFile='${output_file}' --enableFp16='${enableFp16}' --accuracyThreshold='${acc_limit}' --device=GPU' >> "${run_tensorrt_log_file}"
CUDA_VISILE_DEVICE=${cuda_device_id} ./benchmark --modelFile=${model_file} --inputShapes=${input_shapes} --inDataFile=${input_files} --benchmarkDataFile=${output_file} --enableFp16=${enableFp16} --accuracyThreshold=${acc_limit} --device=GPU >> ${run_tensorrt_log_file}
if [ $? = 0 ]; then
run_result='TensorRT: '${model_name}' pass'; echo ${run_result} >> ${run_benchmark_result_file}
# different tensorrt run mode use different cuda command
if [[ ${run_mode_name} == "DIS" ]]; then
export ENABLE_NEW_API=true
echo 'mpirun -np '${run_mode_size}' ./benchmark --modelFile='${model_file}' --inputShapes='${input_shapes}' --inDataFile='${input_files}' --benchmarkDataFile='${output_file}' --enableFp16='${enableFp16}' --accuracyThreshold='${acc_limit}' --device=GPU' >> "${run_tensorrt_log_file}"
mpirun -np ${run_mode_size} ./benchmark --modelFile=${model_file} --inputShapes=${input_shapes} --inDataFile=${input_files} --benchmarkDataFile=${output_file} --enableFp16=${enableFp16} --accuracyThreshold=${acc_limit} --device=GPU >> ${run_tensorrt_log_file}
else
run_result='TensorRT: '${model_name}' failed'; echo ${run_result} >> ${run_benchmark_result_file}; return 1
echo 'CUDA_VISILE_DEVICE='${cuda_device_id}' ./benchmark --modelFile='${model_file}' --inputShapes='${input_shapes}' --inDataFile='${input_files}' --benchmarkDataFile='${output_file}' --enableFp16='${enableFp16}' --accuracyThreshold='${acc_limit}' --device=GPU' >> "${run_tensorrt_log_file}"
CUDA_VISILE_DEVICE=${cuda_device_id} ./benchmark --modelFile=${model_file} --inputShapes=${input_shapes} --inDataFile=${input_files} --benchmarkDataFile=${output_file} --enableFp16=${enableFp16} --accuracyThreshold=${acc_limit} --device=GPU >> ${run_tensorrt_log_file}
fi
if [ $? = 0 ]; then
run_result='TensorRT'${run_mode_name}': '${model_name}' pass'; echo ${run_result} >> ${run_benchmark_result_file}
else
run_result='TensorRT'${run_mode_name}': '${model_name}' failed'; echo ${run_result} >> ${run_benchmark_result_file}; return 1
fi
done < ${models_tensorrt_config}
}
@ -120,5 +139,11 @@ Run_benchmark_status=$?
# Check converter result and return value
echo "Run x86 TensorRT GPU ended on device ${cuda_device_id}"
if [[ ${Run_benchmark_status} != 0 ]];then
echo "Run_TensorRT_PID failed"
cat ${run_tensorrt_log_file}
fi
Print_Benchmark_Result $run_benchmark_result_file
exit ${Run_benchmark_status}