trtexec的一些常用指令

发表于 2024-12-01 分类于 TensorRT 阅读次数：

trtexec —onnx=apa_lidarfreespace_u2nettp_6xdownsample_v9_20240810_scale_simlified_edit.onnx —saveEngine=apa_lidarfreespace_u2nettp_6xdownsample_v9_20240810_scale_simlified_edit_dla.engine —int8 —useDLACore=0 —allowGPUFallback —verbose —dumpLayerInfo —dumpProfile —useSpinWait —separateProfileRun >apa_lidarfreespace_u2nettp_6xdownsample_v9_20240810_scale_simlified_edit_dla.engine.log 2>&1
#转dla控制层的精度，进而控制一层不在dla上运行
trtexec —onnx=apa_lidarfreespace_u2nettp_6xdownsample_v9_20240810_scale_simlified_edit.onnx —saveEngine=apa_lidarfreespace_u2nettp_6xdownsample_v9_20240810_scale_simlified_edit_layerPercision_dla.engine —int8 —useDLACore=0 —allowGPUFallback —verbose —dumpLayerInfo —dumpProfile —useSpinWait —separateProfileRun —precisionConstraints=obey —layerPrecisions=”/Concat”:fp32 >apa_lidarfreespace_u2nettp_6xdownsample_v9_20240810_scale_simlified_edit_layerPercision_dla.engine.log 2>&1
trtexec —onnx=v9_nearest.onnx —saveEngine=v9_nearest_layerPercision_dla.engine —int8 —useDLACore=0 —allowGPUFallback —verbose —dumpLayerInfo —dumpProfile —useSpinWait —separateProfileRun —precisionConstraints=obey —layerPrecisions=”/Concat”:fp32 >v9_nearest_layerPercision_dla.engine.log 2>&1
#构建一个经过qdq_translator后的noqdq onnx 和cache文件生成DLA的engine
trtexec —onnx=v9_nearestqdq_noqdq.onnx —calib=v9_nearestqdq_precision_config_calib.cache —useDLACore=0 —int8 —fp16 —allowGPUFallback —saveEngine=v9_nearestqdq_noqdq.onnx_dlaFP16INT8.engine >v9_nearestqdq_noqdq.onnx_dlaFP16INT8.engine.log 2>&1
trtexec —onnx=v9_nearestqdq_noqdq.onnx —calib=v9_nearestqdq_precision_config_calib.cache —useDLACore=0 —int8 —allowGPUFallback —saveEngine=v9_nearestqdq_noqdq.onnx_dlaINT8.engine >v9_nearestqdq_noqdq.onnx_dlaINT8.engine.log 2>&1
trtexec —onnx=v9_nearestqdq_noqdq.onnx —calib=v9_nearestqdq_precision_config_calib.cache —useDLACore=0 —int8 —fp16 —allowGPUFallback —saveEngine=v9_nearestqdq_noqdq.onnx_dlaFP16INT8_layerPercision.engine —precisionConstraints=obey —layerPrecisions=”/Concat”:fp32 >v9_nearestqdq_noqdq.onnx_dlaFP16INT8_layerPercision.engine.log 2>&1
trtexec —onnx=v9_nearestqdq_noqdq.onnx —calib=v9_nearestqdq_precision_config_calib.cache —useDLACore=0 —fp16 —allowGPUFallback —saveEngine=v9_nearestqdq_noqdq.onnx_dlaINT8_layerPercision.engine —precisionConstraints=obey —layerPrecisions=”/Concat”:fp32 >v9_nearestqdq_noqdq.onnx_dlaINT8_layerPercision.log 2>&1

#精度对比
#Lidar trtexec 给定输入
trtexec —loadEngine=v9_nearest_layerPercision_dla.engine —loadInputs=’input’:lidar_preBin.bin —exportOutput=Lidar_output_dla.json —dumpOutput
CheckDLAOutputScripts$ python3 check_outputs_diff.py Lidar_output_FP32.json Lidar_output_dla.json 0
CheckDLAOutputScripts$ python3 check_cosine_sim.py Lidar_output_FP32.json Lidar_output_dla.json 0

#nsys 分析 dla
/algdata/zyd/nsight_systems/2022.3.2/target-linux-tegra-armv8/nsys profile —accelerator-trace=nvmedia —trace=cuda,nvtx,cublas,cudla,cusparse,cudnn,nvmedia —force-overwrite true -o /algdata/zyd/DLATest/Lidar/apa_lidarfreespace_u2nettp_6xdownsample_v9_20240810_scale_simlified_edit_layerPercision_dla.engine trtexec —loadEngine=/algdata/zyd/DLATest/Lidar/apa_lidarfreespace_u2nettp_6xdownsample_v9_20240810_scale_simlified_edit_layerPercision_dla.engine —iterations=10 —idleTime=50 —duration=0 —useSpinWait

https://github.com/NVIDIA-AI-IOT/cuDLA-samples/blob/main/export/README.md