search

NVIDIA / TensorRT

NVIDIA® TensorRT™ is an SDK for high-performance deep learning inference on NVIDIA GPUs. This repository contains the open source components of TensorRT.
https://developer.nvidia.com/tensorrt
Apache License 2.0
10.63k stars 2.12k forks source link

TensorRT trtexec Layerwise Compute Precision, Sparsity, Tensor DataType and Layout Information #4020

Closed leimao closed 2 months ago

leimao commented 2 months ago

Description

Previously, I remember I can use --exportLayerInfo to dump the comprehensive layerwise info of the engine, including the precision of the layer, and the IO tensor datatype and layouts.

However, for the trtexec from the most recent releases, it seems that these useful information is gone. If this behavior is intended, how do I save the detailed layerwise info, including sparsity, using trtexec?

Environment

NGC Docker container nvcr.io/nvidia/tensorrt:24.06-py3.

Steps To Reproduce

$ docker run -it --rm --gpus all -v $(pwd):/mnt nvcr.io/nvidia/tensorrt:24.06-py3
$ cd /mnt
$ wget https://github.com/onnx/models/raw/main/validated/vision/classification/resnet/model/resnet18-v1-7.onnx
$ trtexec --onnx=resnet18-v1-7.onnx --fp16 --int8 --saveEngine=resnet18-v1-7.engine --exportProfile=resnet18-v1-7.txt --exportLayerInfo=resnet18-v1-7.json
$ cat resnet18-v1-7.json 
{"Layers": ["Reformatting CopyNode for Input Tensor 0 to resnetv15_conv0_fwd + resnetv15_batchnorm0_fwd + resnetv15_relu0_fwd + resnetv15_pool0_fwd"
,"resnetv15_conv0_fwd + resnetv15_batchnorm0_fwd + resnetv15_relu0_fwd + resnetv15_pool0_fwd"
,"resnetv15_stage1_conv0_fwd + resnetv15_stage1_batchnorm0_fwd + resnetv15_stage1_relu0_fwd"
,"resnetv15_stage1_conv1_fwd + resnetv15_stage1_batchnorm1_fwd + resnetv15_stage1__plus0 + resnetv15_stage1_activation0"
,"resnetv15_stage1_conv2_fwd + resnetv15_stage1_batchnorm2_fwd + resnetv15_stage1_relu1_fwd"
,"resnetv15_stage1_conv3_fwd + resnetv15_stage1_batchnorm3_fwd + resnetv15_stage1__plus1 + resnetv15_stage1_activation1"
,"resnetv15_stage2_conv0_fwd + resnetv15_stage2_batchnorm0_fwd + resnetv15_stage2_relu0_fwd"
,"resnetv15_stage2_conv1_fwd + resnetv15_stage2_batchnorm1_fwd"
,"resnetv15_stage2_conv2_fwd + resnetv15_stage2_batchnorm2_fwd + resnetv15_stage2__plus0 + resnetv15_stage2_activation0"
,"resnetv15_stage2_conv3_fwd + resnetv15_stage2_batchnorm3_fwd + resnetv15_stage2_relu1_fwd"
,"resnetv15_stage2_conv4_fwd + resnetv15_stage2_batchnorm4_fwd + resnetv15_stage2__plus1 + resnetv15_stage2_activation1"
,"resnetv15_stage3_conv0_fwd + resnetv15_stage3_batchnorm0_fwd + resnetv15_stage3_relu0_fwd"
,"resnetv15_stage3_conv1_fwd + resnetv15_stage3_batchnorm1_fwd"
,"resnetv15_stage3_conv2_fwd + resnetv15_stage3_batchnorm2_fwd + resnetv15_stage3__plus0 + resnetv15_stage3_activation0"
,"resnetv15_stage3_conv3_fwd + resnetv15_stage3_batchnorm3_fwd + resnetv15_stage3_relu1_fwd"
,"resnetv15_stage3_conv4_fwd + resnetv15_stage3_batchnorm4_fwd + resnetv15_stage3__plus1 + resnetv15_stage3_activation1"
,"resnetv15_stage4_conv0_fwd + resnetv15_stage4_batchnorm0_fwd + resnetv15_stage4_relu0_fwd"
,"resnetv15_stage4_conv1_fwd + resnetv15_stage4_batchnorm1_fwd"
,"resnetv15_stage4_conv2_fwd + resnetv15_stage4_batchnorm2_fwd + resnetv15_stage4__plus0 + resnetv15_stage4_activation0"
,"resnetv15_stage4_conv3_fwd + resnetv15_stage4_batchnorm3_fwd + resnetv15_stage4_relu1_fwd"
,"resnetv15_stage4_conv4_fwd + resnetv15_stage4_batchnorm4_fwd + resnetv15_stage4__plus1 + resnetv15_stage4_activation1"
,"resnetv15_pool1_fwd"
,"Reformatting CopyNode for Input Tensor 0 to resnetv15_dense0_fwd + (Unnamed Layer* 78) [ElementWise]"
,"resnetv15_dense0_fwd + (Unnamed Layer* 78) [ElementWise]"
,"reshape_after_resnetv15_dense0_fwd"
],
"Bindings": ["data"
,"resnetv15_dense0_fwd"
]}
lix19937 commented 2 months ago

Add --profilingVerbosity=detailed in trtexec cmd.

leimao commented 2 months ago

Seems to be working. Thank you.

$ trtexec --onnx=resnet18-v1-7.onnx --fp16 --int8 --sparsity=force --saveEngine=resnet18-v1-7.engine --exportProfile=resnet18-v1-7.txt --exportLayerInfo=resnet18-v1-7.json --profilingVerbosity=detailed
$ cat resnet18-v1-7.json
{"Layers": [{
  "Name": "Reformatting CopyNode for Input Tensor 0 to resnetv15_conv0_fwd + resnetv15_batchnorm0_fwd + resnetv15_relu0_fwd + resnetv15_pool0_fwd",
  "LayerType": "Reformat",
  "Inputs": [
  {
    "Name": "data",
    "Location": "Device",
    "Dimensions": [1,3,224,224],
    "Format/Datatype": "Row major linear FP32"
  }],
  "Outputs": [
  {
    "Name": "Reformatted Input Tensor 0 to resnetv15_conv0_fwd + resnetv15_batchnorm0_fwd + resnetv15_relu0_fwd + resnetv15_pool0_fwd",
    "Location": "Device",
    "Dimensions": [1,3,224,224],
    "Format/Datatype": "Row major Int8 format"
  }],
  "ParameterType": "Reformat",
  "Origin": "REFORMAT",
  "TacticValue": "0x00000000000003e8",
  "StreamId": 0,
  "Metadata": ""
},{
  "Name": "resnetv15_conv0_fwd + resnetv15_batchnorm0_fwd + resnetv15_relu0_fwd + resnetv15_pool0_fwd",
  "LayerType": "CaskConvActPool",
  "Inputs": [
  {
    "Name": "Reformatted Input Tensor 0 to resnetv15_conv0_fwd + resnetv15_batchnorm0_fwd + resnetv15_relu0_fwd + resnetv15_pool0_fwd",
    "Location": "Device",
    "Dimensions": [1,3,224,224],
    "Format/Datatype": "Row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_pool0_fwd",
    "Location": "Device",
    "Dimensions": [1,64,56,56],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "ConvActPool",
  "ConvParameterType": "Convolution",
  "ConvKernel": [7,7],
  "ConvPaddingMode": "kEXPLICIT_ROUND_DOWN",
  "ConvPrePadding": [3,3],
  "ConvPostPadding": [3,3],
  "ConvStride": [2,2],
  "ConvDilation": [1,1],
  "ConvOutMaps": 64,
  "ConvGroups": 1,
  "ConvWeights": {"Type": "Int8", "Count": 9408},
  "ConvBias": {"Type": "Float", "Count": 64},
  "ConvHasBias": 1,
  "ConvHasReLU": 1,
  "ConvHasSparseWeights": 0,
  "ConvHasDynamicFilter": 0,
  "ConvHasDynamicBias": 0,
  "ConvHasResidual": 0,
  "ConvConvXAsActInputIdx": -1,
  "ConvBiasAsActInputIdx": -1,
  "ConvResAsActInputIdx": -1,
  "ConvActivation": "RELU",
  "PoolingParameterType": "Pooling",
  "PoolingPoolingType": "MAX",
  "PoolingWindowSize": [3,3],
  "PoolingPaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PoolingPrePadding": [1,1],
  "PoolingPostPadding": [1,1],
  "PoolingStride": [2,2],
  "PoolingBlendFactor": 0,
  "PoolingAverageCountExcludesPadding": 1,
  "TacticValue": "0xb71c75095873646c",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_conv0_fwd]\u001e[ONNX Layer: resnetv15_batchnorm0_fwd]\u001e[ONNX Layer: resnetv15_relu0_fwd]\u001e[ONNX Layer: resnetv15_pool0_fwd]"
},{
  "Name": "resnetv15_stage1_conv0_fwd + resnetv15_stage1_batchnorm0_fwd + resnetv15_stage1_relu0_fwd",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_pool0_fwd",
    "Location": "Device",
    "Dimensions": [1,64,56,56],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage1_relu0_fwd",
    "Location": "Device",
    "Dimensions": [1,64,56,56],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 64,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 36864},
  "Bias": {"Type": "Float", "Count": 64},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 0,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_avdt_sparse_int8int8_tilesize64x64x64_tapsperload3_threadspercta128_r3s3_u1v1_scalebias_relu",
  "TacticValue": "0x9251329244d2a80b",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage1_conv0_fwd]\u001e[ONNX Layer: resnetv15_stage1_batchnorm0_fwd]\u001e[ONNX Layer: resnetv15_stage1_relu0_fwd]"
},{
  "Name": "resnetv15_stage1_conv1_fwd + resnetv15_stage1_batchnorm1_fwd + resnetv15_stage1__plus0 + resnetv15_stage1_activation0",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage1_relu0_fwd",
    "Location": "Device",
    "Dimensions": [1,64,56,56],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  },
  {
    "Name": "resnetv15_pool0_fwd",
    "Location": "Device",
    "Dimensions": [1,64,56,56],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage1_activation0",
    "Location": "Device",
    "Dimensions": [1,64,56,56],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 64,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 36864},
  "Bias": {"Type": "Float", "Count": 64},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 1,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_avdt_sparse_int8int8_tilesize64x128x32_tapsperload3_threadspercta128_r3s3_u1v1_scalebias_relu",
  "TacticValue": "0x4c3eb07218824fb7",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage1_conv1_fwd]\u001e[ONNX Layer: resnetv15_stage1_batchnorm1_fwd]\u001e[ONNX Layer: resnetv15_stage1__plus0]\u001e[ONNX Layer: resnetv15_stage1_activation0]"
},{
  "Name": "resnetv15_stage1_conv2_fwd + resnetv15_stage1_batchnorm2_fwd + resnetv15_stage1_relu1_fwd",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage1_activation0",
    "Location": "Device",
    "Dimensions": [1,64,56,56],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage1_relu1_fwd",
    "Location": "Device",
    "Dimensions": [1,64,56,56],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 64,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 36864},
  "Bias": {"Type": "Float", "Count": 64},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 0,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_avdt_sparse_int8int8_tilesize64x64x64_tapsperload3_threadspercta128_r3s3_u1v1_scalebias_relu",
  "TacticValue": "0x9251329244d2a80b",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage1_conv2_fwd]\u001e[ONNX Layer: resnetv15_stage1_batchnorm2_fwd]\u001e[ONNX Layer: resnetv15_stage1_relu1_fwd]"
},{
  "Name": "resnetv15_stage1_conv3_fwd + resnetv15_stage1_batchnorm3_fwd + resnetv15_stage1__plus1 + resnetv15_stage1_activation1",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage1_relu1_fwd",
    "Location": "Device",
    "Dimensions": [1,64,56,56],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  },
  {
    "Name": "resnetv15_stage1_activation0",
    "Location": "Device",
    "Dimensions": [1,64,56,56],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage1_activation1",
    "Location": "Device",
    "Dimensions": [1,64,56,56],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 64,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 36864},
  "Bias": {"Type": "Float", "Count": 64},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 1,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_avdt_sparse_int8int8_tilesize64x128x32_tapsperload3_threadspercta128_r3s3_u1v1_scalebias_relu",
  "TacticValue": "0x4c3eb07218824fb7",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage1_conv3_fwd]\u001e[ONNX Layer: resnetv15_stage1_batchnorm3_fwd]\u001e[ONNX Layer: resnetv15_stage1__plus1]\u001e[ONNX Layer: resnetv15_stage1_activation1]"
},{
  "Name": "resnetv15_stage2_conv0_fwd + resnetv15_stage2_batchnorm0_fwd + resnetv15_stage2_relu0_fwd",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage1_activation1",
    "Location": "Device",
    "Dimensions": [1,64,56,56],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage2_relu0_fwd",
    "Location": "Device",
    "Dimensions": [1,128,28,28],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [2,2],
  "Dilation": [1,1],
  "OutMaps": 128,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 73728},
  "Bias": {"Type": "Float", "Count": 128},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 0,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_avdt_sparse_int8int8_tilesize64x128x32_tapsperload3_threadspercta128_r3s3_u2v2_scalebias_relu",
  "TacticValue": "0xc0bbd49d4144045f",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage2_conv0_fwd]\u001e[ONNX Layer: resnetv15_stage2_batchnorm0_fwd]\u001e[ONNX Layer: resnetv15_stage2_relu0_fwd]"
},{
  "Name": "resnetv15_stage2_conv1_fwd + resnetv15_stage2_batchnorm1_fwd",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage2_relu0_fwd",
    "Location": "Device",
    "Dimensions": [1,128,28,28],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage2_batchnorm1_fwd",
    "Location": "Device",
    "Dimensions": [1,128,28,28],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 128,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 147456},
  "Bias": {"Type": "Float", "Count": 128},
  "HasBias": 1,
  "HasReLU": 0,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 0,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "NONE",
  "TacticName": "sm80_xmma_fprop_implicit_gemm_interleaved_i8i8_i8i32_f32_nchw_vect_c_32kcrs_vect_c_32_nchw_vect_c_32_tilesize32x64x64_stage6_warpsize2x2x1_g1_tensor16x8x32_t1r3s3",
  "TacticValue": "0xd3f592fae61c7986",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage2_conv1_fwd]\u001e[ONNX Layer: resnetv15_stage2_batchnorm1_fwd]"
},{
  "Name": "resnetv15_stage2_conv2_fwd + resnetv15_stage2_batchnorm2_fwd + resnetv15_stage2__plus0 + resnetv15_stage2_activation0",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage1_activation1",
    "Location": "Device",
    "Dimensions": [1,64,56,56],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  },
  {
    "Name": "resnetv15_stage2_batchnorm1_fwd",
    "Location": "Device",
    "Dimensions": [1,128,28,28],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage2_activation0",
    "Location": "Device",
    "Dimensions": [1,128,28,28],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [1,1],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [0,0],
  "PostPadding": [0,0],
  "Stride": [2,2],
  "Dilation": [1,1],
  "OutMaps": 128,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 8192},
  "Bias": {"Type": "Float", "Count": 128},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 1,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_implicit_gemm_interleaved_i8i8_i8i32_f32_nchw_vect_c_32kcrs_vect_c_32_nchw_vect_c_32_tilesize32x64x64_stage6_warpsize2x2x1_g1_tensor16x8x32_t1r1s1",
  "TacticValue": "0x596666386c88024b",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage2_conv2_fwd]\u001e[ONNX Layer: resnetv15_stage2_batchnorm2_fwd]\u001e[ONNX Layer: resnetv15_stage2__plus0]\u001e[ONNX Layer: resnetv15_stage2_activation0]"
},{
  "Name": "resnetv15_stage2_conv3_fwd + resnetv15_stage2_batchnorm3_fwd + resnetv15_stage2_relu1_fwd",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage2_activation0",
    "Location": "Device",
    "Dimensions": [1,128,28,28],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage2_relu1_fwd",
    "Location": "Device",
    "Dimensions": [1,128,28,28],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 128,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 147456},
  "Bias": {"Type": "Float", "Count": 128},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 0,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_avdt_sparse_int8int8_tilesize64x64x64_tapsperload3_threadspercta128_r3s3_u1v1_scalebias_relu",
  "TacticValue": "0x9251329244d2a80b",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage2_conv3_fwd]\u001e[ONNX Layer: resnetv15_stage2_batchnorm3_fwd]\u001e[ONNX Layer: resnetv15_stage2_relu1_fwd]"
},{
  "Name": "resnetv15_stage2_conv4_fwd + resnetv15_stage2_batchnorm4_fwd + resnetv15_stage2__plus1 + resnetv15_stage2_activation1",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage2_relu1_fwd",
    "Location": "Device",
    "Dimensions": [1,128,28,28],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  },
  {
    "Name": "resnetv15_stage2_activation0",
    "Location": "Device",
    "Dimensions": [1,128,28,28],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage2_activation1",
    "Location": "Device",
    "Dimensions": [1,128,28,28],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 128,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 147456},
  "Bias": {"Type": "Float", "Count": 128},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 1,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_avdt_sparse_int8int8_tilesize64x64x64_tapsperload3_threadspercta128_r3s3_u1v1_scalebias_relu",
  "TacticValue": "0x9251329244d2a80b",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage2_conv4_fwd]\u001e[ONNX Layer: resnetv15_stage2_batchnorm4_fwd]\u001e[ONNX Layer: resnetv15_stage2__plus1]\u001e[ONNX Layer: resnetv15_stage2_activation1]"
},{
  "Name": "resnetv15_stage3_conv0_fwd + resnetv15_stage3_batchnorm0_fwd + resnetv15_stage3_relu0_fwd",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage2_activation1",
    "Location": "Device",
    "Dimensions": [1,128,28,28],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage3_relu0_fwd",
    "Location": "Device",
    "Dimensions": [1,256,14,14],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [2,2],
  "Dilation": [1,1],
  "OutMaps": 256,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 294912},
  "Bias": {"Type": "Float", "Count": 256},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 0,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_implicit_gemm_interleaved_i8i8_i8i32_f32_nchw_vect_c_32kcrs_vect_c_32_nchw_vect_c_32_tilesize32x32x64_stage6_warpsize2x1x1_g1_tensor16x8x32_t1r3s3",
  "TacticValue": "0x4749124f62d8bd23",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage3_conv0_fwd]\u001e[ONNX Layer: resnetv15_stage3_batchnorm0_fwd]\u001e[ONNX Layer: resnetv15_stage3_relu0_fwd]"
},{
  "Name": "resnetv15_stage3_conv1_fwd + resnetv15_stage3_batchnorm1_fwd",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage3_relu0_fwd",
    "Location": "Device",
    "Dimensions": [1,256,14,14],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage3_batchnorm1_fwd",
    "Location": "Device",
    "Dimensions": [1,256,14,14],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 256,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 589824},
  "Bias": {"Type": "Float", "Count": 256},
  "HasBias": 1,
  "HasReLU": 0,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 0,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "NONE",
  "TacticName": "sm80_xmma_fprop_implicit_gemm_interleaved_i8i8_i8i32_f32_nchw_vect_c_32kcrs_vect_c_32_nchw_vect_c_32_tilesize32x32x64_stage6_warpsize2x1x1_g1_tensor16x8x32_t1r3s3",
  "TacticValue": "0x4749124f62d8bd23",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage3_conv1_fwd]\u001e[ONNX Layer: resnetv15_stage3_batchnorm1_fwd]"
},{
  "Name": "resnetv15_stage3_conv2_fwd + resnetv15_stage3_batchnorm2_fwd + resnetv15_stage3__plus0 + resnetv15_stage3_activation0",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage2_activation1",
    "Location": "Device",
    "Dimensions": [1,128,28,28],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  },
  {
    "Name": "resnetv15_stage3_batchnorm1_fwd",
    "Location": "Device",
    "Dimensions": [1,256,14,14],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage3_activation0",
    "Location": "Device",
    "Dimensions": [1,256,14,14],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [1,1],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [0,0],
  "PostPadding": [0,0],
  "Stride": [2,2],
  "Dilation": [1,1],
  "OutMaps": 256,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 32768},
  "Bias": {"Type": "Float", "Count": 256},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 1,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_implicit_gemm_interleaved_i8i8_i8i32_f32_nchw_vect_c_32kcrs_vect_c_32_nchw_vect_c_32_tilesize32x32x64_stage6_warpsize2x1x1_g1_tensor16x8x32_t1r1s1",
  "TacticValue": "0xcddae68de84cc6ee",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage3_conv2_fwd]\u001e[ONNX Layer: resnetv15_stage3_batchnorm2_fwd]\u001e[ONNX Layer: resnetv15_stage3__plus0]\u001e[ONNX Layer: resnetv15_stage3_activation0]"
},{
  "Name": "resnetv15_stage3_conv3_fwd + resnetv15_stage3_batchnorm3_fwd + resnetv15_stage3_relu1_fwd",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage3_activation0",
    "Location": "Device",
    "Dimensions": [1,256,14,14],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage3_relu1_fwd",
    "Location": "Device",
    "Dimensions": [1,256,14,14],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 256,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 589824},
  "Bias": {"Type": "Float", "Count": 256},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 0,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_avdt_sparse_int8int8_tilesize64x64x64_tapsperload3_threadspercta128_r3s3_u1v1_scalebias_relu",
  "TacticValue": "0x9251329244d2a80b",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage3_conv3_fwd]\u001e[ONNX Layer: resnetv15_stage3_batchnorm3_fwd]\u001e[ONNX Layer: resnetv15_stage3_relu1_fwd]"
},{
  "Name": "resnetv15_stage3_conv4_fwd + resnetv15_stage3_batchnorm4_fwd + resnetv15_stage3__plus1 + resnetv15_stage3_activation1",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage3_relu1_fwd",
    "Location": "Device",
    "Dimensions": [1,256,14,14],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  },
  {
    "Name": "resnetv15_stage3_activation0",
    "Location": "Device",
    "Dimensions": [1,256,14,14],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage3_activation1",
    "Location": "Device",
    "Dimensions": [1,256,14,14],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 256,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 589824},
  "Bias": {"Type": "Float", "Count": 256},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 1,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_avdt_sparse_int8int8_tilesize64x64x64_tapsperload3_threadspercta128_r3s3_u1v1_scalebias_relu",
  "TacticValue": "0x9251329244d2a80b",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage3_conv4_fwd]\u001e[ONNX Layer: resnetv15_stage3_batchnorm4_fwd]\u001e[ONNX Layer: resnetv15_stage3__plus1]\u001e[ONNX Layer: resnetv15_stage3_activation1]"
},{
  "Name": "resnetv15_stage4_conv0_fwd + resnetv15_stage4_batchnorm0_fwd + resnetv15_stage4_relu0_fwd",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage3_activation1",
    "Location": "Device",
    "Dimensions": [1,256,14,14],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage4_relu0_fwd",
    "Location": "Device",
    "Dimensions": [1,512,7,7],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [2,2],
  "Dilation": [1,1],
  "OutMaps": 512,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 1179648},
  "Bias": {"Type": "Float", "Count": 512},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 0,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_implicit_gemm_interleaved_i8i8_i8i32_f32_nchw_vect_c_32kcrs_vect_c_32_nchw_vect_c_32_tilesize32x32x64_stage6_warpsize2x1x1_g1_tensor16x8x32_t1r3s3",
  "TacticValue": "0x4749124f62d8bd23",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage4_conv0_fwd]\u001e[ONNX Layer: resnetv15_stage4_batchnorm0_fwd]\u001e[ONNX Layer: resnetv15_stage4_relu0_fwd]"
},{
  "Name": "resnetv15_stage4_conv1_fwd + resnetv15_stage4_batchnorm1_fwd",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage4_relu0_fwd",
    "Location": "Device",
    "Dimensions": [1,512,7,7],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage4_batchnorm1_fwd",
    "Location": "Device",
    "Dimensions": [1,512,7,7],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 512,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 2359296},
  "Bias": {"Type": "Float", "Count": 512},
  "HasBias": 1,
  "HasReLU": 0,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 0,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "NONE",
  "TacticName": "sm80_xmma_fprop_implicit_gemm_interleaved_i8i8_i8i32_f32_nchw_vect_c_32kcrs_vect_c_32_nchw_vect_c_32_tilesize32x32x64_stage6_warpsize2x1x1_g1_tensor16x8x32_t1r3s3",
  "TacticValue": "0x4749124f62d8bd23",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage4_conv1_fwd]\u001e[ONNX Layer: resnetv15_stage4_batchnorm1_fwd]"
},{
  "Name": "resnetv15_stage4_conv2_fwd + resnetv15_stage4_batchnorm2_fwd + resnetv15_stage4__plus0 + resnetv15_stage4_activation0",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage3_activation1",
    "Location": "Device",
    "Dimensions": [1,256,14,14],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  },
  {
    "Name": "resnetv15_stage4_batchnorm1_fwd",
    "Location": "Device",
    "Dimensions": [1,512,7,7],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage4_activation0",
    "Location": "Device",
    "Dimensions": [1,512,7,7],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [1,1],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [0,0],
  "PostPadding": [0,0],
  "Stride": [2,2],
  "Dilation": [1,1],
  "OutMaps": 512,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 131072},
  "Bias": {"Type": "Float", "Count": 512},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 1,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_implicit_gemm_interleaved_i8i8_i8i32_f32_nchw_vect_c_32kcrs_vect_c_32_nchw_vect_c_32_tilesize32x32x64_stage6_warpsize2x1x1_g1_tensor16x8x32_t1r1s1",
  "TacticValue": "0xcddae68de84cc6ee",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage4_conv2_fwd]\u001e[ONNX Layer: resnetv15_stage4_batchnorm2_fwd]\u001e[ONNX Layer: resnetv15_stage4__plus0]\u001e[ONNX Layer: resnetv15_stage4_activation0]"
},{
  "Name": "resnetv15_stage4_conv3_fwd + resnetv15_stage4_batchnorm3_fwd + resnetv15_stage4_relu1_fwd",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage4_activation0",
    "Location": "Device",
    "Dimensions": [1,512,7,7],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage4_relu1_fwd",
    "Location": "Device",
    "Dimensions": [1,512,7,7],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 512,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 2359296},
  "Bias": {"Type": "Float", "Count": 512},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 0,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_avdt_sparse_int8int8_tilesize64x64x64_tapsperload3_threadspercta128_r3s3_u1v1_scalebias_relu",
  "TacticValue": "0x9251329244d2a80b",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage4_conv3_fwd]\u001e[ONNX Layer: resnetv15_stage4_batchnorm3_fwd]\u001e[ONNX Layer: resnetv15_stage4_relu1_fwd]"
},{
  "Name": "resnetv15_stage4_conv4_fwd + resnetv15_stage4_batchnorm4_fwd + resnetv15_stage4__plus1 + resnetv15_stage4_activation1",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "resnetv15_stage4_relu1_fwd",
    "Location": "Device",
    "Dimensions": [1,512,7,7],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  },
  {
    "Name": "resnetv15_stage4_activation0",
    "Location": "Device",
    "Dimensions": [1,512,7,7],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_stage4_activation1",
    "Location": "Device",
    "Dimensions": [1,512,7,7],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Convolution",
  "Kernel": [3,3],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [1,1],
  "PostPadding": [1,1],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 512,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 2359296},
  "Bias": {"Type": "Float", "Count": 512},
  "HasBias": 1,
  "HasReLU": 1,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 1,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "RELU",
  "TacticName": "sm80_xmma_fprop_avdt_sparse_int8int8_tilesize64x64x64_tapsperload3_threadspercta128_r3s3_u1v1_scalebias_relu",
  "TacticValue": "0x9251329244d2a80b",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_stage4_conv4_fwd]\u001e[ONNX Layer: resnetv15_stage4_batchnorm4_fwd]\u001e[ONNX Layer: resnetv15_stage4__plus1]\u001e[ONNX Layer: resnetv15_stage4_activation1]"
},{
  "Name": "resnetv15_pool1_fwd",
  "LayerType": "CaskPooling",
  "Inputs": [
  {
    "Name": "resnetv15_stage4_activation1",
    "Location": "Device",
    "Dimensions": [1,512,7,7],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_pool1_fwd",
    "Location": "Device",
    "Dimensions": [1,512,1,1],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "ParameterType": "Pooling",
  "PoolingType": "AVERAGE",
  "WindowSize": [7,7],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [0,0],
  "PostPadding": [0,0],
  "Stride": [1,1],
  "BlendFactor": 0,
  "AverageCountExcludesPadding": 1,
  "TacticName": "sm72_xmma_pooling_IMMA_NCxHW32_gap",
  "TacticValue": "0xa3a1a62d21de759d",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_pool1_fwd]"
},{
  "Name": "Reformatting CopyNode for Input Tensor 0 to resnetv15_dense0_fwd + (Unnamed Layer* 78) [ElementWise]",
  "LayerType": "NoOp",
  "Inputs": [
  {
    "Name": "resnetv15_pool1_fwd",
    "Location": "Device",
    "Dimensions": [1,512,1,1],
    "Format/Datatype": "Thirty-two wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "Reformatted Input Tensor 0 to resnetv15_dense0_fwd + (Unnamed Layer* 78) [ElementWise]",
    "Location": "Device",
    "Dimensions": [1,512,1,1],
    "Format/Datatype": "Four wide channel vectorized row major Int8 format"
  }],
  "TacticValue": "0x0000000000000000",
  "StreamId": 0,
  "Metadata": ""
},{
  "Name": "resnetv15_dense0_fwd + (Unnamed Layer* 78) [ElementWise]",
  "LayerType": "CaskConvolution",
  "Inputs": [
  {
    "Name": "Reformatted Input Tensor 0 to resnetv15_dense0_fwd + (Unnamed Layer* 78) [ElementWise]",
    "Location": "Device",
    "Dimensions": [1,512,1,1],
    "Format/Datatype": "Four wide channel vectorized row major Int8 format"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_dense0_fwd_conv_out",
    "Location": "Device",
    "Dimensions": [1,1000,1,1],
    "Format/Datatype": "Row major linear FP32"
  }],
  "ParameterType": "Convolution",
  "Kernel": [1,1],
  "PaddingMode": "kEXPLICIT_ROUND_DOWN",
  "PrePadding": [0,0],
  "PostPadding": [0,0],
  "Stride": [1,1],
  "Dilation": [1,1],
  "OutMaps": 1000,
  "Groups": 1,
  "Weights": {"Type": "Int8", "Count": 512000},
  "Bias": {"Type": "Float", "Count": 1000},
  "HasBias": 1,
  "HasReLU": 0,
  "HasSparseWeights": 1,
  "HasDynamicFilter": 0,
  "HasDynamicBias": 0,
  "HasResidual": 0,
  "ConvXAsActInputIdx": -1,
  "BiasAsActInputIdx": -1,
  "ResAsActInputIdx": -1,
  "Activation": "NONE",
  "TacticName": "sm70_xmma_fprop_conv1x1_i8f32_f32_f32_nchw_vect_c_4kcrs_vect_c_4_nchw_simt_small_batch_bias_relu",
  "TacticValue": "0xc073b0053ce90eac",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_dense0_fwd]"
},{
  "Name": "reshape_after_resnetv15_dense0_fwd",
  "LayerType": "NoOp",
  "Inputs": [
  {
    "Name": "resnetv15_dense0_fwd_conv_out",
    "Location": "Device",
    "Dimensions": [1,1000,1,1],
    "Format/Datatype": "Row major linear FP32"
  }],
  "Outputs": [
  {
    "Name": "resnetv15_dense0_fwd",
    "Location": "Device",
    "Dimensions": [1,1000],
    "Format/Datatype": "Row major linear FP32"
  }],
  "TacticValue": "0x0000000000000000",
  "StreamId": 0,
  "Metadata": "[ONNX Layer: resnetv15_dense0_fwd]"
}],
"Bindings": ["data"
,"resnetv15_dense0_fwd"
]}
leimao commented 2 months ago

BTW, I suggest making --profilingVerbosity=detailed when --exportLayerInfo is turned on. The user will examine the layerwise info only because they are interested in these detailed information.