TensorRT 开始 - HelloWorld开发者社区

TensorRT 是 NVIDIA 自家的高性能推理库，其 Getting Started 列出了各资料入口，如下：

TensorRT 开始

本文基于当前的 TensorRT 8.2 版本，将一步步介绍从安装，直到加速推理自己的 ONNX 模型。

安装

进 TensorRT 下载页选择版本下载，需注册登录。

TensorRT 开始

本文选择了 TensorRT-8.2.2.1.Linux.x86_64-gnu.cuda-11.4.cudnn8.2.tar.gz，可以注意到与 CUDA cuDNN 要匹配好版本。也可以准备 NVIDIA Docker 拉取对应版本的 nvidia/cuda 镜像，再 ADD TensorRT 即可。

# 解压进 $HOME （以免 sudo 编译样例，为当前用户）
tar -xzvf TensorRT-*.tar.gz -C $HOME/
# 软链到 /usr/local/TensorRT （以固定一个路径）
sudo ln -s $HOME/TensorRT-8.2.2.1 /usr/local/TensorRT

之后，编译运行样例，保证 TensorRT 安装正确。

编译样例

样例在 TensorRT/samples，说明见 Sample Support Guide 或各样例目录里的 README.md。

cd /usr/local/TensorRT/samples/

# 设定环境变量，可见 Makefile.config
export CUDA_INSTALL_DIR=/usr/local/cuda
export CUDNN_INSTALL_DIR=/usr/local/cuda
export ENABLE_DLA=
export TRT_LIB_DIR=../lib
export PROTOBUF_INSTALL_DIR=

# 编译
make -j`nproc`

# 运行
export LD_LIBRARY_PATH=/usr/local/TensorRT/lib:$LD_LIBRARY_PATH
cd /usr/local/TensorRT/
./bin/trtexec -h
./bin/sample_mnist -d data/mnist/ --fp16

运行结果参考：

$ ./bin/sample_mnist -d data/mnist/ --fp16
&&&& RUNNING TensorRT.sample_mnist [TensorRT v8202] # ./bin/sample_mnist -d data/mnist/ --fp16
[12/23/2021-20:20:16] [I] Building and running a GPU inference engine for MNIST
[12/23/2021-20:20:16] [I] [TRT] [MemUsageChange] Init CUDA: CPU +322, GPU +0, now: CPU 333, GPU 600 (MiB)
[12/23/2021-20:20:16] [I] [TRT] [MemUsageSnapshot] Begin constructing builder kernel library: CPU 333 MiB, GPU 600 MiB
[12/23/2021-20:20:16] [I] [TRT] [MemUsageSnapshot] End constructing builder kernel library: CPU 468 MiB, GPU 634 MiB
[12/23/2021-20:20:17] [I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +518, GPU +224, now: CPU 988, GPU 858 (MiB)
[12/23/2021-20:20:17] [I] [TRT] [MemUsageChange] Init cuDNN: CPU +114, GPU +52, now: CPU 1102, GPU 910 (MiB)
[12/23/2021-20:20:17] [I] [TRT] Local timing cache in use. Profiling results in this builder pass will not be stored.
[12/23/2021-20:20:33] [I] [TRT] Some tactics do not have sufficient workspace memory to run. Increasing workspace size may increase performance, please check verbose output.
[12/23/2021-20:20:34] [I] [TRT] Detected 1 inputs and 1 output network tensors.
[12/23/2021-20:20:34] [I] [TRT] Total Host Persistent Memory: 8448
[12/23/2021-20:20:34] [I] [TRT] Total Device Persistent Memory: 1626624
[12/23/2021-20:20:34] [I] [TRT] Total Scratch Memory: 0
[12/23/2021-20:20:34] [I] [TRT] [MemUsageStats] Peak memory usage of TRT CPU/GPU memory allocators: CPU 2 MiB, GPU 13 MiB
[12/23/2021-20:20:34] [I] [TRT] [BlockAssignment] Algorithm ShiftNTopDown took 0.01595ms to assign 3 blocks to 8 nodes requiring 57857 bytes.
[12/23/2021-20:20:34] [I] [TRT] Total Activation Memory: 57857
[12/23/2021-20:20:34] [I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +8, now: CPU 1621, GPU 1116 (MiB)
[12/23/2021-20:20:34] [I] [TRT] [MemUsageChange] Init cuDNN: CPU +0, GPU +8, now: CPU 1621, GPU 1124 (MiB)
[12/23/2021-20:20:34] [I] [TRT] [MemUsageChange] TensorRT-managed allocation in building engine: CPU +0, GPU +4, now: CPU 0, GPU 4 (MiB)
[12/23/2021-20:20:34] [I] [TRT] [MemUsageChange] Init CUDA: CPU +0, GPU +0, now: CPU 1622, GPU 1086 (MiB)
[12/23/2021-20:20:34] [I] [TRT] Loaded engine size: 1 MiB
[12/23/2021-20:20:34] [I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +8, now: CPU 1622, GPU 1096 (MiB)
[12/23/2021-20:20:34] [I] [TRT] [MemUsageChange] Init cuDNN: CPU +1, GPU +8, now: CPU 1623, GPU 1104 (MiB)
[12/23/2021-20:20:34] [I] [TRT] [MemUsageChange] TensorRT-managed allocation in engine deserialization: CPU +0, GPU +1, now: CPU 0, GPU 1 (MiB)
[12/23/2021-20:20:34] [I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +8, now: CPU 1485, GPU 1080 (MiB)
[12/23/2021-20:20:34] [I] [TRT] [MemUsageChange] Init cuDNN: CPU +0, GPU +8, now: CPU 1485, GPU 1088 (MiB)
[12/23/2021-20:20:34] [I] [TRT] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +2, now: CPU 0, GPU 3 (MiB)
[12/23/2021-20:20:34] [I] Input:
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@%+-:  =@@@@@@@@@@@@
@@@@@@@%=      -@@@**@@@@@@@
@@@@@@@   :%#@-#@@@. #@@@@@@
@@@@@@*  +@@@@:*@@@  *@@@@@@
@@@@@@#  +@@@@ @@@%  @@@@@@@
@@@@@@@.  :%@@.@@@. *@@@@@@@
@@@@@@@@-   =@@@@. -@@@@@@@@
@@@@@@@@@%:   +@- :@@@@@@@@@
@@@@@@@@@@@%.  : -@@@@@@@@@@
@@@@@@@@@@@@@+   #@@@@@@@@@@
@@@@@@@@@@@@@@+  :@@@@@@@@@@
@@@@@@@@@@@@@@+   *@@@@@@@@@
@@@@@@@@@@@@@@: =  @@@@@@@@@
@@@@@@@@@@@@@@ :@  @@@@@@@@@
@@@@@@@@@@@@@@ -@  @@@@@@@@@
@@@@@@@@@@@@@# +@  @@@@@@@@@
@@@@@@@@@@@@@* ++  @@@@@@@@@
@@@@@@@@@@@@@*    *@@@@@@@@@
@@@@@@@@@@@@@#   =@@@@@@@@@@
@@@@@@@@@@@@@@. +@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@

[12/23/2021-20:20:34] [I] Output:
0:
1:
2:
3:
4:
5:
6:
7:
8: **********
9:

&&&& PASSED TensorRT.sample_mnist [TensorRT v8202] # ./bin/sample_mnist -d data/mnist/ --fp16

快速开始

Quick Start Guide / Using The TensorRT Runtime API

准备教程代码，编译：

git clone --depth 1 https://github.com/NVIDIA/TensorRT.git

export CUDA_INSTALL_DIR=/usr/local/cuda
export CUDNN_INSTALL_DIR=/usr/local/cuda
export TRT_LIB_DIR=/usr/local/TensorRT/lib

# 编译 quickstart
cd TensorRT/quickstart
# Makefile.config
#  INCPATHS += -I"/usr/local/TensorRT/include"
# common/logging.h
#  void log(Severity severity, const char* msg) noexcept override
make

# 运行环境
export PATH=/usr/local/TensorRT/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/TensorRT/lib:$LD_LIBRARY_PATH
cd SemanticSegmentation

获取预训练 FCN-ResNet-101 模型，转成 ONNX：

# 创建本地环境
#  conda create -n torch python=3.9 -y
#  conda activate torch
#  conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch -y
# 不然，容器环境
#  docker run --rm -it --gpus all -p 8888:8888 -v `pwd`:/workspace/SemanticSegmentation -w /workspace nvcr.io/nvidia/pytorch:20.12-py3 bash
$ python export.py
Exporting ppm image input.ppm
Downloading: "https://github.com/pytorch/vision/archive/v0.6.0.zip" to /home/john/.cache/torch/hub/v0.6.0.zip
Downloading: "https://download.pytorch.org/models/resnet101-5d3b4d8f.pth" to /home/john/.cache/torch/hub/checkpoints/resnet101-5d3b4d8f.pth
100%|████████████████████████████████████████| 170M/170M [00:27<00:00, 6.57MB/s]
Downloading: "https://download.pytorch.org/models/fcn_resnet101_coco-7ecb50ca.pth" to /home/john/.cache/torch/hub/checkpoints/fcn_resnet101_coco-7ecb50ca.pth
100%|████████████████████████████████████████| 208M/208M [02:26<00:00, 1.49MB/s]
Exporting ONNX model fcn-resnet101.onnx

再用 trtexec 将 ONNX 转成 TensorRT engine：

$ trtexec --onnx=fcn-resnet101.onnx --fp16 --workspace=64 --minShapes=input:1x3x256x256 --optShapes=input:1x3x1026x1282 --maxShapes=input:1x3x1440x2560 --buildOnly --saveEngine=fcn-resnet101.engine
...
[01/07/2022-20:20:00] [I] Engine built in 406.011 sec.
&&&& PASSED TensorRT.trtexec [TensorRT v8202] ...

随机输入，测试 engine：

$ trtexec --shapes=input:1x3x1026x1282 --loadEngine=fcn-resnet101.engine
...
[01/07/2022-20:20:00] [I] === Performance summary ===
[01/07/2022-20:20:00] [I] Throughput: 12.4749 qps
[01/07/2022-20:20:00] [I] Latency: min = 76.9746 ms, max = 98.8354 ms, mean = 79.5844 ms, median = 78.0542 ms, percentile(99%) = 98.8354 ms
[01/07/2022-20:20:00] [I] End-to-End Host Latency: min = 150.942 ms, max = 188.431 ms, mean = 155.834 ms, median = 152.444 ms, percentile(99%) = 188.431 ms
[01/07/2022-20:20:00] [I] Enqueue Time: min = 0.390625 ms, max = 1.61279 ms, mean = 1.41182 ms, median = 1.46136 ms, percentile(99%) = 1.61279 ms
[01/07/2022-20:20:00] [I] H2D Latency: min = 1.25977 ms, max = 1.53467 ms, mean = 1.27415 ms, median = 1.26514 ms, percentile(99%) = 1.53467 ms
[01/07/2022-20:20:00] [I] GPU Compute Time: min = 75.2869 ms, max = 97.1318 ms, mean = 77.8847 ms, median = 76.3599 ms, percentile(99%) = 97.1318 ms
[01/07/2022-20:20:00] [I] D2H Latency: min = 0.408447 ms, max = 0.454346 ms, mean = 0.425577 ms, median = 0.423004 ms, percentile(99%) = 0.454346 ms
[01/07/2022-20:20:00] [I] Total Host Walltime: 3.2866 s
[01/07/2022-20:20:00] [I] Total GPU Compute Time: 3.19327 s
[01/07/2022-20:20:00] [I] Explanations of the performance metrics are printed in the verbose logs.
[01/07/2022-20:20:00] [I]
&&&& PASSED TensorRT.trtexec [TensorRT v8202] ...

运行教程，使用 engine：

$ ./bin/segmentation_tutorial
[01/07/2022-20:20:34] [I] [TRT] [MemUsageChange] Init CUDA: CPU +322, GPU +0, now: CPU 463, GPU 707 (MiB)
[01/07/2022-20:20:34] [I] [TRT] Loaded engine size: 132 MiB
[01/07/2022-20:20:35] [I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +520, GPU +224, now: CPU 984, GPU 1065 (MiB)
[01/07/2022-20:20:35] [I] [TRT] [MemUsageChange] Init cuDNN: CPU +115, GPU +52, now: CPU 1099, GPU 1117 (MiB)
[01/07/2022-20:20:35] [I] [TRT] [MemUsageChange] TensorRT-managed allocation in engine deserialization: CPU +0, GPU +131, now: CPU 0, GPU 131 (MiB)
[01/07/2022-20:20:35] [I] Running TensorRT inference for FCN-ResNet101
[01/07/2022-20:20:35] [I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +10, now: CPU 966, GPU 1109 (MiB)
[01/07/2022-20:20:35] [I] [TRT] [MemUsageChange] Init cuDNN: CPU +0, GPU +8, now: CPU 966, GPU 1117 (MiB)
[01/07/2022-20:20:35] [I] [TRT] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +722, now: CPU 0, GPU 853 (MiB)

TensorRT 开始

实践

以上给到了官方样例与教程的编译使用。这里，另外找了个 RVM 的模型，从头开始试一试。

准备模型

Robust Video Matting (RVM) 稳定视频抠像，可在任意视频上做实时高清抠像。有 Webcam Demo 可以网页上体验。

准备 ONNX 模型 rvm_mobilenetv3_fp32.onnx，其推断文档给出了模型输入输出：

输入: [src, r1i, r2i, r3i, r4i, downsample_ratio]
- src：输入帧，RGB 通道，形状为 [B, C, H, W]，范围为0~1
- rXi：记忆输入，初始值是是形状为 [1, 1, 1, 1] 的零张量
- downsample_ratio 下采样比，张量形状为 [1]
- 只有 downsample_ratio 必须是 FP32，其他输入必须和加载的模型使用一样的 dtype
输出: [fgr, pha, r1o, r2o, r3o, r4o]
- fgr, pha：前景和透明度通道输出，范围为 0~1
- rXo：记忆输出

准备输入图像 input.jpg 。不用视频，保持代码简单些。

准备环境

conda create -n torch python=3.9 -y
conda activate torch

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch -y

# Requirements
#  https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#requirements
pip install onnx onnxruntime-gpu==1.10

运行 ONNX 模型

rvm_onnx_infer.py:

import onnxruntime as ort
import numpy as np
from PIL import Image

# 读取图像
with Image.open('input.jpg') as img:
    img.load()
#  HWC [0,255] > BCHW [0,1]
src = np.array(img)
src = np.moveaxis(src, -1, 0) .astype(np.float32)
src = src[np.newaxis, :] / 255.

# 载入模型
sess = ort.InferenceSession('rvm_mobilenetv3_fp32.onnx', providers=['CUDAExecutionProvider'])

# 创建 io binding
io = sess.io_binding()

# 在 CUDA 上创建张量
rec = [ ort.OrtValue.ortvalue_from_numpy(np.zeros([1, 1, 1, 1], dtype=np.float32), 'cuda') ] * 4
downsample_ratio = ort.OrtValue.ortvalue_from_numpy(np.asarray([0.25], dtype=np.float32), 'cuda')

# 设置输出项
for name in ['fgr', 'pha', 'r1o', 'r2o', 'r3o', 'r4o']:
    io.bind_output(name, 'cuda')

# 推断
io.bind_cpu_input('src', src)
io.bind_ortvalue_input('r1i', rec[0])
io.bind_ortvalue_input('r2i', rec[1])
io.bind_ortvalue_input('r3i', rec[2])
io.bind_ortvalue_input('r4i', rec[3])
io.bind_ortvalue_input('downsample_ratio', downsample_ratio)

sess.run_with_iobinding(io)

fgr, pha, *rec = io.get_outputs()

# 只将 `fgr` 和 `pha` 回传到 CPU
fgr = fgr.numpy()
pha = pha.numpy()

# 合成 RGBA
com = np.where(pha > 0, fgr, pha)
com = np.concatenate([com, pha], axis=1) # + alpha
#  BCHW [0,1] > HWC [0,255]
com = np.squeeze(com, axis=0)
com = np.moveaxis(com, 0, -1) * 255

img = Image.fromarray(com.astype(np.uint8))
img.show()

运行：

python rvm_onnx_infer.py --model "rvm_mobilenetv3_fp32.onnx" --input-image "input.jpg" --precision float32 --show

结果（背景透明）：

TensorRT 开始

ONNX 转成 TRT 模型

trtexec 将 ONNX 转成 TensorRT engine：

export PATH=/usr/local/TensorRT/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/TensorRT/lib:$LD_LIBRARY_PATH

trtexec --onnx=rvm_mobilenetv3_fp32.onnx --workspace=64 --saveEngine=rvm_mobilenetv3_fp32.engine --verbose

发生问题：

[01/08/2022-20:20:36] [E] [TRT] ModelImporter.cpp:773: While parsing node number 3 [Resize -> "389"]:
[01/08/2022-20:20:36] [E] [TRT] ModelImporter.cpp:774: --- Begin node ---
[01/08/2022-20:20:36] [E] [TRT] ModelImporter.cpp:775: input: "src"
input: "386"
input: "388"
output: "389"
name: "Resize_3"
op_type: "Resize"
attribute {
  name: "coordinate_transformation_mode"
  s: "pytorch_half_pixel"
  type: STRING
}
attribute {
  name: "cubic_coeff_a"
  f: -0.75
  type: FLOAT
}
attribute {
  name: "mode"
  s: "linear"
  type: STRING
}
attribute {
  name: "nearest_mode"
  s: "floor"
  type: STRING
}

[01/08/2022-20:20:36] [E] [TRT] ModelImporter.cpp:776: --- End node ---
[01/08/2022-20:20:36] [E] [TRT] ModelImporter.cpp:779: ERROR: builtin_op_importers.cpp:3608 In function importResize:
[8] Assertion failed: scales.is_weights() && "Resize scales must be an initializer!"

这时，需要动手改动模型了。

首先，安装必要工具：

snap install netron
pip install onnx-simplifier
pip install onnx_graphsurgeon --index-url https://pypi.ngc.nvidia.com

之后，Netron 查看模型 Resize_3 节点：

TensorRT 开始

发现其 scales 输入是依据 downsample_ratio 得到的，即 [1,1,downsample_ratio,downsample_ratio]，可用 ONNX GraphSurgeon 修改成常量。

最后，模型改动步骤如下：

# ONNX 模型简化，并改为静态输入尺寸
python -m onnxsim rvm_mobilenetv3_fp32.onnx rvm_mobilenetv3_fp32_sim.onnx \
--input-shape src:1,3,1080,1920 r1i:1,1,1,1 r2i:1,1,1,1 r3i:1,1,1,1 r4i:1,1,1,1

# ONNX GraphSurgeon 修改模型
python rvm_onnx_modify.py -i rvm_mobilenetv3_fp32_sim.onnx --input-size 1920 1280

# trtexec 将 ONNX 转成 TensorRT engine
trtexec --onnx=rvm_mobilenetv3_fp32_sim_modified.onnx --workspace=64 --saveEngine=rvm_mobilenetv3_fp32_sim_modified.engine

rvm_onnx_modify.py:

def modify(input: str, output: str, downsample_ratio: float = 0.25) -> None:
    print(f'\nonnx load: {input}')
    graph = gs.import_onnx(onnx.load(input))

    _print_graph(graph)

    # update node Resize_3: scales
    resize_3 = [n for n in graph.nodes if n.name == 'Resize_3'][0]
    print()
    print(resize_3)

    scales = gs.Constant('388',
        np.asarray([1, 1, downsample_ratio, downsample_ratio], dtype=np.float32))

    resize_3.inputs = [i if i.name != '388' else scales for i in resize_3.inputs]
    print()
    print(resize_3)

    # remove input downsample_ratio
    graph.inputs = [i for i in graph.inputs if i.name != 'downsample_ratio']

    # remove node Concat_2
    concat_2 = [n for n in graph.nodes if n.name == 'Concat_2'][0]
    concat_2.outputs.clear()

    # remove unused nodes/tensors
    graph.cleanup()

    onnx.save(gs.export_onnx(graph), output)

ONNX 与 TRT 模型输出差异

可用 Polygraphy 查看 ONNX 与 TRT 模型的输出差异。

首先，安装

# 安装 TensorRT Python API
cd /usr/local/TensorRT/python/
pip install tensorrt-8.2.2.1-cp39-none-linux_x86_64.whl

export LD_LIBRARY_PATH=/usr/local/TensorRT/lib:$LD_LIBRARY_PATH
python -c "import tensorrt; print(tensorrt.__version__)"

# 安装 Polygraphy，或者通过 TensorRT/tools/Polygraphy 源码安装
python -m pip install colored polygraphy --extra-index-url https://pypi.ngc.nvidia.com

运行 ONNX 与 TRT 模型，对比输出误差：

# 运行 ONNX 模型，保存输入输出
polygraphy run rvm_mobilenetv3_fp32_sim_modified.onnx --onnxrt --val-range [0,1] --save-inputs onnx_inputs.json --save-outputs onnx_outputs.json
# 运行 TRT 模型，载入 ONNX 输入输出，对比输出的相对误差与绝对误差
polygraphy run rvm_mobilenetv3_fp32_sim_modified.engine --model-type engine --trt --load-inputs onnx_inputs.json --load-outputs onnx_outputs.json --rtol 1e-3 --atol 1e-3

可见 fp32 精度误差在 1e-3 以内，PASSED：

[I]     PASSED | All outputs matched | Outputs: ['r4o', 'r3o', 'r2o', 'r1o', 'fgr', 'pha']
[I] PASSED | Command: /home/john/anaconda3/envs/torch/bin/polygraphy run rvm_mobilenetv3_fp32_sim_modified.engine --model-type engine --trt --load-inputs onnx_inputs.json --load-outputs onnx_outputs.json --rtol 1e-3 --atol 1e-3

也试了 fp16，其精度损失就比较大，FAILED：

[E]     FAILED | Mismatched outputs: ['r4o', 'r3o', 'r2o', 'r1o', 'fgr', 'pha']
[!] FAILED | Command: /home/john/anaconda3/envs/torch/bin/polygraphy run rvm_mobilenetv3_fp16_sim_modified.engine --model-type engine --trt --load-inputs onnx_inputs.json --load-outputs onnx_outputs.json --rtol 1e-3 --atol 1e-3

运行 TRT 模型

这里以 TensorRT C++ runtime APIs 为例，将转出的 RVM TRT 模型运行起来。完整代码见 rvm_infer.cc。

1. 载入模型：创建 runtime，反序列化 TRT 模型文件的数据

static Logger logger{Logger::Severity::kINFO};
auto runtime = std::unique_ptr<nvinfer1::IRuntime>(nvinfer1::createInferRuntime(logger));
auto engine = runtime->deserializeCudaEngine(engine_data.data(), fsize, nullptr);

遍历全部输入输出 bindings：

auto nb = engine->getNbBindings();
for (int32_t i = 0; i < nb; i++) {
  auto is_input = engine->bindingIsInput(i);
  auto name = engine->getBindingName(i);
  auto dims = engine->getBindingDimensions(i);
  auto datatype = engine->getBindingDataType(i);
  // ...
}

Engine
 Name=Unnamed Network 0
 DeviceMemorySize=148 MiB
 MaxBatchSize=1
Bindings
 Input[0] name=src dims=[1,3,1080,1920] datatype=FLOAT
 Input[1] name=r1i dims=[1,1,1,1] datatype=FLOAT
 Input[2] name=r2i dims=[1,1,1,1] datatype=FLOAT
 Input[3] name=r3i dims=[1,1,1,1] datatype=FLOAT
 Input[4] name=r4i dims=[1,1,1,1] datatype=FLOAT
 Output[5] name=r4o dims=[1,64,18,32] datatype=FLOAT
 Output[6] name=r3o dims=[1,40,36,64] datatype=FLOAT
 Output[7] name=r2o dims=[1,20,72,128] datatype=FLOAT
 Output[8] name=r1o dims=[1,16,144,256] datatype=FLOAT
 Output[9] name=fgr dims=[1,3,1080,1920] datatype=FLOAT
 Output[10] name=pha dims=[1,1,1080,1920] datatype=FLOAT

之后，分配好所有 bindings 的 device 内存：

auto nb = engine->getNbBindings();
std::vector<void *> bindings(nb, nullptr);
std::vector<int32_t> bindings_size(nb, 0);
for (int32_t i = 0; i < nb; i++) {
  auto dims = engine->getBindingDimensions(i);
  auto size = GetMemorySize(dims, sizeof(float));
  if (cudaMalloc(&bindings[i], size) != cudaSuccess) {
    std::cerr << "ERROR: cuda memory allocation failed, size = " << size
        << " bytes" << std::endl;
    return false;
  }
  bindings_size[i] = size;
}

到此，准备工作就好了。

2. 前处理：输入数据处理成输入格式，存进输入 bindings

用 OpenCV 读取图像，缩放成 src 的输入尺寸。再把数据从 BGR [0,255] 处理成 RGB [0,1]。因 batch=1，所以处理时可忽略。

// img: HWC BGR [0,255] u8
auto img = cv::imread(input_filename, cv::IMREAD_COLOR);
if (src_h != img.rows || src_w != img.cols) {
  cv::resize(img, img, cv::Size(src_w, src_h));
}

// src: BCHW RGB [0,1] fp32
auto src = cv::Mat(img.rows, img.cols, CV_32FC3);
{
  auto src_data = (float*)(src.data);
  for (int y = 0; y < src_h; ++y) {
    for (int x = 0; x < src_w; ++x) {
      auto &&bgr = img.at<cv::Vec3b>(y, x);
      /*r*/ *(src_data + y*src_w + x) = bgr[2] / 255.;
      /*g*/ *(src_data + src_n + y*src_w + x) = bgr[1] / 255.;
      /*b*/ *(src_data + src_n*2 + y*src_w + x) = bgr[0] / 255.;
    }
  }
}
if (cudaMemcpyAsync(bindings[0], src.data, bindings_size[0],
    cudaMemcpyHostToDevice, stream) != cudaSuccess) {
  std::cerr << "ERROR: CUDA memory copy of src failed, size = "
      << bindings_size[0] << " bytes" << std::endl;
  return false;
}

3. 推理：将 bindings 给到 engine 执行上下文进行推理

auto context = std::unique_ptr<nvinfer1::IExecutionContext>(
    engine->createExecutionContext());
if (!context) {
  return false;
}

bool status = context->enqueueV2(bindings.data(), stream, nullptr);
if (!status) {
  std::cout << "ERROR: TensorRT inference failed" << std::endl;
  return false;
}

4. 后处理：从输出 bindings 取出数据，根据输出格式处理数据

用 cv::Mat 接收输出的前景 fgr 和透明通道 pha：

auto fgr = cv::Mat(src_h, src_w, CV_32FC3);  // BCHW RGB [0,1] fp32
if (cudaMemcpyAsync(fgr.data, bindings[9], bindings_size[9],
    cudaMemcpyDeviceToHost, stream) != cudaSuccess) {
  std::cerr << "ERROR: CUDA memory copy of output failed, size = "
      << bindings_size[9] << " bytes" << std::endl;
  return false;
}
auto pha = cv::Mat(src_h, src_w, CV_32FC1);  // BCHW A [0,1] fp32
if (cudaMemcpyAsync(pha.data, bindings[10], bindings_size[10],
    cudaMemcpyDeviceToHost, stream) != cudaSuccess) {
  std::cerr << "ERROR: CUDA memory copy of output failed, size = "
      << bindings_size[10] << " bytes" << std::endl;
  return false;
}
cudaStreamSynchronize(stream);

再将 fgr pha 合成 RGBA 数据，并复原成原尺寸：

// Compose `fgr` and `pha`
auto com = cv::Mat(src_h, src_w, CV_8UC4);  // HWC BGRA [0,255] u8
{
  auto fgr_data = (float*)(fgr.data);
  auto pha_data = (float*)(pha.data);
  for (int y = 0; y < com.rows; ++y) {
    for (int x = 0; x < com.cols; ++x) {
      auto &&elem = com.at<cv::Vec4b>(y, x);
      auto alpha = *(pha_data + y*src_w + x);
      if (alpha > 0) {
        /*r*/ elem[2] = *(fgr_data + y*src_w + x) * 255;
        /*g*/ elem[1] = *(fgr_data + src_n + y*src_w + x) * 255;
        /*b*/ elem[0] = *(fgr_data + src_n*2 + y*src_w + x) * 255;
      } else {
        /*r*/ elem[2] = 0;
        /*g*/ elem[1] = 0;
        /*b*/ elem[0] = 0;
      }
      /*a*/ elem[3] = alpha * 255;
    }
  }
}
if (dst_h != com.rows || dst_w != com.cols) {
  cv::resize(com, com, cv::Size(dst_w, dst_h));
}

5. 运行得到的抠像结果（背景透明）：

TensorRT 开始