TensorRT C++ API Tutorial

How to use TensorRT C++ API for high performance GPU machine-learning inference.
Supports models with single / multiple inputs and single / multiple outputs with batching.

Project Overview Video . Code Deep-Dive Video

Looking for Maintainers 🚀

This project is actively seeking maintainers to help guide its growth and improvement. If you're passionate about this project and interested in contributing, I’d love to hear from you!

Please feel free to reach out via LinkedIn to discuss how you can get involved.

TensorRT C++ Tutorial

I read all the NVIDIA TensorRT docs so that you don't have to!

This project demonstrates how to use the TensorRT C++ API for high performance GPU inference on image data. It covers how to do the following:

• How to install TensorRT 10 on Ubuntu 20.04 / 22.04.
• How to generate a TensorRT engine file optimized for your GPU.
• How to specify a simple optimization profile.
• How to run FP32, FP16, or INT8 precision inference.
• How to read / write data from / into GPU memory and work with GPU images.
• How to use cuda stream to run async inference and later synchronize.
• How to work with models with static and dynamic batch sizes.
• How to work with models with single or multiple output tensors.
• How to work with models with multiple inputs.
• Includes a Video walkthrough where I explain every line of code.
• The code can be used as a base for any model which takes a fixed size image / images as input, including Insightface ArcFace, YoloV8, SCRFD face detection.
  • You will just need to implement the appropriate post-processing code.
• TODO: Add support for models with dynamic input shapes.
• TODO: Add support for Windows

Getting Started

The following instructions assume you are using Ubuntu 20.04 or 22.04. You will need to supply your own onnx model for this sample code or you can download the sample model (see Sanity Check section below).

Prerequisites

• Tested and working on Ubuntu 20.04 and 22.04 (Windows is not supported at this time)
• Install CUDA 11 or 12, instructions here.
  • Recommended >= 12.0
  • Required >= 11.0
• Install cudnn, instructions here.
  • Required >= 8
  • Required < 9 (OpenCV GPU does not yet support)
• sudo apt install build-essential
• sudo snap install cmake --classic
• sudo apt install libspdlog-dev libfmt-dev (for logging)
• Install OpenCV with cuda support. To compile OpenCV from source, run the build_opencv.sh script provided in ./scripts/.
  • If you use the provided script and you have installed cuDNN to a non-standard location, you must modify the CUDNN_INCLUDE_DIR and CUDNN_LIBRARY variables in the script.
  • Recommended >= 4.8
• Download TensorRT 10 from here.
  • Required >= 10.0
• Navigate to the CMakeLists.txt file and replace the TODO with the path to your TensorRT installation.

Building the Library

• mkdir build
• cd build
• cmake ..
• make -j$(nproc)

Running the Executable

• Navigate to the build directory
• Run the executable and provide the path to your onnx model.
• ex. ./run_inference_benchmark --onnx_model ../models/yolov8n.onnx
  • Note: See sanity check section below for instructions on how to obtain the yolov8n model.
• The first time you run the executable for a given model and options, a TensorRT engine file will be built from your onnx model. This process is fairly slow and can take 5+ minutes for some models (ex. yolo models).
• Alternatively, you can choose to supply your own TensorRT engine file directly:
• ex. ./run_inference_benchmark --trt_model ../models/yolov8n.engine.NVIDIAGeForceRTX3080LaptopGPU.fp16.1.1
  • Note: See V5.0 changelog below for warnings when supply your own TensorRT engine file.

Sanity Check

• To perform a sanity check, download the YOLOv8n model from here.
• Next, convert it from pytorch to onnx using the following script:
  • You will need to run pip3 install ultralytics first.

from ultralytics import YOLO
model = YOLO("./yolov8n.pt")
model.fuse()
model.info(verbose=False)  # Print model information
model.export(format="onnx", opset=12) # Export the model to onnx using opset 12

• Place the resulting onnx model, yolov8n.onnx, in the ./models/ directory.
• Running inference using said model and the image located in ./inputs/team.jpg should produce the following feature vector:
  • Note: The feature vector will not be identical (but very similar) as TensorRT is not deterministic.

3.41113 16.5312 20.8828 29.8984 43.7266 54.9609 62.0625 65.8594 70.0312 72.9531 ...

INT8 Inference

Enabling INT8 precision can further speed up inference at the cost of accuracy reduction due to reduced dynamic range. For INT8 precision, the user must supply calibration data which is representative of real data the model will see. It is advised to use 1K+ calibration images. To enable INT8 inference with the YoloV8 sanity check model, the following steps must be taken:

• Change options.precision = Precision::FP16; to options.precision = Precision::INT8; in main.cpp
• options.calibrationDataDirectoryPath = ""; must be changed in main.cpp to specify path containing calibration data.
  • If using the YoloV8 model, it is advised to used the COCO validation dataset, which can be downloaded with wget http://images.cocodataset.org/zips/val2017.zip
• Make sure the resizing code in the Int8EntropyCalibrator2::getBatch method in engine.cpp (see TODO) is correct for your model.
  • If using the YoloV8 model, the preprocessing code is correct and does not need to be changed.
• Recompile, run the executable.
• The calibration cache will be written to disk (.calibration extension) so that on subsequent model optimizations it can be reused. If you'd like to regenerate the calibration data, you must delete this cache file.
• If you get an "out of memory in function allocate" error, then you must reduce Options.calibrationBatchSize so that the entire batch can fit in your GPU memory.

Benchmarks

Benchmarks run on RTX 3050 Ti Laptop GPU, 11th Gen Intel(R) Core(TM) i9-11900H @ 2.50GHz.

Model	Precision	Batch Size	Avg Inference Time
yolov8n	FP32	1	4.732 ms
yolov8n	FP16	1	2.493 ms
yolov8n	INT8	1	2.009 ms
yolov8x	FP32	1	76.63 ms
yolov8x	FP16	1	25.08 ms
yolov8x	INT8	1	11.62 ms

Sample Integration

Wondering how to integrate this library into your project? Or perhaps how to read the outputs of the YoloV8 model to extract meaningful information? If so, check out my two latest projects, YOLOv8-TensorRT-CPP and YOLOv9-TensorRT-CPP, which demonstrate how to use the TensorRT C++ API to run YoloV8/9 inference (supports object detection, semantic segmentation, and body pose estimation). They make use of this project in the backend!

Project Structure

project-root/
├── include/
│   ├── engine/
│   │   ├── EngineRunInference.inl
│   │   ├── EngineUtilities.inl
│   │   └── EngineBuildLoadNetwork.inl
│   ├── util/...
│   ├── ...
├── src/
|   ├── ...
│   ├── engine.cpp
│   ├── engine.h
│   └── main.cpp
├── CMakeLists.txt
└── README.md

Understanding the Code

• The bulk of the implementation is located in include/engine. I have written lots of comments all throughout the code which should make it easy to understand what is going on.
• The inference code is located in include/engine/EngineRunInference.inl.
• The building and loading of the TensorRT engine file is located in include/engine/EngineBuildLoadNetwork.inl.
• You can also check out my deep-dive video in which I explain every line of code.

How to Debug

• The implementation uses the spdlog library for logging. You can change the log level by setting the environment variable LOG_LEVEL to one of the following values: trace, debug, info, warn, error, critical, off.

• If you have issues creating the TensorRT engine file from the onnx model, consider setting the environment variable LOG_LEVEL to trace and re-run the application. This should give you more information on where exactly the build process is failing.

Show your Appreciation

If this project was helpful to you, I would appreciate if you could give it a star. That will encourage me to ensure it's up to date and solve issues quickly. I also do consulting work if you require more specific help. Connect with me on LinkedIn.

Contributors

Loic Tetrel 💻

thomaskleiven 💻

WiCyn 💻

Changelog

V6.0

• Implementation now requires TensorRT >= 10.0.

V5.0

• Engine class has been modified to take a template parameter which specifies the models output data type. The implementation now supports outputs of type float, __half, int8_t, int32_t, bool, and uint8_t.
• Added support for loading TensorRT engine file directly without needing to compile from onnx model. Howver, it is highly recommended that you use the API provided to build the engine file from the onnx model, instead of loading a TensorRT model directly. If you choose to load a TensorRT model file directly, you must hand-check that the Options have been set correctly for your model (for example, if your model has been compiled for FP32 but you try running FP16 inference, it will fail, potentially without a verbose error).
• Added command line parser.

V4.1

• Added support for fixed batch size > 1.

V4.0

• Added support for INT8 precision.

V3.0

• Implementation has been updated to use TensorRT 8.6 API (ex. IExecutionContext::enqueueV3()).
• Executable has renamed from driver to run_inference_benchmark and now must be passed path to onnx model as command line argument.
• Removed Options.doesSupportDynamicBatchSize. Implementation now auto-detects supported batch sizes.
• Removed Options.maxWorkspaceSize. Implementation now does not limit GPU memory during model constructions, allowing implementation to use as much of memory pool as is available for intermediate layers.

v2.2

• Serialize model name as part of engine file.

V2.1

• Added support for models with multiple inputs. Implementation now supports models with single inputs, multiple inputs, single outputs, multiple outputs, and batching.

V2.0

• Requires OpenCV cuda to be installed. To install, follow instructions here.
• Options.optBatchSizes has been removed, replaced by Options.optBatchSize.
• Support models with more than a single output (ex. SCRFD).
• Added support for models which do not support batch inference (first input dimension is fixed).
• More error checking.
• Fixed a bunch of common issues people were running into with the original V1.0 version.
• Remove whitespace from GPU device name

Contributors ✨

Thanks goes to these wonderful people (emoji key):

This project follows the all-contributors specification. Contributions of any kind welcome!