tensorrt-cpp-api

TensorRT C++ API

A modern, no-throw C++ library for high-performance GPU inference of CNN models with NVIDIA TensorRT — with optional zero-copy Python bindings.

tensorrt_cpp_api turns an ONNX model into a cached, optimized TensorRT engine and runs it with a small, leak-free API: name-keyed tensors at the boundary, caller-owned CUDA streams, explicit host/device transfers, and a Status/Result error model — no exceptions, no OpenCV or TensorRT types in the public headers. It targets TensorRT ≥ 10 (built to the TensorRT 11 surface), CUDA 12, C++20, Linux.

#include 
using namespace trtcpp;

int main() {
    // Build an FP16 engine from ONNX, or load it from the on-disk cache if one is already current.
    BuildOptions opt;
    opt.precision = Precision::kFp16;
    opt.engineCacheDir = "engines";
    auto engine = EngineBuilder{}.buildAndLoad("model.onnx", opt);
    if (!engine) {
        std::fprintf(stderr, "%s\n", engine.status().message().c_str());
        return 1;
    }

    Stream stream;  // owns a CUDA stream — or Stream::wrap(existingHandle) to use yours
    auto input = Tensor::allocate(DType::kFloat32, Shape{1, 3, 640, 640}, Device::kCuda).value();
    // ... fill `input` (e.g. via the fused preproc kernel) ...

    auto output = engine->inferSingle({{engine->inputNames().front(), input.view()}}, stream);
    if (!output) return 1;

    auto host = output->toHost(stream).value();   // explicit D2H + sync; never implicit
    std::span scores = host.as().value();
    // ... post-process `scores` ...
}

Features

• Engine cache that's actually safe. Build-or-load keyed by ONNX content hash + build options
  • TensorRT version + GPU UUID, with a JSON sidecar and atomic writes. A stale cache (changed model, options, driver, or GPU) is detected and rebuilt instead of silently misused.
• Dynamic shapes, done right. Per-input min/opt/max optimization profiles; -1-aware Shape; one optimization profile per execution context for concurrent dynamic-shape inference.
• Concurrency. EnginePool leases execution contexts for multi-stream inference; every call runs on a caller-provided Stream. The engine is thread-compatible; the pool is thread-safe.
• No leaky abstractions. No nvinfer1, OpenCV, or spdlog types in any public header (PImpl + version-gated, generated build_config.h). Consumers need TensorRT at runtime, not compile time.
• Quantization without surprises. Precision::kFp16 / kInt8Qdq / kFp8 …; precision is version-aware and never a silent no-op (it errors clearly when a mode isn't achievable).
• Optional fused preprocessing (tensorrt_cpp_api::preproc): one CUDA kernel does letterbox-resize → BGR↔RGB → per-channel normalize → HWC→NCHW → cast, no intermediate buffers.
• Optional zero-copy Python bindings (trtcpp): feed CuPy / PyTorch / Numba GPU arrays in and get them back via __cuda_array_interface__ / DLPack — no host round-trips, GIL released during inference. See examples/python.
• Installable. cmake --install produces a find_package(tensorrt_cpp_api)-consumable package.

Performance

Single-stream inference latency on an RTX 3080 Laptop GPU (preallocated, zero-copy enqueue loop — examples/benchmark), TensorRT 10:

Inference time is TensorRT-bound — it is the enqueueV3 cost of the engine itself, so the wrapper adds no measurable inference overhead. The library's work is everything around that call: zero-copy name-keyed IO with no per-call allocations or nested-vector copies, a stream-ordered allocator, and the no-throw Status/Result API. The Python bindings run the same path within ~13% of C++ (examples/python/benchmark_parity.py).

Install

TensorRT and CUDA are system/externally provided. In brief:

cmake -S . -B build -DTRT_CPP_API_BUILD_PREPROC=ON        # add -DTensorRT_DIR= for a tarball
cmake --build build -j$(nproc)
cmake --install build --prefix /opt/trtcpp

Then in a downstream project:

find_package(tensorrt_cpp_api REQUIRED)
target_link_libraries(myapp PRIVATE tensorrt_cpp_api::tensorrt_cpp_api tensorrt_cpp_api::preproc)

Python: pip install . (builds the trtcpp wheel via scikit-build-core). Full details — apt vs tarball TensorRT, build options, Python — are in docs/install.md.

Examples

examples/ has four runnable reference programs, each consuming the installed package: classification (ImageNet top-5), detection (YOLOv8n + NMS), segmentation (DeepLabV3), and a zero-copy Python demo with a C++/Python perf-parity benchmark. examples/download_models.sh fetches the models.

Documentation

• Quickstart & core concepts
• Installation
• Upgrading from v6
• API reference: doxygen Doxyfile (HTML in docs/api/)

Sister projects

This library is the inference backend for YOLOv8-TensorRT-CPP and YOLOv9-TensorRT-CPP (object detection, segmentation, pose).

Scope

Linux, CUDA 12, TensorRT ≥ 10, CNN-style vision models. Windows and LLM/transformer-specific features are out of scope.

Contributing

Issues and PRs welcome. Install the hooks with pre-commit install (clang-format + cmake-format); CI runs the build, the CPU test suite, sanitizers, and a Python wheel build. If this project helps you, a ⭐ is appreciated — connect on LinkedIn.

Contributors

Loic Tetrel
💻

thomaskleiven
💻

WiCyn
💻

This project follows the all-contributors specification.

License

See LICENSE. Version history is in CHANGELOG.md.