Getting Started

TensorRtSharp4.0 currently has two verified local validation baselines:

  • CUDA smoke path through CudaSmokeRunner
  • TensorRT vendor-backed minimum path through TensorRtSmokeRunner

Prerequisites

  • .NET SDK 10.0.300 or newer
  • CMake 3.27 or newer
  • Windows x64 or Linux x64

Windows local development currently assumes:

  • CUDA is installed in a standard NVIDIA Toolkit location and is discoverable through CUDA_PATH* or %ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA
  • TensorRT and cuDNN are installed outside the repository and selected with JYPPX_TENSORRT_ROOT and JYPPX_CUDNN_ROOT

First commands

dotnet restore
dotnet build TensorRtSharp.sln -c Debug
powershell -ExecutionPolicy Bypass -File .\eng\Get-Dependencies.ps1
$env:JYPPX_ENABLE_DEVELOPMENT_PROBING = "1"

Verified common examples

MultiStream:

dotnet .\samples\Performance\01.MultiStream\bin\Debug\net8.0\MultiStream.dll

DynamicShape:

dotnet .\samples\Inference\02.DynamicShapes\bin\Debug\net8.0\DynamicShape.dll --tensor-rt-line 10

InferenceBindings:

dotnet .\samples\Inference\01.Bindings\bin\Debug\net8.0\InferenceBindings.dll --tensor-rt-line 10 --batch 2

OnnxToEngine:

dotnet .\applications\OnnxToEngine\bin\Debug\net8.0\OnnxToEngine.dll --tensor-rt-line 10

Classification with your own ONNX classifier assets:

dotnet run --project .\samples\ComputerVision\01.Classification -- --model .\models\classifier.onnx --labels .\models\labels.txt --input-shape 1x3x224x224 --tensor-rt-line 10

YOLO-family detection with your own ONNX detector assets:

dotnet run --project .\applications\YoloVision -- --model .\models\yolo.onnx --labels .\models\coco.names --input-shape 1x3x640x640 --tensor-rt-line 10

Verified smoke commands

CUDA smoke:

dotnet .\smoke\CudaSmokeRunner\bin\Debug\net8.0\CudaSmokeRunner.dll

TensorRT minimum smoke:

dotnet .\smoke\TensorRtSmokeRunner\bin\Debug\net8.0\TensorRtSmokeRunner.dll
Edit this page