### Description * Add std::numeric_limits for MLFloat16 and BFloat16. * Update some comments in csharp ORTFloat16.shared.cs. * Add unit tests (including Clip) Note that the canonical NaN is not consistent in C++ and C#. C# uses negative quiet NaN as canonical NaN, while C++ uses positive quiet NaN. The choice of CSharp Float16.NaN is to be consistent with System.Half.NaN. FP16 data returns from CUDA might have 7FFF as NaN; FP16 data from CPU provider might have 0x7E00 as NaN. Anyway there is no consistent canonical NaN in ORT right now. Because all these NaNs are aligned with IEEE spec, there shall not an issue in downstream. ### Motivation and Context std::numeric_limits is used in codebase but not defined for MLFloat16 and BFloat16. It causes some bugs like https://github.com/microsoft/onnxruntime/issues/21957 introduced by https://github.com/microsoft/onnxruntime/pull/21493. |
||
|---|---|---|
| .. | ||
| ApiDocs | ||
| sample | ||
| src | ||
| test | ||
| testdata | ||
| tools | ||
| .clang-format | ||
| Directory.Build.props.in | ||
| NuGet.CSharp.config | ||
| OnnxRuntime.CSharp.proj | ||
| OnnxRuntime.CSharp.sln | ||
| OnnxRuntime.DesktopOnly.CSharp.sln | ||
| OnnxRuntime.snk | ||
| readme.md | ||
| readme.txt | ||
ORT C# Managed Library
The solution files here are used to produce nuget packages for the C# bindings.
Note that the project naming is currently confusing and needs updating.
- The Microsoft.ML.OnnxRuntime project produces the Microsoft.ML.OnnxRuntime.Managed nuget package.
- The Microsoft.ML.OnnxRuntime nuget package contains the native (i.e. C++) code for various platforms.
Solution files
The main solution file is OnnxRuntime.CSharp.sln. This includes desktop and Xamarin mobile projects. OnnxRuntime.DesktopOnly.CSharp.sln is a copy of that with all the mobile projects removed. This is due to there being no way to selectively exclude a csproj from the sln if Xamarin isn't available.
If changes are required, either update the main solution first and copy the relevant changes across, or copy the entire file and remove the mobile projects (anything with iOS, Android or Droid in the name).
Development setup:
Requirements:
Windows
NOTE: The usage of this solution is primarily for ORT developers creating the managed Microsoft.ML.OnnxRuntime.Managed nuget package. Due to that, the requirements are quite specific.
Visual Studio 2022 v17.2.4 or later, with Xamarin workloads
- v17.2.4 installs dotnet sdk 6.0.301
- in theory you could use an earlier VS version and download dotnet SDK 6.0.300+ from https://dotnet.microsoft.com/en-us/download/dotnet/6.0
- untested
There's no good way to use Visual Studio 2022 17.3 Preview in a CI, so we currently have to build pre-.net6 targets using VS, and .net6 targets using dotnet. We can't build them all using dotnet as the xamarin targets require msbuild. We can't package them using dotnet as that also requires msbuild.
Once the official VS 2022 release supports .net6 and is available in the CI we can revert to the original simple setup of building everything using msbuild.
To test packaging locally you will also need nuget.exe. Download from https://www.nuget.org/downloads. Put in a folder (e.g. C:\Program Files (x86)\NuGet). Add that folder to your PATH.
Linux
- Install .Net SDK.
- Install Mono.
wget http://download.mono-project.com/repo/xamarin.gpg && sudo apt-key add xamarin.gpg && rm xamarin.gpg echo "deb https://download.mono-project.com/repo/ubuntu stable-bionic main" | sudo tee /etc/apt/sources.list.d/mono-official-stable.list sudo apt update -y && apt install -y mono-devel - Install
nupkg.exewget https://dist.nuget.org/win-x86-commandline/latest/nuget.exe && sudo mv nuget.exe /usr/local/bin/nuget.exe echo 'mono /usr/local/bin/nuget.exe $@' | sudo tee /usr/local/bin/nuget chmod a+x /usr/local/bin/nuget
Magic incantations to build the nuget managed package locally:
Windows
If we're starting with VS 2022 17.2.4 we should have dotnet sdk 6.0.301
The IncludeMobileTargets property determines whether the MAUI targets are included. This is set to 'true' for the real package build.
Make sure all the required workloads are installed if the MAUI targets are included.
dotnet workload install maui android ios maccatalyst
- original example from here:
Restore and build the managed package.
msbuild -t:restore .\src\Microsoft.ML.OnnxRuntime\Microsoft.ML.OnnxRuntime.csproj -p:IncludeMobileTargets=true
msbuild -t:build .\src\Microsoft.ML.OnnxRuntime\Microsoft.ML.OnnxRuntime.csproj -p:IncludeMobileTargets=true
Need to run msbuild twice - once to restore which creates some json configs that are needed like Microsoft.ML.OnnxRuntime\obj\project.assets.json, and once to build using those configs.
Create nuget package
msbuild .\OnnxRuntime.CSharp.proj -t:CreatePackage -p:OrtPackageId=Microsoft.ML.OnnxRuntime -p:Configuration=Debug -p:Platform="Any CPU"
Linux
For example, to build a CUDA GPU package, just run:
./build.sh \
--config="Release" \
--cmake_generator Ninja \
--use_cuda \
--cuda_home=/usr/local/cuda \
--cudnn_home=/usr \
--build_nuget \
--msbuild_extra_options \
/p:SelectedTargets=Net6 \
/p:Net6Targets=net6.0 \
/p:TargetFrameworks=netstandard2.0 \
/p:IsLinuxBuild=true
Note: to build a pure CPU development package, you need to add /p:OrtPackageId="Microsoft.ML.OnnxRuntime"
to --msbuild_extra_options. Otherwise, it will try to create Xamarin mobile targets which may not be properly configured on your devbox.
A .nupkg file will be produced at you build root, say, build/Release.
To consume the package, in your .net project,
nuget add <path/to/packages.nuget> -Source ./packages/
dotnet add package microsoft.ml.onnxruntime.managed -s ./packages --prerelease
dotnet add package microsoft.ml.onnxruntime.gpu -s ./packages --prerelease