CpuMath: Sse code is executed when Avx is available, except on .NET Core 3.0

[helloguo]

System information

• Windows 10 Enterprise:
• dotnet Version: 2.1.105:

Issue

• What did you do?
  I followed the tutorial https://www.microsoft.com/net/learn/apps/machine-learning-and-ai/ml-dotnet/get-started/windows to install SDK and create the sample app. I ran this example in release mode dotnet run -c Release and profiled it.

• What happened?
  The profiling shows Sse code is executed as shown below.
  

• What did you expect?
  Since my machine is Avx capable and AddScale has Avx version https://github.com/dotnet/machinelearning/blob/master/src/Microsoft.ML.CpuMath/Avx.cs#L721, I would expect the Avx version is used.

[danmoseley]

This relates to the native code implementation, which is used on .NET Core 2.1 and .NET Framework.

Note Avx.cpp, which as this issue points out was dead, was removed in 263a67b. I do not think there are plans to bring it back and wire it up, although it could be.

[helloguo]

Thank you for explaining. Closing this issue if there is no other concern,

[danmoseley]

I think it's fine to have the issue open, in case someone has the (considerable) motivation to enable AVX support in the native code. (Starting by undeleting the file, then wiring it up as needed, with perf measurements and tests.)

[helloguo]

Adding AVX CPP implementation would benefit Full Framework, which has no C# HW Intrinsic available today. I guess it really depends on the scenarios and how often people use Full Framework vs .NET Core when developing and deploying ML.NET applications.

[codemzs]

closing this issue as there is no plan to enable AVX support in the short-term. In the short term our goals are around ONNX, AutoML and DNN training scenarios.