DirectML v1.4 ARM64 version?

[zorderjohn]

Would it be possible to have an ARM64 version of DirectML? I downloaded the NuGet package but it only includes x86/x64 versions.

My goal is to execute a PyTorch exported ONNX model (yolov5) using the GPU. On x64 it runs fine with DirectML v1.4 but not with the older one included in Windows. That's why I would need to use the newer NuGet version and redistribute it for my ARM64 application (Hololens 2).

Thanks.

[wchao1115]

Thanks @zorderjohn for your feedback. We have more work to do on the ARM platform and isn't ready to commit to it at this time. You will hear from us in this GitHub repo when we have more to share.

[zorderjohn]

Thanks @wchao1115 for your answer, I'm glad you are working on it.
Is there any rough estimation for this to happen? This issue Is a bit blocking for the company I work for and depending on the time frame I would either wait or reconsider an alternative.

[wchao1115]

@zorderjohn DirectML 1.5.1 now supports ARM64 build flavor. Please note that the end-to-end performance when used on an ARM-based GPU still depends on the level of support in the driver on the target hardware which may not be the most up-to-date right now. We are working with the GPU vendors to deliver the maximum support both functionally and on the performance front in their upcoming driver update, so please stay tuned. If you run into an issue with the current in-market driver, please do let us know anyway. Chances are we're already working with the GPU vendor on the fix, but it's great to hear back from the users to confirm.

[emaschino]

Can DirectML run on the HoloLens 2 device? I had a hard time building ONNX Runtime for ARM64 UWP with DirectML EP enabled (microsoft/onnxruntime#19495) and I'm getting catastrophic performances vs running on the CPU only (4x times slower). This is with DirectML 1.10.1 (from NuGet package brought by ONNX Runtime 1.14.1) if it does matter.