Support DirectML for windows

[csukuangfj]

Note: It does not seem to improve the speed when passing --provider=directml from the commandline.

[thewh1teagle]

Note: It does not seem to improve the speed when passing --provider=directml from the commandline.

What's something heavy I can test this on? I wanted to test on whisper medium but this branch does not have it.
Can I use sherpa-onnx.exe binary to test something?

[csukuangfj]

I think we should resolve the conflict and merge this PR.

If I recall correctly, it has not been merged because it does not run faster with DirectlML.

[thewh1teagle]

it has not been merged because it does not run faster with DirectlML.

I would like to check it and compare
Do I need to link onnxruntime.lib when using directml / cuda etc? it doesn't find it

[thewh1teagle]

If I recall correctly, it has not been merged because it does not run faster with DirectlML.

I tried on another c++ project that uses DirectML. it used 95% from the TPU (amd ryzen 5 4500u)
But it wasn't fast. Maybe it does works slow.
The fastest way to use whisper I found is CTranslate2 and even on CPU only!
Do you have idea how can we get similar results with sherpa?

I mean, Nvidia already should work fast, CoreML works fast. the only missing part is wide support for Windows users with AMD / Intel hardware.
I thought that DirectML is the answer, maybe with onnxruntime it will be.

[thewh1teagle]

Any chance you could resolve the conflicts so I can test it with Whisper medium?

[csukuangfj]

Closing via #1153