Where are the C++ kernel implementation of the Winograd convolution algorithm?

[Redtorm]

I would like to know how MIOpen implements the Winograd convolution algorithm, but I only found the convolution kernel implemented in assembly language in the project source code, which is difficult to read and understand. I guess these assembly kernels should be converted from C++ kernels, but I can't find where the related C++ kernel implementations are, could you tell me where these C++ kernel implementation files are?

How can I add my own kernel to MIOpen? Is there a manual for this?

[junliume]

@Kirpich30000 could you comment on the question?

[Kirpich30000]

There is no C++ implementation - kernel were specifically designed for high efficiency and implemented from scratch in assembly.
You could look at this paper for Winograd introduction. And here is a reference code for constructing matrices if you want to try customized Winograd transformation.

[Kirpich30000]

To add your own kernel you will need to add a solver and register it in solver.cpp. Peek any solver and look at the code to see how it is implemented. Some solvers use workspace, others - don't. Some solvers invokes only a single kernel, while the other could invoke multiple kernels to perform the operation.

[atamazov]

@junliume I recommend removing request_for_comments (this is question, RFC is more like a spec-in-progress) and, maybe, closing this.

[DrizztDoUrden]

Also to register the solver for the specific convolution (or not conv) calls it supports like these: https://github.com/ROCmSoftwarePlatform/MIOpen/blob/d19c2575ca918f6b63a022f004d591795afcd85d/src/mlo_dir_conv.cpp#L110

[atamazov]

Intro info about solvers: #866