[codegen][gpu] Adding support to generic op and flexible layout to pad_to_intrinsics on convolution

[jerryyin]

The pad_to_intrinsics pass only support linalg.conv2d op with nhwc_hwcf layout of convolution. This has created inconvenience around taking advantage of other convolution variants for their performance potentials. Once such scenario is the IR from conv_filter_to_channels_last will populate conv2d_nhwc_fhwc represented by linalg.generic.

This PR extend support of the pad_to_intrinsics pass such that other convolution variants including:

• Those that are represented with linalg.generic
• Other layouts such as (filter layout of) fhwc fchw

This PR will unblock #19974, and allow us to continue to use pad_to_intrinsics as igemm padding kernel catch up in performance.

[nirvedhmeshram]

LGTM!

[nirvedhmeshram]

Interesting regression : https://github.com/iree-org/iree/actions/runs/13503269777/job/37727576624?pr=20073#step:9:209

I guess we could land this and #19974 together, but I am not able to come up with a hypothesis on why this by itself is so bad, any thoughts?

EDIT : Just thought of a hypothesis, if its padding more convs now and thats causing a slow down, I dont think merging with #19974 will help either.

[nirvedhmeshram]

Actually I see this on main, so might not be anything in this PR,
https://github.com/iree-org/iree/actions/runs/13501782679/job/37724211700

[jerryyin]

Yes, I agree with you the culprit must be in that I've made this path too flexible and now it can handle any type of convolution (versus in the past it only deal with linalg.conv2d hwcf variant). I'll try to reproduce locally and see what's going on.

Actually I see this on main, so might not be anything in this PR

Oh wow, thanks for point that out. Let me take a second look at main's CI record too.

[jerryyin]

Per discord discussion, the perf degradation is caused by mi300 switching to cpx mode and irrelevant with this PR.

I'll leave this PR open till tomorrow to merge it in case there's other feedbacks.

[nirvedhmeshram]

@jerryyin Since this is an optional (and experimental pass) it is okay the way it is but one thing to consider is if we should have these two cases where we dont do the padding
https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Dialect/LinalgExt/Utils/Utils.cpp#L432-L442

[jerryyin]

@nirvedhmeshram Sounds good, will do.

On a second look, those scenarios has already been blocked by below. Since I preserved this conditional, I don't have to do anything here.

iree/compiler/src/iree/compiler/Preprocessing/Common/PadToIntrinsics.cpp

Lines 192 to 198 in e87dd2e

	if (convolutionDims->outputChannel.size() != 1 ||
	convolutionDims->inputChannel.size() != 1 ||
	convolutionDims->filterLoop.size() < 1 ||
	convolutionDims->outputImage.size() < 1 ||
	convolutionDims->depth.size() != 0) {
	return;
	}

I was imprecise when mentioned that this PR will allow any type of convolution.