[webgpu] support arbitrary input_channel size for im2col

[wenqinI]

Description

This PR supports vec1/vec2 for arbitrary input_channel in im2col kernel, which could bring performance gain to more models.

Like for yolov8n_pose model, there is about ~7% gain for whole model, and ~50% for those conv2d op which input_size are not multiple of 4.

Motivation and Context

[wenqinI]

@guschmue @fs-eire @qjia7 PTAL, thanks!

[wenqinI]

@daijh, PTAL, thanks!

[daijh]

Many comments in shader does not reflect the vec_size changing. Please take a look.

[wenqinI]

@daijh , this PR is updated, PTAL, thanks!

[daijh]

LGTM, please invite more reviewers to check.

[wenqinI]

Hi @qjia7 ,

This PR is ready for review, PTLA, thanks!

[Copilot]

Pull request overview

Extends the WebGPU Conv im2col+matmul path to support non-multiple-of-4 input channel counts by selecting an appropriate vectorization width (vec4/vec2/vec1), enabling broader use of the optimized kernel.

Changes:

• Add vec_size template parameter and logic to support vec1/vec2 loads in im2col_matmul.wgsl.template.
• Select vec_size at runtime based on channel_input and plumb it through shader generation/caching in im2col_matmul.{h,cc}.
• Add two new Conv2D unit tests covering cases with asymmetric padding and C=4 / C=6.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
onnxruntime/test/providers/cpu/nn/conv_op_test.cc	Adds Conv2D test cases that include a non-multiple-of-4 input channel scenario (C=6).
onnxruntime/core/providers/webgpu/nn/im2col_matmul.wgsl.template	Generalizes im2col indexing/tiling and accumulation to support vec_size 1/2/4.
onnxruntime/core/providers/webgpu/nn/im2col_matmul.h	Extends program interface to carry vec_size.
onnxruntime/core/providers/webgpu/nn/im2col_matmul.cc	Computes vec_size, updates program inputs/cache hint, and removes the prior %4 restriction.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.

[daijh]

Description
This PR supports vec1 for arbitrary input_channel in im2col kernel, which could bring performance gain to more models.

Like for yolov8n_pose model, there is about ~7% gain for whole model, and ~50% for those conv2d op which input_size are not multiple of 4.

Motivation and Context

As a last-minute change, the commit message needs to be updated to reflect both vec1 and vec2.

[wenqinI]

Hi @qjia7 and @guschmue,

I tried to rebase this PR, I guess the CI error may be fixed, please help me rerun the CI, PTAL, thanks!

[wenqinI]

Hi @qjia7 and @guschmue,

It seems the nnapi ep will fail in the new test cases (Conv2D_4 and Conv2D_5), I suspect there should be some bugs for this EP, so I create an issue and skip these test in this PR, PTAL, thanks!

[wenqinI]

Hi @qjia7 and @guschmue,

A soft ping...

The CI failed for nnapi is skipped, and all the CI test is passed now, could us push this PR forward? Thanks!

[guschmue]

/azp run Linux QNN CI Pipeline,Win_TRT_Minimal_CUDA_Test_CI,Windows ARM64 QNN CI Pipeline,Windows GPU Doc Gen CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 4 pipeline(s).

[daijh]

@guschmue @fs-eire
Please help to take a look.

[wenqinI]

@guschmue,

A soft ping... PTAL, thank you!