fix(CK): Forward port convolutions update from standalone repo

[geyyer]

Motivation

Forward-port grouped-conv-fwd perf-regression fix (im2col GEMM-size 2GB check) to rocm-libraries mainline

Technical Details

Port this commit ROCm/composable_kernel@248f155#diff-d921922979d6cfc604769c3bc231107ae574375875ecce04bcb1065d43a5ce2e

Introduced by a mainline change for grouped-conv large-tensor / global-load-store CK enablement, which added a LargeTensors template flag to GridwiseGemmMultiD_xdl_cshuffle_v3 and reworked its CheckValidity, but did not exempt the convolution im2col-view path from the pre-existing logical-GEMM-size 2GB check

Test Plan

Regular CI

Test Result

CI should pass

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

JIRA ID : AICK-1818

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	⚠️ Warning	Error: Source/code files changed without an accompanying unit test. Expected: add at least one test file named like test_<name>.py / test_<name>.cpp (or <name>_test.*). Current: code file(s) changed: projects/composablekernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp, projects/composablekernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp; no test file found
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

🎉 All checks passed! This PR is ready for review.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.

[Copilot]

Pull request overview

Forward-ports a composable-kernel fix to avoid incorrectly rejecting grouped-convolution implicit-GEMM (im2col-view) problems due to the logical GEMM matrix size exceeding the existing 2 GiB check, even when the actual underlying tensor footprint fits within int32 addressing.

Changes:

• Add a SkipGemmSizeCheck template flag to GridwiseGemmMultiD_xdl_cshuffle_v3 and gate the existing logical GEMM-size validity check on it.
• Update the grouped-convolution forward device op to instantiate the gridwise GEMM with SkipGemmSizeCheck=true for the convolution/implicit-GEMM path (relying on existing descriptor-span checks instead).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
projects/composablekernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp	Adds SkipGemmSizeCheck and uses it to bypass the logical 2 GiB GEMM-size check when appropriate.
projects/composablekernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp	Enables skipping the logical GEMM-size check for convolution implicit-GEMM instantiations (while keeping descriptor-size validation).

Suppressed comments (2)

projects/composablekernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp:616

• If you rename the macro to a uniquely-prefixed name, the instantiation here should be updated to use the new identifier to avoid relying on a generic global macro name.

        GridwiseGemmV3TemplateParams,

projects/composablekernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp:627

• When using a uniquely-prefixed macro name, make sure the corresponding #undef matches the renamed identifier; otherwise the generic macro name remains a collision risk in translation units that include this header.

#undef GridwiseGemmV3TemplateParams

[AmosLewis]

Have you check the warning, better to add the test.

[geyyer]

Have you check the warning, better to add the test.

@AmosLewis, the added functionality is already covered by existing tests.