Ck tile batched contraction kernel generelizing

[msaffari-amd]

Proposed changes

Extends ck-tile batched contraction kernel to support arbitrary multi-dimensional non-contiguous tensor layouts using descriptors.
Extends example to cover testing this new added feature, user can pass any manual stride for input and outputs.

Checklist

Please put an x into the boxes that apply. You can also fill these out after creating the PR. If you're not sure, please don't hesitate to ask.

• I have added tests relevant to the introduced functionality, and the unit tests are passing locally
• I have added the test to REGRESSION_TESTS list defined at the top of CMakeLists.txt in tests/CMakeLists.txt, IF the test takes more than 30 seconds to run.
• I have added inline documentation which enables the maintainers with understanding the motivation
• I have removed the stale documentation which is no longer relevant after this pull request
• (If this change is user-facing) I have added release notes which provide the end users with a brief summary of the improvement from this pull request
• I have run clang-format on all changed files
• Any dependent changes have been merged

Discussion

If this is a relatively large or complex change, feel free to start a discussion by explaining why you chose the solution you did and what alternatives you considered

[Copilot]

Pull Request Overview

This PR adds comprehensive support for arbitrary multi-dimensional stride patterns and non-contiguous tensor layouts to the batched contraction kernel. Previously, the kernel only supported contiguous row-major layouts with hardcoded strides.

• Introduces TensorDescriptorUtils with vectorization support to create stride-aware tensor descriptors
• Implements custom RunGemm method using descriptor-based tensor views instead of relying solely on UniversalGemmKernel
• Updates reference implementation to use stride-aware indexing for validation
• Adds command-line support for custom stride specifications in examples

Reviewed Changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tensor_descriptor_utils.hpp	Adds vector size template parameters and updates descriptor creation to support vectorized memory access
batched_contraction_kernel.hpp	Implements descriptor-based architecture with custom RunGemm, adds tensor descriptor storage to kernel args
reference_batched_contraction.hpp	Refactors reference computation to use stride-aware offset calculation, changes from std::vector to std::array for D tensors
run_batched_contraction_example.inc	Adds custom stride parsing, implements runtime dispatch for NumDTensor, creates tensors with non-contiguous layouts
contraction_utils.hpp	Updates argument parsing to support stride specifications and adds help documentation
batched_contraction.cpp	Updates dimension case handling (appears to have duplicate case)

---

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

[Copilot]

Duplicate case HANDLE_CASE(2, 1, 1, 1) at lines 219 and 222. The second occurrence at line 222 appears to replace a removed case for (1, 1, 1, 2), which may be intentional removal but the duplicate is incorrect. Remove line 222 or replace it with the intended dimension combination.

Suggested change

HANDLE_CASE(2, 1, 1, 1);

[Copilot]

Loop condition i >= 0 with unsigned type ck_tile::index_t will always be true, causing infinite loop when i underflows. This pattern appears in multiple offset computation lambdas (lines 109, 117, 125, 141, 149, 157, 173, 181, 189, 208, 216, 224). Change loop to use int i or rewrite to avoid decrementing below zero.