fix(ck): [CK] LCOMPILER-2577: Remove erroneous __restrict__ qualifier

[michaelselehov]

Motivation

The gridwise GEMM kernels take a pointer to block-shared LDS memory. Every thread of the block writes into that memory, and other threads then read it after block_sync_lds(). The __restrict__ qualifier does not hold for such a pointer.

Clang lowers __restrict__ to LLVM noalias. LangRef states that noalias also prohibits modifications through other threads, see llvm/llvm-project#211507. The compiler is therefore free to keep an LDS value in a register across the barrier. The kernel then reads stale data and produces wrong results.

JIRA ID: LCOMPILER-2577

Technical Details

This is analogous to #10229 and #9629.

The failing test is TestGroupedConvndBwdData2d/9.Test2D. The kernel that fails the numerical check is DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<256, 64, 16, 16, 4, 4, Default, 16, 16, 1, 1, 4, 1, 1, 1>. It uses gridwise_gemm_multiple_d_xdl_cshuffle.hpp and gridwise_gemm_xdlops_v2r3.hpp. This patch removes __restrict__ from p_shared in both headers, 4 places in total.

I confirmed all sides on MI300X (gfx942):

compiler	source	result
current staging	__restrict__ present	FAILED, max err 130
current staging	__restrict__ removed	PASSED
baseline with the AA change reverted	__restrict__ present	PASSED

The baseline compiler contains a revert of llvm/llvm-project#196923 "Reapply [AA] No synchronization effects for never-escaping identified local". That commit tells alias analysis that a synchronizing operation cannot affect an object that never escapes the function. Together with the invalid __restrict__, this lets the compiler drop the values that other threads wrote into LDS.

Test Plan

Build and run, on MI300X:

test_grouped_convnd_bwd_data --gtest_filter='TestGroupedConvndBwdData2d/9.*'

Test Result

Before:

max err: 130, number of errors: 137150, 22.78% wrong values
[  PASSED  ] 0 tests.
[  FAILED  ] 1 test, listed below:
[  FAILED  ] TestGroupedConvndBwdData2d/9.Test2D, where TypeParam = std::tuple<float,ck::tensor_layout::convolution::NHWGK,ck::tensor_layout::convolution::GKYXC,ck::tensor_layout::convolution::NHWGC>

 1 FAILED TEST

With this patch:

[       OK ] TestGroupedConvndBwdData2d/9.Test2D (3223 ms)
[  PASSED  ] 1 test.

Submission Checklist

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

[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/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp, projects/composablekernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.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.

[ronlieb]

@shumway ok to land ?

[michaelselehov]

I stopped re-running the failed CI jobs. This error repeats over and over again:

subprocess.CalledProcessError: Command '['/opt/python/cp312-cp312/bin/python', '-m', 'pip', 'install', '--force-reinstall', '--pre', 'rocm[libraries,devel]==10.1.0.dev0+f5a34f5ee296a8c235d4a25694aec3befbc657fd']' returned non-zero exit status 1.
Error: Error: failed to run script step: command terminated with non-zero exit code: error executing command [sh -e /__w/_temp/e3bc6a40-97ca-11f1-8415-69b9f438896e.sh], exit code 1

[ronlieb]

@shumway can this be landed now ? we need a code owner to approve