Navi3x test_verify_gemm failure: LLVM ERROR: Cannot select: t152: v2f16,ch = BUFFER_ATOMIC_FADD about amdmigraphx

[ahsan-ca]

Reproduced the error locally.

Did the following:

1. Tried building the test with latest MIGraphX (2b4d328), and used latest rocMLIR changes (ec52b22fa2b68277803190979d39ae7aa1e873d5) and the test case passed successfully.

2. Tried building the test with latest MIGraphX (2b4d328), and used rocMLIR commit that MIGraphX currently uses (5261f9fb8cc5e0109cf59bbf4cb11cc6fb527209) and the test case passed successfully.

3. Tried mainline build 14236 (6.2 http://rocm-ci.amd.com/job/compute-rocm-dkms-no-npi-hipclang/14236/) and test case fails.

4. Now trying mainline build 14273 (6.3 http://rocm-ci.amd.com/job/compute-rocm-dkms-no-npi-hipclang/14273/).

[ahsan-ca]

Passes with mainline build: http://rocm-ci.amd.com/job/compute-rocm-dkms-no-npi-hipclang/14273/

./bin/test_verify "test_gemm_add<migraphx::shape::half_type>"
[   RUN    ] test_gemm_add<migraphx::shape::half_type>
Skipped 86 configs for gpu::mlir_op
[ COMPLETE ] test_gemm_add<migraphx::shape::half_type> (20644.6ms)
[==========] 1 tests ran

[ahsan-ca]

This passes in 6.3. We need to backport the following commit to make it 6.2 pass:

commit a3593f6 (HEAD)
Author: Umang Yadav [email protected]
Date: Sat Jun 8 09:39:43 2024 -0400

update rocMLIR SHA to pull in bug fix for invalid block size on Navi3x (#3159)

This is already a part of https://github.com/ROCm/AMDMIGraphX/pull/3204/commits for backport.