composable_kernel

mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-24 00:45:34 +00:00

Files

Andriy Roshchenko b8440b3aeb [rocm-libraries] ROCm/rocm-libraries#8325 (commit 559eaf6)

[GFX1250][MX GEMM] Unified FLATMM GroupedGemm Implementation
 for MX Data Types (#8325)

## Motivation

Design and test a unified FLATMM GroupedGemm interface so that it
supports all MX FP8, FP6, and FP4 data types on both the gfx950 and
gfx1250 architectures and works seamlessly across these platforms.

## Technical Details

Implementation exposes Grouped Gemm interface for MX FLATMM and MX TDM
FLATMM pipelines.

## Test Plan

Add the following tests:
 - ck_tile/grouped_gemm_mx/test_grouped_gemm_mx_flatmm_non_tdm.cpp
 - ck_tile/grouped_gemm_mx/test_grouped_gemm_mx_flatmm_tdm.cpp
 - ck_tile/flatmm/test_mx_flatmm_persistent.cpp

Verify on the gfx950 and gfx1250 architectures.

## Test Result

All tests pass. Verified on A0 hardware with rocm-7.14.0a20260517

## Submission Checklist

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

2026-06-15 16:12:33 +00:00

CMakeLists.txt

[rocm-libraries] ROCm/rocm-libraries#8325 (commit 559eaf6)

2026-06-15 16:12:33 +00:00

mx_flatmm_arch_traits.hpp

[rocm-libraries] ROCm/rocm-libraries#7359 (commit dd62f9f)

2026-05-29 17:02:45 +00:00

mx_flatmm_instance.cpp.in

[rocm-libraries] ROCm/rocm-libraries#7359 (commit dd62f9f)

2026-05-29 17:02:45 +00:00

mx_flatmm_instance.hpp

[rocm-libraries] ROCm/rocm-libraries#7359 (commit dd62f9f)