[gfx1250] Refactor MoE/GEMM kernels with layout/atom API

[aoli26]

Motivation

Replace legacy gfx1250 MOE/GEMM implementations with focused, layout-based kernels.

Technical Details

• Add A8W4 MXScale, A8W8 PTPC, and A8W8 blockscale kernels.
• Add A8W4 MXScale MoE kernel.
• Support TDM/WMMA pipelining, cluster multicast, ragged M, and strided tensors.
• Remove legacy implementations and simplify correctness tests.

Test Plan

Run:

python3 -m pytest tests/kernels/test_moe_a8w4_mxscale_gfx1250.py
python3 -m pytest tests/kernels/test_gemm_fp8fp4_gfx1250.py

Test Result

All tests passed.

Submission Checklist

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

[Copilot]

Pull request overview

Refactors gfx1250 MoE/GEMM coverage by removing legacy WMMA/TDM implementations and introducing new layout/atom-based kernels (including MXScale/PTPC/blockscale variants), alongside updated correctness/perf tests for the new paths.

Changes:

• Removed legacy gfx1250 WMMA GEMM and WMMA MoE 2-stage kernel implementations plus their large test harnesses.
• Added new gfx1250 kernels: grouped MoE A8W4 MXScale, GEMM A8W4 MXScale, GEMM A8W8 PTPC, and GEMM A8W8 blockscale.
• Simplified/reshaped shared gfx1250 GEMM utilities to support the new kernels’ raw-LDS + fused activation needs.

Reviewed changes

Copilot reviewed 14 out of 16 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/kernels/test_wmma_gemm_gfx1250.py	Removed legacy WMMA GEMM correctness harness.
tests/kernels/test_moe_gemm_wmma_gfx1250.py	Removed legacy WMMA MoE 2-stage test/bench harness.
tests/kernels/test_moe_a8w4_mxscale_gfx1250.py	Added correctness/perf coverage for new grouped A8W4 MXScale MoE path.
kernels/moe/moe_gemm_2stage_wmma_gfx1250.py	Removed legacy WMMA MoE 2-stage kernel implementation.
kernels/moe/moe_a8w4_mxscale_gfx1250.py	Added new grouped contiguous-M MXScale MoE kernel with TDM pipeline and optional cluster multicast.
kernels/gemm/wmma_gemm_gfx1250.py	Removed legacy WMMA GEMM kernel implementation.
kernels/gemm/gemm_common_gfx1250.py	Updated shared gfx1250 GEMM utilities (raw LDS ops + fused silu/swiglu helper).
kernels/gemm/gemm_a8w8_ptpc_gfx1250.py	Added PTPC FP8xFP8 preshuffle GEMM kernel.
kernels/gemm/gemm_a8w8_blockscale_gfx1250.py	Added FP8xFP8 blockscale preshuffle GEMM kernel.
kernels/gemm/gemm_a8w4_mxscale_gfx1250.py	Added A8W4 MXScale GEMM kernel.
kernels/common/mma/pipeline_utils_gfx1250.py	Removed legacy gfx1250 pipeline utilities previously used by WMMA GEMM.

---

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