[gfx1250] Add A8W8/A8W4/A4W4 compute-bound GEMM

[aoli26]

Motivation

Add 256x256 tile specific compute-bound A8W4, A4W4 and A8W8 kernel for gfx1250.

Technical Details

• gemm_{a8w8,a8w4,a4w4}_256x256_gfx1250.py (new): hand-scheduled for exactly one profile — 256x256x128 tile, 2x2 waves. Hand-placed s_wait_dscnt/READY fences, odd/even wave-parity loop clones, operand-major planar LDS with B in its own 64 KiB segment.
• tdm_ops.py: add update_tensor_descriptor_2d_lds_addr (re-point one persistent descriptor per ring slot); widen lds_memref to accept !fly.memref. Both additive.
• Also folds the a8w8 mx32/mx128/ptpc variants into one launcher via is_mxscale / block_size constexprs.

Test Plan

pytest tests/kernels/test_gemm_fp8fp4_gfx1250.py on gfx1250.

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

Adds new hand-scheduled, compute-bound GEMM kernels for gfx1250 (A8W8/A8W4/A4W4 at 256-wide tiles) and extends TDM descriptor utilities to support persistent descriptor reuse and !fly.memref LDS views, with expanded correctness/behavioral tests (including split-K reduction and guard-row OOB detection).

Changes:

• Introduces new gfx1250 compute-bound GEMM launchers for 256x256-family tiles (A8W8/A8W4/A4W4), plus a split-K reduction epilogue kernel.
• Extends tdm_ops.make_tensor_descriptor_2d to accept Fly shared views and adds update_tensor_descriptor_2d_lds_addr for retargeting persistent descriptors.
• Refactors/expands the gfx1250 GEMM test harness to cover new modes, profile constraints, split-K, cluster constraints, determinism, and OOB store clamping.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/kernels/test_gemm_fp8fp4_gfx1250.py	Refactors mode/case construction; adds coverage for new 256-wide kernels, split-K reduce, determinism, and OOB store clamping.
python/flydsl/expr/rocdl/tdm_ops.py	Adds LDS-base extraction for !fly.memref and a helper to update the LDS address field in a 2D descriptor.
kernels/gemm/gemm_a8w8_splitk_reduce_gfx1250.py	New split-K partial-reduction epilogue kernel and cached compiler entrypoint.
kernels/gemm/gemm_a8w8_gfx1250.py	Unifies mx32/mx128/ptpc launcher knobs; adds split-K support and kernel-name sanitization.
kernels/gemm/gemm_a8w8_256x256_gfx1250.py	New hand-scheduled compute-bound A8W8 256-wide kernel with strict tuned-profile assertions.
kernels/gemm/gemm_a8w4_mxscale_gfx1250.py	Adds split-K support and updates wave-size handling.
kernels/gemm/gemm_a8w4_256x256_gfx1250.py	New hand-scheduled compute-bound A8W4 256x256 kernel (fixed tuned profile).
kernels/gemm/gemm_a4w4_256x256_gfx1250.py	New hand-scheduled compute-bound A4W4 256x256 kernel (fixed tuned profile).
kernels/common/kernels_common.py	Adds format_kernel_name() to sanitize kernel symbol names for assembly/linking.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.