[gfx1250][gemm] Optimize FP8 GEMM deep pipeline and add scale loading experiment

[aoli26]

Motivation

Improve gfx1250 MXFP8 GEMM performance and robustness by adding a dedicated FP8 deep-pipeline schedule, tuning LDS/TDM scheduling behavior, and add extra scale-load buff-load path on the main optimized FP8(experiment, default off). It also fixes a split-K accuracy bug.

Technical Details

This PR adds an FP8 deep-pipeline compute schedule for the 256x256x128 gfx1250 MXFP8 configuration, including tuned wait placement, panel scheduling, instruction prefetch behavior, TDM handling, mcast early timeout and cluster fence/signal overlap. It also introduces a segmented LDS layout for the optimized FP8 path(avoid LDS segment conflicts), adds direct buffer-load-to-VGPR scale loading modes, removes older experimental scale staging paths, enables the TDM multicast early-timeout descriptor bit for loads, and improves benchmark timing/graph verification coverage for gfx1250 GEMM.
It also fixes split-K accuracy by switching the cross-workgroup atomic accumulation to device-scoped llvm.atomicrmw fadd.

Test Plan

pytest tests/kernels/test_gemm_fp8fp4_gfx1250.py

Adds test_mxfp8_gemm_splitk covering split_k ∈ {2, 4, 6, 8} for f32 and bf16 outputs.

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

This PR targets gfx1250 MXFP8 GEMM performance/robustness by adding a dedicated FP8 deep-pipeline schedule, introducing a segmented LDS layout for the tuned 256×256×128 configuration, and adding an experimental scale loading path that can bypass TDM/LDS (buffer_load→VGPR). It also updates benchmark/verification utilities and tweaks TDM descriptor behavior.

Changes:

• Add an FP8 deep-pipeline compute schedule (auto-selected when eligible) and a reference segmented LDS layout for the optimized gfx1250 FP8 256×256×128 kernel.
• Introduce experimental scale load modes (vgpr, vgpr_ab_split) that load E8M0 scales directly to VGPRs (plus test/CLI support and new preshuffle layout).
• Improve benchmark timing utilities (single-event fast path, hipGraph capture/replay sanity checks) and enable TDM multicast early-timeout for loads.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
kernels/gemm_fp8fp4_gfx1250.py	Adds FP8 deep-pipeline schedule, segmented LDS layout, VGPR scale-load experiment, and related fencing/scheduling changes.
python/flydsl/expr/rocdl/tdm_ops.py	Enables the TDM descriptor early_timeout bit for load descriptors (kept off for stores).
python/flydsl/expr/rocdl/inline_asm.py	Modifies gfx1250 instruction prefetch wrapper implementation/defaults.
tests/kernels/test_gemm_fp8fp4_gfx1250.py	Adds coalesced scale preshuffle for VGPR scale path, updates tests/bench CLI, adds hipGraph verification and new fill modes.
tests/kernels/benchmark_common.py	Adds a fast timing path using a single event pair when no flush/prep is requested.

---

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