[ASM][flydsl] opt mla kernel for dspr1(gfx1250)

[xiangM99]

Motivation

Update the gfx1250 persistent PS1 (page size 1) FP8 MLA decode

Technical Details

• Continuous KV ring across work items: a work item's last producer steps gather the next
  item's first KV tiles, so its prologue no longer waits on HBM in series with the previous
  item's epilogue.
• Padded output staging: staged output rows carry 16 B of padding so the 16 heads of one
  ds_store do not share LDS banks; the TDM store's column extent drops the pad columns.
• i32 KV tile loop counter (assignment-carried instead of range(init=)), and the
  per-wave page indices of a tile loaded with one 8-dword scalar load.

Test Result

gfx1250, 256 CUs, B=1536, nhead=128, q_seq=1, FP8 Q/KV (--scales poc). Times are per decode
call (stage 1 + reduce).

ctx	old PS1 (µs)	new PS1 (µs)	new vs old	ASM PS64 (µs)	old vs ASM	new vs ASM
1024	180.8	159.6	−11.7%	213.2	1.18x	1.34x
2048	310.3	289.5	−6.7%	320.3	1.03x	1.11x
3072	443.3	418.7	−5.6%	459.8	1.04x	1.10x
4096	575.7	550.6	−4.4%	617.0	1.07x	1.12x
5120	708.1	682.8	−3.6%	686.4	0.97x	1.005x

Submission Checklist

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

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:gfx1250-ffm-triton	Run the five-shard gfx1250 FFM Triton test suite
ci:triton-300x	Run an additional Triton test job on MI300X in PRs (added automatically when gfx942 configs change); main branch always runs both MI35X and MI300X
ci:triton-355	Run the full Triton test suite on MI35X, not only the tests the change affects
multigpu	Aiter multi-GPU tests on the 8-GPU runner
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 6087 --add-label <label>

One backend per PR:
A PR changes one kernel backend: [Triton/Gluon] (Triton and Gluon count as one), [HIP], [ASM], [CK], [OPUS] or [FlyDSL]. If the title ends up with two backend tags, split the PR -- as stacked pull requests when one part cannot merge without the other.

PR title tags & labels:
Component tags ([Triton/Gluon], [HIP], [CK], [ASM], ...) are added to the PR title and as PR labels automatically from the changed files and re-synced on every push — change-type tags like [fix]/[Perf], op tags like [MLA], and human labels (ci:*) are left untouched. Add the no-auto-title label to stop the title rewrites; labels stay in sync either way.