perf(kernels): speed up linear-attention prefill ~2.6x (Q/K LDS tiling + dk-split)

[donghanbj]

Summary

Optimizes the gated-delta chunked linear-attention prefill kernel
(linear_attention_prefill_chunked) on gfx1151, which was latency / LDS-bound.
Two combined, orthogonal changes:

• Q/K LDS tiling + bank-conflict padding — cache the per-chunk K and Q tiles
  in LDS with a +1-float row pad (qk_stride = dk + 1). This removes the 32-way
  LDS bank conflicts (dk=128 is an exact multiple of the 32 banks) and drops the
  repeated global query reads in the P-matrix and output phases.
• dk-split (HIPDNN_EP_LA_PREFILL_DK_SPLIT, default 4) — g_split lanes
  cooperate on each output column and split the dk reduction of the two dv-wide
  phases (B and output), combining via a warp shuffle (block = g_split*dv). This
  fixes the ~50%-idle-thread underutilization (256 threads vs dv=128 columns) and
  multiplies resident wavefronts to hide the latency-bound state reads. Falls back
  to the per-column path (g_split=1) when a shape can't satisfy the split.

The decode path and everything outside the prefill kernel are unchanged.

Results (gfx1151 / Qwen3.5-35B-A3B, seq=1842)

	main	this PR
kernel (standalone, best)	25.3 ms	9.7 ms (2.6x)
TTFT (end-to-end, median)	~3608 ms	~2947 ms (-18%)
prefill throughput	~511 tok/s	~625 tok/s (+22%)
decode	~17.2 ms/tok	~17.5 ms/tok (unchanged)

Kernel occupancy 16 waves/SIMD, no scratch spill. Output is numerically identical
in quality: cosine 0.9999996 vs the per-token decode oracle.

Test plan

• Standalone correctness vs the per-token decode oracle (cosine 0.9999996),
  verified for all g_split in {1, 2, 4, 8}.
• 1_build.bat: LIT 354/356 pass, plugin-registrar + output-allocator unit
  tests pass.
• VLM end-to-end (Qwen3.5-35B-A3B), 3x back-to-back vs main in the same
  thermal window: TTFT / prefill improved as above, generated text correct.
• CI green.

Note: the hand-written GPU kernels under lib/Runtime/Kernels/** are
formatter-exempt per .lintrunner.toml; this PR keeps the existing 4-space
hand style and leaves the decode kernel byte-for-byte unchanged.

[github-actions]

Thanks for opening a PR!

This project follows LLVM's incremental-development and AI-tool-use
guidance. See CONTRIBUTING.md
for the project workflow.

Before requesting review, please check that:

1. The change is focused. Substantial work links the relevant issue
   or design discussion.
2. The PR documents relevant test results and updates affected
   documentation.
3. If AI tools provided substantial assistance, the description
   explains what was assisted and how it was validated, and commit
   trailers identify the tool. The contributor has reviewed and
   understands the result.

Reviewers are assigned through
CODEOWNERS where ownership
is configured.

[github-actions]

L2 Accuracy Results (EP vs CPU)

Model	Combined L2	Total Elems	Skipped NaN/Inf
conv_test_hybrid	4.8668E-07	64	0
GroupQueryAttention_seq256	25.2366	2621440	0
MatMulNBits_o_seq128	259.906	368640	0
QMoE_seq128	34.9552	368640	0

Threshold: 0.01 | Run: 3440 - Commit: a5b51e3

[github-actions]

MorphiZen EP Performance Results

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.51	6.09	366	3	1245
GroupQueryAttention_seq128	4318.11	1.72982	11	6	310
matmul_down_seq128	517.46	2.33	75	3	352

EPContext Export Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.55	45.68	360	3	15590

EPContext Import Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.54	9.61	364	3	15760

OGA Benchmark Results

Model	Warmup	Reps	Prompt Len	Gen Tokens	TTFT (ms)	TPS	Peak Mem (GB)	GPU Mem (GB)
gpt-oss-20b-webgpu-int4-rtn-block-32	1	5	128	128	177.0	80.8	1.33	13.57
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	1	5	128	128	321.7	40.6	1.22	6.43

OGA Wheel Smoke (Python benchmark_e2e.py)

Model	TTFT (ms)	TPS
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	196	39.5

Run: 3440 - Commit: a5b51e3

[donghanbj]

/run-test models=all

[amd-yanjunz]

ROCm OGA CI finished with result: UNSTABLE.

• PR: perf(kernels): speed up linear-attention prefill ~2.6x (Q/K LDS tiling + dk-split) #600
• Source comment: models=all
• TEST_MODE: performance,accuracy
• MODELZOO: LLM_ROCm_dynamic
• TEST_MODELS: (all)
• FORBID_MODELS: google-gemma-4-26B-A4B-it-FP16W4-qmoe,gemma-4-12B-it-fp16-int4-kquant,NVIDIA-Nemotron-3-Super-120B-A12B-BF16-onnx-fp16-int4-rtn-onnx
• NODE: (OGA job default)
• CI_OPTIONS: LLM_RUN_REPETITIONS=5 LLM_WARMUP_ITERATIONS=1 LLM_RUN_EPS=hip OGA_MODEL_BENCHMARK_MAX_LENGTH=0 LLM_TOKENS_MITRIX=128|128,2048|128
• Jenkins build: http://xsjvitisaijenkins:8080/job/daily_hip_ep_oga/1601/
• NTS report: http://xsjnts.xilinx.com/ROCm/hipdnn-detail?database=development&collection=daily_hip_ep_oga&id=6a6ba2fbba0322ef060fad9e