xe2-sdp-hd256

name

xe2-sdp-hd256

description

Use this skill when writing, optimizing, benchmarking, or debugging Flash Attention SDP kernels with head dimension 256 (HD=256) targeting Intel Xe2 (Lunar Lake/LNL, Battlemage/BMG) GPU using SYCL ESIMD. Xe2 is the GPU architecture; LNL and BMG are product names. Covers the S^T (transposed scores) architecture, oneDNN-inspired v2 kernel design, GQA support, softmax optimization, lsc_slm_scatter S transpose elimination, ISA-level analysis, and the complete optimization journey from 64 to 88 TFLOPS. Use whenever the user mentions HD=256 SDP, head_dim=256 attention, rev256, onednn_v2 kernel, S transpose, s_scatter, s_gather, lsc_slm_scatter, lsc_slm_gather, or large head dimension flash attention on Intel GPU.

Xe2 (Lunar Lake/LNL, Battlemage/BMG) ESIMD Flash Attention — HD=256 (S^T Architecture)

Specialized knowledge for GQA flash attention with head_dim=256 on Intel Xe2 (Lunar Lake/LNL, Battlemage/BMG). Current best: 88 TFLOPS (64.9% of 135 TFLOPS peak) at 16Kx16K, 32Q/2KV heads. Reference files hold detailed history; this file holds critical rules and architecture.

---

Quick-Reference Rules (must follow every time)

HD=256 Architecture Differences (vs HD=128)

• S^T approach: QK GEMM produces scores in transposed layout [KV_PER_SG=16][Q_ROWS*Q_GRPS=32] per thread. This is because K is loaded in reversed/transposed format for DPAS efficiency.
• Register transpose required: Before VS GEMM, S must be transposed from [KV×Q] to [Q×KV] and written to SLM as fp16. This is the #1 overhead source (~20% of kernel time).
• Larger SLM footprint: Q_SLM=64KB + S_SLM=32KB + MAX_SLM=4KB + SUM_SLM=4KB = ~104KB (vs ~16-32KB for HD=128).
• 32 threads/WG (8 sg_i × 4 sg_j) — matches oneDNN decomposition. doubleGRF required.

Kernel Constants

HD = 256;              // Head dimension
WG_Q_ROWS = 128;       // Q rows processed per WG
KV_CHUNK = 128;         // KV rows per outer iteration (8 sg_i × 16 KV_PER_SG)
KV_PER_SG = 16;         // KV rows per subgroup per iteration
Q_ROWS = 8;             // Q rows per subgroup tile
Q_PAIRS = 2;            // Q pairs per sg_j (2 × 16Q = 32Q per sg_j)
Q_GRPS = 4;             // Q groups per subgroup (4 × 8 = 32 Q rows per SG)
D_BLKS_PER_SG = 2;      // D blocks per SG (2 × 16D = 32D per sg_i)

Thread Mapping

sg_i = tid & 7;         // 0-7: KV/D dimension
sg_j = tid >> 3;        // 0-3: Q dimension
// sg_i handles: KV rows [sg_i*16 : sg_i*16+16], D cols [sg_i*32 : sg_i*32+32] (VS phase)
// sg_j handles: Q rows [sg_j*32 : sg_j*32+32] (4 groups × 8 rows)

SLM Layout

0x00000 - 0x0FFFF:  Q_SLM    (64 KB) — Q tile [128 Q × 256 D] fp16
0x10000 - 0x17FFF:  S_SLM    (32 KB) — Scores [128 Q × 128 KV] fp16 (after transpose)
0x18000 - 0x18FFF:  MAX_SLM  (4 KB)  — Cross-SG max reduction
0x19000 - 0x19FFF:  SUM_SLM  (4 KB)  — Cross-SG sum reduction

Build Commands

# Full build (correctness + perf benchmark):
icpx test_rev256_onednn_v2.cpp -o test_rev256_onednn_v2.exe `
  -fsycl -fsycl-targets=spir64_gen `
  -Xs "-device bmg -options -doubleGRF" -O2 -w

# Mini build (perf only, faster compile):
icpx rev256_onednn_v2_83tflops_mad_fusion.cpp -o test.exe `
  -fsycl -fsycl-targets=spir64_gen `
  -Xs "-device bmg -options -doubleGRF" -O2 -w

# ISA dump build:
icpx test_rev256_onednn_v2.cpp -o test.exe `
  -fsycl -fsycl-targets=spir64_gen `
  -Xs "-device bmg -options -doubleGRF" -O2 -w -save-temps
# Then: ocloc disasm -file <spv> -device bmg -dump dump_v2_new

S Transpose: Solved via lsc_slm_scatter

The register transpose from [KV=16][Q=16] fp32 → [Q=8][KV=16] fp16 row-major used to generate ~270 mov (4|M0) instructions in ISA. Now eliminated using lsc_slm_scatter<uint32_t, 2, u32, 16>:

Working approach (Approach 2 — s_scatter, +1.7 TFLOPS):

1. Pack adjacent KV rows into u32 via shift+OR: packed = low_kv | (high_kv << 16)
2. lsc_slm_scatter<uint32_t, 2, u32, 16> writes 2 u32 per Q address → 4 KV values per call
3. 4 scatter calls per qp (kv groups 0..3, 4..7, 8..11, 12..15), 8 total per thread
4. SLM layout after scatter is Q-major KV-fast — same as baseline after register transpose
5. Block read unchanged (slm_block_load)

CRITICAL: Element-major data layout. lsc_slm_scatter with NElts > 1 expects element-major layout:

// CORRECT: data[0..15] = elem0 for all addrs, data[16..31] = elem1 for all addrs
data.select<16, 1>(0) = packed0;
data.select<16, 1>(16) = packed1;

// WRONG: address-major interleave (produces ~0.12 rel_rms systematic error)
data.select<16, 2>(0) = packed0;   // DO NOT USE
data.select<16, 2>(1) = packed1;   // DO NOT USE

Failed alternatives:

• Approach 1 (s_gather — block_store + gather read): 73 TFLOPS (-12.9) — SLM gathers too expensive
• SLM scatter NElts=1 (64 D16 messages): 69 TFLOPS — scatter bandwidth too low with single-element writes
• VNNI interleave (stride-2 uint32 pairs + 8×8 transpose): neutral (82.7 TFLOPS) — compiler decomposed stride-8 uint32 gather into even more mov(4|M0)

Performance Testing

• Correctness: 9 test configs (128×128 to 1K×1K, GQA + MHA, multiple data ranges)
• Benchmark sizes: 8K×8K, 8K×16K, 16K×16K (compute-bound regime)
• Warmup: 5 iterations, then 100 timed iterations
• FLOPS formula: 4 * q_len * kv_len * HD * num_heads_q + 2 * q_len * kv_len * num_heads_q
• Peak: 135 TFLOPS (BMG Xe2 XMX fp16)
• Correctness threshold: max_diff < 0.1, check NaN count first

Backup Convention

All backups in SDP/backup/ with naming: rev256_onednn_v2_<TFLOPS>tflops_<technique>.{h,cpp,exe}

---

Performance Ladder (current)

Version	TFLOPS	Roofline%	Technique	Useful?
vnni	64	47%	VNNI-based QK approach	NO — slower than reversed DPAS
xprefetch	68	50%	Extended prefetch hints	NO — marginal gain, added complexity
deferred_comp_70	70	52%	Deferred compensation, cross-phase prefetch	YES — foundation for all later work
nbarrier_71	71	53%	Named barriers instead of regular barriers	YES — small but free gain
best_72	72	53%	Optimal baseline configuration	YES — stable reference point
s_preload_72	72	53%	S-preload in VS phase	NEUTRAL — same perf, more code complexity
fp16_transpose_82	82	61%	Separate F32→FP16 from strided gather	YES — major breakthrough (+14%)
simd_inv_83	82.5	61%	SIMD16 reciprocal instead of scalar div	YES — small gain
mad_fusion_83	82.7	61%	fp32_sum mad fusion (mul+add → mad)	YES — incremental gain
pipelined_qk_84	84	62%	32-thread WG (8×4), pipelined K loads, split barrier B	YES — previous best
s_scatter_88	88	65%	Eliminate S register transpose via lsc_slm_scatter<u32,2>	YES — current best
s_gather (reverted)	73	54%	block_store + lsc_slm_gather<u32,4> read	NO — gather reads too expensive (-12.9 TFLOPS)
SLM scatter NElts=1 (reverted)	69	51%	Replace transpose with 64 SLM scatter D16 ops	NO — major regression, scatter BW too low
VNNI interleave (reverted)	82.7	61%

---

Key Optimization Insights

What Worked

1. lsc_slm_scatter S transpose elimination (+1.7 TFLOPS, 84→88): Replace ~270 register transpose mov instructions with lsc_slm_scatter<uint32_t, 2, u32, 16>. Pack adjacent KV rows into u32 via shift+OR, scatter directly to transposed SLM positions. 8 scatter calls per thread (2 qp × 4 kv_groups). Key discovery: NElts > 1 data must be in element-major layout.
2. Separate F32→FP16 conversion from strided gather (+14%): Convert entire ST_tile to fp16 first, then do strided gather on fp16 data. Halves the register traffic for the transpose.
3. mad instruction fusion: fp32_sum = fp32_sum * delta + local_sum generates SIMD32 mad instruction with 2× throughput on BMG XVE.
4. SIMD16 reciprocal: simd<float,16> inv = 1.0f / sum instead of scalar 1.0f / sum[i] in a loop.
5. Split barrier B (arrive/wait): Named barrier arrive after S SLM writes → V loads + K prefetch + compensation (all independent work) → named barrier wait before VS reads. The arrive/wait window spans V loads + K prefetch + deferred compensation = 100+ cycles of useful work between arrive and wait. See xe2-nbarrier-pipelining skill for the pattern.
6. Deferred compensation: Apply A_tile *= delta before VS GEMM, not interleaved inside VS loop — cleaner DPAS pipeline.
7. Cross-phase prefetch: Prefetch next K tile during VS phase, prefetch V during QK phase.
8. 32-thread WG + pipelined K loads (+1.3 TFLOPS): Expanded from 8×2=16 to 8×4=32 threads matching oneDNN decomposition. K loads pipelined: load K[d+1] after DPAS[d], hiding K load latency behind compute.

What Didn't Work

1. SLM gather S read (Approach 1): slm_block_store of un-transposed data + lsc_slm_gather<uint32_t, 4, u32, 16> to read transposed. 73 TFLOPS (-12.9 from baseline). SLM gather reads are much more expensive than block loads — 4 gathers per S tile vs 1 block load each, and each gather has higher message overhead.
2. SLM scatter NElts=1: 64 lsc_slm_scatter<u16, 1, u16, 8> messages (one per KV × qh). 69 TFLOPS. Per-element scatter bandwidth is severely limited.
3. VNNI interleave transpose: Pairing adjacent KV into uint32 VNNI pairs with stride-2 dest writes — Xe2 compiler decomposes select<8,8>() stride-8 uint32 gather into mov(4|M0), producing even more instructions.
4. Extended prefetch (xprefetch): Adding more prefetch hints beyond the basic cross-phase pattern gave marginal improvement.
5. S-preload optimization: Pre-loading S tiles from SLM before VS GEMM — no measurable improvement.
6. Early V loads: Moving V loads before S transpose to overlap V latency with transpose ALU. Increased register live range, hurt compiler scheduling → 82.7 vs 84.1 baseline.
7. Double-buffered S SLM (83.8 vs 84.1 TFLOPS — neutral): Cross-iteration pipeline with 2 S buffers (S_A + S_B, 32KB each). Pipeline: QK[k+1] → softmax[k+1] → S_store[k+1]→buf_new → compensation[k] → VS[k]→buf_old. Why it failed:
   • barrier_A (full barrier for softmax max reduction) already synchronizes ALL SLM writes including S — the named barrier for S sync was already free (never blocks)
   • Q SLM reduced from 16→15 d_blks (SLM budget: 60KB Q + 32KB S_A + 32KB S_B + 4KB shared = 128KB max) → 16th d_blk loaded from global (extra LSC sends)
   • Extra registers: delta_prev, local_sum_prev carried across iterations
   • Code bloat from prologue/epilogue (3-phase structure duplicates compensation+VS)
   • Key lesson: The split barrier B arrive/wait window (S writes → V loads + K prefetch + compensation → VS reads) already hides S synchronization latency. Double-buffering removes a barrier that was never blocking.
8. S_VNNI + transposed accumulator (abandoned at build): Swap DPAS operands so S_vnni is src1 and V is src2, eliminating the S register transpose entirely. Accumulator becomes [8D×16Q] instead of [8Q×16D], requiring one-time output transpose. Build constraints discovered:
   • lsc_load_2d fp16 transpose: max width=4 (need 2 loads per 8D tile → 8 loads/kv_blk vs 2)

Remaining Gap vs oneDNN (90+ TFLOPS)

1. DPAS density: oneDNN achieves 4 consecutive QK DPAS and 8 consecutive VS DPAS with {Atomic} tags. Ours has 2 consecutive DPAS with intervening sends/movs.
2. V from global: oneDNN loads V via 2D block load from global memory. Ours loads V cooperatively to SLM first.
3. S transpose is now solved: ~270 register transpose movs eliminated via lsc_slm_scatter. Gap narrowed from ~6 to ~3 TFLOPS vs oneDNN.

---

Architecture: oneDNN vs Ours

Aspect	oneDNN SDPA	Our rev256_onednn_v2
Threads/WG	32 (8×4)	32 (8×4)
DPAS/iter	144 (48 QK + 96 VS)	128 (64 QK + 64 VS)
QK DPAS density	4 consecutive	4 consecutive (pipelined K)
VS DPAS density	8 consecutive	8 consecutive
Q storage	SLM (shared by all SGs)	SLM (shared by all SGs)
K access	2D block load	2D block load (pipelined)
V access	2D block load from global	2D block load with VNNI
S storage	SLM (VNNI packed)	SLM (row-major fp16, via lsc_slm_scatter)
Softmax	Cross-SG atomic max	Cross-SG SLM max + barrier
S sync	Split barrier (arrive/wait)	Named barrier (arrive/wait)
Peak measured	~90+ TFLOPS	88 TFLOPS

See analysis/onednn_vs_ours_analysis.md and analysis/kernel_redesign_plan.md for full WG redesign plan.

---

Assets (ready to compile)

icpx <file>.cpp -o <file>.exe -fsycl -fsycl-targets=spir64_gen -Xs "-device bmg -options -doubleGRF" -O2 -w

Asset	Purpose	Perf
assets/rev256_onednn_v2_88tflops_s_scatter.h	Current best — lsc_slm_scatter S transpose elimination. 32-thread WG, pipelined K, split barrier B, all Phase 1-4 optimizations. Primary code reference.	88 TFLOPS (65%)
assets/rev256_onednn_v2_88tflops_s_scatter.cpp	Test harness for 88T — correctness (9 tests) + 3 perf benchmarks. Reproducible.	run directly
assets/rev256_onednn_v2_86tflops_qk_pipeline.h	Pre-scatter — pipelined K loads, 32-thread WG, split barrier B. Diff with 88T to see exactly what s_scatter changed.	86 TFLOPS (64%)
assets/rev256_onednn_v2_86tflops_qk_pipeline.cpp	Test harness for 86T — reproducible perf benchmark.	run directly
assets/rev256_onednn_v2_84tflops_pipelined_qk.h	Pre-pipeline — 32-thread WG with pipelined K loads. Diff with 86T to see K pipeline refinements.	84 TFLOPS (62%)
assets/rev256_onednn_v2_84tflops_pipelined_qk.cpp	Test harness for 84T — reproducible perf benchmark.	run directly
assets/rev256_onednn_v2_83tflops_mad_fusion.h	Pre-32-thread — 16-thread WG with mad fusion, simd_inv, fp16 transpose. Shows the Phase 2 architecture before WG expansion.	83 TFLOPS (61%)
assets/rev256_onednn_v2_83tflops_mad_fusion.cpp	Mini benchmark for 83T — perf-only test harness.	run directly
assets/rev256_onednn_v2_82tflops_fp16_transpose.h	fp16 transpose breakthrough — the single optimization that gave +14.7%. Compare with 83T to see simd_inv + mad fusion changes.	82 TFLOPS (61%)
assets/test_rev256_onednn_v2.cpp	Full test harness — 9 correctness configs (GQA+MHA, multiple ranges/seeds) + 3 perf benchmarks (8K-16K).	run directly
assets/flash.attn.b.mha256.fp16.opt.h	— original HD=256 opt kernel (49 TFLOPS). Non-S^T architecture for contrast.

Key code patterns to study (diff across versions):

• 88T vs 86T: lsc_slm_scatter S transpose elimination (the only diff — S write section)
• 86T vs 84T: Pipelined K loads refinement, split barrier B optimization
• 84T vs 83T: 16→32 thread WG expansion, K load pipelining
• 83T vs 82T: simd_inv (SIMD16 reciprocal) + mad fusion (fused multiply-add)
• 82T: The fp16 transpose breakthrough — separate F32→FP16 from strided gather

---

Workflow

1. Edit kernel in rev256_onednn_v2.h following rules above
2. Build with do_build_full.ps1 (correctness+perf) or do_build.ps1 (perf only)
3. Run exe — check 9/9 correctness PASS, then check TFLOPS at 8K×8K, 16K×16K
4. If improvement: backup to backup/rev256_onednn_v2_<TFLOPS>tflops_<technique>.{h,cpp,exe}
5. If regression: restore from latest backup in backup/
6. ISA analysis: build with -save-temps, disassemble with ocloc, compare mov/DPAS counts

---

Key Files

File	Purpose
rev256_onednn_v2_s_scatter.h	Current best HD=256 kernel (88 TFLOPS, s_scatter)
test_rev256_onednn_v2.cpp	Full test harness (correctness + perf benchmark)
rev256_onednn_v2_83tflops_mad_fusion.cpp	Mini benchmark (perf only, baseline comparison)
flash.attn.b.mha256.fp16.opt.h	HD=256 baseline kernel for perf comparison
do_build_full.ps1	PowerShell build script (full test)
do_build.ps1	PowerShell build script (mini/perf only)
do_isa_new.ps1	PowerShell ISA dump + disassemble script
analysis/onednn_vs_ours_analysis.md	Detailed architecture comparison with oneDNN
analysis/kernel_redesign_plan.md	WG redesign plan (8×4 = 32 threads)
analysis/oneDNN_16kx16k.csv	oneDNN ISA dump for reference

---

Backup Files

Backup	TFLOPS	Description
rev256_onednn_v2_vnni_64tflops.h	64	Initial VNNI approach
rev256_onednn_v2_xprefetch_68tflops.h	68	Extended prefetch experiment
rev256_onednn_v2_deferred_comp_70tflops.h	70	Deferred compensation baseline
rev256_onednn_v2_70tflops_best.h	70	Cross-phase prefetch optimized
rev256_onednn_v2_71tflops_nbarrier.h	71	Named barriers
rev256_onednn_v2_72tflops_best.h	72	Best pre-transpose-opt config
rev256_onednn_v2_72tflops_s_preload.h	72	S-preload experiment
rev256_onednn_v2_82tflops_fp16_transpose.h	82	F32→FP16 before transpose (breakthrough)
rev256_onednn_v2_83tflops_simd_inv.h	82.5	SIMD16 reciprocal
rev256_onednn_v2_83tflops_mad_fusion.h	82.7	mad fusion
rev256_onednn_v2_84tflops_pipelined_qk.h	84	32-thread WG, pipelined K loads, split barrier B
rev256_onednn_v2_88tflops_s_scatter.h	88	lsc_slm_scatter S transpose elimination (current best)
rev256_onednn_v2_double_buffer.h	83.8	Double-buffered S cross-iteration pipeline (reverted)

---

Related Skills

Skill	When to use
xe2-nbarrier-pipelining	Named barrier arrive/wait patterns for pipelining independent GPU operations
xe2-sdp-kernels	HD=128 flash attention — different architecture, higher roofline%
xe2-sdp-bf16	BF16 variants — bf16 ALU limits, V conversion interleaving
xe2-gtpin-profiling	GTPin setup and ISA analysis methodology
intel-gpu-vtune-profiling	VTune GPU Compute/Media Hotspots — capture and analyze HW metrics (XMX%, stalls, L3, occupancy)
intel-gpu-kernel-opt	General Xe2 optimization patterns
sycl-esimd-build	Compilation flags, doubleGRF, spill detection
xe2-dpas-patterns	DPAS tiling, VNNI packing, systolic pipeline

---

Reference Files

File	Contents
references/optimization-history.md	Full chronological journey from 64→88 TFLOPS with root cause analysis
references/vtune-analysis.md	VTune GPU hotspot analysis across all 13 versions — HW metrics, stall breakdown, phase-by-phase explanation
references/vtune-summary.csv	Raw VTune metrics CSV (13 rows × 30 columns) for programmatic analysis
assets/rev256_onednn_v2_88tflops_s_scatter.h	Current best kernel (88T) — lsc_slm_scatter S transpose elimination
assets/rev256_onednn_v2_88tflops_s_scatter.cpp	Test harness for 88T — reproducible
assets/rev256_onednn_v2_86tflops_qk_pipeline.h	Pre-scatter kernel (86T) — pipelined K, split barrier B
assets/rev256_onednn_v2_86tflops_qk_pipeline.cpp	Test harness for 86T — reproducible
assets/rev256_onednn_v2_84tflops_pipelined_qk.h	Pipelined QK kernel (84T) — 32-thread WG
assets/rev256_onednn_v2_84tflops_pipelined_qk.cpp	Test harness for 84T — reproducible
assets/rev256_onednn_v2_83tflops_mad_fusion.h	MAD fusion kernel (83T) — pre-32-thread
assets/rev256_onednn_v2_83tflops_mad_fusion.cpp	Mini benchmark for 83T
assets/rev256_onednn_v2_82tflops_fp16_transpose.h	fp16 transpose breakthrough (82T)
assets/test_rev256_onednn_v2.cpp	Full correctness + perf test harness
assets/flash.attn.b.mha256.fp16.opt.h	Baseline kernel (49T) for comparison