perf(q6_k): SIMD-fy dequantQ6_KBlock via ByteVector ql + qh extraction

[michalharakal]

Summary

Replaces the scalar 32-iteration inner loop in dequantQ6_KBlock — the dequant step under matmulQ6_KVec — with a fused ByteVector pipeline. Per floatStep-wide chunk of l:

1. Load three byte slices: ql[qlBase + l], ql[qlBase + l + 32], qh[qhBase + l]
2. Assemble four 6-bit codes per lane:
   • q1 = (ql0 & 0x0F) | ((qh & 0x03) << 4) − 32
   • q2 = (ql32 & 0x0F) | ((qh >>> 2 & 0x03) << 4) − 32
   • q3 = (ql0 >>> 4) | ((qh >>> 4 & 0x03) << 4) − 32
   • q4 = (ql32 >>> 4) | (qh >>> 6 << 4) − 32
3. Widen to FloatVector, multiply by per-sub-block d·scale, store to four 32-element regions of the scratch FloatArray.

Inline replacement — matmulQ6_KVec's outer structure (scratch + SIMD dot product) is unchanged.

Why this matters

Q6_K is the dominant format for the embedding, lm_head, and FFN matrices of Gemma 4 E2B Q4_K_M. The scalar dequant was the hot path inside the per-cell, per-block loop under matmulQ6_KVec. Closing it brings Q6_K up to the SIMD pipeline standard set by Q4_K (#562/#563) and Q8_0 (already SIMD).

Scope

• Inline replacement in JvmQuantizedVectorKernels.dequantQ6_KBlock. No new SPI surface (full Q6KMatmulKernel SPI is a fair follow-up if a native FFM provider needs to register here).
• matmulQ6_KVec outer loop / parallelism / scratch-allocation strategy unchanged.

Out of scope (next M5 follow-ups)

• Full fused Q6_K matmul (eliminate scratch FloatArray entirely) — more involved because sub-block scales differ across l ∈ 0..15 vs 16..31, requiring 16 parallel accumulators per output cell per block.
• Q6KMatmulKernel sibling SPI (similar shape to Q4KMatmulKernel from feat(kernel): SIMD-fused Q4_K matmul kernel + Q4KMatmulKernel SPI #562).
• Q4_0 SIMD — interleaved nibble layout makes a clean SIMD pipeline harder than Q4_K/Q6_K.
• Native (FFM) Q4_K / Q6_K kernel — priority 100, calls hand-tuned NEON/AVX2 via MemorySegment. Closes M5.

Test plan

• ./gradlew :skainet-backends:skainet-backend-cpu:jvmTest — 218/218 tests pass, including Q6KMatmulTest's parity vs the canonical ggml DequantOps.dequantQ6KFromBytes reference (covers single output row + multi-row × multi-block within 1e-4 relative tolerance).

🤖 Generated with Claude Code