perf: Q4_0/Q8_0 dequant throughput 5x below memcpy ceiling — missing SIMD vectorization

[noahgift]

Benchmark data (aprender-bench-compute)

Operation	Size	Throughput	vs memcpy
Q8_0 dequant	262K	1.22 Gelem/s (4.88 GB/s)	5.5x below
Q4_0 dequant	262K	1.25 Gelem/s (5.00 GB/s)	5.4x below
Q8_0 quantize	262K	302 Melem/s (1.21 GB/s)	22x below
Q4_0 quantize	262K	284 Melem/s (1.14 GB/s)	24x below
memcpy (f32 clone)	262K	25 GiB/s (26.8 GB/s)	baseline

Key observations

1. Q4_0 and Q8_0 dequant are nearly identical speed (~1.2 Gelem/s). Q4_0 reads half the data, so it should be ~2x faster if memory-bandwidth-limited. Equal speed means compute-limited on the unpacking logic.

2. Quantize is 4x slower than dequant. Quantize needs to find block scales (reduction), but 4x overhead is high.

3. Dequant at 4.88 GB/s output vs 26.8 GB/s memcpy — 5.5x gap. With AVX2 SIMD unpacking, Q8_0 dequant should approach memcpy speed (just multiply int8 × scale). Q4_0 needs nibble extraction but should still be 3-4x faster with SIMD.

Impact

For fused dequant+matvec in realizar, dequant overhead is a significant fraction. A 4096×11008 Q4_0 matrix = 22.5M Q4 values → ~18ms dequant time at current throughput. With SIMD: ~3.3ms.

Suggested fix

• AVX2/AVX-512 vectorized Q8_0 dequant: _mm256_cvtepi8_epi32 + _mm256_mul_ps
• AVX2 Q4_0 dequant: nibble extraction via shift+mask, then cvt + mul
• Process 32 elements per iteration (one Q4_0/Q8_0 block = 32 elements)

Reproduce

cargo bench -p aprender-bench-compute --bench quantization

[noahgift]

Required Provable Contract

Fix must include contracts/aprender/dequantize-kernel-v1.yaml:

metadata:
  version: "1.0.0"
  description: "Dequantization throughput contract (Q4_0, Q8_0)"
  references:
    - "contracts/aprender/quantized-dot-product-v1.yaml (format registry)"

kernel_contract:
  q4_0:
    equation: "x_i = scale * (nibble_i - 8)"
    block_size: 32
    superblock_bytes: 18
    
  q8_0:
    equation: "x_i = scale * q_i  (signed i8)"
    block_size: 32
    superblock_bytes: 34
    
  performance_bounds:
    # memcpy ceiling on this machine: ~25 GiB/s
    # Target: dequant should approach memcpy (memory-bound)
    q8_0_min_throughput: "5 GiB/s"   # Currently 1.22 GiB/s
    q4_0_min_throughput: "5 GiB/s"   # Currently 1.23 GiB/s
    
falsification_tests:
  - id: FALSIFY-DEQUANT-001
    rule: "Q8_0 dequant matches scalar reference"
    prediction: "bit-exact for all valid Q8_0 blocks"
    test: "proptest with random scale and i8 values"
    if_fails: "SIMD dequant introduced precision error"
    
  - id: FALSIFY-DEQUANT-002
    rule: "Q4_0 dequant matches scalar reference"
    prediction: "bit-exact for all valid Q4_0 blocks"
    test: "proptest with random scale and nibble values"
    if_fails: "Nibble extraction bug (hi vs lo nibble swap)"
    
  - id: FALSIFY-DEQUANT-003
    rule: "Dequant throughput gate"
    prediction: "Q8_0 dequant >= 5 GiB/s for 262144 elements"
    test: "Compare criterion results from bench_dequant"
    if_fails: "Not SIMD-vectorized — check for autovectorization blockers"
    
  - id: FALSIFY-DEQUANT-004
    rule: "Roundtrip fidelity"
    prediction: "max |dequant(quant(x)) - x| <= max_quantization_error for format"
    test: "Random f32 inputs through quantize → dequantize pipeline"
    if_fails: "Quantization or dequantization has a systematic bias"