SIMD OCC PLAN

Goal:

Accelerate FM rank/Occ layer.

Motivation:

Correctness and contracts are stabilized.

Next leverage point:

Occ()

Current:

scalar implementation

Future:

scalar ↓

AVX2 ↓

AVX512

Benchmark isolation requirements:

measure:

• Occ only
• rank only

exclude:

• mmap startup
• HTTP
• IPC
• backend startup
• shard orchestration

Benchmark outputs:

p50 p95 p99

Machine profile mandatory:

• CPU
• compiler
• SIMD capability
• corpus size

Success metric:

lower rank latency

Future dispatch layer:

capability probe ↓

runtime dispatch ↓

scalar / avx2 / avx512

Principle:

benchmark = contract

performance follows correctness