Module gather_bench 

Gather throughput microbenchmark dispatch.

Provides two kernels for measuring KV cache read throughput:

• gather_bench_nibble — Simulates TurboQuant SDPA: unpack 4-bit nibble indices then gather from a 16-entry centroid table.
• gather_bench_f16_seq — Baseline: sequential F16 read + widen to F32.

The throughput ratio between the two kernels determines whether nibble-gather meets the ADR-007 gate of ≥ 50% of sequential F16 throughput.

Statics

GATHER_BENCH_SHADER_SOURCE

MSL source for the gather benchmark kernels (embedded at compile time).

Functions

dispatch_gather_f16_seq

Dispatch the sequential F16 read kernel.

dispatch_gather_nibble

Dispatch the nibble-gather kernel.

register

Register both gather benchmark kernels with the given registry.