MPS coherent SSB dispatch, 2026-10-18

Historical: describes the code as of 2026-10-18; the phase-variance fit trajectory it checked is retired. See the current fit methods.

The coherent Auto reconstruction now uses a 16 × 8 Metal threadgroup instead of 32 × 8. This changes scheduling only. The optical calculation, BF contribution order, float32 precision and phase estimator are unchanged. The native fitting and loss paths are unchanged.

Measured result

Two measured acquisitions retained their full 512 × 512 × 192 × 192 inputs, all calibrated BF pixels, fitted tilt and depth, and automatic output sampling. Median resident reconstruction time in five interleaved C10 trials changed from approximately 93 to 88 ms for a 1536 × 1536 output, and from 152 to 146 ms for a 2048 × 2048 output on an Apple M5 Max.

In the consuming browser application, median C10 response changed from 122 to 119 ms and from 191 to 187 ms respectively. These measurements include request scheduling, reconstruction, transfer and the application’s frame presentation marker. They are not physical scanout measurements or a claim of sustained 10 FPS. Timing varies with other GPU activity.

Validation

• Complete phase arrays matched frozen same-device results exactly for 122 application calls, covering ordinary and tilt/depth-corrected reconstruction.

• Existing MPS reconstruction and infrastructure checks: 77 passed; two optional simulation-fixture tests were skipped and are not counted as passes.

• The frozen GPU fit trajectory, repeat determinism and native saved-result workflow passed without updating reference pins or tolerances.

• Hardware browser checks exercised C10 and tilt with Auto sampling and full BF support. The remaining cost is primarily optical correction.

CUDA launch variants did not show a consistent improvement under a shared GPU workload; CUDA kernels remain unchanged. Private commands, timings, arrays and application evidence are retained as ssb-explorer-100ms-20261018.