MPS SSB correction and reduction fusion — 17 October 2026

Experiment: ssb-parameter-reuse-20261017. Acquisition identities, full-array comparisons, profiler traces and application evidence remain private.

Change

The coherent upsampling kernel now corrects and accumulates 64 existing bright-field groups per launch. Each frequency reuses its coordinates and tilt shift across groups. Partial group spectra stay in registers instead of being written to memory and read by a second reduction kernel.

The BF group boundaries, within-group additions and sequential group additions are unchanged. So are aperture support, Hermitian handling, complex64 source spectra, float32 output, the phase-of-mean estimator and native loss evaluation. Cancellation checkpoints remain between bounded launches. No BF pixels are removed, and output sampling is unchanged.

Measurements

Two full 512 × 512 scans with 192 × 192 detectors were measured on an Apple M5 Max with 128 GiB memory. Each row contains 20 alternating reference/candidate pairs in one process, after preparation. Every call changed parameters. These are complete prepared phase evaluations, including the NumPy output handoff; they exclude loading, fitting, transport and browser drawing.

BF pixels / output	Change	Reference median	Candidate median
8,807 / 1536²	C10, joint model	123.3 ms	121.8 ms
8,807 / 1536²	Tilt, joint model	124.0 ms	123.7 ms
8,807 / 1536²	C10, ordinary control	110.8 ms	108.8 ms
8,930 / 2048²	C10, joint model	299.7 ms	246.9 ms
8,930 / 2048²	Tilt, joint model	294.8 ms	239.4 ms
8,930 / 2048²	C10, ordinary control	271.4 ms	211.9 ms

The smaller output has no meaningful speed gain. The larger output improves by 18–22% in this paired run. Tail latency remains variable: its C10 p95/max were 478/891 ms before and 393/460 ms after. These measurements do not establish a hardware limit or a 60 FPS interaction rate.

Temporary peak MLX allocation falls from 545 to 187 MB for the smaller output and from 555 to 332 MB for the larger output. The latter previously wrote about 17.1 GB of logical partial-spectrum data per redraw; fusion writes about 0.27 GB of accumulated spectra. Those are array-size calculations, not DRAM counter readings, and exclude source reads and subsequent FFT buffers.

Controlled non-scientific ablations on the larger case gave median times of 276 ms for the reference, 256 ms without source-spectrum reads, and 149 ms without optical transfer calculations. These variants diagnose cost only; their results are not reconstructions. Optical arithmetic and intermediate traffic both matter. Removing source reads alone did not remove most of the cost. The Metal trace did not provide hardware ALU or DRAM counters.

Rejected alternatives and validation

• A separable geometry cache added about 497 MB, ran slower and changed rounding. It is not used.

• Full-array addressing with offsets was slower than the existing sliced inputs in a paired test. The existing addressing is retained.

• All 120 real full-image comparisons were bit-identical to the frozen GPU implementation, across ordinary and joint C10/tilt changes.

• The MPS upsampling suite passed 43 tests, including incomplete BF groups across multiple launches, higher-order aberrations, cancellation, anisotropic sampling and saved-result behavior.

• Native reconstruction, the 200-trial native fit and saved-run round trip passed. The separate external frozen fit-trajectory gate could not run because its acquisition fixture was unavailable. No pin was changed.

The change targets reconstruction redraw. It does not change the fitting objective or promise faster optimization, loading, CUDA or browser rendering.