Higher-order SSB: group alias-edge work on CUDA

Workload and change

The target is a new scientific phase image every 16.7 ms at 2× output sampling, with full 512 × 512 × 192 × 192 data, fifth-order aberrations, sample tilt and depth retained. This is parameter interaction after calibration, not refitting.

Fine-grid paired correction has special native FFT alias columns that require independent mirror reads. Previously those lanes were scattered through the ordinary interior. They split warps inside every bright-field iteration. The new index mapping groups the special columns after the ordinary interior. It changes thread ownership only: all output pixels, independent Nyquist corrections, arithmetic, BF accumulation order and precision are retained. There is no new cache or workspace and no per-size kernel implementation.

The existing guard restricts this path to higher-order fine-grid reconstruction from a 512-square real FFT. Native output, other scan layouts, fitting, Newton refinement, loss diagnostics and MPS are unchanged.

Matched reconstruction measurements

RTX PRO 6000 Blackwell Max-Q Workstation Edition. Reference and candidate alternate in one prepared session over twelve rounds, excluding two warmups. The parameter sequence changes C10, C12, astigmatism angle, C30, a higher-order azimuth, and sample tilt/depth. Synchronization bounds reconstruction; host copy and browser rendering are excluded. The acquisition retains all 8,807 active BF pixels and its saved joint fit.

Output sampling	Reference median	Candidate median	Time reduction
1×	18.21 ms	18.20 ms	unchanged
2×	41.98 ms	37.68 ms	10.2%
3×	46.57 ms	42.36 ms	9.0%
4×	51.11 ms	47.20 ms	7.6%

The installed run_ssb_preview runner separately measured complete warm phase reconstruction plus host delivery at 16.72, 37.68, 41.93 and 46.34 ms for 1×, 2×, 3× and 4×. These sequential measurements are not the causal speedup comparison and are not notebook frame rates.

A second full acquisition, with 9,275 active BF pixels and different sampling, matched exactly at 1×, 2×, 3×, 4× and 8× over five rotation/depth settings. Its initial matched 2× measurements were 143.10 versus 126.76 ms. Another GPU was active during that interval; later quieter measurements were substantially faster for both variants. Do not interpret those different runtime intervals as an additional software gain or rank acquisitions by data dimensions alone.

Live notebook interaction

The CUDA notebook was executed in a fresh kernel and its real widget was driven in the in-app browser. Apple Metal WebGPU rendered phase and FFT side by side. Both counters measured newly painted scientific data frames, using the elapsed time between the first and last recorded paint. The target sweep rate was 60; the measured rates below are not the control setting or animation-only FPS. The saved joint fit retained twelve nonzero higher-order magnitudes, tilt and depth. The source scan and detector dimensions remained unchanged.

Sampling and sweep	Reference phase FPS	Candidate phase FPS	Candidate FFT FPS
1× C10	not measured	50.48	50.50
2× C10	22.31	24.87	24.86
2× C30	not measured	24.36	24.35

The 2× C10 rate improved approximately 11.5%. Each row is one live observation interval, not a long-duration reliability result. The previous kernel and the candidate were selected within the same prepared session for the 2× comparison. No parameters were re-fitted, no saved calibration was overwritten, and the saved fit was restored after the sweeps.

The 60 FPS target is not met. At 2×, roughly 38 ms of reconstruction already exceeds the 16.7 ms frame budget. Browser delivery adds overhead, but cannot explain away this backend limit. This is a measured limit of the current code and workload, not a theoretical hardware bound. MPS interaction was not changed or re-measured in this experiment.

Verification

All 48 matched first-acquisition outputs and all 25 second-acquisition joint parameter cases are bit-identical to the frozen reference. Existing CUDA wave-average and thick-sample tests passed: 31 tests, including independent unpaired references, anisotropic detectors, and alias/Nyquist edges. Five public-repository checks passed.

Aggressive register caps were rejected: the 48- and 56-register variants changed the phase by up to 6.61e-6 radians. Smaller shared-memory tiles did not improve the measured workload. A 64-register cap was exact in its probe but added little benefit on top of the layout change and is not enabled.

The frozen reference kernel file has SHA-256 026f16815ffe7e8312a42437863e939fcfd45546b469019e210591fce9d97822. Raw evidence and before/candidate sources are retained privately under 20261023-ssb-alias-layout. No package publication is claimed.

The candidate kernel file has SHA-256 3197bf5fd16049d8876d412bbae2d5375aa9eca4019b6dd3da64054e3fe64bb5.