CUDA SSB compact support and boundary dispatch

The fine-grid correction still dominated interactive reconstruction after the closed-sideband optimization. This pass retains the full 512 × 512 × 192 × 192 input, all 8,807 bright-field pixels, 14 aberration terms, sample tilt and depth. It changes scratch storage and launch ownership, not the correction formulas, BF order, precision, fit objective or output sampling.

Implementation

A GPU-generated frequency map retains every frequency that can intersect the full soft-aperture support. Its radius is rounded outward; the correction still uses its original support predicate. The map is cached by grid and calibration, not fitted aberrations. It stores the ordinary frequency pairs first and the independent native-alias/Nyquist entries last.

Two specializations of the same correction source process those two ranges. Ordinary pairs no longer carry the boundary branch through every BF iteration. The existing compensated sums and complex128 group reductions operate on the compact scratch plane. Results are scattered into a full complex128 spectrum before the unchanged full-size inverse FFT. No interpolation, binning, detector subsampling or reduced-resolution interaction is introduced.

At 4×, reduction scratch storage decreased from 1,024 MiB to 264.95 MiB. This is one buffer, not total resident acquisition memory. At 2× it decreased from 256 MiB to 236.52 MiB. The coordinate map adds a small cached index allocation. The BF group count still follows the previous full-plane workspace budget so that reduction grouping and rounding do not change.

Matched warm reconstruction

Physical device: RTX PRO 6000 Blackwell Max-Q Workstation Edition, 300 W. A fresh process alternated frozen reference and candidate, with the same saved joint model and changing defocus/higher-order coefficients. Each entry uses 18 synchronized warm observations; it includes the native output copy but excludes host transfer, notebook delivery and browser rendering.

Sampling	Before median / p90, ms	After median / p90, ms
1×, unchanged dispatch	18.98 / 19.02	18.99 / 19.03
2×	31.50 / 31.59	29.59 / 29.64
3×	35.46 / 35.54	30.55 / 30.66
4×	39.51 / 39.55	31.02 / 31.05

Hardware counters

Nsight Compute 2025.4.1 used its default clock control, near 1 GHz. It captured the first dense launch before and the first compact interior launch after; the latter is followed by a small boundary launch. These counters identify the main-kernel change, not total application latency. End-to-end reconstruction above includes both launches and the full inverse FFT.

Metric	2× before / after	4× before / after
Profiled launch duration, ms	11.87 / 10.18	13.11 / 10.46
Compute throughput, %	75.03 / 83.25	73.10 / 82.82
DRAM throughput, %	24.34 / 27.85	33.98 / 27.72
Registers per thread	69 / 70	69 / 70
Achieved occupancy, %	47.76 / 47.52	47.47 / 47.15
Local spill requests	0 / 0	0 / 0

The remaining correction is still primarily arithmetic work, not saturated DRAM traffic. These counters do not establish a theoretical hardware ceiling.

Actual notebook image updates

The existing notebook ran the saved full-field higher-order model. The C10 sweep requested 60 updates/s, with phase and FFT visible. Rates below count new scientific canvas paints, not requested updates or browser animation callbacks. The browser used an Apple Metal WebGPU adapter; reconstruction ran on the remote CUDA device above. Intervals were 17–21 seconds after startup.

Sampling	Before phase / FFT, fps	After phase / FFT, fps
2×	28.93 / 28.93	30.56 / 30.54
3×	Not repeated in this pass	27.92 / 27.93
4×, first pass	21.03 / 21.03	22.28 / 22.24
4×, passive callback timers	-	24.41 / 24.40
4×, final uninstrumented repeat	-	21.91 / 21.92

The 4× variation is retained rather than reporting only the fastest interval. Across 522 warm instrumented 4× requests, preview including host transfer was 29.46 ms median / 29.98 ms p90; the full Python widget callback was 31.52 / 32.41 ms. Delivery and browser updates contribute to the remaining interval. The phase and FFT retain 2048² pixels at 4×; the existing 3× FFT uses 2048² padding for the 1536² phase image. No FFT quality change was made.

Direct C10 keyboard changes, FFT histogram adjustment, sampling switches and restoring the saved joint fit were verified. No browser warnings or errors were observed. Temporary timing hooks were removed, and the notebook was left paused at 4× with the saved fit restored. The calibration file was not saved over during testing.

Validation and limits

Eighty full-acquisition comparisons across 1×, 2×, 3×, 4× and 8×, varied rotation, defocus, astigmatism, higher-order terms, tilt and depth matched the frozen GPU reference bit-for-bit. The native fit and loss path are unchanged. Dedicated CUDA tests also cover anisotropic sampling, alias/Nyquist preservation, compact workspace reuse, output ownership and changing output dimensions. The final CUDA reconstruction, joint-fit regression, preview-ownership and repository privacy run passed all 41 tests in 145.94 seconds.

Rejected experiments include cooperative sideband lanes (exact but slower), staged polynomial evaluation (changed rounding without a useful speed gain), and register caps/normalization shortcuts (small or inconsistent gains). The private prototype needed an argument-list correction before its acceptance run; only the corrected, independently tested production path was used in the browser.

60 fps remains unmet. These results do not qualify another CUDA architecture, MPS, an installed wheel, or a release. No package publication is part of this pass.

Experiment: 20261024-ssb-cuda-single-sideband. Raw counter reports, frozen sources, rejected variants, parity records and notebook frame counts are in the private evidence archive. Source checkout base: e8750f3e3652, with pre-existing uncommitted work preserved.

Kernel source SHA-256 before: 32225632c94df818da972fe8718b6467d6a9ae9cd5534cd312e10d1215df3677. After: d38d2eb354011403790e000dd4dca9c680da2cf11e55c312de77f248292295aa.