Exact group support in CUDA SSB

Exact group support in CUDA SSB#

Question and change#

Can fine-grid reconstruction avoid optical calculations whose contribution is exactly zero? Each BF group now caches its coordinate bounds on the GPU. If neither shifted soft aperture reaches that box, the entire group contributes zero at that frequency. An outward float32 margin protects the aperture edge. All BF pixels, nonzero corrections, summation order and precision are retained. The native-grid and fitting paths keep their existing kernels.

Measured result#

Two full 512 × 512 × 192 × 192 acquisitions used Auto sampling, all calibrated BF pixels, and their fitted tilt and depth. On an RTX PRO 6000 Blackwell Max-Q, ten changed-parameter application calls after the first call gave:

Output

C10 reconstruction median, before → after

Full response median, before → after

After response p95 / max

1536²

56.3 → 48.8 ms

63.6 → 57.1 ms

57.9 / 57.9 ms

2048²

97.8 → 84.0 ms

119.6 → 110.7 ms

116.2 / 116.7 ms

Tilt updates showed the same reconstruction benefit. These are resident reconstruction and local HTTP response timings, not loading, fitting or displayed FPS. Other GPU activity can increase latency.

Nsight Compute measured 88.9% compute throughput and 14.6% DRAM throughput in a representative pre-change correction launch, with no register spilling. The principal remaining cost is optical arithmetic. This measurement does not establish a theoretical minimum or 60 FPS.

Validation and rejected alternatives#

  • All 132 complete application phase arrays were bit-identical to frozen pre-change outputs, covering native/Auto and ordinary/tilt-corrected modes.

  • 57 CUDA tests passed, including unpruned GPU controls after rotation changes, higher-order optics, native diagnostic loss and engine cleanup.

  • Frozen fit and native Metal workflow gates passed without updating pins.

  • Larger CUDA BF groups saved reduction traffic but slowed the optical step.

  • A new Metal support-mask cache gave little joint-reconstruction benefit relative to its added memory. The existing Metal kernel was retained.

Private timing distributions, profiler data and application checks are archived under experiment ssb-kernel-limits-20261018.