The ring sigma is the cancelling difference sum2/n - m^2, and both sums were
float accumulated by atomics whose order is arbitrary. Two costs: the
cancellation left only ~4 digits in the variance, and the ordering moved the
resulting threshold by ~0.05 counts between runs - enough to flip a pixel
sitting on the hard "value >= threshold" test, and with it a connected
component's size. So the GPU engine did not reproduce the CPU one and did not
reproduce itself.
Only the accumulators that span blocks are widened. The per-block staging stays
float, because a block contributes a few dozen similar-magnitude pixels to a
ring and there is nothing to lose there - that also keeps the shared-memory
footprint of the hot loop, and hence its occupancy, exactly as it was: measured
on a 4.5 MP frame, 0.960 vs 0.966 ms/frame (40.9x over the CPU path, unchanged).
finalize_rings now does the cancellation in double and rounds to float last,
which is what AdaptiveSpotFinderCPU::AccumulateRings does.
The device properties are also read from the current device rather than device
0; callers round-robin engines across GPUs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
leonarski_fandClaude Opus 5
04450eb618