The per-ring sums were floats reduced by atomics, so the ring sigma - and with it the
detection threshold - depended on the order the blocks happened to arrive in. Detection
compares an INTEGER pixel value against that threshold, so a threshold that drifts
across an integer flips every pixel of that value in the ring at once, which is how a
last-bit difference turned into a different spot list.
A preprocessed pixel is an exact int32 and the masked and saturated sentinels are
skipped, so v and v*v are exact in 64 bits, and integer addition is associative: the
sums no longer care about arrival order. Both engines now accumulate the same way, so
they agree exactly rather than approximately, and the GPU spot list is bit-identical
across runs. The corrected sums that feed the reported azimuthal profile stay float -
a pixel value times a float correction has no exact integer form - but they do not
enter the detection decision.
Cost: the ring reduction needs 28 bytes per bin instead of 20 in the plain pass, which
drops it from eight co-resident blocks per SM to seven and costs about 11% of that
kernel (0.582 -> 0.650 ms/frame on a 4.5 Mpx frame). End to end it does not show:
alternating runs on three rotation crystals came out the same or slightly faster, and
the battery is unchanged in every number. The CPU engine got 30% faster (32.2 -> 22.6
ms/frame), integers being cheaper than doubles.
Tests: exact CPU/GPU agreement on the spot list, and 50 repeats of bit-identical output
where there were four.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
leonarski_fandClaude Opus 5
abb94ca450