The de-novo indexer projects every spot onto each of 16384 search directions and
bins the projections; the peak of each direction's spectrum is a reciprocal
lattice row spacing. One thread owned a whole direction, so neighbouring lanes
wrote 12.6 kB apart and every warp instruction touched 32 separate sectors of a
206 MB buffer with no chance of staying in a 4 MB L2. 160 million scattered
global read-modify-writes, at about 14% of the card's bandwidth.
One block per direction now, with the bins in shared memory. They are counts, so
they are held as integers: an integer atomicAdd is a real shared-memory
instruction where the float one compiles to a compare-and-swap retry loop, and a
count below 2^24 converts to float exactly, so the output is bit for bit what the
repeated += 1.0 produced. Above 48 kB of bins the old kernel still runs.
245.75 ms per launch -> 4.08 ms, so 0.98 s of the run -> 0.016 s. This machine
runs two of them at once on two cards, so it is worth about half a second here
and about a second on the single-GPU machines the viewer and the broker run on.
The FFT it feeds takes 3.3 ms; preparing its input took 70x longer than
transforming it.
Alongside it, the per-frame scale fit divided by k^2 once per observation per
IRLS iteration, and a loop-invariant divisor does not get hoisted out of a double
division - ptxas emits the whole Newton refinement of the reciprocal every time.
Hoisted, as 1/sigma already is a few lines above; the same expression in the
three CPU scale paths went with it so the two stay algebraically identical.
54.92 ms per launch -> 44.56 ms, 1.65 s -> 1.34 s.
That one is not bit-identical - a multiply by a rounded reciprocal differs from a
correctly rounded quotient in the last place - so it can move a frame that sits
on the convergence tolerance. Battery: 21/24 space groups, no failures, and 17 of
24 crystals identical to the previous run, against a floor of 13 of 24 for the
same binary run twice.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
