The image loop on a 16 Mpx detector is two thirds of the run and both cards are busy for essentially
all of it, so card time removed is wall time removed. Of the six milliseconds a frame costs, two and
a half were spent decompressing it - and not because the card was short of bandwidth. The LZ4 pass
moved 53 GB/s where the strong-pixel flagger, reading the same image and the same bin table, gets
276. It is latency, not bandwidth: the copy loop moves 32 bytes per warp iteration with a syncwarp
after each one, and for a match copy the source and the destination both derive from the same
pointer, so nothing pipelines. The warp spends its time waiting for global memory, one dependent
round trip at a time.
So decode where the waiting is cheap. One CUDA block now owns one bitshuffle block: its first warp
decodes the payload into shared memory, and the whole block then un-transposes and preprocesses out
of shared and writes finished pixels. A shared round trip is tens of cycles rather than hundreds,
and the 72 MB shuffled intermediate never reaches DRAM at all - the pair of kernels moved about 238
MB a frame and the fused one moves 93.
The parser is lifted into a device function that both kernels call over the same bytes, so the
standalone path and the fused one cannot decode a chunk differently. The statistics reduction had to
change with it: 48 bytes of static shared on top of a full bitshuffle block costs a whole resident
block per multiprocessor, so the counts now reduce through a warp shuffle and one integer atomic per
warp. Blocks larger than 16 kB keep the two-kernel path, and the beam stop's own decoder is
untouched.
What this costs is decoder parallelism: a block that holds 16 kB of shared is one of four resident
per multiprocessor on this card, where the old kernel fitted thirty-two warps each decoding on its
own. The trade is favourable here and should be better on the production cards, which have half
again as much shared memory per multiprocessor. Measured on a 16 Mpx rotation set at the production
GPU count, with the indexing work of the next commit: 37.2 s -> 32.9 s, and the merged output is
byte-identical.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011n8riB6X59oRjkrSHzNPAU
