Profiling showed the per-space-group "group hkl" step (~0.30s/2-pass on lyso) is not
gemmi-bound (the ASU keying is ~6ms) but memory-bandwidth-bound: stamping the group id
onto, and reading it back from, the `group` field scattered across the 56-byte Obs
struct touches the whole ~350MB partials array twice per pass.
Precompute the per-obs AcceptReflection finiteness once (immutable) into a flat 1-byte
array, then stamp the ASU-group id from rawrun_group + that flat array into a flat
group_ids vector for the GPU, and build the group CSR (a stable counting sort, now
parallel) from group_ids - all sequential/flat reads. The Obs.group field is written
only when a CPU stage will read it (no GPU: scaling/CC/combine otherwise use
group_ids / rawrun_group, never partials.group), so the default path skips the strided
Obs pass entirely. group hkl ~0.31 -> ~0.20 s/2-pass on lyso.
Output is bit-identical (group_ids values and the obs-index-ordered gperm are unchanged),
so the merged results are unchanged; validated across the battery (15/15 deterministic
crystals bit-identical to the CPU path, only EP_cs_01-24 noise keeps its benign wobble).
Non-CUDA build unaffected (need_obs_group is always true there).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
leonarski_fandClaude Opus 4.8
87c70195f8