ApplyCellSurface (detector modulation, time x detector, crystal-frame SH and
goniometer-frame absorption) spends most of its time in two passes per round:
the per-group reference sums and the per-(block, cell) fit sums. With a GPU
both now run on the device (RotationScaleMergeGPU::Surface*) over the same
terms in the same order:
- reference: one thread per ASU group, walking a group-order permutation of
the terms in fulls order;
- fit: each subset cut into the host's reduction blocks with every block's
terms ordered by cell (stable counting sort, host); a per-term kernel forms
w*Is, w*Iref, Iref and one thread per (block, cell) runs the two fma chains;
the block slots are added per cell in block order.
Every rounding is spelled out (__dmul_rn/__dadd_rn/fma) to be the one the host
build makes: GCC at -march=x86-64-v3 fuses swI's multiply-add only in the
parity-filtered copy of the reference loop, and both fit sums.
Host side, exact on both paths: the 19 serial nth_element selections of the
shell edges become one parallel sort (same order statistics), the per-term shell
lookup runs on all threads, and the gate's per-shell CC is one walk over the
groups instead of one per shell.
Exact: p.mtz md5 identical to the oracle on myob/cytc/thau x10sa, GPU build
(all CUDA architectures) and CPU build. CorrectionSurfaceGPU test checks the
device sums bit for bit against an explicitly rounded host loop.
Measured (cytc/thau, two interleaved A/B pairs, box at load 13-25 from sibling
work): ApplyCellSurface host core-seconds -83% on cytc; SG adoption -> writing
reflections 4.58->3.75 and 3.85->2.54 s (cytc), 3.15->2.28 and 2.29->2.03 s
(thau); RSM final merge -0.7..-0.9 s and P1 cross-check -1.5..-1.8 s on cytc.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K5K8jvPPbmCrbqnWkddTuB
