A port of the fast-feedback indexer's CUDA kernels to plain C++/Eigen: the same half-sphere and
rotation samples, the same scores and the same 32 output cells for Refine(). The loops are
reorganised so the inner one runs over samples and vectorises (GCC/Clang, no intrinsics), and the
cell search separates the rotation about the candidate vector out of the per-spot terms. The
vector refinement reproduces the library's warp sums - partial sums its shuffle reduction leaves in
the other lanes carry into the next round - and its float arithmetic (FMA contraction, the
Fibonacci points in float), because the candidates searched depend on it: on the library's 27 test
spot lists and 4 synthetic stills the CPU and GPU best cells agree to ~1e-6 relative, the top-32
sets in 30 of 31 cases completely.
FFBIDX now resolves to the GPU library where a GPU is visible and to the CPU port otherwise (no
GPU, --no-gpu, or a build without CUDA); Auto takes FFBIDX for stills whenever a cell is known,
also without a GPU, instead of FFTW. With a GPU nothing changes (a stills set and a rotation set
reproduce the baseline exactly). Without a GPU the stills indexer pool gets one indexer per worker
instead of eight, since the CPU indexer holds no card memory.
Single core, synthetic still, loaded 5950X: ~8 ms per call at 30 spots, ~15 ms at 80, ~35 ms at
250; 16 threads ~1300/700/330 calls/s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015CeEGgpKdfkCTJb3YNMJcu
(cherry picked from commit 9a8eb8669d1be6c3ec16db6786280affd6830ee5)
leonarski_f
ea383f9e83