perf: optimize every custom periodogram implementation across all backends - #15

Merged

tjayasinghe merged 9 commits into

release/v1.1.0from

feat/perf-optimizations

Jul 2, 2026

Commits on Jul 1, 2026

perf: fast scatter shims + PDM fine-bin regrouping (CPU/torch/CUDA)

tjayasinghe
and
claude
committed
abcdd08
perf(mhaov): assemble the Gram from harmonic trig sums, not a design tensor

tjayasinghe
and
claude
committed
a6bb69d
perf(gls): one trig evaluation for all six direct-path sums

tjayasinghe
and
claude
committed
b7caafd
perf(gls): batch the NUFFT base-grid pair; assemble one-shot cufinufft on device

tjayasinghe
and
claude
committed
16c1036
perf: extend the multicore numba CPU tier to PDM, CE, SL, MHAOV, and TLS

tjayasinghe
and
claude
committed
559fcdf
perf(cuda): precision-templated RawKernels, CE int atomics, BLS device cache

tjayasinghe
and
claude
committed
dc75b4f
perf(string-length): in-block CUDA sort kernel + torch.sort fast path

tjayasinghe
and
claude
committed
27a1a09
test+docs: fast-backend parity suite, changelog, backend docs

tjayasinghe
and
claude
committed
6153d76
ci: install EGL/OpenGL system libraries for the Linux GUI tests

tjayasinghe
and
claude
committed
51c92c6