wwu-mmll / confound_corrected_cpm / 28545012926

85%

main: 85%

push

github

Batch edge-statistic computation over runs in _select_edges

The direct edge-selection path recomputed the correlation r/p from scratch
once per run (per permutation), even though the statistic depends only on
the data and the fixed correlation type — not on the per-run selector
params. For a 5-fold run with 200 permutations that is 1005 calls into
correlations_and_pvalues; on GPU each call pays kernel-launch and
host->device transfer overhead, which is why GPU showed no speedup over
CPU. Compute r/p once per fold in a single batched pass over all target
columns, then apply the (unchanged, possibly per-run) thresholding to each
column. Call count for that scenario drops 1005 -> 10.

This mirrors what the inner-CV path (inner_fold.py) already did, so the two
paths are now consistent. Results are identical up to float32 rounding: an
end-to-end run (edges, permutation edges, predictions, cv results,
permutation p-values) is bit-for-bit unchanged on the golden dataset; the
batched vs per-column r/p agree to float32 precision (guarded by the new
test_batched_edge_statistics_match_per_column).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

4 of 4 new or added lines in 1 file covered. (100.0%)

2113 of 2916 relevant lines covered (72.46%)

0.72 hits per line