2 months ago
9f742cc
Davy* perf(separator): fixed per-chunk working memory for the spectral session (#172 PR 3) Eliminate every steady-state heap allocation in the spectral transform and composition path. Before: each chunk allocated ~33 MB (FourStem: 22 MB forward-transform output + 4 x 2.75 MB composed waveforms) to ~52 MB (TwoStem: + 24.75 MB spectral/time premix copies). After: the first chunk sizes a persistent SpectralSessionState (plans + forward buffer + one shared composition buffer + premix accumulators) and every later chunk reuses it — pointer-stability tests pin this down. - spectral: spec_into / ispec_into write into caller-owned buffers; the allocating spec / ispec become thin wrappers (golden-vector tests unchanged); byte-exact equivalence + no-reallocation tests added - spectral_session: SpectralSessionState owns all chunk-loop buffers; composition is factored into compose_outputs_into_buffers, pure over the extracted core outputs, so the explicit FourStem/TwoStem layouts are unit-tested without an ORT session (correctness vs per-stem composition, buffer reuse across chunks) - ORT inputs remain borrowed views over the reused buffers; ring feeding and OLA are untouched (StemMode semantics and 50% overlap unchanged; FFT implementation selection stays realfft f64 pending #170 measurements per the issue) Verified against the published candidate artifact (model-spectral-v1.0.0) and the dev waveform model: identical equivalence numbers to the PR 2 baseline (single-window max-abs <= 1.2e-5 all stems + premix; streaming end-to-end passes both modes). cargo test --lib 1039 passed; clippy -D warnings and fmt clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(separator): make the forward-buffer restore genuinely unconditional The take/restore of the reused forward-transform buffer had three ? early-returns between them (borrowed tensor construction, session lock), contradicting the 'even on error' comment. Move every fallible step into the result closure so the restore cannot be skipped. Behavior after an error is unchanged (the run aborts); the comment is now true. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Parent0a785a5