2069578 - Move the mel buffer into the single-item MelBatch in vendored parakeet.cpp

Status: RESOLVED FIXED

(Core :: Audio/Video, task)

Description

[Sylvestre Ledru [:Sylvestre]]

Coverity CIDs 1700839 and 1700842 (COPY_INSTEAD_OF_MOVE) flag the same statement, duplicated in Model::transcribe_16k (third_party/parakeet.cpp/src/model.cpp:153) and Model::transcribe_16k_with_timestamps (:405):

mb1.data = feats;   // feats is [n_mels,T] = the B=1 batch buffer

feats is a local std::vector<float> that is not read again on that branch, so std::move(feats) is correct.

This is not a micro-optimisation: the sub_tile > 0 branch is the long-audio path, where feats is n_mels * T floats - roughly 30 MB for ten minutes of audio at 128 mels / 100 fps. The copy doubles peak mel-buffer footprint for exactly the inputs where memory is already tightest.

I will fw this upstream

Comment 1

[Sylvestre Ledru [:Sylvestre]]

Attachment: Bug 2069578 - Move the mel buffer into the 1-item MelBatch in vendored parakeet.cpp (Coverity CIDs 1700839, 1700842) r?#media-playback-reviewers

transcribe_16k and transcribe_16k_with_timestamps both copied feats into
mb1.data on the long-audio path, where feats is n_mels * T floats (tens of MB
for minutes of audio) and is dead after the assignment.

Carried as parakeet-move-mel-batch.patch so it survives re-vendoring; it should
go upstream to mudler/parakeet.cpp and be dropped once it lands there.

Comment 2

[Sylvestre Ledru [:Sylvestre]]

upstream https://github.com/mudler/parakeet.cpp/pull/67/changes

Comment 3

[Pulsebot]

Pushed by [e-mail]:
https://github.com/mozilla-firefox/firefox/commit/8e67ec05999d
https://hg.mozilla.org/integration/autoland/rev/eae40bd23d4f
Move the mel buffer into the 1-item MelBatch in vendored parakeet.cpp (Coverity CIDs 1700839, 1700842) r=media-playback-reviewers,padenot

Comment 4

[Cristian Tuns]

https://hg.mozilla.org/mozilla-central/rev/eae40bd23d4f