fix: ensure frame-aligned spans in queue to prevent channel misalignment #818
Conversation
|
oh this one is fun! Thanks for all the work. I kinda gave up on this :) |
|
@yara-blue friendly nudge: do you intend to review this and other PRs, or do you expect me to merge it myself when happy? I understand when you're busy, but would just like to know. |
|
This is super difficult code I tried a while back and gave up on it (could not get it performant and working). I'd love to review it but indeed do not have the time. A quick peek shows nothing that worries me. If you are happy with it feel free to go ahead and merge! Little update on what I've been cooking: |
Fixes two related issues with frame alignment in the queue: 1. Frame alignment for non-power-of-2 channel counts: The queue used a fixed threshold of 512 samples which could cause channel misalignment when the channel count doesn't evenly divide 512 (e.g., 6 channels: 512 % 6 = 2 sample offset). Replaced with dynamic threshold function that rounds to nearest frame boundary. 2. Mid-span ending detection: Sources that end before their promised current_span_len() would cause the queue to immediately transition to the next source, potentially mid-frame. Added sample tracking and silence padding to complete frames before transitioning. Changes: - Replace THRESHOLD constant with threshold(channels) helper function that ensures frame-aligned span lengths - Add sample tracking to detect mid-span source endings - Inject silence padding to complete incomplete frames - Add span_ending_mid_frame test to verify padding behavior Supersedes #684
roderickvd
force-pushed
the
fix/frame-aligned-spans
branch
from
1514256 to
b11c8b9
Compare
Thanks, will do. Slightly related question: before I fork have you by any chance changed your stance on #786? Echoing what I said there: not feeling bad about it myself nor want anyone else to feel bad about it, but need that refactoring for my projects so will have to fork it otherwise.
Sounds interesting. Will that also set us up to remove spans entirely? |
Fixes two related issues with frame alignment in the queue:
Frame alignment for non-power-of-2 channel counts: The queue used a fixed threshold of 512 samples which could cause channel misalignment when the channel count doesn't evenly divide 512 (e.g., 6 channels: 512 % 6 = 2 sample offset). Replaced with dynamic threshold function that rounds to nearest frame boundary.
Mid-span ending detection: Sources that end before their promised
current_span_len()would cause the queue to immediately transition to the next source, potentially mid-frame. Added sample tracking and silence padding to complete frames before transitioning.Supersedes #684, carrying over the test suite that was added there.
Branched off of #817 which fixes another span boundary problem.