Repository navigation
sidecar: hand a state input's earlier rows at the times they arrived - #10
Merged
Merged
Conversation
An instance of a frame-parallel node with a state input is handed the rows of the ticks it did not process. Each tick's rows were recorded at the tick's pts, in the clock's time base, and a node reads them in the rows stream's own time base, as it reads the messages of its own tick. When the two bases differ, the rows another instance rendered were folded at the wrong time. blitz's compose over a picture decoded in the region (1/90000 from MPEG-TS) with change rows from a producer clocked by the same input's raw picture (1/61440 on the pipe) scheduled every row it was handed that way about 1.46 times later than it arrived, so each worker showed a different, older document, while --jobs 1 was right. A row stamped between two ticks was moved to the tick's time even on one base. Each row is now recorded at its own pts: a message's, or the frame's it rode, in its stream's time base, in the order the node reads them. shape_state now also says when its newest word arrived, and a fleet test reads cues on 1/48000 that fall between its ticks on 1/10, the same at -jobs 1, 2 and 4. An exec test runs blitz's emit_changes into compose over an MPEG-TS input at --jobs 1 and the default, frame for frame, where a blitz checkout sits beside this one. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
A node spread over several workers is handed the rows of the ticks it did not render, so the output is the same at any worker count. On ffrwd 0.30.1 that broke for blitz's
composefed by a rows producer on the GPU path: the frames flipped between older and newer documents every few seconds, while--jobs 1was right.Cause. The replay recorded each handed-on row at its tick's pts in the clock's time base, while the node reads rows in the rows stream's own time base. On the GPU path the two differ: compose's picture is the one the region decodes from the packets (1/90000 for MPEG-TS), and the producer reads the same input's raw picture off the pipe (1/61440), whose base its rows take. A replayed row was therefore read about 1.46 times later than it arrived, blitz held it as future, and every instance that had not rendered that row's own tick showed an older document. A row stamped between two ticks was also moved to its tick's time even on one base.
Fix. The replay records each row at its own pts in its stream's time base, a message's pts or the pts of the frame it rode on, in the order the SDK folds them. The
earlier-rowsdoc in the world and NODE-CLI.md name the base; no world bump. Neither blitz nor a producer needs a change.Measured on the 4090. A 20 s MPEG-TS input into
composewith blitz'semit_changes, default workers against--jobs 1: 521 of 600 frames differed before, 0 after;--jobs 1unchanged. Through a realtime input: 34 differed before, 0 after.Tests. A fleet test in
node_shapesruns a state node at-jobs1, 2 and 4 with cues falling between its ticks and compares the bytes and each tick's newest-row time, failing 5 of 5 without the fix. An exec test runs blitz'semit_changesintocomposeover a generated MPEG-TS file at--jobs 1and the default and compares framemd5 frame for frame, skipping where the blitz checkout is absent.Checks. cargo fmt and clippy
-D warningsclean with default features and--no-default-features; node_shapes 33; ruff and mypy clean; unit tier 5143; the exec tests touched pass.🤖 Generated with Claude Code