Reduce memory use for progressive lossless images#795
Open
hjanuschka wants to merge 5 commits into
Open
Conversation
Only retain progressive render snapshots in the CLI when they can be written to an output.
Inverse squeeze steps read neighbor grids (next average and previous decoded) that the transform graph counts as buffer uses, but the per-step code never released them, so those intermediate modular buffers stayed allocated for the whole frame. Mark them used on the final render so they are freed once consumed.
Modular frames never reclaim center group buffers via get_buffer, so the scratch pool grew to a full-frame copy that was retained for the pipeline's lifetime. Cap it to the few buffers sequential rendering can actually reuse.
veluca93
approved these changes
Jun 8, 2026
veluca93
left a comment
veluca93
left a comment
Member
There was a problem hiding this comment.
I intend to revisit how modular transform processing works, but this is a good fix in the meantime.
Collaborator
Author
|
started working on the modular transforms, it was just to big! so this PR is a temp. improvement! |
Member
Performance Summary (Commit
|
| Machine | Threading | Base MP/s | PR MP/s | Avg Improvement |
|---|---|---|---|---|
| desktop | Single | 80.80 | 81.01 | +0.61% ± 0.39% |
| desktop | Multi | 80.82 | 80.82 | +0.49% ± 0.33% |
| framework-desktop | Single | 94.43 | 93.71 | +0.13% ± 0.45% |
| framework-desktop | Multi | 94.79 | 94.65 | +0.20% ± 0.44% |
| pixel7a | Single (Fast) | 28.58 | 28.90 | +0.44% ± 0.40% |
| pixel7a | Single (Mid) | 20.83 | 21.10 | +0.39% ± 0.38% |
| pixel7a | Multi | 29.08 | 29.06 | -0.07% ± 0.42% |
Member
|
I think I know where the speed regressions are coming from - can you remove the part to limit the scratch buffers for now? |
This reverts commit 6489571.
Collaborator
Author
|
done, also did #796 that has the refactored interleave! and completely wins in terms of RSS (could we bench this?) |
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
3 participants
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.
Fixes #782.
The inverse squeeze transforms read two neighbor grids per step (the next average and the previously decoded output), which the transform graph counts as buffer uses, but
do_runonly released the primary inputs. So those intermediate modular buffers stayed allocated for the whole frame. This releases the neighbor grids on the final render, mirroring how the graph registers them (with a dedup guard for when a coarser average channel maps the neighbor onto the same grid index).jxl_clialso no longer stores progressive render snapshots when there's no output path to write them to.Peak RSS on the repro:
-s)--render-interval 1000--render-interval 250For full decode that's now below djxl (~1516 MiB) and ~1.5x jxl-oxide (~792 MiB), down from 3.4x.