Rebalanced Faster Decoding and fixed Progressive for Lossless #4201
Conversation
Wouldn't build.
Flyby fix to reinstate libjxl#627 after being reverted in libjxl#3420
Code order was changed making it infeasible
Code order was changed making it infeasible
|
|
|
Congrats! Unfortunately I concentrated on other things from #4028, and it stroke back. |
|
Have you checked also influence of this fix on density? #4135 |
|
Thanks! You gave me a possible lead on fixing the |
|
Going to hold on #4207 before merging with main again, to hopefully pass all checks. |
WASM is failing too, hmm... |
Last I checked it was a cache error, so should be fine next time it runs |
|
@Melirius As it turns out, #4154 is the culprit of the small regression for
|
|
Just increase the number of buckets per property, this is the only sane way. |
Fixes #4584 A minor edit with large consequences. Previously the setting would encode uncompressed files due to a bug, causing the much faster encoding as stated in the old description. In the upcoming v0.12 libjxl release, [my PR](libjxl/libjxl#4201) overhauls the setting, to correctly scale decode speed at the cost of density. --------- Signed-off-by: Jonathan Brown <jonathanbr30@gmail.com>
|
When setting |
|
Ah, I thought we'd done that. Thanks for catching it. Probably use Gradient instead as that's closest. |
…ftwareFoundation#4812) Fixes AcademySoftwareFoundation#4584 A minor edit with large consequences. Previously the setting would encode uncompressed files due to a bug, causing the much faster encoding as stated in the old description. In the upcoming v0.12 libjxl release, [my PR](libjxl/libjxl#4201) overhauls the setting, to correctly scale decode speed at the cost of density. --------- Signed-off-by: Jonathan Brown <jonathanbr30@gmail.com>
…ftwareFoundation#4812) Fixes AcademySoftwareFoundation#4584 A minor edit with large consequences. Previously the setting would encode uncompressed files due to a bug, causing the much faster encoding as stated in the old description. In the upcoming v0.12 libjxl release, [my PR](libjxl/libjxl#4201) overhauls the setting, to correctly scale decode speed at the cost of density. --------- Signed-off-by: Jonathan Brown <jonathanbr30@gmail.com>
Description
This PR changes the encoding parameters for Lossless when
--faster_decodingis specified, resulting in up to 80% smaller files and 25% faster decoding. Also includes a flyby fix for lossy Delta Palette encoding on images above 2048*2048.Progressive Lossless is also improved. Up to 40% smaller files and 20% faster decoding, with an additional 800% encode speed increase, thanks to enabling local MA trees.
These results change depending on encoding parameters, image content, CPU and decoder.
Only libjxl and jxl-oxide were tested, due to jxlatte being too unstable. Oxide proved to be considerably faster for Progressive Lossless and our new level 4 Faster Decoding, suggesting libjxl could be improved by examining it.
Tested on a Ryzen 5600H with a 10 MP image at effort 7.
BPP is now much more inline with the base effort 7 encode, following a logical progression of less density as you increase level.
Levels 1-4 now make sense for decoding speed as well, with 3 and 4 now faster than effort 1. Progressive also got a boost.
A bonus major speed up for progressive lossless encoding too, using Local MA trees but still buffering the full image.
Fixes #2812
Fixes #3532
Pull Request Checklist
./ci.sh lintfor automatic code formatting.Please review the full contributing guidelines for more details.