improve JPEG encode_block perf by about 25% with faster integer encoding #1912
Conversation
|
Looks like NonZeroU32::unsigned_abs was stabilized in 1.64... let's see if there's a workaround |
|
We can also bump MSRV for a perf win, it's not a set-in-stone restriction. Bump here and as indicated in the comment to do it with the PR: https://github.com/image-rs/image/blob/master/Cargo.toml#L7-L8 |
|
I see MSRV has been bumped to 1.67 since, which is higher than this PR's requirement of 1.64. I understand there is nothing blocking this, and it can be merged? The change is simple, covered by tests, and the performance gains are tantalizing. |
|
Will look into reviving this PR once |
…ed_abs. This has a significant speed impact on leading zero calc on x86
|
Rebased since this change had sat there for a while |
|
There's still a CI test_toolchain with 1.63 that needs to be upgraded before this can be integrated |
There was a problem hiding this comment.
Thanks for this PR!
I think it is kind of doing a lot and would be better split into a bunch of PRs. Perhaps:
- Adding benchmarks to give a baseline performance number and justify subsequent PRs
- Optimizations to
BitWriter - New version of
write_block - New version of
encode_coefficient
| @@ -765,7 +855,8 @@ fn build_quantization_segment(m: &mut Vec<u8>, precision: u8, identifier: u8, qt | |||
| } | |||
| } | |||
|
|
|||
| fn encode_coefficient(coefficient: i32) -> (u8, u16) { | |||
| #[cfg(feature = "benchmarks")] | |||
| fn encode_coefficient_old(coefficient: i32) -> (u8, u16) { | |||
There was a problem hiding this comment.
I don't think we should keep around an old implementation simply to benchmark against
I license past and future contributions under the dual MIT/Apache-2.0 license,
allowing licensees to choose either at their option.
Also add benches for tracking performance of JPEG block encoding.
pre:
test codecs::jpeg::encoder::tests::bench_encode_block ... bench: 548 ns/iter (+/- 131) = 135 MB/spost:
test codecs::jpeg::encoder::tests::bench_encode_block ... bench: 410 ns/iter (+/- 140) = 180 MB/s