zstd: x86 assembler implementation of sequenceDecs.executeSimple#531
Merged
klauspost merged 3 commits intoklauspost:masterfrom Mar 18, 2022
Merged
zstd: x86 assembler implementation of sequenceDecs.executeSimple#531klauspost merged 3 commits intoklauspost:masterfrom
klauspost merged 3 commits intoklauspost:masterfrom
Conversation
klauspost
reviewed
Mar 14, 2022
Owner
klauspost
left a comment
klauspost
left a comment
There was a problem hiding this comment.
Just a few initial thoughts.
Contributor
Author
|
I fixed the code, so the tests pass. Now going to prepare an |
WojciechMula
force-pushed
the
asm-seqdec-execute
branch
from
9d30996 to
7643f8e
Compare
Contributor
Author
|
I'll fix the CI error once pushing to github will be working. The fresh results from IceLake: I also tried copying via AVX registers, but there is no big difference. IIRC I observed better speedup in my early decodeSync implementation. So, maybe for that function, we'll see a nicer AVX impact. |
klauspost
reviewed
Mar 17, 2022
Contributor
Author
|
Re: performing all memcpy inside asm -- overall, it's better, few minimal regressions. Comparison with the version with a threshold. |
Owner
|
@WojciechMula Looks good 👍🏼 |
Contributor
Author
@klauspost sorry, I copied a wrong file - but the actual results are still quite good IMHO |
Contributor
Author
|
Re: copying in 32-byte chunks - 2xSSE reg. There are some nice speed-ups, but I feel like there are more regressions. |
Owner
Yeah, it will over-copy quite a bit, since most matches/lits will be <16 bytes. We could add some logic that uses one or the other, based on |
Contributor
Author
I fully agree. So, I'm going to revert this change, squash the commits and it'll be ready to merge. I'd like to add the history support in another PR, as I see there's some more work and likely we'd need another specialisation. Does it sound fine? |
Contributor
Author
BTW this is on my TODO list: check if having a separate path for |
Method sequenceDecs.executeSimple is simplified sequenceDecs.execute for cases when both the history and dictionary are empty. These cases are 83% of all calls to `execute` when run `go test`. In benchmarks it is 99%.
klauspost
reviewed
Mar 18, 2022
WojciechMula
force-pushed
the
asm-seqdec-execute
branch
from
2cb26c5 to
3295600
Compare
- allocate padded out buffer - remove not needed Go fallback code - add missing checks
Contributor
Author
I checked this (see: https://github.com/WojciechMula/compress/tree/asm-seqdec-execute-small-ll-ml), but the results are not too promising. |
klauspost
approved these changes
Mar 18, 2022
Owner
Took this out, since it isn't true, since non-stream benchmarks hits decodeSync, not decode+execute. The code is largely unused for now, but a foundation we can build on. |
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
2 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.
This is plain x86 and x86 with BMI2 implementation of
sequenceDecs.executeSimple. Part of #515.I extracted function
executeSimpleto handle cases when no history nor dictionary is used. My quick check showed that forgo testsuch cases is 83% of all calls, while forgo test -bench .it's 99%. Thus, it's the vast majority of cases. Of course, we may consider handling all cases in another PR (but after completing #529).As always, I'm marking it as a draft, as some tests fail. I will figure out what's wrong, likely as usual I missed something silly.[fixed (I was right, it was silly)]Below are preliminary benchmark results from IceLake machine: it's noasm vs GOARM64=v3. Currently, the branch is built on top of #528, thus we see the combined performance boost from x86 BMI use in both
decodeandexecute.