Repository navigation
Conversation
|
@Cyan4973 has exported this pull request. If you are a Meta employee, you can view the originating Diff in D127402945. |
Cyan4973
added a commit
to Cyan4973/openzl
that referenced
this pull request
Nov 3, 2026
Summary: X-link: meta#1200 The widths `BITSPLIT_TOP8` picks match none of the 4 float patterns, so they ran in the generic kernels, one element at a time through a jump table per field. These kernels now process one field at a time over blocks of 2048 elements, so each loop vectorizes. The top-bits check and the top8 max scan vectorize too. The wire format is unchanged. Kernels, GB/s, in cache (clang, AVX2): | layout | encode | decode | |---|---|---| | u32 {24,8} | 1.4 -> 26.8 | 1.6 -> 30.8 | | u32 {23,8} | 1.4 -> 25.5 | 1.6 -> 30.0 | | u16 {8,8} | 0.5 -> 14.8 | 0.7 -> 21.2 | | u64 {56,8} | 2.4 -> 23.3 | 2.4 -> 26.4 | | u8 {5} | 0.5 -> 15.7 | 0.6 -> 36.9 | | scan | GB/s | |---|---| | top-bits check, u32 | 8.8 -> 79 | | top8 max, u16 | 14 -> 116 | End to end, top8 then FSE, on 964 float32 files (compression / decompression, MB/s): | files | before | after | `FLOAT32_DECONSTRUCT` | size vs deconstruct | |---|---|---|---|---| | all >= 0, max >= 2: {23,8} (415) | 439 / 690 | 662 / 1113 | 703 / 1073 | -0.40% | | all >= 0, max < 2: {22,8} (141) | 434 / 682 | 635 / 1078 | 656 / 1000 | +0.06% | | signed: {24,8} (408) | 514 / 768 | 757 / 1285 | 755 / 1123 | +0.05% | Trade-offs: - The 4 float patterns keep their specialized kernels, which stay faster in cache. - gcc vectorizes these loops only at -O3. At -O2, {24,8} only reaches 5.2 / 5.6 GB/s. Reviewed By: kevinjzhang Differential Revision: D127402945
Cyan4973
force-pushed
the
export-D127402945
branch
from
November 3, 2026 21:23
d23da32 to
59ccfe7
Compare
) Summary: X-link: meta#1200 The widths `BITSPLIT_TOP8` picks match none of the 4 float patterns, so they ran in the generic kernels, one element at a time through a jump table per field. These kernels now process one field at a time over blocks of 2048 elements, so each loop vectorizes. The top-bits check and the top8 scan vectorize too. The wire format is unchanged. Kernels, GB/s, in cache (clang, AVX2): | layout | encode | decode | |---|---|---| | u32 {24,8} | 1.4 -> 26.8 | 1.6 -> 30.8 | | u32 {23,8} | 1.4 -> 25.5 | 1.6 -> 30.0 | | u16 {8,8} | 0.5 -> 14.8 | 0.7 -> 21.2 | | u64 {56,8} | 2.4 -> 23.3 | 2.4 -> 26.4 | | u8 {5} | 0.5 -> 15.7 | 0.6 -> 36.9 | | scan | GB/s | |---|---| | top-bits check, u32 | 8.8 -> 79 | | top8 scan, u16 | 14 -> 116 | End to end, top8 then FSE, on 964 float32 files (compression / decompression, MB/s): | files | before | after | `FLOAT32_DECONSTRUCT` | size vs deconstruct | |---|---|---|---|---| | all >= 0, max >= 2: {23,8} (415) | 439 / 690 | 662 / 1113 | 703 / 1073 | -0.40% | | all >= 0, max < 2: {22,8} (141) | 434 / 682 | 635 / 1078 | 656 / 1000 | +0.06% | | signed: {24,8} (408) | 514 / 768 | 757 / 1285 | 755 / 1123 | +0.05% | Trade-offs: - The 4 float patterns keep their specialized kernels, which stay faster in cache. - gcc vectorizes these loops only at -O3. At -O2, {24,8} only reaches 5.2 / 5.6 GB/s. Reviewed By: kevinjzhang Differential Revision: D127402945
meta-codesync
Bot
changed the title
Cyan4973
force-pushed
the
export-D127402945
branch
from
November 4, 2026 00:22
59ccfe7 to
2923d3d
Compare
|
This pull request has been merged in 8724633. |
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
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.
Summary:
X-link: #1200
The widths
BITSPLIT_TOP8picks match none of the 4 float patterns, so they ran in the generic kernels, one element at a time through a jump table per field. These kernels now process one field at a time over blocks of 2048 elements, so each loop vectorizes. The top-bits check and the top8 scan vectorize too. The wire format is unchanged.Kernels, GB/s, in cache (clang, AVX2):
End to end, top8 then FSE, on 964 float32 files (compression / decompression, MB/s):
FLOAT32_DECONSTRUCTTrade-offs:
Reviewed By: kevinjzhang
Differential Revision: D127402945