perf: add AVX512-optimized is_ascii to fix native CPU regression#22
Merged
perf: add AVX512-optimized is_ascii to fix native CPU regression#22
Conversation
When compiling with `-C target-cpu=native` on AVX512 CPUs, LLVM generates
poor code for stdlib's is_ascii(), causing 30x slowdowns. This adds a custom
implementation using AVX512BW intrinsics with runtime feature detection.
- Add src/simd.rs with is_ascii() using _mm512_movepi8_mask
- Use masked loads for safe tail handling (<64 bytes)
- Runtime detection via is_x86_feature_detected!("avx512bw")
- Falls back to stdlib is_ascii() on non-AVX512 systems
Benchmarks vs master:
- ASCII workloads: 1.3-1.8x faster (default compilation)
- ASCII workloads: 30-35x faster (with -C target-cpu=native)
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
Yore Benchmarks
Details
| Benchmark suite | Current: a6dcb41 | Previous: dd07916 | Ratio |
|---|---|---|---|
decode_checked/mostly_ascii/8 |
33 ns/iter (± 12) |
34 ns/iter (± 13) |
0.97 |
decode_checked/mostly_ascii/64 |
66 ns/iter (± 7) |
70 ns/iter (± 14) |
0.94 |
decode_checked/mostly_ascii/256 |
139 ns/iter (± 13) |
143 ns/iter (± 12) |
0.97 |
decode_checked/mostly_ascii/512 |
232 ns/iter (± 19) |
228 ns/iter (± 21) |
1.02 |
decode_checked/mostly_ascii/1024 |
415 ns/iter (± 24) |
423 ns/iter (± 28) |
0.98 |
decode_checked/mostly_ascii/2048 |
879 ns/iter (± 53) |
888 ns/iter (± 47) |
0.99 |
decode_checked/mostly_ascii/4096 |
1964 ns/iter (± 74) |
2004 ns/iter (± 75) |
0.98 |
decode_checked/ascii/8 |
9 ns/iter (± 0) |
8 ns/iter (± 0) |
1.13 |
decode_checked/ascii/64 |
8 ns/iter (± 0) |
8 ns/iter (± 0) |
1 |
decode_checked/ascii/256 |
10 ns/iter (± 0) |
10 ns/iter (± 0) |
1 |
decode_checked/ascii/512 |
16 ns/iter (± 0) |
15 ns/iter (± 0) |
1.07 |
decode_checked/ascii/1024 |
26 ns/iter (± 3) |
25 ns/iter (± 0) |
1.04 |
decode_checked/ascii/2048 |
51 ns/iter (± 4) |
50 ns/iter (± 11) |
1.02 |
decode_checked/ascii/4096 |
91 ns/iter (± 0) |
91 ns/iter (± 3) |
1 |
decode_checked/extended/8 |
35 ns/iter (± 1) |
36 ns/iter (± 1) |
0.97 |
decode_checked/extended/64 |
84 ns/iter (± 3) |
85 ns/iter (± 2) |
0.99 |
decode_checked/extended/256 |
259 ns/iter (± 3) |
260 ns/iter (± 6) |
1.00 |
decode_checked/extended/512 |
484 ns/iter (± 19) |
485 ns/iter (± 89) |
1.00 |
decode_checked/extended/1024 |
904 ns/iter (± 5) |
902 ns/iter (± 11) |
1.00 |
decode_checked/extended/2048 |
1758 ns/iter (± 48) |
1760 ns/iter (± 100) |
1.00 |
decode_checked/extended/4096 |
3464 ns/iter (± 392) |
3478 ns/iter (± 32) |
1.00 |
decode_lossy/all_bad/8 |
48 ns/iter (± 1) |
39 ns/iter (± 0) |
1.23 |
decode_lossy/all_bad/64 |
63 ns/iter (± 0) |
63 ns/iter (± 1) |
1 |
decode_lossy/all_bad/256 |
178 ns/iter (± 10) |
173 ns/iter (± 17) |
1.03 |
decode_lossy/all_bad/512 |
336 ns/iter (± 42) |
323 ns/iter (± 7) |
1.04 |
decode_lossy/all_bad/1024 |
647 ns/iter (± 21) |
622 ns/iter (± 78) |
1.04 |
decode_lossy/all_bad/2048 |
1256 ns/iter (± 7) |
1220 ns/iter (± 36) |
1.03 |
decode_lossy/all_bad/4096 |
2482 ns/iter (± 22) |
2394 ns/iter (± 35) |
1.04 |
decode_lossy/mostly_ascii/8 |
47 ns/iter (± 21) |
39 ns/iter (± 15) |
1.21 |
decode_lossy/mostly_ascii/64 |
65 ns/iter (± 6) |
65 ns/iter (± 15) |
1 |
decode_lossy/mostly_ascii/256 |
155 ns/iter (± 16) |
148 ns/iter (± 17) |
1.05 |
decode_lossy/mostly_ascii/512 |
245 ns/iter (± 21) |
236 ns/iter (± 23) |
1.04 |
decode_lossy/mostly_ascii/1024 |
429 ns/iter (± 24) |
414 ns/iter (± 31) |
1.04 |
decode_lossy/mostly_ascii/2048 |
806 ns/iter (± 40) |
789 ns/iter (± 39) |
1.02 |
decode_lossy/mostly_ascii/4096 |
1583 ns/iter (± 130) |
1545 ns/iter (± 55) |
1.02 |
encode_checked/mostly_ascii/8 |
42 ns/iter (± 17) |
42 ns/iter (± 17) |
1 |
encode_checked/mostly_ascii/64 |
126 ns/iter (± 3) |
128 ns/iter (± 20) |
0.98 |
encode_checked/mostly_ascii/256 |
456 ns/iter (± 6) |
457 ns/iter (± 17) |
1.00 |
encode_checked/mostly_ascii/512 |
905 ns/iter (± 70) |
911 ns/iter (± 180) |
0.99 |
encode_checked/mostly_ascii/1024 |
1921 ns/iter (± 46) |
1801 ns/iter (± 103) |
1.07 |
encode_checked/mostly_ascii/2048 |
3693 ns/iter (± 192) |
3541 ns/iter (± 331) |
1.04 |
encode_checked/mostly_ascii/4096 |
7053 ns/iter (± 294) |
7027 ns/iter (± 206) |
1.00 |
encode_checked/ascii/8 |
9 ns/iter (± 0) |
8 ns/iter (± 0) |
1.13 |
encode_checked/ascii/64 |
8 ns/iter (± 0) |
8 ns/iter (± 0) |
1 |
encode_checked/ascii/256 |
10 ns/iter (± 0) |
11 ns/iter (± 0) |
0.91 |
encode_checked/ascii/512 |
14 ns/iter (± 0) |
14 ns/iter (± 0) |
1 |
encode_checked/ascii/1024 |
25 ns/iter (± 0) |
24 ns/iter (± 0) |
1.04 |
encode_checked/ascii/2048 |
50 ns/iter (± 1) |
50 ns/iter (± 6) |
1 |
encode_checked/ascii/4096 |
91 ns/iter (± 2) |
89 ns/iter (± 1) |
1.02 |
encode_checked/extended/8 |
54 ns/iter (± 1) |
52 ns/iter (± 1) |
1.04 |
encode_checked/extended/64 |
206 ns/iter (± 3) |
208 ns/iter (± 8) |
0.99 |
encode_checked/extended/256 |
716 ns/iter (± 6) |
705 ns/iter (± 83) |
1.02 |
encode_checked/extended/512 |
1400 ns/iter (± 15) |
1382 ns/iter (± 283) |
1.01 |
encode_checked/extended/1024 |
2749 ns/iter (± 26) |
2747 ns/iter (± 162) |
1.00 |
encode_checked/extended/2048 |
5513 ns/iter (± 45) |
5425 ns/iter (± 338) |
1.02 |
encode_checked/extended/4096 |
10854 ns/iter (± 165) |
10785 ns/iter (± 218) |
1.01 |
encode_lossy/all_bad/8 |
55 ns/iter (± 1) |
52 ns/iter (± 0) |
1.06 |
encode_lossy/all_bad/64 |
251 ns/iter (± 4) |
227 ns/iter (± 2) |
1.11 |
encode_lossy/all_bad/256 |
890 ns/iter (± 11) |
797 ns/iter (± 5) |
1.12 |
encode_lossy/all_bad/512 |
1747 ns/iter (± 72) |
1559 ns/iter (± 10) |
1.12 |
encode_lossy/all_bad/1024 |
3484 ns/iter (± 54) |
3094 ns/iter (± 20) |
1.13 |
encode_lossy/all_bad/2048 |
6797 ns/iter (± 1146) |
6138 ns/iter (± 129) |
1.11 |
encode_lossy/all_bad/4096 |
13522 ns/iter (± 602) |
12228 ns/iter (± 90) |
1.11 |
This comment was automatically generated by workflow using github-action-benchmark.
There was a problem hiding this comment.
⚠️ Performance Alert ⚠️
Possible performance regression was detected for benchmark 'Yore Benchmarks'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 1.10.
| Benchmark suite | Current: a6dcb41 | Previous: dd07916 | Ratio |
|---|---|---|---|
decode_checked/ascii/8 |
9 ns/iter (± 0) |
8 ns/iter (± 0) |
1.13 |
decode_lossy/all_bad/8 |
48 ns/iter (± 1) |
39 ns/iter (± 0) |
1.23 |
decode_lossy/mostly_ascii/8 |
47 ns/iter (± 21) |
39 ns/iter (± 15) |
1.21 |
encode_checked/ascii/8 |
9 ns/iter (± 0) |
8 ns/iter (± 0) |
1.13 |
encode_lossy/all_bad/64 |
251 ns/iter (± 4) |
227 ns/iter (± 2) |
1.11 |
encode_lossy/all_bad/256 |
890 ns/iter (± 11) |
797 ns/iter (± 5) |
1.12 |
encode_lossy/all_bad/512 |
1747 ns/iter (± 72) |
1559 ns/iter (± 10) |
1.12 |
encode_lossy/all_bad/1024 |
3484 ns/iter (± 54) |
3094 ns/iter (± 20) |
1.13 |
encode_lossy/all_bad/2048 |
6797 ns/iter (± 1146) |
6138 ns/iter (± 129) |
1.11 |
encode_lossy/all_bad/4096 |
13522 ns/iter (± 602) |
12228 ns/iter (± 90) |
1.11 |
This comment was automatically generated by workflow using github-action-benchmark.
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
1 participant
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
-C target-cpu=nativeon AVX512 CPUsis_asciiimplementation using AVX512BW intrinsicsProblem
When compiling with
-C target-cpu=nativeon AVX512-capable CPUs, LLVM generates poor code for stdlib'sis_ascii(), causing 30x slowdowns in ASCII-heavy workloads.Solution
Add
src/simd.rswith AVX512-optimizedis_asciiusing:_mm512_movepi8_maskto extract MSB of 64 bytes at once_mm512_maskz_loadu_epi8for safe tail handlingis_x86_feature_detected!("avx512bw")Benchmarks
-C target-cpu=nativeTest plan
cargo test)