Tokenizing UTF-8 with SIMD ㊗️ #283
Merged
+7,289
−486
Conversation
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
Benchmarks suggest, for xlsum.csv: sz_find_newline_utf8_serial: 7.49 GB/s sz_find_whitespace_utf8_serial: 5.53 GB/s sz_find_newline_utf8_ice: 17.54 GB/s - 2.3x sz_find_whitespace_utf8_ice: 6.10 GB/s - 1.1x
sz_find_newline_utf8_serial: 0.9 GB/s sz_find_whitespace_utf8_serial: 0.7 GB/s sz_find_newline_utf8_ice: 14.5 GB/s - 16x sz_find_whitespace_utf8_ice: 1.3 GB/s - 1.85x
Benchmarking `sz_find_newline_utf8_serial`: > Throughput: 880.96 MiB/s @ 72.65 ms/call Benchmarking `sz_find_whitespace_utf8_serial`: > Throughput: 594.65 MiB/s @ 107.63 ms/call Benchmarking `sz_find_newline_utf8_ice`: > Throughput: 19.20 GiB/s @ 3.25 ms/call > + 22.3 x against `sz_find_newline_utf8_serial` Benchmarking `sz_find_whitespace_utf8_ice`: > Throughput: 1.32 GiB/s @ 47.44 ms/call > + 2.3 x against `sz_find_whitespace_utf8_serial`
Benchmarking `sz_utf8_count_serial`: > Throughput: 2.71 GiB/s @ 23.02 ms/call Benchmarking `sz_utf8_count_haswell`: > Throughput: 45.78 GiB/s @ 1.37 ms/call > + 16.9 x against `sz_utf8_count_serial` Benchmarking `sz_utf8_count_ice`: > Throughput: 41.79 GiB/s @ 1.50 ms/call > + 15.4 x against `sz_utf8_count_serial`
Benchmarking `sz_utf8_find_whitespace_serial`: > Throughput: 264.52 MiB/s @ 15.48 s/call Benchmarking `sz_utf8_find_whitespace_neon`: > Throughput: 4.98 GiB/s @ 803.59 ms/call > + 19.3 x against `sz_utf8_find_whitespace_serial` Benchmarking `sz_utf8_find_newline_serial`: > Throughput: 6.36 GiB/s @ 628.70 ms/call Benchmarking `sz_utf8_find_newline_neon`: > Throughput: 10.15 GiB/s @ 394.23 ms/call > + 59.5 % against `sz_utf8_find_newline_serial`
We now have a separate pack/unaligned state and a `sz_hash_state_internal_t_` for inner logic.
PyTest now downloads the Unicode speicifcation and compares all valid UTF-8 encoded characters via StringZilla's new `utf8_case_fold` API
Different bindings used different design for skipping or including empty segments or trailing newlines in input strings. The new design is consistent across C++, Python, and Rust.
Performance in SVE2 still looks pretty bad. Benchmarking `sz_utf8_find_newline_serial`: > Throughput: 579.69 MiB/s @ 7.07 s/call Benchmarking `sz_utf8_find_whitespace_serial`: > Throughput: 495.51 MiB/s @ 8.27 s/call Benchmarking `sz_utf8_find_newline_neon`: > Throughput: 37.80 GiB/s @ 105.83 ms/call > + 66.8 x against `sz_utf8_find_newline_serial` Benchmarking `sz_utf8_find_whitespace_neon`: > Throughput: 400.16 MiB/s @ 10.24 s/call > - 19.2 % against `sz_utf8_find_whitespace_serial` Benchmarking `sz_utf8_find_newline_sve2`: > Throughput: 28.04 GiB/s @ 142.67 ms/call > + 49.5 x against `sz_utf8_find_newline_serial` Benchmarking `sz_utf8_find_whitespace_sve2`: > Throughput: 389.96 MiB/s @ 10.50 s/call > - 21.3 % against `sz_utf8_find_whitespace_serial`
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.
No description provided.