zstd: Branchless getBits for amd64 w/o BMI2 #640
Merged
Conversation
This file contains 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
This produces the same number of instructions, while requiring less generating code. Benchmarks on the Intel Core i7-3770K show a tiny speedup: name old speed new speed delta Decoder_DecoderSmall/kppkn.gtb.zst-8 430MB/s ± 1% 437MB/s ± 1% +1.60% (p=0.000 n=10+9) Decoder_DecoderSmall/geo.protodata.zst-8 1.11GB/s ± 1% 1.13GB/s ± 0% +1.37% (p=0.000 n=9+9) Decoder_DecoderSmall/plrabn12.txt.zst-8 334MB/s ± 1% 339MB/s ± 1% +1.41% (p=0.000 n=9+10) Decoder_DecoderSmall/lcet10.txt.zst-8 392MB/s ± 2% 404MB/s ± 1% +3.05% (p=0.000 n=10+10) Decoder_DecoderSmall/asyoulik.txt.zst-8 355MB/s ± 2% 357MB/s ± 1% ~ (p=0.315 n=10+9) Decoder_DecoderSmall/alice29.txt.zst-8 344MB/s ± 1% 350MB/s ± 1% +1.69% (p=0.000 n=10+10) Decoder_DecoderSmall/html_x_4.zst-8 2.34GB/s ± 1% 2.37GB/s ± 1% +1.10% (p=0.000 n=10+10) Decoder_DecoderSmall/paper-100k.pdf.zst-8 3.75GB/s ± 0% 3.76GB/s ± 1% ~ (p=0.182 n=9+10) Decoder_DecoderSmall/fireworks.jpeg.zst-8 8.59GB/s ± 1% 8.58GB/s ± 1% ~ (p=0.842 n=10+9) Decoder_DecoderSmall/urls.10K.zst-8 561MB/s ± 1% 556MB/s ± 1% -0.82% (p=0.019 n=10+10) Decoder_DecoderSmall/html.zst-8 900MB/s ± 1% 913MB/s ± 1% +1.42% (p=0.000 n=10+9) Decoder_DecoderSmall/comp-data.bin.zst-8 399MB/s ± 1% 395MB/s ± 1% -0.99% (p=0.000 n=10+10) Decoder_DecodeAll/kppkn.gtb.zst-8 518MB/s ± 0% 526MB/s ± 0% +1.52% (p=0.000 n=10+9) Decoder_DecodeAll/geo.protodata.zst-8 1.28GB/s ± 0% 1.27GB/s ± 2% ~ (p=0.739 n=10+10) Decoder_DecodeAll/plrabn12.txt.zst-8 427MB/s ± 1% 433MB/s ± 1% +1.24% (p=0.000 n=10+10) Decoder_DecodeAll/lcet10.txt.zst-8 480MB/s ± 1% 490MB/s ± 1% +2.06% (p=0.000 n=10+10) Decoder_DecodeAll/asyoulik.txt.zst-8 435MB/s ± 0% 447MB/s ± 0% +2.70% (p=0.000 n=7+9) Decoder_DecodeAll/alice29.txt.zst-8 422MB/s ± 0% 438MB/s ± 1% +3.96% (p=0.000 n=8+9) Decoder_DecodeAll/html_x_4.zst-8 1.60GB/s ± 0% 1.61GB/s ± 0% +0.99% (p=0.000 n=9+10) Decoder_DecodeAll/paper-100k.pdf.zst-8 4.55GB/s ± 1% 4.44GB/s ± 1% -2.42% (p=0.000 n=10+10) Decoder_DecodeAll/fireworks.jpeg.zst-8 9.52GB/s ± 1% 9.47GB/s ± 2% ~ (p=0.143 n=10+10) Decoder_DecodeAll/urls.10K.zst-8 678MB/s ± 1% 684MB/s ± 0% +0.83% (p=0.000 n=10+10) Decoder_DecodeAll/html.zst-8 1.05GB/s ± 0% 1.07GB/s ± 1% +2.11% (p=0.000 n=10+10) Decoder_DecodeAll/comp-data.bin.zst-8 397MB/s ± 1% 391MB/s ± 1% -1.37% (p=0.000 n=10+10) Decoder_DecodeAllFiles/Mark.Twain-Tom.Sawyer.txt/fastest-8 437MB/s ± 0% 436MB/s ± 1% -0.21% (p=0.025 n=9+9) Decoder_DecodeAllFiles/Mark.Twain-Tom.Sawyer.txt/default-8 448MB/s ± 0% 451MB/s ± 0% +0.70% (p=0.000 n=9+9) Decoder_DecodeAllFiles/Mark.Twain-Tom.Sawyer.txt/better-8 478MB/s ± 0% 475MB/s ± 0% -0.53% (p=0.000 n=10+10) Decoder_DecodeAllFiles/Mark.Twain-Tom.Sawyer.txt/best-8 461MB/s ± 0% 470MB/s ± 0% +2.07% (p=0.000 n=8+9) Decoder_DecodeAllFiles/e.txt/fastest-8 9.62GB/s ± 3% 9.62GB/s ± 2% ~ (p=1.000 n=10+10) Decoder_DecodeAllFiles/e.txt/default-8 391MB/s ± 0% 406MB/s ± 0% +3.81% (p=0.000 n=10+8) Decoder_DecodeAllFiles/e.txt/better-8 438MB/s ± 0% 448MB/s ± 0% +2.39% (p=0.000 n=8+10) Decoder_DecodeAllFiles/e.txt/best-8 500MB/s ± 0% 500MB/s ± 0% ~ (p=0.119 n=9+9) Decoder_DecodeAllFiles/fse-artifact3.bin/fastest-8 1.07GB/s ± 1% 1.04GB/s ± 1% -2.61% (p=0.000 n=10+10) Decoder_DecodeAllFiles/fse-artifact3.bin/default-8 1.21GB/s ± 1% 1.19GB/s ± 1% -1.33% (p=0.000 n=10+10) Decoder_DecodeAllFiles/fse-artifact3.bin/better-8 994MB/s ± 0% 990MB/s ± 0% -0.42% (p=0.002 n=10+9) Decoder_DecodeAllFiles/fse-artifact3.bin/best-8 389MB/s ± 0% 381MB/s ± 0% -2.00% (p=0.000 n=8+10) Decoder_DecodeAllFiles/gettysburg.txt/fastest-8 274MB/s ± 1% 274MB/s ± 1% ~ (p=1.000 n=10+10) Decoder_DecodeAllFiles/gettysburg.txt/default-8 224MB/s ± 1% 223MB/s ± 1% -0.64% (p=0.015 n=10+10) Decoder_DecodeAllFiles/gettysburg.txt/better-8 228MB/s ± 1% 227MB/s ± 1% -0.40% (p=0.041 n=10+10) Decoder_DecodeAllFiles/gettysburg.txt/best-8 225MB/s ± 1% 223MB/s ± 0% -0.52% (p=0.008 n=10+6) Decoder_DecodeAllFiles/html.txt/fastest-8 599MB/s ± 1% 614MB/s ± 1% +2.41% (p=0.000 n=10+10) Decoder_DecodeAllFiles/html.txt/default-8 601MB/s ± 0% 613MB/s ± 0% +2.01% (p=0.000 n=8+9) Decoder_DecodeAllFiles/html.txt/better-8 626MB/s ± 1% 638MB/s ± 0% +1.99% (p=0.000 n=10+10) Decoder_DecodeAllFiles/html.txt/best-8 601MB/s ± 0% 612MB/s ± 0% +1.87% (p=0.000 n=10+10) Decoder_DecodeAllFiles/pi.txt/fastest-8 9.64GB/s ± 2% 9.66GB/s ± 1% ~ (p=0.529 n=10+10) Decoder_DecodeAllFiles/pi.txt/default-8 390MB/s ± 0% 403MB/s ± 0% +3.48% (p=0.000 n=10+10) Decoder_DecodeAllFiles/pi.txt/better-8 439MB/s ± 0% 451MB/s ± 0% +2.65% (p=0.000 n=10+10) Decoder_DecodeAllFiles/pi.txt/best-8 500MB/s ± 0% 499MB/s ± 0% -0.27% (p=0.009 n=7+10) Decoder_DecodeAllFiles/pngdata.bin/fastest-8 1.70GB/s ± 1% 1.69GB/s ± 1% -0.63% (p=0.013 n=10+9) Decoder_DecodeAllFiles/pngdata.bin/default-8 1.52GB/s ± 1% 1.51GB/s ± 0% -0.75% (p=0.000 n=10+9) Decoder_DecodeAllFiles/pngdata.bin/better-8 1.92GB/s ± 0% 1.90GB/s ± 0% -1.02% (p=0.000 n=10+10) Decoder_DecodeAllFiles/pngdata.bin/best-8 1.47GB/s ± 0% 1.46GB/s ± 0% -0.88% (p=0.000 n=10+9) Decoder_DecodeAllFiles/sharnd.out/fastest-8 9.60GB/s ± 1% 9.67GB/s ± 1% +0.67% (p=0.029 n=10+10) Decoder_DecodeAllFiles/sharnd.out/default-8 9.65GB/s ± 2% 9.71GB/s ± 1% ~ (p=0.353 n=10+10) Decoder_DecodeAllFiles/sharnd.out/better-8 9.67GB/s ± 1% 9.66GB/s ± 0% ~ (p=0.549 n=10+9) Decoder_DecodeAllFiles/sharnd.out/best-8 9.70GB/s ± 1% 9.61GB/s ± 0% -0.91% (p=0.010 n=10+9) [Geo mean] 935MB/s 940MB/s +0.57%
|
Nice! I will do a few tests and merge if no problems show up. |
|
I remember adding this, but abandoning it, since it wasn't a clear result. I reran the test with this patch, and the results are up and down: With BMI it is a win, but without it seems worse for quite a few cases. Maybe using this for BMI and use the old method for x64 would be the best? |
|
This patch doesn't change the BMI2 path. So using this for BMI2 is just dropping the patch :) I see similar results on my machine: better on some benchmarks, worse on others. I find it hard to tell how these benchmarks relate to the more end-to-end Decode* benchmarks. Also, code size still goes down. |
|
I wonder if something else is affecting the benchmarks, since all of the ones you posted are with bmi. Only |
|
I should maybe have said this earlier, but my CPU does not have the BMI2 instructions. |
|
@greatroar Ah, ok :) Actually that is great since it is much more relevant that there is a speedup on your cpu than mine.
|
klauspost
approved these changes
Jul 12, 2022
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.
This produces the same number of instructions, while requiring less generating code. Benchmarks on the Intel Core i7-3770K show a tiny speedup: