tang-hi opened a new issue, #12396: URL: https://github.com/apache/lucene/issues/12396
### Description Since the introduction of Vector API into Lucene via #12311, I have found it to be an interesting tool. As a result, I have attempted to use it to rewrite the [ForUtil.java](lucene/core/src/java/org/apache/lucene/codecs/lucene90/ForUtil.java) file. Initially, I attempted to port [simdcomp](https://github.com/lemire/simdcomp/blob/0c900190bbc0970f4342f2ea778564c95ca725cf/src/simdbitpacking.c) and implement pack and unpack for bitPerValue values that were less than or equal to 8. The results of my efforts can be seen below. | Benchmark | Mode | Cnt | Score | Error | Units | |-----------------------------|------|-------------|----------------|---------------|-------| | VectorizeBenchmark.decode1 | thrpt | 15 | 58293945.447 | ± 1524549.587 | ops/s | | VectorizeBenchmark.decode2 | thrpt | 15 | 55598229.538 | ± 4516920.135 | ops/s | | VectorizeBenchmark.decode3 | thrpt | 15 | 57163871.965 | ± 1566246.490 | ops/s | | VectorizeBenchmark.decode4 | thrpt | 15 | 55128874.528 | ± 4752397.170 | ops/s | | VectorizeBenchmark.decode5 | thrpt | 15 | 53822335.729 | ± 4599217.489 | ops/s | | VectorizeBenchmark.decode6 | thrpt | 15 | 48155246.120 | ± 7519360.551 | ops/s | | VectorizeBenchmark.decode7 | thrpt | 15 | 50253799.192 | ± 820075.648 | ops/s | | VectorizeBenchmark.decode8 | thrpt | 15 | 68849728.856 | ± 1818468.973 | ops/s | | VectorizeBenchmark.encode1 | thrpt | 15 | 33998510.772 | ± 2924618.992 | ops/s | | VectorizeBenchmark.encode2 | thrpt | 15 | 43238190.552 | ± 810373.966 | ops/s | | VectorizeBenchmark.encode3 | thrpt | 15 | 36613553.485 | ± 483115.838 | ops/s | | VectorizeBenchmark.encode4 | thrpt | 15 | 45675726.831 | ± 1081153.655 | ops/s | | VectorizeBenchmark.encode5 | thrpt | 15 | 33591855.278 | ± 1084009.112 | ops/s | | VectorizeBenchmark.encode6 | thrpt | 15 | 36110726.127 | ± 767075.709 | ops/s | | VectorizeBenchmark.encode7 | thrpt | 15 | 34754339.379 | ± 275025.123 | ops/s | | VectorizeBenchmark.encode8 | thrpt | 15 | 55075742.358 | ± 991165.320 | ops/s | | VectorizeBenchmark.vectorizedDecode1 | thrpt | 15 | 43878020.796 | ± 7148545.623 | ops/s | | VectorizeBenchmark.vectorizedDecode2 | thrpt | 15 | 103091446.773 | ± 44115190.011 | ops/s | | VectorizeBenchmark.vectorizedDecode3 | thrpt | 15 | 83168059.373 | ± 24930903.852 | ops/s | | VectorizeBenchmark.vectorizedDecode4 | thrpt | 15 | 63156089.355 | ± 15039408.293 | ops/s | | VectorizeBenchmark.vectorizedDecode5 | thrpt | 15 | 96567546.695 | ± 37142784.493 | ops/s | | VectorizeBenchmark.vectorizedDecode6 | thrpt | 15 | 73897063.180 | ± 11549757.437 | ops/s | | VectorizeBenchmark.vectorizedDecode7 | thrpt | 15 | 79716185.567 | ± 29990852.039 | ops/s | | VectorizeBenchmark.vectorizedDecode8 | thrpt | 15 | 92621676.617 | ± 29702056.667 | ops/s | | VectorizeBenchmark.vectorizedEncode1 | thrpt | 15 | 51140300.852 | ± 139758.385 | ops/s | | VectorizeBenchmark.vectorizedEncode2 | thrpt | 15 | 82646100.574 | ± 1289600.954 | ops/s | | VectorizeBenchmark.vectorizedEncode3 | thrpt | 15 | 88124485.953 | ± 742170.198 | ops/s | | VectorizeBenchmark.vectorizedEncode4 | thrpt | 15 | 91029285.467 | ± 5594858.437 | ops/s | | VectorizeBenchmark.vectorizedEncode5 | thrpt | 15 | 96843051.648 | ± 8024430.836 | ops/s | | VectorizeBenchmark.vectorizedEncode6 | thrpt | 15 | 98596724.128 | ± 10068466.227 | ops/s | | VectorizeBenchmark.vectorizedEncode7 | thrpt | 15 | 85885746.715 | ± 6031740.563 | ops/s | | VectorizeBenchmark.vectorizedEncode8 | thrpt | 15 | 117139889.194 | ± 8721517.095 | ops/s | However, I noticed that the compression format used in [ForUtil.java](lucene/core/src/java/org/apache/lucene/codecs/lucene90/ForUtil.java) was different. It employed some tricks to speed up the process, such as simd. Therefore, I attempted to vectorize it while maintaining the compression format. The results can be seen below. | Benchmark | Mode | Cnt | Score | Error | Units | |-----------------------------|------|-----|----------------|----------------|-------| | Benchmark.encode1 | thrpt | 15 | 38017254.040 | ± 3905466.628 | ops/s | | Benchmark.encode2 | thrpt | 15 | 45170109.395 | ± 1539203.478 | ops/s | | Benchmark.encode3 | thrpt | 15 | 38757256.653 | ± 1044709.221 | ops/s | | Benchmark.encode4 | thrpt | 15 | 49307206.891 | ± 799168.007 | ops/s | | Benchmark.encode5 | thrpt | 15 | 35130626.548 | ± 792210.817 | ops/s | | Benchmark.encode6 | thrpt | 15 | 38326892.073 | ± 981865.963 | ops/s | | Benchmark.encode7 | thrpt | 15 | 37372342.721 | ± 1177478.683 | ops/s | | Benchmark.encode8 | thrpt | 15 | 60757390.416 | ± 458876.638 | ops/s | | Benchmark.vectorizedEncode1 | thrpt | 15 | 56413094.655 | ± 435917.201 | ops/s | | Benchmark.vectorizedEncode2 | thrpt | 15 | 88770400.646 | ± 11183716.176 | ops/s | | Benchmark.vectorizedEncode3 | thrpt | 15 | 39932842.378 | ± 2366921.921 | ops/s | | Benchmark.vectorizedEncode4 | thrpt | 15 | 85888128.739 | ± 5499354.172 | ops/s | | Benchmark.vectorizedEncode5 | thrpt | 15 | 34402027.732 | ± 1414839.159 | ops/s | | Benchmark.vectorizedEncode6 | thrpt | 15 | 35794303.501 | ± 782940.005 | ops/s | | Benchmark.vectorizedEncode7 | thrpt | 15 | 33845690.180 | ± 2586648.353 | ops/s | | Benchmark.vectorizedEncode8 | thrpt | 15 | 97914288.675 | ± 8971857.035 | ops/s | I have only implemented bitPerValue values of 1, 2, 4, and 8. I am curious if it is possible to change the compression format. Additionally, do you have any best practices for integrating vectorized code into Lucene? Any suggestions would be appreciated. Currently, I am working on my own [repo](https://github.com/tang-hi/forutil). However, the code is still in a rough state and lacks documentation. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
