benwtrent opened a new pull request, #12694:
URL: https://github.com/apache/lucene/pull/12694
{DRAFT}
After finalizing work and merging:
https://github.com/apache/lucene/pull/12582
Investigation on if adding unsigned vector operations should occur.
Quantizing within `[0-255]` can reduce error. However, panama vector operations
over unsigned bytes is slightly more expensive (see JMH benchmarks below). Need
to benchmark recall vs. latency over some data sets to verify if this is worth
it or not.
<details>
<summary> M1 (AMD 128 NEON) </summary>
```
Benchmark (size) Mode Cnt
Score Error Units
VectorUtilBenchmark.binaryCosineScalar 128 thrpt 5
8.369 ± 0.208 ops/us
VectorUtilBenchmark.binaryCosineScalar 207 thrpt 5
5.124 ± 0.210 ops/us
VectorUtilBenchmark.binaryCosineScalar 256 thrpt 5
4.193 ± 0.014 ops/us
VectorUtilBenchmark.binaryCosineScalar 1024 thrpt 5
1.043 ± 0.002 ops/us
VectorUtilBenchmark.binaryCosineUnsignedScalar 128 thrpt 5
8.359 ± 0.100 ops/us
VectorUtilBenchmark.binaryCosineUnsignedScalar 207 thrpt 5
5.193 ± 0.025 ops/us
VectorUtilBenchmark.binaryCosineUnsignedScalar 256 thrpt 5
4.194 ± 0.015 ops/us
VectorUtilBenchmark.binaryCosineUnsignedScalar 1024 thrpt 5
1.043 ± 0.002 ops/us
VectorUtilBenchmark.binaryCosineUnsignedVector 128 thrpt 5
21.068 ± 0.072 ops/us
VectorUtilBenchmark.binaryCosineUnsignedVector 207 thrpt 5
12.901 ± 0.041 ops/us
VectorUtilBenchmark.binaryCosineUnsignedVector 256 thrpt 5
11.595 ± 0.128 ops/us
VectorUtilBenchmark.binaryCosineUnsignedVector 1024 thrpt 5
3.197 ± 0.007 ops/us
VectorUtilBenchmark.binaryCosineVector 128 thrpt 5
23.552 ± 0.081 ops/us
VectorUtilBenchmark.binaryCosineVector 207 thrpt 5
14.358 ± 0.077 ops/us
VectorUtilBenchmark.binaryCosineVector 256 thrpt 5
13.165 ± 0.053 ops/us
VectorUtilBenchmark.binaryCosineVector 1024 thrpt 5
3.681 ± 0.027 ops/us
VectorUtilBenchmark.binaryDotProductScalar 128 thrpt 5
25.125 ± 0.043 ops/us
VectorUtilBenchmark.binaryDotProductScalar 207 thrpt 5
15.512 ± 0.061 ops/us
VectorUtilBenchmark.binaryDotProductScalar 256 thrpt 5
12.557 ± 0.044 ops/us
VectorUtilBenchmark.binaryDotProductScalar 1024 thrpt 5
3.110 ± 0.029 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedScalar 128 thrpt 5
25.115 ± 0.082 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedScalar 207 thrpt 5
15.518 ± 0.039 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedScalar 256 thrpt 5
12.554 ± 0.037 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedScalar 1024 thrpt 5
3.112 ± 0.011 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedVector 128 thrpt 5
38.071 ± 0.060 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedVector 207 thrpt 5
25.039 ± 0.120 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedVector 256 thrpt 5
20.578 ± 0.062 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedVector 1024 thrpt 5
5.465 ± 0.008 ops/us
VectorUtilBenchmark.binaryDotProductVector 128 thrpt 5
45.923 ± 0.150 ops/us
VectorUtilBenchmark.binaryDotProductVector 207 thrpt 5
30.516 ± 0.053 ops/us
VectorUtilBenchmark.binaryDotProductVector 256 thrpt 5
25.510 ± 0.053 ops/us
VectorUtilBenchmark.binaryDotProductVector 1024 thrpt 5
6.744 ± 0.046 ops/us
```
</details>
<details>
<summary> GCP AVX512 </summary>
```
Benchmark (size) Mode Cnt
Score Error Units
VectorUtilBenchmark.binaryCosineScalar 128 thrpt 5
7.290 ± 0.003 ops/us
VectorUtilBenchmark.binaryCosineScalar 207 thrpt 5
4.236 ± 0.015 ops/us
VectorUtilBenchmark.binaryCosineScalar 256 thrpt 5
3.452 ± 0.015 ops/us
VectorUtilBenchmark.binaryCosineScalar 1024 thrpt 5
0.885 ± 0.003 ops/us
VectorUtilBenchmark.binaryCosineUnsignedScalar 128 thrpt 5
7.304 ± 0.007 ops/us
VectorUtilBenchmark.binaryCosineUnsignedScalar 207 thrpt 5
4.225 ± 0.013 ops/us
VectorUtilBenchmark.binaryCosineUnsignedScalar 256 thrpt 5
3.431 ± 0.026 ops/us
VectorUtilBenchmark.binaryCosineUnsignedScalar 1024 thrpt 5
0.879 ± 0.006 ops/us
VectorUtilBenchmark.binaryCosineUnsignedVector 128 thrpt 5
29.931 ± 0.049 ops/us
VectorUtilBenchmark.binaryCosineUnsignedVector 207 thrpt 5
17.284 ± 0.018 ops/us
VectorUtilBenchmark.binaryCosineUnsignedVector 256 thrpt 5
19.145 ± 0.067 ops/us
VectorUtilBenchmark.binaryCosineUnsignedVector 1024 thrpt 5
6.109 ± 0.004 ops/us
VectorUtilBenchmark.binaryCosineVector 128 thrpt 5
32.736 ± 0.027 ops/us
VectorUtilBenchmark.binaryCosineVector 207 thrpt 5
18.272 ± 0.640 ops/us
VectorUtilBenchmark.binaryCosineVector 256 thrpt 5
21.435 ± 0.051 ops/us
VectorUtilBenchmark.binaryCosineVector 1024 thrpt 5
7.029 ± 0.011 ops/us
VectorUtilBenchmark.binaryDotProductScalar 128 thrpt 5
16.971 ± 0.053 ops/us
VectorUtilBenchmark.binaryDotProductScalar 207 thrpt 5
9.508 ± 0.091 ops/us
VectorUtilBenchmark.binaryDotProductScalar 256 thrpt 5
8.121 ± 0.059 ops/us
VectorUtilBenchmark.binaryDotProductScalar 1024 thrpt 5
2.501 ± 0.011 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedScalar 128 thrpt 5
16.977 ± 0.056 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedScalar 207 thrpt 5
10.448 ± 0.045 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedScalar 256 thrpt 5
8.352 ± 0.042 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedScalar 1024 thrpt 5
2.502 ± 0.042 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedVector 128 thrpt 5
69.663 ± 0.079 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedVector 207 thrpt 5
44.077 ± 0.059 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedVector 256 thrpt 5
41.963 ± 0.030 ops/us
VectorUtilBenchmark.binaryDotProductUnsignedVector 1024 thrpt 5
11.856 ± 0.020 ops/us
VectorUtilBenchmark.binaryDotProductVector 128 thrpt 5
85.247 ± 0.175 ops/us
VectorUtilBenchmark.binaryDotProductVector 207 thrpt 5
48.486 ± 0.055 ops/us
VectorUtilBenchmark.binaryDotProductVector 256 thrpt 5
50.560 ± 0.045 ops/us
VectorUtilBenchmark.binaryDotProductVector 1024 thrpt 5
14.697 ± 0.010 ops/us
```
</details>
--
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
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
