On Fri, 5 Sep 2025 10:12:35 GMT, Aleksey Shipilev <[email protected]> wrote:
>> Intrinsic support for `VectorMask.laneIsSet` with a **variable** input index
>> was introduced in PR #14200, but was inadvertently broken by PR #25673. This
>> PR restores the intrinsic functionality and adds some JTReg tests.
>>
>> Benchmarks on Nvidia Grace machine with 128-bit SVE:
>>
>> Benchmark Unit Before Score
>> Error After Score Error Uplift
>> microMaskLaneIsSetByte128_var ops/ms 21702.14415 91.902159
>> 103472.9391 36.057447 4.767867
>> microMaskLaneIsSetByte64_var ops/ms 21468.51868 107.94177
>> 103365.6561 69.47736 4.814754
>> microMaskLaneIsSetDouble128_var ops/ms 77489.32791 153.242699
>> 413499.4127 311.854079 5.336211
>> microMaskLaneIsSetFloat128_var ops/ms 41034.95204 399.421823
>> 206840.0988 74.702234 5.040583
>> microMaskLaneIsSetFloat64_var ops/ms 77607.40268 175.938921
>> 413745.3001 149.716794 5.33126
>> microMaskLaneIsSetInt128_var ops/ms 41452.48893 76.143208
>> 206845.9754 59.371129 4.989953
>> microMaskLaneIsSetInt64_var ops/ms 77726.2542 173.180518
>> 413427.8838 363.575023 5.319024
>> microMaskLaneIsSetLong128_var ops/ms 77646.11218 177.496587
>> 413403.4404 236.609314 5.3242
>> microMaskLaneIsSetShort128_var ops/ms 21374.93265 48.13101
>> 103417.4618 34.827021 4.838259
>> microMaskLaneIsSetShort64_var ops/ms 41066.19395 353.320621
>> 206801.109 106.408938 5.035799
>>
>>
>> Benchmarks on Intel 6444y machine with 512-bit avx3:
>>
>> Benchmark Unit Before Score
>> Error After Score Error Uplift
>> microMaskLaneIsSetByte128_var ops/ms 57658.45497 240.209309
>> 211643.8406 29.214532 3.670647
>> microMaskLaneIsSetByte256_var ops/ms 57451.68169 116.994128
>> 211609.4652 160.48513 3.683259
>> microMaskLaneIsSetByte512_var ops/ms 57530.22411 311.63868
>> 199802.8084 408.144015 3.473005
>> microMaskLaneIsSetByte64_var ops/ms 57642.2672 161.406221
>> 205252.4464 196.86852 3.560797
>> microMaskLaneIsSetDouble256_var ops/ms 114401.3789 231.797375
>> 361400.344 565.593984 3.159055
>> microMaskLaneIsSetDouble512_var ops/ms 57379.27882 159.699503
>> 211476.1138 136.980026 3.685583
>> microMaskLaneIsSetFloat128_var ops/ms 113943.9512 141.062663
>> 360855.3915 494.471996 3.166955
>> microMaskLaneIsSetFloat256_var ops/ms 57682.78182 138.142053
>> 211659.5098 30.167972 3.66937
>> microMaskLaneIsSetFloat512_var ops/ms 57617.66405 301.748599
>> 211246.8588 597.18949 3.666355
>> microMaskLaneIsSetInt128_var ops/ms 113914.5062 118.681382
>> 360856.4465 555.097397 3.167783
>> microMaskLaneIsSetInt256_var ops/ms 57681.79883 112.391639
>> 211555.6742 217.556981 3.667633
>> microMaskLaneIsSetInt512_var ops/ms 573...
>
> test/micro/org/openjdk/bench/jdk/incubator/vector/VectorExtractBenchmark.java
> line 34:
>
>> 32: @Warmup(iterations = 5, time = 1)
>> 33: @Measurement(iterations = 5, time = 1)
>> 34: @Fork(value = 1, jvmArgs = {"--add-modules=jdk.incubator.vector"})
>
> Don't do 1 fork, do at least 3.
The test results show that this test is stable, so I think forking once is
enough? We have many JMH benchmarks that fork once.
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/27113#discussion_r2328949227
