RFR: 8265128: [REDO] Optimize Vector API slice and unslice operations

Sandhya Viswanathan sviswanathan at openjdk.java.net
Fri Apr 30 00:17:24 UTC 2021


On Thu, 29 Apr 2021 21:29:03 GMT, Sandhya Viswanathan <sviswanathan at openjdk.org> wrote:

> All the slice and unslice variants that take more than one argument can benefit from already intrinsic methods on similar lines as slice(origin) and unslice(origin).
> 
> Changes include:
>  * Rewrite Vector API slice/unslice using already intrinsic methods
>  * Fix in library_call.cpp:inline_preconditions_checkIndex() to not modify control if intrinsification fails
>  * Vector API conversion tests thresholds adjustment
>  
> Base Performance:
> Benchmark (size) Mode Cnt Score Error Units
> TestSlice.vectorSliceOrigin 1024 thrpt 5 11763.372 ± 254.580 ops/ms
> TestSlice.vectorSliceOriginVector 1024 thrpt 5 599.286 ± 326.770 ops/ms
> TestSlice.vectorSliceUnsliceOrigin 1024 thrpt 5 6627.601 ± 22.060 ops/ms
> TestSlice.vectorSliceUnsliceOriginVector 1024 thrpt 5 401.858 ± 220.340 ops/ms
> TestSlice.vectorSliceUnsliceOriginVectorPart 1024 thrpt 5 421.993 ± 231.703 ops/ms
> 
> Performance with patch:
> Benchmark (size) Mode Cnt Score Error Units
> TestSlice.vectorSliceOrigin 1024 thrpt 5 11792.091 ± 37.296 ops/ms
> TestSlice.vectorSliceOriginVector 1024 thrpt 5 8388.174 ± 115.886 ops/ms
> TestSlice.vectorSliceUnsliceOrigin 1024 thrpt 5 6662.159 ± 8.203 ops/ms
> TestSlice.vectorSliceUnsliceOriginVector 1024 thrpt 5 5206.300 ± 43.637 ops/ms
> TestSlice.vectorSliceUnsliceOriginVectorPart 1024 thrpt 5 5194.278 ± 13.376 ops/ms

@PaulSandoz I have implemented your review comments. Please take a look.

-------------

PR: https://git.openjdk.java.net/jdk/pull/3804


More information about the hotspot-compiler-dev mailing list