RFR: 8264109: Add vectorized implementation for VectorMask.andNot()

Paul Sandoz psandoz at openjdk.java.net
Wed Mar 31 16:45:26 UTC 2021


On Fri, 26 Mar 2021 01:50:33 GMT, Xiaohong Gong <xgong at openjdk.org> wrote:

> Currently "VectorMask.andNot()" is not vectorized:
>     public VectorMask<E> andNot(VectorMask<E> m) {
>         // FIXME: Generate good code here.
>         return bOp(m, (i, a, b) -> a && !b);
>     }
> This can be implemented with` "and(m.not())" `directly. Since `"VectorMask.and()/not()" `have been vectorized in hotspot, `"andNot"`
> can be vectorized as well by calling them.
> 
> The performance gain is >100% for such a simple JMH:
>   @Benchmark
>   public Object andNot(Blackhole bh) {
>     boolean[] mask = fm.apply(SPECIES.length());
>     boolean[] r = fmt.apply(SPECIES.length());
>     VectorMask<Byte> rm = VectorMask.fromArray(SPECIES, r, 0);
> 
>     for (int ic = 0; ic < INVOC_COUNT; ic++) {
>       for (int i = 0; i < mask.length; i += SPECIES.length()) {
>         VectorMask<Byte> vmask = VectorMask.fromArray(SPECIES, mask, i);
>           rm = rm.andNot(vmask);
>         }
>     }
>     return rm;
>   }

Would you mind updating the existing `AbstractMask.andNot` implementation? rather than changing `VectorMask.andNot`. That fits in with the current implementation pattern.

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

PR: https://git.openjdk.java.net/jdk/pull/3211


More information about the core-libs-dev mailing list