[vectorIntrinsics] RFR: Improve mask reduction operations on AVX

[Mai Đặng Quân Anh]

On Tue, 2 Nov 2021 10:58:19 GMT, Mai Đặng Quân Anh <duke at openjdk.java.net> wrote:

> Hi,
> This patch improves the logic of vector mask reduction operations on AVX, especially int, float, long, double, by using vmovmskpd and vmovmskps instructions. I also do a little refactoring to reduce duplication in toLong. The patch temporarily disables these operations on Neon, though.
> Thank you very much.

Microbenchmark shows significant improvement on my AVX2 machine.

    Before:
    Benchmark               Mode  Cnt   Score   Error  Units
    MaskReduction.byte128   avgt   25   0.773 ± 0.006  ns/op
    MaskReduction.byte256   avgt   25   0.778 ± 0.007  ns/op
    MaskReduction.int128    avgt   25   1.061 ± 0.008  ns/op
    MaskReduction.int256    avgt   25   1.553 ± 0.010  ns/op
    MaskReduction.long128   avgt   25  43.008 ± 0.354  ns/op
    MaskReduction.long256   avgt   25   1.271 ± 0.006  ns/op
    MaskReduction.short128  avgt   25   0.989 ± 0.006  ns/op
    MaskReduction.short256  avgt   25   0.919 ± 0.005  ns/op

    After:
    Benchmark               Mode  Cnt   Score   Error  Units
    MaskReduction.byte128   avgt   25   0.566 ± 0.001  ns/op
    MaskReduction.byte256   avgt   25   0.556 ± 0.003  ns/op
    MaskReduction.int128    avgt   25   0.553 ± 0.003  ns/op
    MaskReduction.int256    avgt   25   0.828 ± 0.002  ns/op
    MaskReduction.long128   avgt   25  41.618 ± 0.241  ns/op
    MaskReduction.long256   avgt   25   0.552 ± 0.001  ns/op
    MaskReduction.short128  avgt   25   0.775 ± 0.004  ns/op
    MaskReduction.short256  avgt   25   0.834 ± 0.006  ns/op

The benchmark simply loads 2 constant vector, compares them and returns the toLong of the result as follow:

    return BYTE_128_1.eq(BYTE_128_2).toLong();

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

PR: https://git.openjdk.java.net/panama-vector/pull/158