[vectorIntrinsics] Withdrawn: (Draft) Perofrmance improvements for polluted cases

[Radoslaw Smogura]

On Fri, 6 Aug 2021 18:22:35 GMT, Radoslaw Smogura <duke at openjdk.org> wrote:

> Hi all,
> 
> I would like to submit this piece of work, for byte buffers and polluted cases. It resolves some performance issues related to mem barriers when in scope are both on- and off-heap buffer.
> 
> This is not the most ideal as I could not connect it with loop unrolling (other PR). However, it's still 3x faster than the normal case.
> 
> The approach uses subtype check to switch loop over on / off heap kind, and void true nullable base in byte buffer.
> 
> After
> 
> Benchmark                                     (size)  Mode  Cnt   Score   Error  Units
> ByteBufferVectorAccess.arrayCopy                1024  avgt   30  19.647 ? 0.235  ns/op
> ByteBufferVectorAccess.directBuffers            1024  avgt   30  19.384 ? 0.099  ns/op
> ByteBufferVectorAccess.heapBuffers              1024  avgt   30  17.020 ? 1.273  ns/op
> ByteBufferVectorAccess.heapBuffersNotInlined    1024  avgt   30  17.755 ? 1.814  ns/op
> ByteBufferVectorAccess.pollutedBuffers2         1024  avgt   30  46.248 ? 0.685  ns/op
> ByteBufferVectorAccess.pollutedBuffers3         1024  avgt   30  92.144 ? 2.336  ns/op
> ByteBufferVectorAccess.pollutedBuffers4         1024  avgt   30  92.182 ? 3.180  ns/op
> ByteBufferVectorAccess.pollutedBuffers5         1024  avgt   30  94.900 ? 2.619  ns/op
> 
> 
> Before
> 
> Benchmark                                (size)  Mode  Cnt    Score   Error  Units
> ByteBufferVectorAccess.arrayCopy           1024  avgt   30   14.772 ? 0.115  ns/op
> ByteBufferVectorAccess.directBuffers       1024  avgt   30   18.967 ? 0.036  ns/op
> ByteBufferVectorAccess.heapBuffers         1024  avgt   30   14.783 ? 0.128  ns/op
> ByteBufferVectorAccess.pollutedBuffers2    1024  avgt   30  109.559 ? 1.164  ns/op
> ByteBufferVectorAccess.pollutedBuffers3    1024  avgt   30  219.787 ? 2.055  ns/op
> ByteBufferVectorAccess.pollutedBuffers4    1024  avgt   30  218.920 ? 2.123  ns/op
> ByteBufferVectorAccess.pollutedBuffers5    1024  avgt   30  219.689 ? 1.874  ns/op

This pull request has been closed without being integrated.

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

PR: https://git.openjdk.org/panama-vector/pull/109