[vectorIntrinsics] RFR: Optimize mem barriers for ByteBuffer cases

[Andrew Haley]

On 8/2/21 5:30 PM, Radosław Smogura wrote:
> 
> Here are my results from cloud without unrolling fix and with unrolling fix cherry picked (1 fork, 0.5 sec test):
> 
> Benchmark                                (size)   Mode  Cnt      Score     Error  Units
> ByteBufferVectorAccess.directBuffers     524288  thrpt   10  35105.776 ? 166.178  ops/s
> ByteBufferVectorAccess.heapBuffers       524288  thrpt   10  36401.373 ? 102.513  ops/s
> 
> 
> Benchmark                                (size)   Mode  Cnt      Score     Error  Units
> ByteBufferVectorAccess.directBuffers     524288  thrpt   10  70723.249 ? 548.895  ops/s
> ByteBufferVectorAccess.heapBuffers       524288  thrpt   10  69522.365 ? 891.461  ops/s
> 
> I don't know capacity of this offering.

That's fab. Looks like we're hitting the bandwidth of the memory system! Good job.

-- 
Andrew Haley  (he/him)
Java Platform Lead Engineer
Red Hat UK Ltd. <https://www.redhat.com>
https://keybase.io/andrewhaley
EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671