8215262: PPC64: FMA Vectorization on PPC64
gromero at linux.vnet.ibm.com
Wed Dec 12 22:39:06 UTC 2018
Hi Michi and Martin,
On 12/12/2018 01:53 PM, Doerr, Martin wrote:
> this looks good to me. We should be able to test it since it is for Power8. Maybe Gustavo would like to test in addition.
Michi, the change looks good to me too.
I've tested it on POWER8 and on POWER9 and observed a reduction of ~30% on my
POWER8, but on my POWER9 (Witherspoon) it was ~80%, so it can be even better on
a Boston (POWER9 too) machine. Thus, thumbs up from my side.
I can sponsor that change but I'll wait an ack. from SAP that it's ok for
pushing. Maybe Martin wants to perform additional tests after his first review
or others might want to review too... I'm wondering if it would be reasonable
to run JCK/TCK against it. Would it test Math.fma() since it's a 9+ method and
needs a warmed up JVM? Unfortunately I can't run the JCK/TCK against it myself
if it applies here.
> Best regards,
> *From:*Michihiro Horie <HORIE at jp.ibm.com>
> *Sent:* Mittwoch, 12. Dezember 2018 15:50
> *To:* ppc-aix-port-dev at openjdk.java.net; hotspot compiler <hotspot-compiler-dev at openjdk.java.net>
> *Cc:* Gustavo Romero <gromero at linux.vnet.ibm.com>; Doerr, Martin <martin.doerr at sap.com>; Simonis, Volker <volker.simonis at sap.com>
> *Subject:* RFR: 8215262: PPC64: FMA Vectorization on PPC64
> Dear all,
> Would you please review the following change?
> This change supports FMA vectorization on PPC64.
> Bug: https://bugs.openjdk.java.net/browse/JDK-8215262
> Webrev: http://cr.openjdk.java.net/~mhorie/8215262/webrev.00/ <http://cr.openjdk.java.net/%7Emhorie/8215262/webrev.00/>
> Microbenchmark I used is http://cr.openjdk.java.net/~mhorie/8215262/VFma.java <http://cr.openjdk.java.net/%7Emhorie/8215262/VFma.java>
> It showed ~50% reduction in the execution time.
> I also tested with JTREG and there was no difference in the results compared to the original.
> Best regards,
> IBM Research - Tokyo
More information about the hotspot-compiler-dev