RFR: 8321010: RISC-V: C2 RoundVF [v2]

Ludovic Henry luhenry at openjdk.org
Mon Sep 2 08:40:21 UTC 2024


On Mon, 22 Apr 2024 03:09:16 GMT, Fei Yang <fyang at openjdk.org> wrote:

>>> Hello Hamlin, I recall you had licheepi board. I would be nice if you can try to measure rvv performance gain with this https://github.com/syntacore/syntaj21/tree/rvv0.7.1
>>> 
>>> This PR showed it's not always easy to win perf just by using rvv - #17413
>>> 
>>> I understand it might not be possible, but would be nice to give it a try (I can share hsdis with support for 0.7.1 if needed)
>> 
>> Had a internal discussion about your suggestion, seems 0.7.1 is not incompatible with 1.0/2.0, and for this simple intrinsic, we think a better path is to have it first, then re-visit it when we have real hardware to measure the performance later.
>
> @Hamlin-Li: Thanks for the quick update. Considering that saving/restoring for FRM could be expensive, I do wonder if we could gather some performance numbers before we go. I see people are now testing on RVV-1.0 hardwares [1] and I am also trying to get one (AFAIK, more powerful RVV-1.0 hardwares are also coming later this year, SG2044, SG2380, etc.). Also from discussion on [2], I see there are also other approaches available there without flipping the FP rounding mode. But I am not sure if they make sense for our case or work better without actual testing.
> 
> [1] https://github.com/openjdk/jdk/pull/18382#issuecomment-2045145255
> [2] https://github.com/openjdk/jdk/pull/8204

@RealFYang @turbanoff could we please have another review? Thank you!

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

PR Comment: https://git.openjdk.org/jdk/pull/17745#issuecomment-2324153121


More information about the hotspot-compiler-dev mailing list