[vectorIntrinsics] Issue to VectorAPI "selectFrom" for byte type with Arm SVE 2048-bits
Xiaohong Gong
Xiaohong.Gong at arm.com
Thu Oct 29 06:52:10 UTC 2020
Hi Paul,
> > Thanks very much for your detailed explanation!
> >
> >> 1) the selectFrom(Vector ) methods will never be able to index
> anything greater than the 127th lane element, if supported by the
> species. I think we would require the notion of a Vector whose lane
> elements are unsigned bytes, which I think requires Valhalla and
> additional numeric types.
> > This limitation would obviously become even more apparent if say
> a future SVE specification supported vector shapes of 4096 bits.
> >
> > Yes, I agree with you. And it also makes sense to me that only
> the unsigned bytes can be the valid elements for this API, even the
> larger vector shapes might be supported like SVE 2048-bits.
> > So regarding to the test failures, I think we'd better to limit
> the lane elements to the valid values to make the tests pass. I
> will raise a patch for it if it's ok to you.
> >
>
> Yes, clamp to Byte.MAX_VALUE for a maximum valid index on relevant
> platforms.
>
OK, here is the PR to the tests: https://github.com/openjdk/panama-vector/pull/19.
It's a great pleasure if you could have a look at it. Thanks!
> >> Because of 3) we cannot directly create a VectorShuffle for a
> 2048:byte species that supports indexes > 127. That seems fixable
> by modifying the implementation of ByteMaxShuffle to hold short[]
> instead of byte[], or byte[] twice the length etc. That may require
> some care with regards to the intrinsic. This seems possible.
> >
> > Yes, exactly. Currently we have created a patch to fix this
> issue. And the main idea is the same with https://mail.openjdk.java.net/pipermail/panama-dev/2020-August/010384.html.
> >
>
> Ok, I’ll wait for a PR, but I wonder if its possible to focus the
> code more on the ByteMax implementations.
>
> Paul;.
Thanks for your suggestion! Yeah, it's better to make the codes focus more on the
ByteMax related parts. And we will also try to find a better way for it.
Thanks,
Xiaohong Gong
More information about the panama-dev
mailing list