Re: [Rapp-dev] What's the delivery requirements on new vectorized functions regarding benchmarking and tuning?

[Hans-Peter Nilsson]

> From: Willie Betschart <address@hidden>
> Date: Fri, 11 Mar 2016 10:42:32 +0000

> If a patch with new vectorization functionality is submitted,
> how many targets/backends needs to be tuned and benchmarked in
> the patch.

I think you should make an effort to do that for as many targets
as you have available.

> Is it enough with one backend?

Answering in general, I think two is a minimum when adding a new
vectorization feature, even one using existing target macros.
You have that and more even on any non-museum-grade PC, and I
think most developers with an interest in RAPP and vectorization
have access to more than that.

(Answering for you, I think *you* have easy access to all but
three in the current sources, and there's even an old G4 Mac
under my desk for AltiVec if you want to save me some time.
It's probably better than not having tuning at all for recent
PowerPC machines and it's a good sanity-check.)

> Will rapp-dev take
> responsibility to do benchmark tuning on all backends when
> releasing the patch?

You mean will the RAPP project have that as a stop-gap before a
new release of RAPP?  No.

Anyone is welcome to help.  In a release, whatever platforms
have been tuned will have updated tune-files.  There are
fall-backs in place so the release will be functional for the
new feature on the non-tuned platforms even for configurations
where auto-tuning will not work; the new functionality will
simply use the generic function.

brgds, H-P