Re: [Qemu-devel] [PATCH] tcg/ppc: Add vector opcodes

[David Gibson]

On Fri, Feb 22, 2019 at 05:13:37PM +0000, Mark Cave-Ayland wrote:
> On 22/02/2019 05:59, Richard Henderson wrote:
> 
> > This requires VSX, not just Altivec, so Power7 or later.
> > 
> > Signed-off-by: Richard Henderson <address@hidden>
> > ---
> > 
> > At present there are no tunables that can avoid the 64-bit element
> > load/store requirement.  As with requiring AVX1 for x86 hosts, I'm
> > not sure it's worth inventing such a tunable for pre-power7 hosts.
> > 
> > Tested vs aarch64 risu test cases.  It's probably worth testing
> > this vs Mark's target/ppc conversion.
> 
> Oooh this looks really exciting! However... I only have a G4 Mac Mini around 
> that I
> use for testing which is Altivec-only :(  Is it much work to support
> non-VSX hosts?

I have access to POWER8 and POWER9 machines, but I haven't worked with
RISU before.  If you can give me a straightforward recipe I can try
running the tests.

> 
> This leads me to a related point that came up when Howard and I were testing 
> the PPC
> vector patches - how do we know at runtime which optimisations were being 
> used, e.g.
> what is the value of have_avx2 on a particular CPU running QEMU?
> 
> Under Linux this isn't too bad since you can just do "cat /proc/cpuinfo | 
> grep avx2"
> but it becomes more tricky when getting bug reports from Windows users who 
> aren't
> particularly technical...
> 
> 
> ATB,
> 
> Mark.
> 

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

signature.asc
Description: PGP signature