Re: [PATCH] powerpc: add libmvec implementations of log and logf

[Shawn Landden <shawn@git.icu>]

On Tue, May 7, 2019 at 10:21 AM Joseph Myers <joseph@codesourcery.com> wrote:
>
> On Mon, 6 May 2019, Shawn Landden wrote:
>
> > When fed numbers in the range of 0 to 2^32-1 (as doubles) the vector log
> > is about 75% faster than scalar log.
>
> Since you and Bert Tenjy have now posted rival patch versions implementing
> vector log, a key performance comparison would be against the version
> posted by Bert Tenjy
> <https://sourceware.org/ml/libc-alpha/2019-05/msg00054.html> - and in
> general, you should compare and contrast the two versions of log to
> justify why one might be better than another.  It's also important in such
> cases to be very clear about where exactly the patch you are posting fits
> into exactly what other patch series.
>

When I first looked at Bert's patch I saw all the scalar code for the
special path and thought it wasn't a real vector implementation.
Otherwise I wouldn't have tried to compete with him.

HOWEVER, once I fixed the bugs in my benchmark programs, there is no
performance advantage of doing either log or logf with vector
operation, at least on this power 9 VPS---either mine or Bert's
patches. I think this is because they require too many table lookups
for the constants, and Power ISA 3.0 lacks both a gather instruction,
and a permutexvar instruction, both of which could be used to
vectorize these table lookups. (shufflevector can't do it)
> --
> Joseph S. Myers
> joseph@codesourcery.com