Re: [PATCH v2 4/6] x86-64: Add vector log/logf to libmvec microbenchmark

[Joseph Myers <joseph@codesourcery.com>]

On Fri, 12 Nov 2021, Sunil Pandey via Libc-alpha wrote:

> On Fri, Nov 12, 2021 at 3:18 PM Joseph Myers <joseph@codesourcery.com>
> wrote:
> 
> > On Fri, 12 Nov 2021, Sunil K Pandey via Libc-alpha wrote:
> >
> > > +# Random inputs in [a=1.10,b=100.00]
> >
> > It seems odd not to have any inputs in (0, 1.10), or larger than 100 (in
> > the absence of actual evidence that such numbers are rare as arguments to
> > log in practice).  All positive numbers are reasonable arguments to log.
> >
> On some cpu there may be a slow path around 0 and 1, hence range 
> selected from 1.10 to avoid slow path which may distort microbenchmark 
> number on some cpu. There is no such issue for numbers larger than 100, 
> we have to pick some number and choose 100 for microbenchmark.

I think the point of such benchmarks should be to provide typical inputs 
to the function such as might be found in ordinary programs using the 
function, and provide a sensible way in which completely different 
implementations of the same function can be compared.  It should not be 
based on the details of where slow paths in a particular implementation 
are; that artificially advantages the current implementation against any 
alternative implementation that might be proposed.

-- 
Joseph S. Myers
joseph@codesourcery.com