Re: [PATCH v3 2/5] benchtests: Add memset zero fill benchtest

["Lucas A. M. Magalhaes via Libc-alpha" <libc-alpha@sourceware.org>]

Hi Wilco,
> > By doing this you are measuring just the call it self and accumulating
> > the results. This is indeed not measurable for really small lengths.
> > You could try moving the memset and the timing out of the loop and
> > measure the time spent in multiple runs. To fix the memset you could
> > memset a bigger buffer and move the s pointer on each loop. I guess this
> > will reduce the variations Wilco mentioned.
> 
> That would basically end up the same as bench-memset-walk.c given that
> you need a huge buffer to get reasonable accuracy (bench-memset does
> 8192 iterations by default, and that is still inaccurate for small sizes).
> In that case it would be easier to improve bench-memset-walk.c rather than
> adding yet another benchmark that is too inaccurate to be useful.
Yeah, I agree with you.
> 
> Alternatively we could use the timing loop I suggested which allows any
> pattern of zero/non-zero to be tested accurately:
> 
>       TIMING_NOW (start);
>       for (j = 0; j < iters; j++)
>         CALL (impl, s, memset_value[j & MASK], n);
>       TIMING_NOW (stop);
> 
Sorry but I suppose don't understood your suggestion completely.  The
memset_value array will hold patterns like [0,0], [0,1] or [1,1],
right?  If so, this will not work to measure the zero-to-one pattern for
example, as it will be mixing zero-to-one with one-to-zero calls. In
order to measure just an specific patter the buffer must be loaded
previously of the timing loop.

---
Lucas A. M. Magalhães