[PATCH v2 2/5] benchtests: Add memset zero fill benchtest

[naohirot@fujitsu.com]

Hi Noah,

> I see. I think 16 for the inner loop makes sense. From the x86_64
> perspective this
> will keep the loop from running out of the LSD which is necessary for
> accurate
> benchmarking. I guess then somewhere between [2, 8] is reasonable for the
> outer
> loop?
> 
> 
> > #define START_SIZE (16 * 1024)
> > ...
> > static void
> > __attribute__((noinline, noclone))
> > do_one_test (json_ctx_t *json_ctx, impl_t *impl, CHAR *s,
> >              int c1 __attribute ((unused)), int c2 __attribute ((unused)),
> >              size_t n)
> > {
> >   size_t i, j, iters = INNER_LOOP_ITERS; // 32;
> >   timing_t start, stop, cur, latency = 0;
> >
> >   for (i = 0; i < 512; i++) // for (i = 0; i < 2; i++)
> >     {
> >
> >       CALL (impl, s, c1, n * 16);
> >       TIMING_NOW (start);
> >       for (j = 0; j < 16; j++)
> >         CALL (impl, s + n * j, c2, n);
> >       TIMING_NOW (stop);
> >       TIMING_DIFF (cur, start, stop);
> >       TIMING_ACCUM (latency, cur);
> >     }
> >
> This looks good. But as you said, a much smaller value for outer loop.

I made one improvement that replaced 
  CALL (impl, s, c1, n * 16);
to
  __builtin_memset (s, c1, n * 16);
and tentatively chose outer loop two times such as the followings:

-----
static void
__attribute__((noinline, noclone))
do_one_test (json_ctx_t *json_ctx, impl_t *impl, CHAR *s,
             int c1 __attribute ((unused)), int c2 __attribute ((unused)),
             size_t n)
{
  size_t i, j, iters = 32;
  timing_t start, stop, cur, latency = 0;

  for (i = 0; i < 2; i++)
    {
      __builtin_memset (s, c1, n * 16);
      TIMING_NOW (start);
      for (j = 0; j < 16; j++)
        CALL (impl, s + n * j, c2, n);
      TIMING_NOW (stop);
      TIMING_DIFF (cur, start, stop);
      TIMING_ACCUM (latency, cur);
    }

  json_element_double (json_ctx, (double) latency / (double) iters);
}
-----

In case of __memset_generic on a64fx, execution of outer loop 8times
and 2times took as follows:

8times
real    0m26.236s
user    0m18.806s
sys     0m6.562s

2times
real    0m12.956s
user    0m5.081s
sys     0m6.594s

The performance difference is shown in a comparison graph [1],
there is a difference at 16KB.
This difference would not be critical if we use the performance data
mainly to compare "before" with "after" such as master version of
memset with patched version of memset.


This graph[1] can be drawn as the following:

$ cat 2times/bench-memset-zerofill.out 8times/bench-memset-zerofill.out | \
> merge_strings4graph.sh __memset_generic 2times 8times | \
> plot_strings.py -l -p thru -v -


In order to use __builtin_memset() and create the comparison graph [1],
I submitted two ground work patches [2][3].

[1] https://drive.google.com/file/d/1vD1VE3pdHLoYdaAMWXtImvDlGFDHYkyx/view?usp=sharing
[2] https://sourceware.org/pipermail/libc-alpha/2021-July/129459.html
[3] https://sourceware.org/pipermail/libc-alpha/2021-July/129460.html

Thanks.
Naohiro