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

["naohirot@fujitsu.com" <naohirot@fujitsu.com>]

Hi Wilco, Noah,

> From: Tamura, Naohiro/田村 直広 <naohirot@fujitsu.com>
> Sent: Wednesday, July 28, 2021 4:28 PM
> 
> Taking Noah's comment [1] into account, the final code should be like
> the below. Can we agree with this code?
> 
> Two results, two loop version in the mail [1] and one loop version
> below, are almost same in case of __memset_generic on a64fx as
> shown in the graph [2].
> 
> -----
> 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, iters = 32;
>   timing_t start, stop, cur, latency = 0;
> 
>   CALL (impl, s, c2, n); // warm up
> 
>   for (i = 0; i < iters; i++)
>     {
>       memset (s, c1, n); // alternation
> 
>       TIMING_NOW (start);
> 
>       CALL (impl, s, c2, n);
> 
>       TIMING_NOW (stop);
>       TIMING_DIFF (cur, start, stop);
>       TIMING_ACCUM (latency, cur);
>     }
> 
>   json_element_double (json_ctx, (double) latency / (double) iters);
> }
> -----

I'd like to share an interesting insight which was found when
START_SIZE was changed to smaller size 256 from 16KB.
Currently DC ZVA is called if size is more than 256B and value is zero
in __memset_generic (sysdeps/aarch64/memset.S).
However DC ZVA is slower than store instruction if size is less than
16KB on A64FX[3].
So this would indicate that the appropriate DC ZVA start size might
be different on each CPU.
It would be interesting to see how other CPU behaves.

The code is below, which measures 4 patterns, zero-over-zero,
zero-over-one, one-over-zero and one-over-one from 256B to 64MB.
In the graph [3], 4 patterns are abbreviated 0o0, 0o1, 1o0 and 1o1.


#define START_SIZE 256
#define MIN_PAGE_SIZE (getpagesize () + 64 * 1024 * 1024)

  for (c1 = 0; c1 < 2; c1++)
    for (c2 = 0; c2 < 2; c2++)
      for (i = START_SIZE; i <= MIN_PAGE_SIZE; i <<= 1)
        {
          do_test (&json_ctx, 0, c1, c2, i);
          do_test (&json_ctx, 3, c1, c2, i);
        }

I'd like to submit V3 patch incorporating above change too.

[3] https://drive.google.com/file/d/1fonjDDlF4LPLfZY9-z22DGn-yaSpGN4g/view?usp=sharing

Thanks.
Naohiro

> [1] https://sourceware.org/pipermail/libc-alpha/2021-July/129486.html
> [2] https://drive.google.com/file/d/1bptHqg5vvFAGoYgoR3w_pvclXFSP8Sr0/view?usp=sharing
> 
> Thanks.
> Naohiro