[PATCH v1 5/5] x86: Double size of ERMS rep_movsb_threshold in dl-cacheinfo.h

[H.J. Lu]

On Fri, Nov 5, 2021 at 9:39 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>
> On Fri, Nov 5, 2021 at 9:32 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> >
> > On Mon, Nov 01, 2021 at 12:49:52AM -0500, Noah Goldstein wrote:
> > > No bug.
> > >
> > > This patch doubles the rep_movsb_threshold when using ERMS. Based on
> > > benchmarks the vector copy loop, especially now that it handles 4k
> > > aliasing, is better for these medium ranged.
> > >
> > > On Skylake with ERMS:
> > >
> > > Size,   Align1, Align2, dst>src,(rep movsb) / (vec copy)
> > > 4096,   0,      0,      0,      0.975
> > > 4096,   0,      0,      1,      0.953
> > > 4096,   12,     0,      0,      0.969
> > > 4096,   12,     0,      1,      0.872
> > > 4096,   44,     0,      0,      0.979
> > > 4096,   44,     0,      1,      0.83
> > > 4096,   0,      12,     0,      1.006
> > > 4096,   0,      12,     1,      0.989
> > > 4096,   0,      44,     0,      0.739
> > > 4096,   0,      44,     1,      0.942
> > > 4096,   12,     12,     0,      1.009
> > > 4096,   12,     12,     1,      0.973
> > > 4096,   44,     44,     0,      0.791
> > > 4096,   44,     44,     1,      0.961
> > > 4096,   2048,   0,      0,      0.978
> > > 4096,   2048,   0,      1,      0.951
> > > 4096,   2060,   0,      0,      0.986
> > > 4096,   2060,   0,      1,      0.963
> > > 4096,   2048,   12,     0,      0.971
> > > 4096,   2048,   12,     1,      0.941
> > > 4096,   2060,   12,     0,      0.977
> > > 4096,   2060,   12,     1,      0.949
> > > 8192,   0,      0,      0,      0.85
> > > 8192,   0,      0,      1,      0.845
> > > 8192,   13,     0,      0,      0.937
> > > 8192,   13,     0,      1,      0.939
> > > 8192,   45,     0,      0,      0.932
> > > 8192,   45,     0,      1,      0.927
> > > 8192,   0,      13,     0,      0.621
> > > 8192,   0,      13,     1,      0.62
> > > 8192,   0,      45,     0,      0.53
> > > 8192,   0,      45,     1,      0.516
> > > 8192,   13,     13,     0,      0.664
> > > 8192,   13,     13,     1,      0.659
> > > 8192,   45,     45,     0,      0.593
> > > 8192,   45,     45,     1,      0.575
> > > 8192,   2048,   0,      0,      0.854
> > > 8192,   2048,   0,      1,      0.834
> > > 8192,   2061,   0,      0,      0.863
> > > 8192,   2061,   0,      1,      0.857
> > > 8192,   2048,   13,     0,      0.63
> > > 8192,   2048,   13,     1,      0.629
> > > 8192,   2061,   13,     0,      0.627
> > > 8192,   2061,   13,     1,      0.62
> > > ---
> > >  sysdeps/x86/dl-cacheinfo.h | 9 ++++++---
> > >  1 file changed, 6 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/sysdeps/x86/dl-cacheinfo.h b/sysdeps/x86/dl-cacheinfo.h
> > > index e6c94dfd02..712b7c7fd0 100644
> > > --- a/sysdeps/x86/dl-cacheinfo.h
> > > +++ b/sysdeps/x86/dl-cacheinfo.h
> > > @@ -871,7 +871,7 @@ dl_init_cacheinfo (struct cpu_features *cpu_features)
> > >    if (CPU_FEATURE_USABLE_P (cpu_features, AVX512F)
> > >        && !CPU_FEATURE_PREFERRED_P (cpu_features, Prefer_No_AVX512))
> > >      {
> > > -      rep_movsb_threshold = 2048 * (64 / 16);
> > > +      rep_movsb_threshold = 4096 * (64 / 16);
> >
> > Please also update the default of x86_rep_stosb_threshold in
>
> Do you know what to set it at?

Oops.  I meant

    x86_rep_movsb_threshold {
      type: SIZE_T
      # Since there is overhead to set up REP MOVSB operation, REP MOVSB
      # isn't faster on short data.  The memcpy micro benchmark in glibc
      # shows that 2KB is the approximate value above which REP MOVSB
      # becomes faster than SSE2 optimization on processors with Enhanced
      # REP MOVSB.  Since larger register size can move more data with a
      # single load and store, the threshold is higher with larger register
      # size.  Note: Since the REP MOVSB threshold must be greater than 8
      # times of vector size and the default value is 2048 * (vector size

       ^^^^^^^
      # / 16), the default value and the minimum value must be updated at
      # run-time.  NB: Don't set the default value since we can't tell if
      # the tunable value is set by user or not [BZ #27069].
      minval: 1
    }

> I haven't tested recently but last time I checked stosb was significantly
> better even for smaller values than movsb. Think it warrants another patch
> as the numbers in this commit are for movsb and I don't think the two are
> necessarily 1-1.
>
> >
> > sysdeps/x86/dl-tunables.list
> >
> > >  #if HAVE_TUNABLES
> > >        minimum_rep_movsb_threshold = 64 * 8;
> > >  #endif
> > > @@ -879,14 +879,14 @@ dl_init_cacheinfo (struct cpu_features *cpu_features)
> > >    else if (CPU_FEATURE_PREFERRED_P (cpu_features,
> > >                                   AVX_Fast_Unaligned_Load))
> > >      {
> > > -      rep_movsb_threshold = 2048 * (32 / 16);
> > > +      rep_movsb_threshold = 4096 * (32 / 16);
> > >  #if HAVE_TUNABLES
> > >        minimum_rep_movsb_threshold = 32 * 8;
> > >  #endif
> > >      }
> > >    else
> > >      {
> > > -      rep_movsb_threshold = 2048 * (16 / 16);
> > > +      rep_movsb_threshold = 4096 * (16 / 16);
> > >  #if HAVE_TUNABLES
> > >        minimum_rep_movsb_threshold = 16 * 8;
> > >  #endif
> > > @@ -896,6 +896,9 @@ dl_init_cacheinfo (struct cpu_features *cpu_features)
> > >    if (CPU_FEATURE_USABLE_P (cpu_features, FSRM))
> > >      rep_movsb_threshold = 2112;
> > >
> > > +
> > > +
> > > +
> >
> > Please don't add these blank lines.
> Fixed.
>
>
> >
> > >    unsigned long int rep_movsb_stop_threshold;
> > >    /* ERMS feature is implemented from AMD Zen3 architecture and it is
> > >       performing poorly for data above L2 cache size. Henceforth, adding
> > > --
> > > 2.25.1
> > >
> >
> > Thanks.
> >
> > H.J.



-- 
H.J.