Re: Re: [PATCH v1 3/6] x86: Remove mem{move|cpy}-ssse3

[Noah Goldstein <goldstein.w.n@gmail.com>]

On Wed, Mar 30, 2022 at 10:34 PM Mayshao-oc <Mayshao-oc@zhaoxin.com> wrote:
>
> On Thur, Mar 31, 2022 at 12:45 AM Noah Goldstein<goldstein.w.n@gmail.com> wrote:
>
>
> > On Wed, Mar 30, 2022 at 4:57 AM Mayshao-oc <Mayshao-oc@zhaoxin.com> wrote:
> > >
> > > On Tue, Mar 29, 2022 at 10:57 AM Noah Goldstein<goldstein.w.n@gmail.com> wrote:
> > >
> > >
> > > >On Mon, Mar 28, 2022 at 9:51 PM Mayshao-oc <Mayshao-oc@zhaoxin.com> wrote:
> > > > >
> > > > > On Mon, Mar 28, 2022 at  9:07 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> > > > >
> > > > >
> > > > > > On Mon, Mar 28, 2022 at 1:10 AM Mayshao-oc <Mayshao-oc@zhaoxin.com> wrote:
> > > > > > >
> > > > > > > On Fri, Mar 25, 2022 at 6:36 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
> > > > > > >
> > > > > > > > With SSE2, SSE4.1, AVX2, and EVEX versions very few targets prefer
> > > > > > > > SSSE3. As a result its no longer with the code size cost.
> > > > > > > > ---
> > > > > > > > sysdeps/x86_64/multiarch/Makefile          |    2 -
> > > > > > > > sysdeps/x86_64/multiarch/ifunc-impl-list.c |   15 -
> > > > > > > > sysdeps/x86_64/multiarch/ifunc-memmove.h   |   18 +-
> > > > > > > > sysdeps/x86_64/multiarch/memcpy-ssse3.S    | 3151 --------------------
> > > > > > > > sysdeps/x86_64/multiarch/memmove-ssse3.S   |    4 -
> > > > > > > > 5 files changed, 7 insertions(+), 3183 deletions(-)
> > > > > > > > delete mode 100644 sysdeps/x86_64/multiarch/memcpy-ssse3.S
> > > > > > > > delete mode 100644 sysdeps/x86_64/multiarch/memmove-ssse3.S
> > > > > > >
> > > > > > > On some platforms, such as Zhaoxin, the memcpy performance of SSSE3
> > > > > > > is better than that of AVX2, and the current computer system has sufficient
> > > > > > > disk capacity and memory capacity.
> > > > > >
> > > > > > How does the SSSE3 version compare against the SSE2 version?
> > > > >
> > > > > On some Zhaoxin processors, the overall performance of SSSE3 is about
> > > > > 10% higher than that of SSE2.
> > > > >
> > > > >
> > > > > Best Regards,
> > > > > May Shao
> > > >
> > > > Any chance you can post the result from running `bench-memset` or some
> > > > equivalent benchmark? Curious where the regressions are. Ideally we would
> > > > fix the SSE2 version so its optimal.
> > >
> > > Bench-memcpy on Zhaoxin KX-6000 processor shows that, when length <=4 or
> > > length >= 128, memcpy SSSE3 can achieve an average performance improvement
> > > of 25% compared to SSSE2.
> >
> > Thanks
> >
> > The size <= 4 regression is expected as profiles of SPEC show the [5, 32] sized
> > copies to significantly hotter.
> >
> > Regarding the large sizes, it seems to be because the SSSE3 version avoids
> > unaligned loads/stores much more aggressively.
>
> Agree.
>
> > For now we will keep the function. Will look into a replacement that isn't so
> > costly to code size.
>
> Thanks very much for your support.

Will SSE4.1 be an issue for you? I think the only reasonable way to fix this is
with `pshufb`.
>
> > Out of curiosity, is bench-memcpy-random performance also improved with
> > SSSE3? The jump table / branches generally look really nice in micro-benchmarks
> > but that may not be fully indicative of how it will performance in an
> > application.
>
> Bench-memcpy-random shows about a 5% performance drop for SSSE3:

Thanks.

>         __memcpy_sse2_unaligned   __memcpy_ssse3  Improvement(ssse3 over sse2)
> length=32768 805982          874585         -8.51%
> length=65536 885317         940458        -6.23%
> length=131072 929177         979173         -5.38%
> length=262144 980083        1033130         -5.41%
> length=524288 1042590 1095560 -5.08%
> length=1048576 1078020 1127990 -4.64%
>
>
> > >
> > > I have attached the test results, hope this is what you want to see.
> > >
> > > > > > > It is strongly recommended to keep the SSSE3 version.
> > > > > > >
> > > > > >
> > > > > >
> > > > > > --
> > > > > > H.J.