Re: About the memcpy optimization patch

7 Jul 2015


      On 07/07/15 04:38, Yingshiuan Pan wrote:
...
aarch64, CA57*2 + CA53*4
well, I think I should rerun the benchmarks with 32-bit android
memcpy() performance can be very sensitive to CPU micro-architecture.
Out of interest how many 32-bit Android reference platforms do you work 
with (and do any use C-A7 cores)?
Daniel.
...
--
潘穎軒 Yingshiuan (Peter) Pan
2015-07-07 11:33 GMT+08:00 yfw <fengwei.yin@linaro.org
mailto:fengwei.yin@linaro.org>:
Hi Yingshiuan,
What kind of cpu arch is Juno? Here is the info shared from Spreadtrum
about their core:
     Cortex-A7 with NEON.
     # cat /proc/cpuinfo
     Processor     : ARMv7 Processor rev 5 (v7l)
     processor     : 0
     model name    : ARMv7 Processor rev 5 (v7l)
     BogoMIPS      : 1548.16
     Features      : swp half thumb fastmult vfp edsp neon vfpv3 tls
vfpv4 idiva idivt vfpd32
     CPU implementer       : 0x41
     CPU architecture: 7
     CPU variant   : 0x0
     CPU part      : 0xc07
     CPU revision  : 5
     ... ...

Regards
Yin, Fengwei

On 2015年07月07日 11:11, Yingshiuan Pan wrote:

    Hi Fengwei,

    I did not see such big performance regression on Juno,
    followings are my
    result:

    geekbench 3.3.2:
                          single    multi
    w/o opt. memcpy        857     2221
    with opt. memcpy       855     2263
    performance gain       -0.2%      1.9%

    Quadrant 2.1
    w/o opt. memcpy    9662
    with opt. memcpy   9563
    performance gain    -1.0%


    --
    潘穎軒 Yingshiuan (Peter) Pan


    2015-07-06 11:16 GMT+08:00 yfw <fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>
    <mailto:fengwei.yin@linaro.org <mailto:fengwei.yin@linaro.org>>>:

         Hi Yingshiuan,
         That's fine. Do we have updated patch for memcpy? Did you
    get the
         Quadrant/Geekbench3 data in your side? Thanks.

         Regards
         Yin, Fengwei

         On 2015年07月06日 09:24, Yingshiuan Pan wrote:

             Hi Fengwei
             Sorry, my carelessness, you're right.


             --
             潘穎軒 Yingshiuan (Peter) Pan


             2015-07-03 19:51 GMT+08:00 yfw <fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>
             <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>>
             <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org> <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>>>>:


                  Hi Yingshiuan,
    https://android-review.googlesource.com/#/c/147915/ is for memchr.
                  while
    http://review.android.git.linaro.org/#/c/15529/ is for memcpy.
             Could you
                  please double check whether you have correct link?
    Thanks.

                  Regards
                  Yin, Fengwei

                  在 2015/7/3 14:04, Yingshiuan Pan 写道:

                      Hi Fengwei,

                      Thanks for the report, there is a new revision
    here:
    https://android-review.googlesource.com/#/c/147915/
                      Although only single instruction is changed,
    someone
             said he did see
                      measurable improvement.
                      Would you give a try? I'll also run those
    benchmarks to
             see if I can
                      reproduce the result.

                      btw, what is the platform you are using?

                      --
                      潘穎軒 Yingshiuan (Peter) Pan


                      2015-07-03 13:59 GMT+08:00 yfw
    <fengwei.yin@linaro.org <mailto:fengwei.yin@linaro.org>
             <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>>
                      <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>
             <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>>>
                      <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>
             <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>> <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>
             <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>>>>>:

                            From their report, they used:
                           Quadrant Professional(v2.1.1)
                           GeekBench3 (v3.3.2)

                           I suppose the methodology to run these
    benchmarks
             is very
                      known.


                           Regards
                           Yin, Fengwei

                           在 2015/7/3 11:53, Tom Gall 写道:

                               Hi Fengwei,

                               Thanks for the report.  I think we
    would be
             interested in
                               knowing the
                               nature of the test being used.

                               I've added Bero to the CC as I'm sure
    he'll be
                      interested as well.

                               On Thu, Jul 2, 2015 at 9:35 PM, yfw
                      <fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org> <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>>
             <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org> <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>>>
                               <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>
             <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>>

                      <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>
             <mailto:fengwei.yin@linaro.org
    <mailto:fengwei.yin@linaro.org>>>>> wrote:

                                   Hi,
                                   About the patch
    http://review.android.git.linaro.org/#/c/15529/, we got
                                   an issue report from member. It could
             improve the micro
                                   benchmark but
                                   hurt general benchmark actually.
    Here is
             their test
                      result:

                                         bionic-benchmarks32
    BM_string_memcpy:
                                         Size  AOSP    Linaro  Perf
    Improvement
                                         8     450.7   329.5   -37%
                                         64    1715.7  2225.3  23%
                                         512   3100.7  4253.0  27%
                                         1K    3271.8  4652.3  30%
                                         8K    2460.9  3174.7  22%
                                         16K   2267.4  2953.2  23%
                                         32K   2222.9  2800.7  21%
                                         64K   2204.3  2640.1  17%

                                                         without
    patch   with
             patch
                      perf delta
                                         Quadrant mem    3400
             2100        -38%
                                         GeekBench3
                                           single-core   436
          260
                     -40%
                                           multi-core    434
          358
                     -18%

                                   I am wondering whether this is an
    known
             issue or
                      not for
                                   this list.
                                   Thanks in advance.

                                   Regards
                                   Yin, Fengwei

               _______________________________________________
                                   linaro-android mailing list
    linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>
             <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>>
                      <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>
             <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>>>

    <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>
             <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>>
                      <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>
             <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>>>>
    https://lists.linaro.org/mailman/listinfo/linaro-android


    _______________________________________________
                           linaro-android mailing list
    linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>
             <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>>
                      <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>
             <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>>>
                      <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>
             <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>>
                      <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>
             <mailto:linaro-android@lists.linaro.org
    <mailto:linaro-android@lists.linaro.org>>>>
    https://lists.linaro.org/mailman/listinfo/linaro-android


linaro-android mailing list
linaro-android@lists.linaro.org
https://lists.linaro.org/mailman/listinfo/linaro-android

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

Re: About the memcpy optimization patch