[dpdk-dev] [PATCH 0/4] Optimize memcpy for AVX512 platforms

[Wang, Zhihong]

> -----Original Message-----
> From: Stephen Hemminger [mailto:stephen at networkplumber.org]
> Sent: Friday, January 15, 2016 12:49 AM
> To: Wang, Zhihong <zhihong.wang at intel.com>
> Cc: dev at dpdk.org; Ananyev, Konstantin <konstantin.ananyev at intel.com>;
> Richardson, Bruce <bruce.richardson at intel.com>; Xie, Huawei
> <huawei.xie at intel.com>
> Subject: Re: [PATCH 0/4] Optimize memcpy for AVX512 platforms
> 
> On Thu, 14 Jan 2016 01:13:18 -0500
> Zhihong Wang <zhihong.wang at intel.com> wrote:
> 
> > This patch set optimizes DPDK memcpy for AVX512 platforms, to make full
> > utilization of hardware resources and deliver high performance.
> >
> > In current DPDK, memcpy holds a large proportion of execution time in
> > libs like Vhost, especially for large packets, and this patch can bring
> > considerable benefits.
> >
> > The implementation is based on the current DPDK memcpy framework, some
> > background introduction can be found in these threads:
> > http://dpdk.org/ml/archives/dev/2014-November/008158.html
> > http://dpdk.org/ml/archives/dev/2015-January/011800.html
> >
> > Code changes are:
> >
> >   1. Read CPUID to check if AVX512 is supported by CPU
> >
> >   2. Predefine AVX512 macro if AVX512 is enabled by compiler
> >
> >   3. Implement AVX512 memcpy and choose the right implementation based
> on
> >      predefined macros
> >
> >   4. Decide alignment unit for memcpy perf test based on predefined macros
> >
> > Zhihong Wang (4):
> >   lib/librte_eal: Identify AVX512 CPU flag
> >   mk: Predefine AVX512 macro for compiler
> >   lib/librte_eal: Optimize memcpy for AVX512 platforms
> >   app/test: Adjust alignment unit for memcpy perf test
> >
> >  app/test/test_memcpy_perf.c                        |   6 +
> >  .../common/include/arch/x86/rte_cpuflags.h         |   2 +
> >  .../common/include/arch/x86/rte_memcpy.h           | 247
> ++++++++++++++++++++-
> >  mk/rte.cpuflags.mk                                 |   4 +
> >  4 files changed, 255 insertions(+), 4 deletions(-)
> >
> 
> This really looks like code that could benefit from Gcc
> function multiversioning. The current cpuflags model is useless/flawed
> in real product deployment


I've tried gcc function multi versioning, with a simple add() function
which returns a + b, and a loop calling it for millions of times. Turned
out this mechanism adds 17% extra time to execute, overall it's a lot
of extra overhead.

Quote the gcc wiki: "GCC takes care of doing the dispatching to call
the right version at runtime". So it loses inlining and adds extra
dispatching overhead.

Also this mechanism works only for C++, right?

I think using predefined macros at compile time is more efficient and
suits DPDK more.

Could you please give an example when the current CPU flags model
stop working? So I can fix it.