[dpdk-dev] [PATCH v2 5/7] examples/l3fwd: add neon support for l3fwd

[Jianbo Liu]

On 11 May 2017 at 12:27, Sekhar, Ashwin <Ashwin.Sekhar at cavium.com> wrote:
>
> On Thu, 2017-05-11 at 04:14 +0000, Sekhar, Ashwin wrote:
> ...
>> > > Combining all the above comments, I made some changes on top of
>> > > your
>> > > patch. These changes are giving 3-4% improvement over your
>> > > version.
>> > >
>> > > You may find the changes at
>> > > https://gist.github.com/ashwinyes/34cbdd999784402c859c71613587faf
>> > > c
>> > >
>> > Is the correct in Line 103/104, you only process one packets in the
>> > last FWDSTEP packets?
>> Its doing processx4_* there. So its processing 4 packets.
>>
>> >
>> > Actually, I don't like your change in l3fwd_lpm_send_packets,
>> > making
>> > the simple logic complicated. And I don't think it can help to
>> > improve
>> > performance. :-)
>> Its not making it complicated. The number of lines of code may be
>> higher by may be 10 lines, but the conditions of the loops are
>> simplified which reduces the number of branch instructions and helps
>> the processor to go through them faster.

I suspected not much improvement we can get.

>>
>> If possible, please try it out on your machine.

OK, I'll test. If no performance regression, I'll adopt your suggestion in v3.

>
> Missed out one point.
> Since 2 loops are form "for (i = 0; i < FWDSTEP; i++)" i.e. looping for
> constant number of iterations, compiler will easily unroll them.
>
> Thanks
> Ashwin
>> >
>> >
>> > >
>> > >
>> > > Please check it out and let me know your comments.
>> > >
>> > > Thanks
>> > > Ashwin