DPDK patches and discussions
 help / color / mirror / Atom feed
From: "Varghese, Vipin" <Vipin.Varghese@amd.com>
To: Konstantin Ananyev <konstantin.ananyev@huawei.com>,
	Stephen Hemminger <stephen@networkplumber.org>
Cc: "Yigit, Ferruh" <Ferruh.Yigit@amd.com>,
	"bruce.richardson@intel.com" <bruce.richardson@intel.com>,
	"konstantin.v.ananyev@yandex.ru" <konstantin.v.ananyev@yandex.ru>,
	"aman.deep.singh@intel.com" <aman.deep.singh@intel.com>,
	"dev@dpdk.org" <dev@dpdk.org>
Subject: RE: [PATCH v2 1/3] app/testpmd: add register keyword
Date: Fri, 6 Sep 2024 13:02:01 +0000	[thread overview]
Message-ID: <PH7PR12MB859639DA1EDC9494B9703817829E2@PH7PR12MB8596.namprd12.prod.outlook.com> (raw)
In-Reply-To: <0ae233fb72ce49cea5186e1f924db76b@huawei.com>

[-- Attachment #1: Type: text/plain, Size: 2244 bytes --]

[AMD Official Use Only - AMD Internal Distribution Only]


<snipped>

> > > >> --- a/app/test-pmd/macswap_sse.h
> > > >> +++ b/app/test-pmd/macswap_sse.h
> > > >> @@ -16,13 +16,13 @@ do_macswap(struct rte_mbuf *pkts[], uint16_t
> nb,
> > > >>        uint64_t ol_flags;
> > > >>        int i;
> > > >>        int r;
> > > >> -     __m128i addr0, addr1, addr2, addr3;
> > > >> +     register __m128i addr0, addr1, addr2, addr3;
> > > > Some compilers treat register as a no-op. Are you sure? Did you check
> with godbolt.
> > >
> > > Thank you Stephen, I have tested the code changes on Linux using GCC
> > > and Clang compiler.
> > >
> > > In both cases in Linux environment, we have seen the the values
> > > loaded onto register `xmm`.
> > >
> > > ```
> > > registerconst__m128i shfl_msk = _mm_set_epi8(15, 14, 13, 12, 5, 4,
> > > 3, 2, 1, 0, 11, 10, 9, 8, 7, 6); vmovdqaxmm0, xmmwordptr[rip+
> > > .LCPI0_0]
>
> Yep, that what I would probably expect: one time load before the loop starts,
> right?
> Curious  what exactly it would generate then if 'register' keyword is missed?
> BTW, on my box,  gcc-11  with '-O3 -msse4.2 ...'  I am seeing expected
> behavior without 'register' keyword.
> Is it some particular compiler version that misbehaves?

Thank you, Konstantin, for this pointer. I have been trying this understand this a bit more internally. Here are my observations

1. shuf simd ISA works on XMM register only.
2. Any values from variables has to be loaded to `xmm` register before processing.
3. when compiled for `-march=native` with compiler not aware (SoC Arch gcc weights) without patch might have generating with ` movzx   eax, BYTE PTR [rbp-48]`
4. when register keyword is applied for both shufl_mask and addr, the compiler generates trying to get the variables directly into xmm using ` vmovdqu (%rsi),%xmm1`

So, I think you are right, from gcc12.3 and gcc 13.1 which supports `-march=znver4` this problem will not come.

>
> > >
> > > ```
> > >
> > > Both cases we have performance improvement.
> > >
> > >
> > > Can you please help us understand if we have missed out something?
> >
> > Ok, not sure why compiler would not decide to already use a register here?

[-- Attachment #2: Type: text/html, Size: 7828 bytes --]

  reply	other threads:[~2024-09-06 13:02 UTC|newest]

Thread overview: 18+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2024-07-16  6:37 [PATCH] app/testpmd: improve sse based macswap Vipin Varghese
2024-07-23 16:45 ` Ferruh Yigit
2024-07-23 17:12   ` Bruce Richardson
2024-07-25 12:47     ` Varghese, Vipin
2024-07-25 12:52       ` Bruce Richardson
2024-08-21 14:38 ` [PATCH v2 0/3] " Vipin Varghese
2024-08-21 14:38   ` [PATCH v2 1/3] app/testpmd: add register keyword Vipin Varghese
2024-08-21 14:55     ` Stephen Hemminger
2024-08-27 15:32       ` Varghese, Vipin
2024-08-27 17:39         ` Stephen Hemminger
2024-08-29  8:14           ` Varghese, Vipin
2024-09-03 11:52           ` Konstantin Ananyev
2024-09-06 13:02             ` Varghese, Vipin [this message]
2024-10-04  5:04               ` Ferruh Yigit
2024-08-21 14:38   ` [PATCH v2 2/3] app/testpmd: move offload update Vipin Varghese
2024-08-21 14:38   ` [PATCH v2 3/3] app/testpmd: interleave SSE SIMD Vipin Varghese
2024-10-04  5:08   ` [PATCH v2 0/3] app/testpmd: improve sse based macswap Ferruh Yigit
2024-10-08  1:12     ` Ferruh Yigit

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=PH7PR12MB859639DA1EDC9494B9703817829E2@PH7PR12MB8596.namprd12.prod.outlook.com \
    --to=vipin.varghese@amd.com \
    --cc=Ferruh.Yigit@amd.com \
    --cc=aman.deep.singh@intel.com \
    --cc=bruce.richardson@intel.com \
    --cc=dev@dpdk.org \
    --cc=konstantin.ananyev@huawei.com \
    --cc=konstantin.v.ananyev@yandex.ru \
    --cc=stephen@networkplumber.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).