Re: [dpdk-dev] [PATCH v1 3/3] net/i40e: auto-vectorization to speed up Tx free

[Jerin Jacob <jerinjacobk@gmail.com>]

On Sat, Mar 7, 2020 at 8:34 PM Gavin Hu <Gavin.Hu@arm.com> wrote:
>
> Hi Jerin,
>
> > -----Original Message-----
> > From: Jerin Jacob <jerinjacobk@gmail.com>
> > Sent: Friday, March 6, 2020 3:45 PM
> > To: Gavin Hu <Gavin.Hu@arm.com>
> > Cc: dpdk-dev <dev@dpdk.org>; nd <nd@arm.com>; David Marchand
> > <david.marchand@redhat.com>; thomas@monjalon.net;
> > jerinj@marvell.com; Ye, Xiaolong <xiaolong.ye@intel.com>; Honnappa
> > Nagarahalli <Honnappa.Nagarahalli@arm.com>; Ruifeng Wang
> > <Ruifeng.Wang@arm.com>; Phil Yang <Phil.Yang@arm.com>; Joyce Kong
> > <Joyce.Kong@arm.com>; Steve Capper <Steve.Capper@arm.com>
> > Subject: Re: [dpdk-dev] [PATCH v1 3/3] net/i40e: auto-vectorization to
> > speed up Tx free
> >
> > On Fri, Mar 6, 2020 at 10:35 AM Gavin Hu <gavin.hu@arm.com> wrote:
> > >
> > > Tx mbuf free is a hotspot for i40e on aarch64, as there are no
> > > inter-loop dependencies, it is safe to enable auto-vectorization
> > > to speed up.
> > >
> > > This patch showed 2~3% performance lift on ThunderX2 and no
> > degradation
> > > on Arm N1SDP. The test case is single core RFC2544 zero-loss test.
> > >
> > > Signed-off-by: Gavin Hu <gavin.hu@arm.com>
> > > Reviewed-by: Steve Capper <steve.capper@arm.com>
> > > ---
> > >  drivers/net/i40e/i40e_rxtx_vec_common.h | 5 +++++
> > >  1 file changed, 5 insertions(+)
> > >
> > > diff --git a/drivers/net/i40e/i40e_rxtx_vec_common.h
> > b/drivers/net/i40e/i40e_rxtx_vec_common.h
> > > index 0e6ffa007..fc0fa45d4 100644
> > > --- a/drivers/net/i40e/i40e_rxtx_vec_common.h
> > > +++ b/drivers/net/i40e/i40e_rxtx_vec_common.h
> > > @@ -98,6 +98,11 @@ i40e_tx_free_bufs(struct i40e_tx_queue *txq)
> > >         if (likely(m != NULL)) {
> > >                 free[0] = m;
> > >                 nb_free = 1;
> > > +#if defined(__clang__)
> > > +#pragma clang loop vectorize(assume_safety)
> > > +#elif defined(__GNUC__)
> > > +#pragma GCC ivdep
> > > +#endif
> >
> > IMO, It is better to abstract the compiler features  (above compiler
> > feature and __restrict__) as macros in
> > rte_common.h or so. It will help to support other compilers(ICC or
> > Windows) and enable them to have "changes" in one place.
>
> How about defining RTE_LOOP_AUTO_VECTORIZATION in the rte_common.h?

Other compiler stuff in rte_common.h are starting with __rte in small
letter(__rte_packed, __rte_unused) etc.
I think, a better name would be __rte_loop_auto_vectorize or so.
No strong opinion for the name though.

# Probably it is worth checking and add performance result of x86
testing in git commit as well as it
is common code.


> #if defined(__clang__)
>         define RTE_LOOP_AUTO_VECTORIZATION  \
>                 #pragma clang loop vectorize(assume_safety)
> #elif defined(__GNUC__)
>         define RTE_LOOP_AUTO_VECTORIZATION  \
>                 #pragma GCC ivdep
> #else
>         define RTE_LOOP_AUTO_VECTORIZATION
> #endif
> If you agree, I will submit a v2. Thanks for your comments!
> /Gavin
> >
> >
> >
> > >                 for (i = 1; i < n; i++) {
> > >                         m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
> > >                         if (likely(m != NULL)) {
> > > --
> > > 2.17.1
> > >