> From: Neil Horman [mailto:nhorman at tuxdriver.com]
> Sent: Wednesday, December 17, 2014 3:33 PM
> To: Ananyev, Konstantin
> Cc: dev at dpdk.org
> Subject: Re: [dpdk-dev] [PATCH 10/17] librte_acl: add AVX2 as new
> rte_acl_classify() method
>
> On Tue, Dec 16, 2014 at 04:16:48PM +0000, Ananyev, Konstantin wrote:
> >
> >
> > > -----Original Message-----
> > > From: Neil Horman [mailto:nhorman at tuxdriver.com]
> > > Sent: Monday, December 15, 2014 8:21 PM
> > > To: Ananyev, Konstantin
> > > Cc: dev at dpdk.org
> > > Subject: Re: [dpdk-dev] [PATCH 10/17] librte_acl: add AVX2 as new
> > > rte_acl_classify() method
> > >
> > > On Mon, Dec 15, 2014 at 04:33:47PM +0000, Ananyev, Konstantin wrote:
> > > > Hi Neil,
> > > >
> > > > > -----Original Message-----
> > > > > From: Neil Horman [mailto:nhorman at tuxdriver.com]
> > > > > Sent: Monday, December 15, 2014 4:00 PM
> > > > > To: Ananyev, Konstantin
> > > > > Cc: dev at dpdk.org
> > > > > Subject: Re: [dpdk-dev] [PATCH 10/17] librte_acl: add AVX2 as new
> > > > > rte_acl_classify() method
> > > > >
> > > > > On Sun, Dec 14, 2014 at 06:10:52PM +0000, Konstantin Ananyev wrote:
> > > > > > Introduce new classify() method that uses AVX2 instructions.
> > > > > > From my measurements:
> > > > > > On HSW boards when processing >= 16 packets per call,
> > > > > > AVX2 method outperforms it's SSE counterpart by 10-25%,
> > > > > > (depending on the ruleset).
> > > > > > At runtime, this method is selected as default one on HW that
> > > > > > supports AVX2.
> > > > > >
> > > > > > Signed-off-by: Konstantin Ananyev <konstantin.ananyev at intel.com>
> > > > > > ---
> > > > > > lib/librte_acl/Makefile | 9 +
> > > > > > lib/librte_acl/acl.h | 4 +
> > > > > > lib/librte_acl/acl_run.h | 2 +-
> > > > > > lib/librte_acl/acl_run_avx2.c | 58 +++++
> > > > > > lib/librte_acl/acl_run_avx2.h | 305 ++++++++++++++++++++++++
> > > > > > lib/librte_acl/acl_run_sse.c | 537
> > > > > > +-----------------------------------------
> > > > > > lib/librte_acl/acl_run_sse.h | 533
> > > > > > +++++++++++++++++++++++++++++++++++++++++
> > > > > > lib/librte_acl/rte_acl.c | 5 +-
> > > > > > lib/librte_acl/rte_acl.h | 2 +
> > > > > > 9 files changed, 917 insertions(+), 538 deletions(-)
> > > > > > create mode 100644 lib/librte_acl/acl_run_avx2.c
> > > > > > create mode 100644 lib/librte_acl/acl_run_avx2.h
> > > > > > create mode 100644 lib/librte_acl/acl_run_sse.h
> > > > > >
> > > > > > diff --git a/lib/librte_acl/Makefile b/lib/librte_acl/Makefile
> > > > > > index 65e566d..223ec31 100644
> > > > > > --- a/lib/librte_acl/Makefile
> > > > > > +++ b/lib/librte_acl/Makefile
> > > > > > @@ -45,8 +45,17 @@ SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_bld.c
> > > > > > SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_gen.c
> > > > > > SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_run_scalar.c
> > > > > > SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_run_sse.c
> > > > > > +SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_run_avx2.c
> > > > > >
> > > > > > CFLAGS_acl_run_sse.o += -msse4.1
> > > > > > +ifeq ($(CC), icc)
> > > > > > +CFLAGS_acl_run_avx2.o += -march=core-avx2
> > > > > > +else ifneq ($(shell \
> > > > > > +test $(GCC_MAJOR_VERSION) -le 4 -a $(GCC_MINOR_VERSION) -le 6 &&
> > > > > > echo 1), 1)
> > > > > > +CFLAGS_acl_run_avx2.o += -mavx2
> > > > > > +else
> > > > > > +CFLAGS_acl_run_avx2.o += -msse4.1
> > > > > > +endif
> > > > > >
> > > > > This seems broken. You've unilaterally included acl_run_avx2.c in
> > > > > the build
> > > > > list above, but only enable -mavx2 if the compiler is at least gcc
> > > > > 4.6.
> > > >
> > > > Actually 4.7 (before that version, as I know, gcc doesn't support avx2)
> > > >
> > > > > Unless
> > > > > you want to make gcc 4.6 a requirement for building,
> > > >
> > > > I believe DPDK is required to be buildable by gcc 4.6
> > > > As I remember, we have to support it all way down to gcc 4.3.
> > > >
> > > > > you need to also exclude
> > > > > the file above from the build list.
> > > >
> > > > That means that for gcc 4.6 and below rte_acl_classify_avx2() would
> > > > not be defined.
> > > > And then at runtime, I have to check for that somehow and (re)populate
> > > > classify_fns[].
> > > > Doesn't seems like a good way to me.
> > > There are plenty of ways around that.
> > >
> > > At a minimum you could make the classify_fns array the one place that you
> > > need
> > > to add an ifdef __AVX__ call.
> > >
> > > You could also create a secondary definition of rte_acl_classify_avx2,
> > > and mark
> > > it as a weak symbol, which only returns -EOPNOTSUPP. That would be good,
> > > since
> > > the right thing will just automatically happen then if you don't build the
> > > actual avx2 classification code
> > >
> > > > Instead, I prefer to always build acl_run_avx2.c,
> >
> >
> > > But you can't do that. You just said above that you need to support down
> > > to gcc
> > > 4.3. I see you've worked around that with some additional ifdef __AVX__
> > > instructions, but in so doing you ignore the possibiity that sse isn't
> > > supported, so you need to add __SSE__ checks now as well. ifdeffing that
> > > much
> > > just isn't scalable.
> >
> > We don't need to worry about compiler without SSE4.1 support.
> > I believe that all compilers that DDPDK has to build with, do support
> > SSE4.1.
> > So for SSE4.1 we only has to worry about situation when target CPU doesn't
> > support it
> > We manage it by runtime selection.
> > For AVX2 - situation is a bit different: it could be both compiler and
> > target CPU that don't support it.
> >
> > > And for your effort, you get an AVX2 classification path
> > > that potentially doesn't actually do vectorized classification.
> > >
> > > It really seems better to me to not build the code if the compiler doesn't
> > > support the instruction set it was meant to enable, and change the
> > > classification function pointer to something that informs the user of the
> > > lack
> > > of support at run time.
> > >
> > > > but for old compilers that don't support AVX2 -
> > > > rte_acl_classify_avx2() would simply be identical to
> > > > rte_acl_classify_sse().
> > > >
> > > That doesn't make sense to me, for two reasons:
> > >
> > > 1) What if the machine being targeted doesn't support sse either?
> > >
> >
> > Exactly the same what is happening now on the machine with now SSE4.1
> > support.
> > There is absolutely no difference here.
> >
> > > 2) If an application selects an AVX2 classifier, I as a developer expect
> > > to
> > > either get AVX2 based classification, or an error indicating that I can't
> > > do
> > > AVX2 classification, not a silent performance degradation down to scalar
> > > classification.
> >
> > In fact I was considering both variants for compilers not supporting AVX2:
> > 1. silently degrade to SSE method.
> > 2. create a dummy function rte_acl_classify_error() and put it into
> > classify_fns[RTE_ACL_CLASSIFY_AVX2].
> >
> > I choose #1 because it seems like a less distraction for the user -
> > all would keep working as before, user just wouldn't see any improvement
> > comparing to SSE method.
> > Again didn't want to spread "ifdef __AVX2__" into rte_acl.c
> > Though I don't have any strong opinion here.
> > So if you can provide some good reason why #2 is preferable, I am ok to
> > switch to #2.
> >
> Because 2 doesn't require any ifdeffing. As you note above the problem here
> is
> that AVX2 support is both compiler and machine dependent. If you make a weak
> symbol version of rte_acl_classify_avx2 that always gets built, then you've
> reduced the problem to just being compiler support, which you can check in the
> makefile.
I don't think we'll get rid of ifdefing with #2.
We'll remove 2 ifdefs in acl_run_avx2.h, but then we have to introduce 2 new
in rte_acl.c instead.
>From my understanding, we we'll need something like that:
static const rte_acl_classify_t classify_fns[] = {
[RTE_ACL_CLASSIFY_DEFAULT] = rte_acl_classify_scalar,
[RTE_ACL_CLASSIFY_SCALAR] = rte_acl_classify_scalar,
[RTE_ACL_CLASSIFY_SSE] = rte_acl_classify_sse,
+#if (defined __GNUC__ && __GNUC__ <= 4 && __GNUC_MINOR__ < 7)
+ [RTE_ACL_CLASSIFY_AVX2] = rte_acl_classify_error,
+#else
[RTE_ACL_CLASSIFY_AVX2] = rte_acl_classify_avx2,
+#endif
};
static void __attribute__((constructor))
rte_acl_init(void)
{
enum rte_acl_classify_alg alg = RTE_ACL_CLASSIFY_DEFAULT;
+#if (defined __GNUC__ && __GNUC__ <= 4 && __GNUC_MINOR__ < 7)
if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
alg = RTE_ACL_CLASSIFY_AVX2;
else if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE4_1))
+#else
+ if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE4_1))
alg = RTE_ACL_CLASSIFY_SSE;
+#endif
rte_acl_set_default_classify(alg);
}
Correct?
Konstantin
>
> > >
> > > > > That in turn I think allows you to remove a
> > > > > bunch of the ifdeffing that you've done in some of the avx2 specific
> > > > > files.
> > > >
> > > > Actually there are not many of them.
> > > > One in acl_run_avx2.h and another in acl_run_avx2.c.
> > > >
> > > 2 in acl_run_avx2.h and 1 in rte_acl_osdep_alone.h, which is really 3
> > > more than
> > > you need if you just do an intellegent weak classifier function defintion.
> >
> > grep -n __AVX2__ lib/librte_acl/*.[c,h] | grep -v endif
> > lib/librte_acl/acl_run_avx2.c:45:#ifdef __AVX2__
> > lib/librte_acl/acl_run_avx2.h:36:#ifdef __AVX2__
> >
> > rte_acl_osdep_alone.h - is a different story.
> > It needs to be there anyway, as in rte_common_vect.h.
> > In fact rte_acl_osdep_alone.h is only needed for cases when
> > RTE_LIBRTE_ACL_STANDALONE=y.
> > That comes from the old days, when we had to to support building librte_acl
> > library without the rest of DPDK.
> > I think we don't need it anymore and plan to remove it.
> > Just thought it should be in a separate patch.
> > Konstantin
> >
> > >
> > > Neil
> >
