From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mga03.intel.com (mga03.intel.com [134.134.136.65]) by dpdk.org (Postfix) with ESMTP id 6CC4E18F for ; Wed, 17 Dec 2014 20:23:09 +0100 (CET) Received: from orsmga003.jf.intel.com ([10.7.209.27]) by orsmga103.jf.intel.com with ESMTP; 17 Dec 2014 11:19:57 -0800 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.04,691,1406617200"; d="scan'208";a="500389182" Received: from irsmsx103.ger.corp.intel.com ([163.33.3.157]) by orsmga003.jf.intel.com with ESMTP; 17 Dec 2014 11:17:50 -0800 Received: from irsmsx105.ger.corp.intel.com ([169.254.7.144]) by IRSMSX103.ger.corp.intel.com ([169.254.3.113]) with mapi id 14.03.0195.001; Wed, 17 Dec 2014 19:22:06 +0000 From: "Ananyev, Konstantin" To: Neil Horman Thread-Topic: [dpdk-dev] [PATCH 10/17] librte_acl: add AVX2 as new rte_acl_classify() method Thread-Index: AQHQGIBe6R2bzzgwb0yDNvzVCW5u5ZyQ0+9ggABERICAAUFi0IABksMAgAAvUEA= Date: Wed, 17 Dec 2014 19:22:06 +0000 Message-ID: <2601191342CEEE43887BDE71AB977258213C1AB4@IRSMSX105.ger.corp.intel.com> References: <1418580659-12595-1-git-send-email-konstantin.ananyev@intel.com> <1418580659-12595-11-git-send-email-konstantin.ananyev@intel.com> <20141215160009.GC3803@hmsreliant.think-freely.org> <2601191342CEEE43887BDE71AB977258213C0D9C@IRSMSX105.ger.corp.intel.com> <20141215202043.GD3803@hmsreliant.think-freely.org> <2601191342CEEE43887BDE71AB977258213C12AD@IRSMSX105.ger.corp.intel.com> <20141217153232.GA6618@localhost.localdomain> In-Reply-To: <20141217153232.GA6618@localhost.localdomain> Accept-Language: en-IE, en-US Content-Language: en-US X-MS-Has-Attach: X-MS-TNEF-Correlator: x-originating-ip: [163.33.239.180] Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: quoted-printable MIME-Version: 1.0 Cc: "dev@dpdk.org" Subject: Re: [dpdk-dev] [PATCH 10/17] librte_acl: add AVX2 as new rte_acl_classify() method X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.15 Precedence: list List-Id: patches and discussions about DPDK List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Wed, 17 Dec 2014 19:23:18 -0000 > From: Neil Horman [mailto:nhorman@tuxdriver.com] > Sent: Wednesday, December 17, 2014 3:33 PM > To: Ananyev, Konstantin > Cc: dev@dpdk.org > Subject: Re: [dpdk-dev] [PATCH 10/17] librte_acl: add AVX2 as new rte_acl= _classify() method >=20 > On Tue, Dec 16, 2014 at 04:16:48PM +0000, Ananyev, Konstantin wrote: > > > > > > > -----Original Message----- > > > From: Neil Horman [mailto:nhorman@tuxdriver.com] > > > Sent: Monday, December 15, 2014 8:21 PM > > > To: Ananyev, Konstantin > > > Cc: dev@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@tuxdriver.com] > > > > > Sent: Monday, December 15, 2014 4:00 PM > > > > > To: Ananyev, Konstantin > > > > > Cc: dev@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 wrot= e: > > > > > > Introduce new classify() method that uses AVX2 instructions. > > > > > > From my measurements: > > > > > > On HSW boards when processing >=3D 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 s= upports AVX2. > > > > > > > > > > > > Signed-off-by: Konstantin Ananyev > > > > > > --- > > > > > > 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-$(CONFIG_RTE_LIBRTE_ACL) +=3D acl_bld.c > > > > > > SRCS-$(CONFIG_RTE_LIBRTE_ACL) +=3D acl_gen.c > > > > > > SRCS-$(CONFIG_RTE_LIBRTE_ACL) +=3D acl_run_scalar.c > > > > > > SRCS-$(CONFIG_RTE_LIBRTE_ACL) +=3D acl_run_sse.c > > > > > > +SRCS-$(CONFIG_RTE_LIBRTE_ACL) +=3D acl_run_avx2.c > > > > > > > > > > > > CFLAGS_acl_run_sse.o +=3D -msse4.1 > > > > > > +ifeq ($(CC), icc) > > > > > > +CFLAGS_acl_run_avx2.o +=3D -march=3Dcore-avx2 > > > > > > +else ifneq ($(shell \ > > > > > > +test $(GCC_MAJOR_VERSION) -le 4 -a $(GCC_MINOR_VERSION) -le 6 = && echo 1), 1) > > > > > > +CFLAGS_acl_run_avx2.o +=3D -mavx2 > > > > > > +else > > > > > > +CFLAGS_acl_run_avx2.o +=3D -msse4.1 > > > > > > +endif > > > > > > > > > > > This seems broken. You've unilaterally included acl_run_avx2.c i= n the build > > > > > list above, but only enable -mavx2 if the compiler is at least gc= c 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() woul= d not be defined. > > > > And then at runtime, I have to check for that somehow and (re)popul= ate 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 g= ood, since > > > the right thing will just automatically happen then if you don't buil= d 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 __AV= X__ > > > 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 S= SE4.1. > > So for SSE4.1 we only has to worry about situation when target CPU does= n'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 do= esn'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_classi= fy_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 s= upport. > > There is absolutely no difference here. > > > > > 2) If an application selects an AVX2 classifier, I as a developer exp= ect to > > > either get AVX2 based classification, or an error indicating that I c= an't do > > > AVX2 classification, not a silent performance degradation down to sca= lar > > > classification. > > > > In fact I was considering both variants for compilers not supporting AV= X2: > > 1. silently degrade to SSE method. > > 2. create a dummy function rte_acl_classify_error() and put it into c= lassify_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 improvemen= t 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 h= ere 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 i= n 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 n= ew in rte_acl.c instead. >>From my understanding, we we'll need something like that: static const rte_acl_classify_t classify_fns[] =3D { [RTE_ACL_CLASSIFY_DEFAULT] =3D rte_acl_classify_scalar, [RTE_ACL_CLASSIFY_SCALAR] =3D rte_acl_classify_scalar, [RTE_ACL_CLASSIFY_SSE] =3D rte_acl_classify_sse, +#if (defined __GNUC__ && __GNUC__ <=3D 4 && __GNUC_MINOR__ < 7) + [RTE_ACL_CLASSIFY_AVX2] =3D rte_acl_classify_error, +#else =20 [RTE_ACL_CLASSIFY_AVX2] =3D rte_acl_classify_avx2, +#endif }; static void __attribute__((constructor)) rte_acl_init(void) { enum rte_acl_classify_alg alg =3D RTE_ACL_CLASSIFY_DEFAULT; +#if (defined __GNUC__ && __GNUC__ <=3D 4 && __GNUC_MINOR__ < 7) if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2)) alg =3D 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 =3D RTE_ACL_CLASSIFY_SSE; +#endif rte_acl_set_default_classify(alg); } Correct? Konstantin >=20 > > > > > > > > That in turn I think allows you to remove a > > > > > bunch of the ifdeffing that you've done in some of the avx2 speci= fic 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 defin= tion. > > > > 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=3Dy. > > 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 > >