From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ravi Kerur Subject: Re: [PATCH] Implement memcmp using AVX/SSE instructio Date: Thu, 23 Apr 2015 15:26:40 -0700 Message-ID: References: <1429716828-19012-1-git-send-email-rkerur@gmail.com> <1429716828-19012-2-git-send-email-rkerur@gmail.com> <55389E44.8030603@intel.com> <20150423081138.GA8592@bricha3-MOBL3> <2601191342CEEE43887BDE71AB97725821420FC7@irsmsx105.ger.corp.intel.com> <20150423140042.GA7248@bricha3-MOBL3> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Cc: "dev-VfR2kkLFssw@public.gmane.org" To: Bruce Richardson Return-path: In-Reply-To: <20150423140042.GA7248@bricha3-MOBL3> List-Id: patches and discussions about DPDK List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces-VfR2kkLFssw@public.gmane.org Sender: "dev" On Thu, Apr 23, 2015 at 7:00 AM, Bruce Richardson < bruce.richardson-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org> wrote: > On Thu, Apr 23, 2015 at 06:53:44AM -0700, Ravi Kerur wrote: > > On Thu, Apr 23, 2015 at 2:23 AM, Ananyev, Konstantin < > > konstantin.ananyev-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org> wrote: > > > > > > > > > > > > -----Original Message----- > > > > From: dev [mailto:dev-bounces-VfR2kkLFssw@public.gmane.org] On Behalf Of Bruce > Richardson > > > > Sent: Thursday, April 23, 2015 9:12 AM > > > > To: Wodkowski, PawelX > > > > Cc: dev-VfR2kkLFssw@public.gmane.org > > > > Subject: Re: [dpdk-dev] [PATCH] Implement memcmp using AVX/SSE > instructio > > > > > > > > On Thu, Apr 23, 2015 at 09:24:52AM +0200, Pawel Wodkowski wrote: > > > > > On 2015-04-22 17:33, Ravi Kerur wrote: > > > > > >+/** > > > > > >+ * Compare bytes between two locations. The locations must not > > > overlap. > > > > > >+ * > > > > > >+ * @note This is implemented as a macro, so it's address should > not > > > be taken > > > > > >+ * and care is needed as parameter expressions may be evaluated > > > multiple times. > > > > > >+ * > > > > > >+ * @param src_1 > > > > > >+ * Pointer to the first source of the data. > > > > > >+ * @param src_2 > > > > > >+ * Pointer to the second source of the data. > > > > > >+ * @param n > > > > > >+ * Number of bytes to compare. > > > > > >+ * @return > > > > > >+ * true if equal otherwise false. > > > > > >+ */ > > > > > >+static inline bool > > > > > >+rte_memcmp(const void *src_1, const void *src, > > > > > >+ size_t n) __attribute__((always_inline)); > > > > > You are exposing this as public API, so I think you should follow > > > > > description bellow or not call this _memcmp_ > > > > > > > > > > int memcmp(const void *s1, const void *s2, size_t n); > > > > > > > > > > The memcmp() function returns an integer less than, equal to, or > > > greater > > > > > than > > > > > zero if the first n bytes of s1 is found, > respectively, > > > to be > > > > > less than, to > > > > > match, or be greater than the first n bytes of s2. > > > > > > > > > > > > > +1 to this point. > > > > > > > > Also, if I read your quoted performance numbers in your earlier mail > > > correctly, > > > > we are only looking at a 1-4% performance increase. Is the additional > > > code to > > > > maintain worth the benefit? > > > > > > Yep, same thought here, is it really worth it? > > > Konstantin > > > > > > > > > > > /Bruce > > > > > > > > > -- > > > > > Pawel > > > > > > > I think I haven't exploited every thing x86 has to offer to improve > > performance. I am looking for inputs. Until we have exhausted all > avenues I > > don't want to drop it. One thing I have noticed is that bigger key size > > gets better performance numbers. I plan to re-run perf tests with 64 and > > 128 bytes key size and will report back. Any other avenues to try out > > please let me know I will give it a shot. > > > > Thanks, > > Ravi > > Hi Ravi, > > are 128 byte comparisons realistic? An IPv6 5-tuple with double vlan tags > is still > only 41 bytes, or 48 with some padding added? > While for a memcpy function, you can see cases where you are going to copy > a whole > packet, meaning that sizes of 128B+ (up to multiple k) are realistic, it's > harder > to see that for a compare function. > > In any case, we await the results of your further optimization work to see > how > that goes. > > Hi Bruce, Couple of things I am planning to try 1. Use _xor_ and _testz_ instructions for comparison instead of _cmpeq_ and _mask_. 2. I am using unaligned loads, not sure about the penalty, I plan to try with aligned loads if address is aligned and compare results. Agreed that with just L3 or even if we go with L2 + L3 + L4 tuples it will not exceed 64 bytes, 128 bytes is just a stretch for some weird MPLSoGRE header formats. My focus is currently on improving performance for < 64 bytes and < 128 bytes key lengths only. Thanks, Ravi Regards, > /Bruce >