From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (qmail 4156 invoked by alias); 2 Sep 2016 18:53:52 -0000 Mailing-List: contact gcc-patches-help@gcc.gnu.org; run by ezmlm Precedence: bulk List-Id: List-Archive: List-Post: List-Help: Sender: gcc-patches-owner@gcc.gnu.org Received: (qmail 3991 invoked by uid 89); 2 Sep 2016 18:53:50 -0000 Authentication-Results: sourceware.org; auth=none X-Virus-Found: No X-Spam-SWARE-Status: No, score=1.0 required=5.0 tests=BAYES_20,KAM_LAZY_DOMAIN_SECURITY autolearn=no version=3.3.2 spammy=shave, successive, __d, 37726 X-HELO: blaine.gmane.org Received: from Unknown (HELO blaine.gmane.org) (195.159.176.226) by sourceware.org (qpsmtpd/0.93/v0.84-503-g423c35a) with ESMTP; Fri, 02 Sep 2016 18:53:49 +0000 Received: from list by blaine.gmane.org with local (Exim 4.84_2) (envelope-from ) id 1bftax-0006hd-PC for gcc-patches@gcc.gnu.org; Fri, 02 Sep 2016 20:53:43 +0200 To: gcc-patches@gcc.gnu.org From: Eelis Subject: Re: [patch, libstdc++] std::shuffle: Generate two swap positions at a time if possible Date: Fri, 02 Sep 2016 19:27:00 -0000 Message-ID: <57C9CAAE.50202@eelis.net> References: <5728B8CC.7030405@eelis.net> <20160831124502.GH3342@redhat.com> <57C9C304.90607@eelis.net> Mime-Version: 1.0 Content-Type: multipart/mixed; boundary="------------000304080604060001070806" User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:38.0) Gecko/20100101 Thunderbird/38.7.0 In-Reply-To: <57C9C304.90607@eelis.net> Cc: libstdc++@gcc.gnu.org X-IsSubscribed: yes X-SW-Source: 2016-09/txt/msg00117.txt.bz2 Message-ID: <20160902192700.v0SPL1L0Lmbcubiako4xXpU0eL1aDguNNHl36p9LU_4@z> This is a multi-part message in MIME format. --------------000304080604060001070806 Content-Type: text/plain; charset=windows-1252; format=flowed Content-Transfer-Encoding: 7bit Content-length: 1093 On 2016-09-02 20:20, Eelis van der Weegen wrote: > On 2016-08-31 14:45, Jonathan Wakely wrote: >> Is this significantly faster than just using >> uniform_int_distribution<_IntType>{0, __bound - 1}(__g) so we don't >> need to duplicate the logic? (And people maintaining the code won't >> reconvince themselves it's correct every time they look at it :-) >> >> [..] >> >> Could we hoist this test out of the loop somehow? >> >> If we change the loop condition to be __i+1 < __last we don't need to >> test it on every iteration, and then after the loop we can just do >> the final swap if (__urange % 2). > > Reusing std::uniform_int_distribution seems just as fast, so I've removed __generate_random_index_below. > > I've hoisted the (__i + 1 == __last) check out of the loop (in a slightly different way), and it seems to shave off a couple more cycles, yay! > > Updated patch attached. > Please ignore that patch, I used __g()&1 but that's invalid (the new "UniformRandomBitGenerator" name is misleading). Updated patch (which uses a proper distribution even for the [0,1] case) attached. --------------000304080604060001070806 Content-Type: text/x-patch; name="newer-double-step-shuffle.patch" Content-Transfer-Encoding: 7bit Content-Disposition: attachment; filename="newer-double-step-shuffle.patch" Content-length: 1965 Index: libstdc++-v3/include/bits/stl_algo.h =================================================================== --- libstdc++-v3/include/bits/stl_algo.h (revision 239895) +++ libstdc++-v3/include/bits/stl_algo.h (working copy) @@ -3772,6 +3772,47 @@ typedef typename std::make_unsigned<_DistanceType>::type __ud_type; typedef typename std::uniform_int_distribution<__ud_type> __distr_type; typedef typename __distr_type::param_type __p_type; + + typedef typename std::remove_reference<_UniformRandomNumberGenerator>::type _Gen; + typedef typename std::common_type::type __uc_type; + + const __uc_type __urngrange = __g.max() - __g.min(); + const __uc_type __urange = __uc_type(__last - __first); + + if (__urngrange / __urange >= __urange) + // I.e. (__urngrange >= __urange * __urange) but without wrap issues. + { + _RandomAccessIterator __i = __first + 1; + + // Since we know the range isn't empty, an even number of elements + // means an uneven number of elements /to swap/, in which case we + // do the first one up front: + + if ((__urange % 2) == 0) + { + __distr_type __d{0, 1}; + std::iter_swap(__i++, __first + __d(__g)); + } + + // Now we know that __last - __i is even, so we do the rest in pairs, + // using a single distribution invocation to produce swap positions + // for two successive elements at a time: + + while (__i != __last) + { + const __uc_type __swap_range = __uc_type(__i - __first) + 1; + const __uc_type __comp_range = __swap_range * (__swap_range + 1); + + std::uniform_int_distribution<__uc_type> __d{0, __comp_range - 1}; + const __uc_type __pospos = __d(__g); + + std::iter_swap(__i++, __first + (__pospos % __swap_range)); + std::iter_swap(__i++, __first + (__pospos / __swap_range)); + } + + return; + } + __distr_type __d; for (_RandomAccessIterator __i = __first + 1; __i != __last; ++__i) --------------000304080604060001070806--