From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from us-smtp-delivery-124.mimecast.com (us-smtp-delivery-124.mimecast.com [63.128.21.124]) by sourceware.org (Postfix) with ESMTP id DB96D3857C4E for ; Mon, 5 Oct 2020 23:40:37 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.3.2 sourceware.org DB96D3857C4E Received: from mimecast-mx01.redhat.com (mimecast-mx01.redhat.com [209.132.183.4]) (Using TLS) by relay.mimecast.com with ESMTP id us-mta-308-OekykyMxMkWGk0V8_J-WzA-1; Mon, 05 Oct 2020 19:40:33 -0400 X-MC-Unique: OekykyMxMkWGk0V8_J-WzA-1 Received: from smtp.corp.redhat.com (int-mx03.intmail.prod.int.phx2.redhat.com [10.5.11.13]) (using TLSv1.2 with cipher AECDH-AES256-SHA (256/256 bits)) (No client certificate requested) by mimecast-mx01.redhat.com (Postfix) with ESMTPS id A2325101FFA5; Mon, 5 Oct 2020 23:40:32 +0000 (UTC) Received: from localhost (unknown [10.33.37.1]) by smtp.corp.redhat.com (Postfix) with ESMTP id 520677AEC4; Mon, 5 Oct 2020 23:40:32 +0000 (UTC) Date: Tue, 6 Oct 2020 00:40:31 +0100 From: Jonathan Wakely To: Daniel Lemire Cc: libstdc++@gcc.gnu.org, gcc-patches@gcc.gnu.org Subject: Re: [PATCH, libstdc++] Improve the performance of std::uniform_int_distribution (fewer divisions) Message-ID: <20201005234031.GF7004@redhat.com> References: <20201005232515.GD7004@redhat.com> MIME-Version: 1.0 In-Reply-To: <20201005232515.GD7004@redhat.com> X-Clacks-Overhead: GNU Terry Pratchett X-Scanned-By: MIMEDefang 2.79 on 10.5.11.13 X-Mimecast-Spam-Score: 0 X-Mimecast-Originator: redhat.com Content-Type: text/plain; charset=us-ascii; format=flowed Content-Disposition: inline X-Spam-Status: No, score=-8.7 required=5.0 tests=BAYES_00, DKIMWL_WL_HIGH, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, DKIM_VALID_EF, RCVD_IN_DNSWL_NONE, RCVD_IN_MSPIKE_H5, RCVD_IN_MSPIKE_WL, SPF_HELO_NONE, SPF_PASS, TXREP autolearn=unavailable autolearn_force=no version=3.4.2 X-Spam-Checker-Version: SpamAssassin 3.4.2 (2018-09-13) on server2.sourceware.org X-BeenThere: libstdc++@gcc.gnu.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Libstdc++ mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Mon, 05 Oct 2020 23:40:39 -0000 On 06/10/20 00:25 +0100, Jonathan Wakely wrote: >I'm sorry it's taken a year to review this properly. Comments below ... > >On 27/09/19 14:18 -0400, Daniel Lemire wrote: >>(This is a revised patch proposal. I am revising both the description >>and the code itself.) >> >>Even on recent processors, integer division is relatively expensive. >>The current implementation of std::uniform_int_distribution typically >>requires two divisions by invocation: >> >> // downscaling >> const __uctype __uerange = __urange + 1; // __urange can be zero >> const __uctype __scaling = __urngrange / __uerange; >> const __uctype __past = __uerange * __scaling; >> do >> __ret = __uctype(__urng()) - __urngmin; >> while (__ret >= __past); >> __ret /= __scaling; >> >>We can achieve the same algorithmic result with at most one division, >>and typically no division at all without requiring more calls to the >>random number generator. >>This was recently added to Swift (https://github.com/apple/swift/pull/25286) >> >>The main challenge is that we need to be able to compute the "full" >>product. E.g., given two 64-bit integers, we want the 128-bit result; >>given two 32-bit integers we want the 64-bit result. This is fast on >>common processors. >>The 128-bit product is not natively supported in C/C++ but can be >>achieved with the >>__int128 extension when it is available. The patch checks for >>__int128 support; when >>support is lacking, we fallback on the existing approach which uses >>two divisions per >>call. >> >>For example, if we replace the above code by the following, we get a substantial >>performance boost on skylake microarchitectures. E.g., it can >>be twice as fast to shuffle arrays of 1 million elements (e.g., using >>the followingbenchmark: https://github.com/lemire/simple_cpp_shuffle_benchmark ) >> >> >> unsigned __int128 __product = (unsigned >>__int128)(__uctype(__urng()) - __urngmin) * uint64_t(__uerange); >> uint64_t __lsb = uint64_t(__product); >> if(__lsb < __uerange) >> { >> uint64_t __threshold = -uint64_t(__uerange) % uint64_t(__uerange); >> while (__lsb < __threshold) >> { >> __product = (unsigned __int128)(__uctype(__urng()) - >>__urngmin) * (unsigned __int128)(__uerange); >> __lsb = uint64_t(__product); >> } >> } >> __ret = __product >> 64; >> >>Included is a patch that would bring better performance (e.g., 2x gain) to >>std::uniform_int_distribution in some cases. Here are some actual numbers... >> >>With this patch: >> >>std::shuffle(testvalues, testvalues + size, g) : 7952091 >>ns total, 7.95 ns per input key >> >>Before this patch: >> >>std::shuffle(testvalues, testvalues + size, g) : >>14954058 ns total, 14.95 ns per input key >> >> >>Compiler: GNU GCC 8.3 with -O3, hardware: Skylake (i7-6700). >> >>Furthermore, the new algorithm is unbiased, so the randomness of the >>result is not affected. >> >>I ran both performance and biases tests with the proposed patch. >> >> >>This patch proposal was improved following feedback by Jonathan >>Wakely. An earlier version used the __uint128_t type, which is widely >>supported but not used in the C++ library, instead we now use unsigned >>__int128. Furthermore, the previous patch was accidentally broken: it >>was not computing the full product since a rhs cast was missing. These >>issues are fixed and verified. > >After looking at GCC's internals, it looks like __uint128_t is >actually fine to use, even though we never currently use it in the >library. I didn't even know it was supported for C++ mode, sorry! > >>Reference: Fast Random Integer Generation in an Interval, ACM Transactions on >>Modeling and Computer Simulation 29 (1), 2019 https://arxiv.org/abs/1805.10941 > >>Index: libstdc++-v3/include/bits/uniform_int_dist.h >>=================================================================== >>--- libstdc++-v3/include/bits/uniform_int_dist.h (revision 276183) >>+++ libstdc++-v3/include/bits/uniform_int_dist.h (working copy) >>@@ -33,7 +33,8 @@ >> >>#include >>#include >>- >>+#include >>+#include >>namespace std _GLIBCXX_VISIBILITY(default) >>{ >>_GLIBCXX_BEGIN_NAMESPACE_VERSION >>@@ -239,18 +240,61 @@ >> = __uctype(__param.b()) - __uctype(__param.a()); >> >> __uctype __ret; >>- >>- if (__urngrange > __urange) >>+ if (__urngrange > __urange) >> { >>- // downscaling >>- const __uctype __uerange = __urange + 1; // __urange can be zero >>- const __uctype __scaling = __urngrange / __uerange; >>- const __uctype __past = __uerange * __scaling; >>- do >>- __ret = __uctype(__urng()) - __urngmin; >>- while (__ret >= __past); >>- __ret /= __scaling; >>- } >>+ const __uctype __uerange = __urange + 1; // __urange can be zero >>+#if _GLIBCXX_USE_INT128 == 1 >>+ if(sizeof(__uctype) == sizeof(uint64_t) and >>+ (__urngrange == numeric_limits::max())) >>+ { >>+ // 64-bit case >>+ // reference: Fast Random Integer Generation in an Interval >>+ // ACM Transactions on Modeling and Computer Simulation 29 (1), 2019 >>+ // https://arxiv.org/abs/1805.10941 >>+ unsigned __int128 __product = (unsigned __int128)(__uctype(__urng()) - __urngmin) * uint64_t(__uerange); > >Is subtracting __urngmin necessary here? > >The condition above checks that __urngrange == 2**64-1 which means >that U::max() - U::min() is the maximum 64-bit value, which means >means U::max()==2**64-1 and U::min()==0. So if U::min() is 0 we don't >need to subtract it. > >Also, I think the casts to uint64_t are unnecessary. We know that >__uctype is an unsigned integral type, and we've checked that it has >exactly 64-bits, so I think we can just use __uctype. It's got the >same width and signedness as uint64_t anyway. > >That said, the uint64_t(__uerange) above isn't redundant, because it >should be (unsigned __int128)__uerange, I think. Ah yes, you pointed out that last bit in your Sept 28 2019 email.