From owner-freebsd-arm@FreeBSD.ORG Sat Sep 7 08:29:51 2013 Return-Path: Delivered-To: freebsd-arm@freebsd.org Received: from mx1.freebsd.org (mx1.freebsd.org [IPv6:2001:1900:2254:206a::19:1]) (using TLSv1 with cipher ADH-AES256-SHA (256/256 bits)) (No client certificate requested) by hub.freebsd.org (Postfix) with ESMTP id 291252D9 for ; Sat, 7 Sep 2013 08:29:51 +0000 (UTC) (envelope-from andrew@fubar.geek.nz) Received: from nibbler.fubar.geek.nz (nibbler.fubar.geek.nz [199.48.134.198]) by mx1.freebsd.org (Postfix) with ESMTP id 0C94A238D for ; Sat, 7 Sep 2013 08:29:50 +0000 (UTC) Received: from bender.Home (97e5e46b.skybroadband.com [151.229.228.107]) by nibbler.fubar.geek.nz (Postfix) with ESMTPSA id 04B365DFF7; Sat, 7 Sep 2013 08:29:43 +0000 (UTC) Date: Sat, 7 Sep 2013 09:29:37 +0100 From: Andrew Turner To: Warner Losh Subject: Re: stream benchmarking on RPi Message-ID: <20130907092937.0aee2245@bender.Home> In-Reply-To: References: Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Cc: freebsd-arm@freebsd.org X-BeenThere: freebsd-arm@freebsd.org X-Mailman-Version: 2.1.14 Precedence: list List-Id: Porting FreeBSD to the StrongARM Processor List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Sat, 07 Sep 2013 08:29:51 -0000 On Thu, 5 Sep 2013 14:49:19 -0600 Warner Losh wrote: > > On Sep 5, 2013, at 2:02 PM, Jia-Shiun Li wrote: > > > Hi all, > > > > just did a brief test using stream, the memory bandwidth benchmark, > > on RPi with Raspbian and FreeBSD. Share these info to see if someone > > might think of it useful. > > > > FreeBSD is faster at copying. I guess that must be attributed to > > recent VM and/or superpage commits. I remembered it to be under > > 300MB/s months before. On the other hand, scale, add, and triad are > > significantly slower. Anyone have clues or any wild guesses? > > Soft float? This looks to be the case, stream uses floating point arithmetic for the scale, add, and triad tests. The copy test is a straight copy from one array of doubles to another. It looks like most of the speed up in the copy test is from the move from gcc to clang. I am planning on investigating moving to softfp on armv6 for 10.1. I don't expect this to cause any major issues as it is compatible with the current float ABI in that both pass floating point values in ARM registers. I think this will improve the test results, but may not be at the same level as Raspbian appears to use the hard-float ABI. This is also planned, however I am unsure when this will appear. Andrew