the accuracy is controlled by round_off etc, O2, O3, .. etc are tied with roundoff, iEEE internal flags the fp-accuracy is just a gcc option
On Sat, Aug 13, 2011 at 6:07 AM, Min Zhao <mzhao...@gmail.com> wrote: > Hi Ram, > > I was going to use -fp-accuracy option. I found it is very > counter-initutive when "-fp-accuracy=relaxed" means less fp freedom > but more fp accuracy. "relaxed" does not sound like improving accuracy > at all. > > Can we change the setting of -fp-accuracy to be more easily > understandable, like "strict|strict-fma|limited|relaxed"? > > Also, the default -O3 setting does not map to any of -fp-accuracy > value, which seems strange to me. > > Thanks, > > Min > > > On Wed, Apr 27, 2011 at 5:47 AM, Ramanarayanan, Ramshankar > <ramshankar.ramanaraya...@amd.com> wrote: >> Sun, >> >> I think my choice of words contradict with the keywords used in the flags. >> >> The main change here is enabling -OPT:roundoff=2 at -O3 instead of >> -OPT:roundoff=1. This change enables fast-math functions, aggressive loop >> nest optimizations, reassociation on floating point expressions and more >> aggressive round-off settings. These aggressive floating point optimizations >> improve performance when using the -O3 flag but may affect floating point >> accuracy. The use of -fp-accuracy=relaxed in addition to -O3 is recommended >> for cases which need more floating point accuracy. -fp-accuracy=relaxed >> automatically sets -OPT:roundoff=1. User may also use -OPT:roundoff=1. >> >> Please let me know if you have questions. >> >> Ram >> >> -----Original Message----- >> From: Sun Chan [mailto:sun.c...@gmail.com] >> Sent: Wednesday, April 27, 2011 5:02 PM >> To: Ramanarayanan, Ramshankar >> Cc: open64-devel@lists.sourceforge.net >> Subject: Re: [Open64-devel] code review request for update to O3 flag >> >> you are really saying that to get back previous behavior, one needs to >> -fp_accuracy=not_relaxed or something to that effect? I don't follow >> your message. >> Sun >> >> On Wed, Apr 27, 2011 at 6:37 PM, Ramanarayanan, Ramshankar >> <ramshankar.ramanaraya...@amd.com> wrote: >>> Could a gate keeper approve this patch? >>> >>> >>> >>> This update enhances performance of the compiled code on X8664 when using >>> the O3 flag. Improvements come mainly from relaxing the floating point >>> accuracy setting at O3. This enables a wide range of optimizations including >>> loop nest optimizations and associative redundancy elimination >>> optimizations. Given this change, users will need to use >>> -fp-accuracy=relaxed flag in addition to -O3 if they require the earlier >>> floating point precision. During subsequent tuning we found that the bad >>> reference bias heuristic affects the computed cache costs and leads to >>> incorrect choice of inner loops and is thus ignored. >>> >>> >>> >>> Following tests have been conducted with this change. >>> >>> >>> >>> 1: No compiler time failure for x86 build >>> >>> 2: SPEC CPU 2006 validated with AMD flags and with O3 flag >>> >>> 3: The gcc regression suite has no new failures on x86/Linux >>> >>> >>> >>> Best regards, >>> >>> Ram >>> >>> >>> >>> Ramshankar Ramanarayanan >>> >>> Member of Technical Staff >>> >>> Open Source Compiler Engineering >>> >>> Advanced Micro Devices, Bangalore >>> >>> >>> >>> >>> >>> ------------------------------------------------------------------------------ >>> WhatsUp Gold - Download Free Network Management Software >>> The most intuitive, comprehensive, and cost-effective network >>> management toolset available today. Delivers lowest initial >>> acquisition cost and overall TCO of any competing solution. >>> http://p.sf.net/sfu/whatsupgold-sd >>> _______________________________________________ >>> Open64-devel mailing list >>> Open64-devel@lists.sourceforge.net >>> https://lists.sourceforge.net/lists/listinfo/open64-devel >>> >>> >> >> >> >> ------------------------------------------------------------------------------ >> WhatsUp Gold - Download Free Network Management Software >> The most intuitive, comprehensive, and cost-effective network >> management toolset available today. Delivers lowest initial >> acquisition cost and overall TCO of any competing solution. >> http://p.sf.net/sfu/whatsupgold-sd >> _______________________________________________ >> Open64-devel mailing list >> Open64-devel@lists.sourceforge.net >> https://lists.sourceforge.net/lists/listinfo/open64-devel >> > ------------------------------------------------------------------------------ FREE DOWNLOAD - uberSVN with Social Coding for Subversion. Subversion made easy with a complete admin console. Easy to use, easy to manage, easy to install, easy to extend. Get a Free download of the new open ALM Subversion platform now. http://p.sf.net/sfu/wandisco-dev2dev _______________________________________________ Open64-devel mailing list Open64-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/open64-devel
