Date: Fri, 3 Sep 1999 09:53:05 +0100 (BST) From: Doug Rabson <dfr@nlsystems.com> To: Tony Griffiths <tonyg@onthenet.com.au> Cc: John Polstra <jdp@polstra.com>, alpha@FreeBSD.ORG Subject: Re: relative alpha speed Message-ID: <Pine.BSF.4.10.9909030951240.2081-100000@salmon.nlsystems.com> In-Reply-To: <37CF883C.B81B8FF5@OntheNet.com.au>
next in thread | previous in thread | raw e-mail | index | archive | help
On Fri, 3 Sep 1999, Tony Griffiths wrote: > Doug Rabson wrote: > > > On Thu, 2 Sep 1999, John Polstra wrote: > > > > > In article <Pine.BSF.4.10.9909012240130.706-100000@salmon.nlsystems.com>, > > > Doug Rabson <dfr@nlsystems.com> wrote: > > > > On Wed, 1 Sep 1999 jon@cops.com wrote: > > > > > > > > > I am a little perplexed about all of this. If the performance and > > > > > disk space usage are better on intel what benefit do I have using > > > > > an alpha instead of an intel... besides just being cooler than all > > > > > of my intel friends? > > > > > > > > Floating point performance rocks compared to intel. > > > > > > ... until you make the Alpha conform to the IEEE FP rules and handle > > > the whole range of numbers it is supposed to handle (-mieee). Then it > > > becomes much slower than ix86 once again. I've been disappointed in > > > the performance of the Alphas, given that performance is supposed to > > > be their strong point. > > > > I think the problem may be with our compiler. I don't think the penalty > > for using -mieee ought to be so high. I wonder how well the Compaq > > compiler deals with this. > > I've been lead to believe that the main problem with ieee format FP is the > method of handling exceptions (NANs, etc). Normally, with 'native' FP, > exceptions are handled imprecisely while I belive that with ieee FP the > default is precise exceptions. Of course with a pipelined architecture, > there is a considerable penalty handling precise exceptions (ie. the pipeline > has to be drained after each FP operation that could possibly generate an > exception!!!). > > Check to see if there is a conpiler switch for this... The compiler switch is what we are discussing. In egcs/gcc-2.95, -mieee forces the compiler to restrict its use of floating point in a way which allows the kernel to emulate operations involving or producing NaNs, INFs and denorms. This tends to slow the code down somewhat, possibly because gcc is being extra conservative. > > And yes, compiling for Alpha object code is going to be considerably slower > than for a CISC environment. True. -- Doug Rabson Mail: dfr@nlsystems.com Nonlinear Systems Ltd. Phone: +44 181 442 9037 To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-alpha" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Pine.BSF.4.10.9909030951240.2081-100000>