Author: Vincent Diepeveen
Date: 12:02:35 10/02/02
Go up one level in this thread
On October 01, 2002 at 15:22:45, Jeremiah Penery wrote: >On October 01, 2002 at 09:21:35, Vincent Diepeveen wrote: > >>On September 30, 2002 at 00:09:28, Jeremiah Penery wrote: >> >>>On September 29, 2002 at 23:31:36, Robert Hyatt wrote: >>> >>>>I don't know what this means. I have several dozen programs (Crafty >>>>is only one) that we have run using intel's compiler and gcc, and in >>>>_every_ case, Intel's compiler is faster. On P2's, on P3's and on >>>>P4's... Of course I wouldn't use intel's compiler for an AMD chip, >>>>why would they want to optimize for a competitor's chip??? >>> >>>They don't have to optimize specifically for the competitor's chip, as Intel >>>compiler still produces probably the fastest binaries for AMD machines. Any >>>general optimizations (P2, P3, and even P4 optimizations (excluding SSE2 stuff >>>or whatever)) are just as helpful for AMD processors as they are for Intel ones. >> >>Oh well very old intel c++ compilers i already found out i couldn't >>use P4 optimizations. I never would *assume* in fact that P4 optimizations >>are downwards compatible. >> >>So i never went further than P3 optimizations for all other hardware than >>P4. >> >>And it's definitely *not* the case that optimizations that work for P3 >>also are fast for AMD. A good example is doing 2 vector instructions >>after each other at the AMD, which is very bad for it. There are of course >>thousands of other combinations which hurt AMD more than P3 or P2. > >And there are thousands of combinations that probably help AMD more than P3/P2, >but they still help P3/P2(P4), so the Intel compiler may be doing them. > >>I hope you realize a chip is a very complex thing. It's millions of >>transistors, so there is plenty of combinations to execute instructions >>which are getting some kind of penalty on cpu A and not on cpu I. >>In short there is plenty of room to make your own cpu look better >>than a competing cpu. > >So please tell me why Intel compiler is still the fastest one for AMD >processors. :) Exactly, it isn't the fastest one. default msvc compile is way faster for DIEP. intel c++ 5.01 is faster on K7 than intel c++ 6.0 is for DIEP. For very obvious reasons... but still both are hell slower than gcc 3.2. MSVC is 2% slower than intel c++ 5.01 is for DIEP (using profile info) so obviously the 5% or so that the profile info gives to extra speed for intel c++ means in fact that it is 3% slower. In short the intel c++ is way way slower than msvc *everywhere* if we take into account a version that didn't use profile info. in case of intel c++ 6.0 even with. gcc 3.2 amazingly beats them all, not because it is faster than msvc, no it isn't. but it gets 20% faster using the profile info giving it 5% faster speed at a k7 than the fastest possible msvc version does (which is msvc 6.0 sp4 with processor pack; that's 2% faster than any other msvc version is for me). Best regards, Vincent
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.