Author: Matt Taylor
Date: 22:21:23 12/20/02
Go up one level in this thread
On December 20, 2002 at 16:43:20, Vincent Diepeveen wrote: >On December 19, 2002 at 16:21:19, Matt Taylor wrote: > >>On December 19, 2002 at 12:31:15, Robert Hyatt wrote: >> >>>On December 19, 2002 at 12:00:32, Vincent Diepeveen wrote: >>> >>>>On December 19, 2002 at 11:43:22, Robert Hyatt wrote: >>>> >>>>>On December 19, 2002 at 09:43:54, Vincent Diepeveen wrote: >>>>> >>>>>>On December 19, 2002 at 07:53:32, Frederic Louguet wrote: >>>>>> >>>>>>>My program Chess Wizard has always run much faster on Athlon than on P4, but >>>>>>>maybe using default optimizations is not enough. So I have a question for >>>>>>>compiler gurus : what are the best compiler optimizations you can use under >>>>>>>Microsoft Visual C++ 6.0 with processor pack, running under Windows XP, for >>>>>>>Athlon XP/MP and for Pentium 4. >>>>>>> >>>>>>>Before I try everything myself, maybe someone has made experiments and has >>>>>>>interesting results to share ? >>>>>> >>>>>>Yes. >>>>>> >>>>>>Try mingw with msys. >>>>>>it works under windows. >>>>>> >>>>>>download the latest gcc compiler. >>>>>>then compile a version with profile info. produce profile >>>>>>info. then recompile with that and branches get optimized. >>>>>> >>>>>>should get your chess wizard, assuming it's a non-bitboard (gcc >>>>>>is horrible optimizing 64 bits code for 32 bits processors), a lot >>>>>>faster. >>>>>> >>>>>>It sure works for diep very fast. >>>>> >>>>> >>>>>Also he should try the Intel C/C++ compiler v 7.0 download from Intel's >>>>>web site. It produces faster code for me than any version of gcc that is >>>>>available. >>>> >>>>you have a P4. he has an AMD K7. I bet even old visual c++ 6.0 sp4 procpack >>>>beats the intel 7.0 on that hardware ;) >>> >>> >>>I suppose that is the _big_ difference between us. I prefer to _test_ and >>>produce data that can be used to make a decision. Rather than "guesswork". >> >>The least he could do is read the optimization manuals and make a clever >>observation. You might as well call the Athlon optimization manual the P4 Jr. >>optimization manual. The bulk of the manuals are identical concepts with >>different wording and style. >> >>I have seen a number of tests which pretty much confirm that P4-optimized code >>is going to fly on Athlon, too. It won't be *best* on Athlon, but it's going to >>beat Visual C++ 6.0 SP4 with Processor Pack. When you think about it, it makes >>sense, too. All superscalar processors are suffering mostly from the same >>effects. >> >>-Matt > >Bob is wrong about the guess work. Testing shows simply that at athlon K7 >DIEP is a lot faster with gcc with fbranch-probabilities, than it is with >intel c++ (of course not even counting bugs in the code intel c++ >generates). > >Your assumption is wrong too. With regard to single cpu applications >(and specint which includes crafty is run single cpu as far as i know) >the P4 and the P3 and the K7 are not much different from the old PentiumPro. > >The weak chain of the pentiumpro and the PII definitely came back into the >K7 and the P4. The weak chain being clearly branches (i am not using >64 bits bitboards so all these problems crafty has i do not have with >that vector instructions might be faster or slower in some way). > >The penalty for branches at the P6 and P2 were already so huge in 1996, >that by now in 2002, every chessprogrammer who did a lot of optimization >work, definitely optimized this a lot. > >It is a gross misunderstanding guessing that i could change lossless >something with regard to single cpu DIEP to get it faster at the P4. > >Add to that, that the P4 optimizations from the intel compiler do not >seem to run on the K7 at all, apart from that they generate bugs. > >Blaming that onto the K7 is no good idea. 50% of the computerchess users >at least have a K7. If something is broken, then work around it, don't >blame it on a processor. Or the fact that the K7 and P7 do not completely share the same instruction set. >Yet the intel c++ lemma seems to be that the specint number must be increased. >Rest of world can drop dead from them seemingly when they can win 0.01% for >that goal. > >intel c++ 6.0 or 7.0 optimized executables will *never* run fast on K7 >when compared to speed other compilers or old 5.0 generates for it. Ironic that. I've run Intel C 6/7 binaries on my dual-K7. They work fine. They're faster than their gcc equivalents. I'm really not sure what you're talking about since my experiences have been quite good. -Matt
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.