Martin Albrecht wrote: >> It would appear this cache program is not particularly good at guess cache >> sizes, but what is probably more important is that it is able to optimise >> the code so it works best on a particular processor. It is pretty >> irrelevant what it thinks that processor may or may not be. It's more >> important it optimises itself to run well on that processor. > > Agreed, it would be irrelevant what the cache size is if this program would > run code that is quite similar to the actual code that it optimises for. > However, the code in cache.c is quite different from the actual code. I > should > investigate how the values suggested by cache.c modify the running time of > the > actual code. > > Martin >
You should also check if the library actually works best when given the true cache sizes. You may find it does not, in which case it would be a waste of time to get a program that was better able to determine the cache size. My guess is that as processors get more and more complex, with multiple cores, multiple cache levels, perhaps common or multiple caches across different cores, hyperthreading .. etc etc, it may be more and more difficult to optimise the library even if given all the characteristics of the processor(s). ATLAS library tunes itself to the processor(s). Perhaps your code could do likewise, although hopefully in less time than what ATLAS does, as that is taking an excessive amount of time on processors it does not know about. It seems to me, that any attempt to keep a list of processors (as ATLAS does), is almost doomed, as they get introduced so often. It's much better if the code can dynamically tune itself. Of course, that is easier said than it is done! dave --~--~---------~--~----~------------~-------~--~----~ To post to this group, send an email to sage-devel@googlegroups.com To unsubscribe from this group, send an email to sage-devel-unsubscr...@googlegroups.com For more options, visit this group at http://groups.google.com/group/sage-devel URL: http://www.sagemath.org -~----------~----~----~----~------~----~------~--~---
