Am 31.10.2014 um 16:46 schrieb Krzysztof Czarnowski: > Cachegrind's cache mode is not very general and I wonder if it makes > sense to use it on Quark to get at least approximate results on an > application's cache performance. The Quark's cache is integrated (both > instruction and data) single level 16kB size with16B line. > > My best bet that is to run > $ valgrind --tool=cachegrind --I1=32,2,16 --D1=32,2,16 --LL=16384,2,16 prog > and look at LL statistics. Sure it's not the real thing but seems the > closest I can get.
That should do it, yes. The misses of the LL should be a good estimation for the misses of the L1 unified cache of the Quark, provided that the Quark does something similar to LRU replacement (?). Allowing for more flexible cache hierarchy designs (e.g. just one unified L1) in cachegrind may be possible, but must be done with care. Naive solutions which dynamically check for cache configuration parameters easily could slow down the simulation on every platform significantly... To not forget about this, can you enter a bug report? Josef > > But maybe I miss something and there is some better way... Any > suggestions welcome, > Krzysztof > > > > > ------------------------------------------------------------------------------ > > > > _______________________________________________ > Valgrind-users mailing list > Valgrind-users@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/valgrind-users > ------------------------------------------------------------------------------ _______________________________________________ Valgrind-users mailing list Valgrind-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/valgrind-users
