On Fri, Aug 11, 2006 at 02:29:32AM -0700, Nathaniel Smith wrote: > At a cursory glance, gcrypt's implementation looks just the same as > all the others; I don't know what makes it so fast.
It looks like it is in the expansion code (and actually if you look at the SIMD implementation, the only place SSE is actually used is in the message expansion, the rest of it is basically a standard C implementation). Just out of curiosity, I commented out the message expansion loop in the Botan code, and it ran about 40% faster (just under 100 MiB/s on a 1.3 GHz Athlon). I'm not sure how to make that loop much faster using only portable code, though, and I don't have a lot of free time at the moment to try. An off the wall idea: run multiple SHA-1 compression functions in parallel using SSE2/other SIMD. That would require changes to how calculate_ident is called, in addition to probably being a lot of work in general, but it seems like an interesting avenue. -Jack _______________________________________________ Monotone-devel mailing list [email protected] http://lists.nongnu.org/mailman/listinfo/monotone-devel
