Also. If you have GPU implementation of a matrix, or use another type of a vector processor, there is no way you can program that in if you force vector operations to use a visitor patterns.
On Dec 29, 2012, at 6:43 PM, Konstantin Berlin <kber...@gmail.com> wrote: > That's a good point about the compiler. I never tested the performance of > visitors vs. sequential array access. I just don't want the vector operations > to be tied to any particular implementation detail. > > On Dec 29, 2012, at 6:30 PM, Ted Dunning <ted.dunn...@gmail.com> wrote: > >> Actually, the visitor pattern or variants thereof can produce very >> performant linear algebra implementations. You can't usually get quite >> down to optimized BLAS performance, but you get pretty darned fast code. >> >> The reason is that the visitor is typically a very simple class which is >> immediately inlined by the JIT. Then it is subject to all of the normal >> optimizations exactly as if the code were written as a single concrete >> loop. For many implementations, the bounds checks will be hoisted out of >> the loop so you get pretty decent code. >> >> More importantly in many cases, visitors allow in place algorithms. >> Combined with view operators that limit visibility to part of a matrix, >> and the inlining phenomenon mentioned above, this can have enormous >> implications to performance. >> >> A great case in point is the Mahout math library. With no special efforts >> taken and using the visitor style fairly ubiquitously, I can get about 2 G >> flops from my laptop. Using Atlas as a LINPAK implementation gives me >> about 5 G flops. >> >> I agree with the point that linear algebra operators should be used where >> possible, but that just isn't feasible for lots of operations in real >> applications. Getting solid performance with simple code in those >> applications is a real virtue. >> >> On Sat, Dec 29, 2012 at 3:22 PM, Konstantin Berlin <kber...@gmail.com>wrote: >> >>> While visitor pattern is a good abstraction, I think it would make for >>> terrible linear algebra performance. All operations should be based on >>> basic vector operations, which internally can take advantage of sequential >>> memory access. For large problems it makes a difference. The visitor >>> pattern is a nice add on, but it should not be the engine driving the >>> package under the hood, in my opinion. >>> > --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org For additional commands, e-mail: dev-h...@commons.apache.org
