Comments on: Matrix Multiplication http://matrixprogramming.com/2008/01/matrixmultiply Compiling numerical libraries Tue, 15 May 2012 02:26:33 +0000 hourly 1 http://wordpress.org/?v=3.2.1 By: Evgenii Rudnyi http://matrixprogramming.com/2008/01/matrixmultiply/comment-page-1#comment-855 Evgenii Rudnyi Tue, 06 Dec 2011 22:10:03 +0000 http://1.rudnyi.peterhost.ru/?p=312#comment-855 You are right that it would be good to check all possible optimization flags. I have limited myself in this study by -O3 only. In my view though, it is unlikely that by pure optimization at the compiler level one will achieve the performance of the optimized BLAS. Well, if you try it, please report your results here. You are right that it would be good to check all possible optimization flags. I have limited myself in this study by -O3 only.

In my view though, it is unlikely that by pure optimization at the compiler level one will achieve the performance of the optimized BLAS. Well, if you try it, please report your results here.

]]> By: bit http://matrixprogramming.com/2008/01/matrixmultiply/comment-page-1#comment-854 bit Tue, 06 Dec 2011 21:28:54 +0000 http://1.rudnyi.peterhost.ru/?p=312#comment-854 I potentially see one flaw here that is not addressed. When you compiled your programs did you enable auto-vectorization and auto-parallelization? I noticed in your compiler flags you did not include any cpu optimizations or anythign like that. I know from experience that enabling sse2 optimizations and autovectorization of these kinds of loops decreases compiler time significantly. The python modules would have certainly compiled with this in mind and in order to do a fair comparison you would need to ensure that you enable these types of optimizations. If you do this I think you will find that they all run about the same speed. If I can find it I have a list of compiler flags to do this and then the test could be reapproached with them I potentially see one flaw here that is not addressed. When you compiled your programs did you enable auto-vectorization and auto-parallelization? I noticed in your compiler flags you did not include any cpu optimizations or anythign like that. I know from experience that enabling sse2 optimizations and autovectorization of these kinds of loops decreases compiler time significantly. The python modules would have certainly compiled with this in mind and in order to do a fair comparison you would need to ensure that you enable these types of optimizations. If you do this I think you will find that they all run about the same speed.

If I can find it I have a list of compiler flags to do this and then the test could be reapproached with them

]]>