Comments on: Computational performance — a beginner’s case study http://lbrandy.com/blog/2009/07/computational-performance-a-beginners-case-study/ { on programming and the internets, every monday } Fri, 03 Feb 2012 14:14:15 +0000 hourly 1 http://wordpress.org/?v=3.0.1 By: Ilmari Heikkinen http://lbrandy.com/blog/2009/07/computational-performance-a-beginners-case-study/comment-page-1/#comment-26245 Ilmari Heikkinen Wed, 24 Feb 2010 18:45:02 +0000 http://lbrandy.com/blog/?p=837#comment-26245 Looking at the ASM output for gcc -O3 -msse2 I noticed that horiz_add() and fast() use the SIMD 4-float vector instructions already (movaps, addps and friends.) And there's some funky business going on as the fast() and a manual SSE version written with SSE intrinsics are slower than horiz_add() on my Pentium M. Looking at the ASM output for gcc -O3 -msse2 I noticed that horiz_add() and fast() use the SIMD 4-float vector instructions already (movaps, addps and friends.) And there’s some funky business going on as the fast() and a manual SSE version written with SSE intrinsics are slower than horiz_add() on my Pentium M.

]]> By: Casey http://lbrandy.com/blog/2009/07/computational-performance-a-beginners-case-study/comment-page-1/#comment-26243 Casey Wed, 24 Feb 2010 17:37:23 +0000 http://lbrandy.com/blog/?p=837#comment-26243 Your conclusion also depends on the language you're working in if I recall my High Performance Computing book correctly. I don't see it specified, but I believe you're using C or C++. Fortran, I think, would be faster if you go vertical instead of horizontal. Your conclusion also depends on the language you’re working in if I recall my High Performance Computing book correctly. I don’t see it specified, but I believe you’re using C or C++. Fortran, I think, would be faster if you go vertical instead of horizontal.

]]> By: Evil teach http://lbrandy.com/blog/2009/07/computational-performance-a-beginners-case-study/comment-page-1/#comment-26237 Evil teach Wed, 24 Feb 2010 15:14:21 +0000 http://lbrandy.com/blog/?p=837#comment-26237 You should also look into the effect of page faulting on the problem. If your arrays are larger than your working set, that some of your memory can be faulted in/out. The resulting disk io can add significant costs to your result. You should also look into the effect of page faulting on the problem.

If your arrays are larger than your working set, that some of your memory can be faulted in/out. The resulting disk io can add significant costs to your result.

]]> By: Tom Finnigan http://lbrandy.com/blog/2009/07/computational-performance-a-beginners-case-study/comment-page-1/#comment-14705 Tom Finnigan Tue, 14 Jul 2009 20:29:21 +0000 http://lbrandy.com/blog/?p=837#comment-14705 One big takeaway from all this is that you need a good profiler. There are so many levels of optimization (compiler, OS, CPU, other hardware) that it's very hard to get enough intuition to predict performance just from the code. To get the best performance, you need to experiment with different methods, and choose the one that performs the best. But that takes a lot of time, and to justify that time investment you need to know that it's the bottle neck of your program. So, you need a good profiler. The other thing I wanted to mention is that when I develop on windows I miss the OSX performance tools. Shark is great, but so are the libraries, like the Accelerate framework. This example would be one function call, and the framework uses SIMD and threading as appropriate. Amazing to work with. One big takeaway from all this is that you need a good profiler.

There are so many levels of optimization (compiler, OS, CPU, other hardware) that it’s very hard to get enough intuition to predict performance just from the code. To get the best performance, you need to experiment with different methods, and choose the one that performs the best. But that takes a lot of time, and to justify that time investment you need to know that it’s the bottle neck of your program. So, you need a good profiler.

The other thing I wanted to mention is that when I develop on windows I miss the OSX performance tools. Shark is great, but so are the libraries, like the Accelerate framework. This example would be one function call, and the framework uses SIMD and threading as appropriate. Amazing to work with.

]]> By: louis http://lbrandy.com/blog/2009/07/computational-performance-a-beginners-case-study/comment-page-1/#comment-13816 louis Wed, 08 Jul 2009 13:59:54 +0000 http://lbrandy.com/blog/?p=837#comment-13816 I merged those changes. Thanks guys. I merged those changes. Thanks guys.

]]> By: Andy Kish http://lbrandy.com/blog/2009/07/computational-performance-a-beginners-case-study/comment-page-1/#comment-13786 Andy Kish Wed, 08 Jul 2009 06:36:54 +0000 http://lbrandy.com/blog/?p=837#comment-13786 Naixn: The weird output is fixed in my github fork: http://github.com/Kobold/simple-optimization-test/tree/master Louis: Check your fork queue. ;-) Naixn: The weird output is fixed in my github fork: http://github.com/Kobold/simple-optimization-test/tree/master

Louis: Check your fork queue. ;-)

]]> By: anonymous http://lbrandy.com/blog/2009/07/computational-performance-a-beginners-case-study/comment-page-1/#comment-13704 anonymous Tue, 07 Jul 2009 11:19:19 +0000 http://lbrandy.com/blog/?p=837#comment-13704 What about interlacing the arrays? What about interlacing the arrays?

]]>