http://danbystrom.se/2008/12/14/improving-performance/ Blog without an interesting name Fri, 06 Jan 2012 17:43:21 +0000 hourly 1 http://wordpress.com/ http://danbystrom.se/2008/12/14/improving-performance/#comment-114 Fri, 08 May 2009 10:48:37 +0000 http://danbystrom.se/?p=117#comment-114 Ah, that’s really smart!

]]> http://danbystrom.se/2008/12/14/improving-performance/#comment-110 Tue, 05 May 2009 00:18:07 +0000 http://danbystrom.se/?p=117#comment-110 Actually, it’s not quite based on picture size http://starcalc.110mb.com/evolisa.zip

]]> http://danbystrom.se/2008/12/14/improving-performance/#comment-108 Mon, 04 May 2009 08:58:12 +0000 http://danbystrom.se/?p=117#comment-108 Yes, if you’re willing to have two different loops depending on picture size, then you can have one version with uint and one with double. Have you tried using generics to avoid code duplication? The assembler version is definetly not written for a 64-bit processor, on the contrary: if you do run your C# version on a 64-bit processor, then the JIT compiler will produce code optimized for your processor and so you may get better performance than my 32-bit code.

]]> http://danbystrom.se/2008/12/14/improving-performance/#comment-106 Sun, 03 May 2009 15:42:59 +0000 http://danbystrom.se/?p=117#comment-106 I used your improvement, and I have an improvement of my own. Using ints instead of doubles for pictures with under 10,922 pixels (using an if statement to call the correct method) will improve the speed by around 25%. You may be able to increase that maximum pixel number to 21,845 by using unsigned ints. Also, your assembly version is obviously made for 64-bit processors, since on my 32-bit machine it runs around 20% slower than your normal improvement. (Also, 64-bit machines could use a long instead of an int for a speed increase for all pictures, but 32-bit machines would suffer greatly from that.)

]]> http://danbystrom.se/2008/12/14/improving-performance/#comment-60 Fri, 19 Dec 2008 01:04:37 +0000 http://danbystrom.se/?p=117#comment-60 Can someone please upload the changed sourcecode?

]]> http://danbystrom.se/2008/12/14/improving-performance/#comment-56 Wed, 17 Dec 2008 12:11:39 +0000 http://danbystrom.se/?p=117#comment-56 Det är precis sånt här som jag vill läsa innan julen så att jag har något att laborera med i mörkret när barnen och frun gått och lagt sig. Riktigt snyggt jobbat Dan!

]]> http://danbystrom.se/2008/12/14/improving-performance/#comment-54 Tue, 16 Dec 2008 21:21:42 +0000 http://danbystrom.se/?p=117#comment-54 [...] Dan Byström provided an optimization that resulted in a 25 times performance improvement for the fitness function, completely crazy stuff [...]

]]> http://danbystrom.se/2008/12/14/improving-performance/#comment-53 Tue, 16 Dec 2008 11:59:58 +0000 http://danbystrom.se/?p=117#comment-53 [...] Filed under: Programming — danbystrom @ 13:57 Follow-up on Improving performance… I just tested that I can optimize this [...]

]]> http://danbystrom.se/2008/12/14/improving-performance/#comment-52 Tue, 16 Dec 2008 10:14:16 +0000 http://danbystrom.se/?p=117#comment-52 Checking every other pixel is a special case of doing fitness evaluations on source image pyramid.
Using n-level pyramid, and properly setting fitness steps:
if(fitness<pyramid_threshold[current_level])
current_level–;
would make this even faster

]]> http://danbystrom.se/2008/12/14/improving-performance/#comment-51 Mon, 15 Dec 2008 10:20:06 +0000 http://danbystrom.se/?p=117#comment-51 True. At least not until the very end.

]]>