Conversation
Notices
-
Friday, 14-Jun-13 01:35:32 UTC from web
-
I need to know why the floating point (double) precision of my Atom N270 is almost perfect around values of 0. I tested 7 different systems and the others are worse. If I restructure my test such that no optimization is possible (separating the computation in different volatile variables) then it is as bad as the others. !coderponies
-
@broniebrown This is the code I used http://pastebin.com/GKChMveB I visualize the sign of the doubles in a 256x256 image (blue negative, red positive, yellow zero).
-
@broniebrown i actually ran into this issue a couple times, it has to do with the way your processor manages the variable, double being 8bits. the variable is stored in a algorithm state which gets interpreted. some programming languages manage this better then others but my language of that time period was C++ :/
-
@ryanjjjj I found out what this reason is. On 32 bit systems the GCC uses the 387 instructions (for floating point calculations) that actually increase the double/float size for intermediate double/float calculations. On 64 bit systems it is disabled by default because SSE2 is faster and the mixture with 387 is experimental. Depending on how you wrote the code the compiler optimizes in such a way that you have the best accuracy. I made a test so that the compiler doesn't optimize this part and I got a result like in the other cases.
-
@broniebrown but will it blend? this is the question.
-
Light likes this.
-
@broniebrown not "Recommended" means do it but we are not responsible! *puts laptop processor into blender.*
-
