Page 1 of 1

x64 vs win32, MSVC 2010 vs MSVC 2013

Posted: Fri Oct 17, 2014 10:23 am
by theBulletGuy
Hello,
Numerical output of my demo is very important, but I get inconsistent results when building for win32 and x64, and when building in MSVC 2010 and MSVC 2013. Although results for the same build (same platform and MSVC version) are deterministic and consistent both in release and debug modes.

All built in Double Precision.

A single object with forces on it gives consistent results, but when objects start colliding numbers start to differ. And I am talking about significant differences, not round off errors.
----------------|Win32------| X64---
|MSVC 2010 | Output 1a | Output 2
|MSVC 2013 | Output 1b | Output 3

Output 1a == output 1b
Output 1 <> output 2 <> output 3

Any help is appreciated.

Re: x64 vs win32, MSVC 2010 vs MSVC 2013

Posted: Fri Oct 17, 2014 12:54 pm
by Basroil
How much is "not roundoff error"? When compiling for x64 on newer chips, you might be using AVX instead of SSE, and while they should give the same results that's not always the case if the compiler takes shortcuts.