To view parent comment, click here.
To read all comments associated with this story, please click here.
renox,
Floating point used to be far more accurate than ints when cpu registers were limited to 32bit. Now that we have 64bit registers and SSE's FP mantissa is only 52bits, fixed point integers can give us more accuracy as well.
I wonder if there are any game worlds that are big enough such that this makes an observable difference?
If each object were tracked to millimetre resolution, then SSE's 64bit FP is accurate up to half a light year in any direction. 64bit signed fixed point would cover 970 light years in any direction.
So I guess there's probably not a very compelling reason to switch unless your simulating a large universe.
Nice chart showing breakdown of FP formats.
http://www.monash.edu.au/policy/gpnumacc.htm




Member since:
2011-01-28
renox,
"Sigh, and people still look at me as if I was a madman when I say that floating point suck and that we should use interval arithmetic.."
Very insightful. Beyond rounding issues, it's always bugged me a bit too that the resolution of game coordinates become worse and worse the further a player gets away from the origin. Therefor if your running precise mathematical computations in a game (ie object projections), the results inherently depend on which arbitrary digits the mantissa is representing. Using integral/fixed point arithmetic would offer equal precision at all coordinates.