RESOLVED: Float 16 textures do math wrong?

Bugs
1

Hi there,

I’ve simplified a bug down to what I hope is a reproducible level. I have 2 TOP glsls and their format is float 16 rgba and when I try to add the r,g,b channels in this single pixel each is off by something close to 10. Here you can see that the b channel result is about 14 less than what 222+29164 would equal. Similar issue with the red channel 56320 + 141 != 56448

image
image1349×678 48.7 KB

Almost always it is smaller than it is supposed to be and I can’t for the life of me figure out why this is happening. What I do know is that when I change the pixel format to rgba32 the addition works out. Is this me not understanding something about binary representations of numbers?

2

16-bit floats can only represent ±1024 perfectly. Anything beyond that is an approximation and subject to floating point roundoff.

3

oh gosh so to put it simply

Is this me not understanding something about binary representations of numbers?

Yuppers =D. Thanks for the help!