16 bit floats

[PhilipLB]

Hi,

I'm currently writing a serializer for the volume component to save voxeldata to disc. And actually, there's no need to be super-exact with the density values so to save some space, I'd like to save the data as 16 bit floats. But... I don't see a direct datatype for this like float16 or so. How would you do that? An easy one would be to use fixed point numbers and just save shorts.

[Kojack]

C++ and the cpu itself have no support for half floats (the gpu does). You'll need to get a library (or at least a class) that supports them.
Fixed point is probably going to be easier.
What's the range of values you are expecting for density? It may be easier to just use the short as 0-65535 representing the scaled range of density values (such as 0-1 or whatever it is). If the density range can change, then store the min and max as well.

[PhilipLB]

That's what I supposed, thanks. Yep, going for shorts now. Else, everything could be really slow due to the lack of CPU support.

[bstone]

Ogre::Bitwise has floatToHalf() and halfToFloat(). You might try them and see how big the performance hit would be compared to the 16bit fixed point.

[PhilipLB]

Ah, good to know, thx.

[Kojack]

I didn't know Bitwise had half conversions.
Then again, I've never heard of Ogre::Bitwise.
(Hey, it's a big library)

[bstone]

(Hey, it's a big library)

And evolving!

[Zonder]

reading this made me read the 80286 architecture docs geek but I actually though the 386 had 16bit float registers.... never mine lol. What surprised me most was the 386 stopped manufacture in 2006!!!

[mkultra333]

Be nice if float16 was available on the CPU. They're handy for so many things. I'm just working on some metaballs at the moment based on some BSD license code I found. It's SSE optimized, and I'm just thinking how great it would be to process 8 float16 at a time instead of 4 floats.