[Openexr-devel] Re: Conversion between half-float and float

[Derek Ledbetter]

On Jun 25, 2004, at 10:17 PM, Kennant Tom wrote:

> I have read through that piece of code before. I have thought about
> implementing the conversion using the SIMD2 intrinsics. It seems that
> the conversion involves more than a few logical shift operations and
> the while-loop for denormalizing normalized values is difficult to
> implement neatly with SIMD2 intrinsics. Do you know the actual logic
> behind the implementation using the AltiVec functions.

Here's the part of the code that converts a denormalized half to a 
float:
    vector float tiny_result
        = vec_ctf(
            vec_and(
                half_vector,
                half_m_mask),
            14+10);

First it masks out the sign and exponent.  (The exponent is already 
zero, but the sign may not be.  The sign is copied to the result at the 
very end of the function.)  Then it treats the mantissa as an integer, 
multiplies it by 2^(-24), and stores the result as floating-point.  
This is pretty fast on AltiVec, because it uses the same registers for 
both integers and floats.  This algorithm might be slower with scalar 
PowerPC instructions, because on the PowerPC, the only way to move a 
value between an integer register and a floating-point register is to 
write it to memory and read it again.

--
Derek Ledbetter
address@hidden