half_float | 16 bit floating-point type mimicking | 3D Animation library

This isn't complete as it doesn't address all edge-cases, infinities or rounding modes, it also doesn't support subnormals. However, this seems to work:

1. Check if multiplying by zero, return zero if so.
2. Check if multiplying by NaN, return zero if so.
3. Calculate an interim exponent by adding the two exponents. We need to subtract the bias from this to avoid it being double-added.
4. Prepend the inferred 1s to the mantissae.
5. Multiply the mantissae
6. Truncate this mantissae to the length of your stored mantissa plus 2. (x.yyyy... * y.zzzz... always equals aa.bbbb... if yyyy... and zzzz... are the same length)
7. The 2 high bits will always be in {11, 10, 01} so we check the high bit and right shift by 1 if it's set. We must also increment the exponent if we do this.
8. We then take the low n-2 bits as our new mantissa.
9. The resultant sing is gained with l.sign xor r.sign.
10. Concatenate the whole lot to get the result.

In VHDL, this looks like: