canbula / ieee754 Goto Github PK

For 0.5 by using this tool, you can find the accurate IEEE754 Half Precision representation for 0.5 as follows:

But in this repo, we have:

This is inaccurate by 2 bits.

If you choose Quadruple precision, the mantissa portion of the number gets approximated after a certain number of bits.

As an example, the results for Quadruple precision from this converter (https://babbage.cs.qc.cuny.edu/IEEE-754/) are different from the one using the IEEE754 class.

Digging a bit in the methods, the approximation is related to the fact that building the mantissa multiplying by 2 makes the error due to the initial approximation of the floating point number larger and larger, until it affects the result of the mantissa computation.

I read that python and numpy do not handle particularly well high precision real numbers.

I had to employ gmpy2 for my needs, maybe it's worth having a look at it for future modifications to the class.

something is wrong with small numbers lile -0.5687.

IEEE754(-0.5687) = 10111111110010001100101100101001 (= 0xBFC8CB29) while the resultt should be 0xbf119653.

Using list[str] as a return type requires python3.9 or greater. But I was able to install this on python3.6.