Floating Point

[rp181]

Does anyone know exactly how much slower operations with floating point numbers are than with fixed point?

[Bianco]

You can't say that exactly as it depends on the implementation.

[lilltroll]

If it doesn't need to be IEEE 754 compliant, and you can skip a software exception handling for floating points, many operations can be saved. E.g. instead of using double prec. use {63 bit + sign} or 64 bit (2 ALU registers) for the digits and 32 or {31bit + sign} (1 ALU register) for the exponent.

Using directed rounding mode, you can calculate FADD and FMUL in about 10 operations or a little more, (converting from integer to float vice-versa not included)

[yzoer]

Floating point is ALWAYS going to be slower than fixed point because of normalization. If you *really* need extra precision, you can use 64-bit integers using assembly.

-Yvo