Fixed vs Floating point - is faster possible?

evsource · Joined: 21 Nov 2006 Posts: 129

Hi,

I know the benefits of fixed point vs. floating point calculations. I have code that needs to run as fast as possible, so I try to write the fastest code possible. I'm wondering if anyone can improve on the following example (let's assume a 18F chip at 40mHz).

Let's say I need to do this calculation:

(160 / 360) * 636

First, floating point:

mskala · Joined: 06 Mar 2007 Posts: 100 Location: Massachusetts, USA

Two comments:
1) This type of thing is very dependent on what fits into the variable size
in intermediate steps, so it is not so general that a future user could
just send any old values into a function using it.
2) Faster would be to identify what values are 'constants' of the calculations,
condense those down to a single operation. If you have a hardware
multiplier in your pic, then you're pretty much done. If not, then see
what bit shifts can be done to accomplish without any multiply or divide.
Meaning, you only multiply or divide by powers of 2 which are then
added or subtracted. (Check in a spreadsheet if the error is
acceptable over entire range of values, because it can deal with round-
off errors).

Mark S.

Ttelmah · Guest

Also, you have already identified one big speed difference in float. Multiplication is always a lot faster than division. So, can you re-design what you want to do to use multiplication by a reciprocal value?. Obvously with the constant values shown, this is easy, but in a real operation, with variables, it may still be possible.

Best Wishes