Re: [avr-gcc-list] linear regression on 128L

[Ben L. Titzer]

On Apr 3, 2005, at 3:41 PM, intiha Ho gai wrote:

> Hi,
> Having found out that the avr-gcc lib supports floats and handles them
> quite gracefully, I tried porting my linear regression code from
> desktop on to the atmega 128L, more or less as is. For testing i used
> the same table of X and Y values as sample inputs to the algorithm.
>
>  Now I am using microsecond granularity as input to the regression
> engine and since the values are fairly large - and since there are sum
> of products- I get fairly large numbers (If i do the same on my
> desktop, the biggest number i get is sumXSq=7260107358192.000000). Now
> I know that doubles are 4 bytes - and i was thinking that would not be
> enough - Any way around this?.

Double precision floating point numbers should be 64 bits--8 bytes, on 
all desktop platforms. Did you declare your variables as double or as 
float? It might be a limitation in the floating point emulation library 
of gcc--it may only support floats, when you in fact want doubles. Ask 
yourself if you need that much precision in your results. A standard 
floating point number gives something like 22 or 24 bits of precision 
(about 7 decimal places). If you're guiding a missile with a 600 mile 
range, the representation of distance is accurate down to 1 foot 
resolution.

Also, if precision is an issue, you may want to carefully review your 
code to avoid potential roundoff problems (e.g. adding or subtracting 3 
numbers in the wrong order can amplify round off problems).

> Btw when i do print the same result over the serial port (from my
> embedded port) I get something like sumXSq=7260108000000.000000 - same
> but with 6 of the last significant decimal points lost after
> rounding). Any suggestions? Please note that currently optimization
> isnt an issue, I just want to get it to run and even if it monstrously
> slow, I just have to do this once every boot up!
>
> I would appreciate any help.
>
> Best Regards.
> Affan.
>
>
> _______________________________________________
> AVR-GCC-list mailing list
> address@hidden
> http://lists.nongnu.org/mailman/listinfo/avr-gcc-list
========================================================
Whenever anyone says, "theoretically," they really mean, "not really."
--Dave Parnas