[CLN-list] floating point underflow - a possible explanation

[Michael Schuster]

Hi again, 

when thinking of my problem, I found the following (it's a bit talking to 
myself, but perhaps you can find it useful).

+ The problem with the underflow happens (at least in my case) because the 
declaration in fnCalc3() 
  Real y2 = Real(string("-6.6061s14"));y Real wrapper class. 
makes a cl_SF out of y2. The other variable is cl_LF. It seems this causes the 
underflow. 

+ Actually I thought, this problem can not occur - that's why I use the 
wrapper class, as I always use the double initialisation-

But I found this:
    valarray<Real> c;
    c.resize(3);
    Real dd=c[0];	
valarray will call default constructor which is Real():cl_F() and therefore 
produce a c[0] with type cl_SF. Ok, didnot think of that. My problem.
Anyway with the following I get a lot of conversions then
    dd=Real(1.0/3.0);                   <<= cl_LF
    dd=Real(1.0/3.0)+c[0];		  <<= back cl_SF again
    dd=c[0]+Real(1.0/3.0);            <<= still cl_SF

So once I have cl_SF in my system, it can be anywhere. So now I changed 
default constructor to 
	Real():cl_F(cl_float(cl_I(0),float_format(_PRECISION_CLN_NEWMAT))) {};
and everthing seems ok now. 
To be absolutely sure I can change also 
	Real(const cl_F &f):cl_F(cl_float(f,float_format(_PRECISION_CLN))) {}
in my wrapper. (The string initialization is intentional.)

Michael 

Watch this!! 1.49557e-52
Am Montag 29 Oktober 2007 schrieb Michael Schuster:
> Hi Richy & Cln-List,
>
> Am Samstag 27 Oktober 2007 schrieb Richard B. Kreckel:
> > In order for anyone to be able to help you it is essential that you
> > write a reduced test case where the problem appears and can be
> > investigated.
>
> Attachted I made a little demo programm which shows the problem. My output
> is :
> fnCalc1:1.49557e-52
> fnCalc2:1.49557e-52
> fnCalc2 + 0 :1.49557e-52
> fnCalc3:0
> Watch this!! 0
> Watch this!! 1.49557e-52
>
> The line >>Watch this!! 0<< should be  >>Watch this!! 1.49557e-52<< at
> least in my mathematical opion.
>
> Ok, I donot know if my construct with the class Real is a possibile good
> design. I wanted to have an easy possibility to switch from built-in-double
> type to a class type with higher precision.
>
> Another problem is why I come (in my original programm) to a Real
> value-type of cl_SF while I only use the Real base class with cl_float()
> function as in the demo. But this is for next mailings here.
>
> My system:
> $gcc --version
> gcc (GCC) 4.0.2 20050901 (prerelease) (SUSE Linux)
> Copyright (C) 2005 Free Software Foundation, Inc.
> Dies ist freie Software; die Kopierbedingungen stehen in den Quellen. Es
> gibt KEINE Garantie; auch nicht für VERKAUFBARKEIT oder FÜR SPEZIELLE
> ZWECKE.
>
> $rpm -q cln
> cln-1.1.9-5
>
> Thanks for any hint.
>
> Michael