U++ forum: Welcome to the forum

Search on this site

Search in forums

Home » U++ Library support » U++ Core » U++ can't handle float to string and back for large numbers

Show: Today's Messages :: Show Polls :: Message Navigator
E-mail to friend

U++ can't handle float to string and back for large numbers [message #47765]

Tue, 21 March 2017 09:57

cbpporter
Messages: 1427
Registered: September 2007

Ultimate Contributor

I'm working on a library that tries to do a lot for things, including float work.

I real Maths work, where you do more than 1.0 + 2.0, you often encounter large numbers.

I noticed that U++ can't handle the extremes at all, both by debugging the float values and by using CParser extensively. I looked over CParser code extensively and have not yet found the problem. Most if not all the float to string bits form Core have this problem.

Even something as simple as:

	double d = 1.79769e+302;
	printf("%e\n", d);   // good
	Cout() << d << "\n"; // wrong

Is is wrong at it gets worse as worse as you go up to e303 and up to e308.

I am currently investigating this, since U++ problems with "atof" are also my libraries problems by inheritance.

PS: My official release day was planned for Wednesday, but it will probably be moved back to Friday.

Why do such weird problems only happen before release day? Laughing

Report message to a moderator

Re: U++ can't handle float to string and back for large numbers [message #47766 is a reply to message #47765]

Tue, 21 March 2017 10:28

cbpporter
Messages: 1427
Registered: September 2007

Ultimate Contributor

Better test case: 1e308.

After further investigations, it looks like CParser is good. So "atof" is pretty much good I think.

DblStr is to blame, the "ftoa" part. I'm investigating that right now.

Report message to a moderator

Re: U++ can't handle float to string and back for large numbers [message #47767 is a reply to message #47766]

Tue, 21 March 2017 10:52

cbpporter
Messages: 1427
Registered: September 2007

Ultimate Contributor

Still under investigation, but:

double normalize(double d, int& exp)
{
	if(IsNull(d) || d == 0)
	{
		exp = Null;
		return d;
	}
	bool sign = (d < 0);
	if(sign) d = -d;
	exp = minmax<int>(ilog10(d), -308, +308); // 8-byte double!
	d /= ipow10(exp);
	if(d >= 10) { d /= 10; exp++; }
	if(d < 1)   { d *= 10; exp--; }
	return sign ? -d : d;
}

Looks like fixes the problem. This, and adding special case to suppress the inherent U++ IsNull mechanic. And some custom work to handle infinite better.

Report message to a moderator

Re: U++ can't handle float to string and back for large numbers [message #47768 is a reply to message #47765]

Tue, 21 March 2017 21:35

mr_ped
Messages: 826
Registered: November 2005
Location: Czech Republic - Praha

Experienced Contributor

I'm not IEEE expert, but why "exp = minmax<int>(ilog10(d), -308, +308);", I mean why decimal exp, when the native is binary, if I understand the IEEE format correctly. The whole code looks weird, all the computations instead of of working directly with bits, this will be very likely prone to some rounding/cutoffs artefacts. Also doing it directly from bits with decimal conversions only in the last phase would be probably faster, not just more accurate.

Report message to a moderator

Re: U++ can't handle float to string and back for large numbers [message #47769 is a reply to message #47768]

Wed, 22 March 2017 07:27

cbpporter
Messages: 1427
Registered: September 2007

Ultimate Contributor

Yeah, I'm no expert either. And probably 308 is not enough to handle subnormal values.

But it is enough for my needs for now. Without this increase, I can't print a single double and neither can't all of U++, including the debugger.

So the debugger fails, DblStr fails and I need to ftoa the values both in the production code and for debugging.

Report message to a moderator

Re: U++ can't handle float to string and back for large numbers [message #47870 is a reply to message #47767]

Sat, 15 April 2017 11:25

mirek
Messages: 14257
Registered: November 2005

Ultimate Member

cbpporter wrote on Tue, 21 March 2017 10:52

Still under investigation, but:
double normalize(double d, int& exp)
{
	if(IsNull(d) || d == 0)
	{
		exp = Null;
		return d;
	}
	bool sign = (d < 0);
	if(sign) d = -d;
	exp = minmax<int>(ilog10(d), -308, +308); // 8-byte double!
	d /= ipow10(exp);
	if(d >= 10) { d /= 10; exp++; }
	if(d < 1)   { d *= 10; exp--; }
	return sign ? -d : d;
}
Looks like fixes the problem. This, and adding special case to suppress the inherent U++ IsNull mechanic. And some custom work to handle infinite better.

Thanks, fix applied. This code is ancient, I guess at that time, Tom has not considered handling 'exteremes' so important.

Report message to a moderator

Re: U++ can't handle float to string and back for large numbers [message #47898 is a reply to message #47870]

Thu, 20 April 2017 10:44

cbpporter
Messages: 1427
Registered: September 2007

Ultimate Contributor

Thank you Mirek!

Now I only need to get rid of http://ultimatepp.org/redmine/issues/1650 and I can update U++ again Smile

Report message to a moderator

Previous Topic:	new operator declared twice in MinGW TDM
Next Topic:	How to bind parameter to PostCallback with Function, Event object

Goto Forum:

-=] Back to Top [=-

[ Syndicate this forum (XML) ] [

] [

]

Current Time: Fri May 09 17:55:34 CEST 2025

Total time taken to generate the page: 0.00413 seconds