World: r3wp

For the floating point format experts around: 


I am investigating the simplest and most accurate way to support 
floating point comparison by implementing an `almost-equal` operator 
in Red/System. Due to the intricacies of such task, I have selected 
the two algorithms that look the best to me:


1) Tolerance-based: http://realtimecollisiondetection.net/blog/?p=89

2) Ulp-based: http://www.cygnus-software.com/papers/comparingfloats/Comparingfloating point numbers.htm


Of course, in both approaches, there are some accuracy parameters 
to be set by the user to adjust the comparison formula (good defaults 
will be provided anyway).


Are there better algorithms or better implementations of methods 
1) or 2)?

I have x-posted that message to the Red ML also: http://groups.google.com/group/red-lang/browse_thread/thread/527afaf181eee8c7?hl=en

2) is now used in R3

Do you know if it uses the implementation from above link or a custom 
one?

ad 2) This seems to be dated, and new stuff by Bruce Dawson's here 
- http://www.cygnus-software.com/papers/comparingfloats/comparingfloats.htm

Well, the implementation may be dated, but the principle remains 
the same. Moreover, we use it for doubles, not for floats, but as 
said, the principle remains the same.

Well, I just pointed to the new stuff, can't comment on validity, 
as I don't understand  the topics discussed properly, so you might 
be right of course ...

Doc, the last part of your 2nd choise (the ulp-based) named "Comparing 
using integers" is the method, I would look at. It's simple and I 
would guess, performs well. Tests are needed to say, if the other 
methods are faster. Could it be CPU dependent, which is faster?

In order to implement the algorithm for doubles did you need to have 
64-bit integers available?

I guess, you could split the 64-bit in 2 times 32-bit. If the most 
significant 32-bits are the same, you do the comparison the the least 
significant ones. This should work.

*on* the

Is that assuming you can convert a double to two 32-bit integers 
or that you jsut split the 64 bits into 2x32 bits?

jsut -> just

No conversion, just split.


The IEEE float and double formats were designed so that the numbers 
are Òlexicographically orderedÓ, which Ð in the words of IEEE architect 
William Kahan means Òif two floating-point numbers in the same format 
are ordered ( say x < y ), then they are ordered the same way when 
their bits are reinterpreted as Sign-Magnitude integers.Ó

Thanks.

Pekr: the link you're giving is the exact same article I've pointed 
to in 2), it just have a left menu added for quicker navigation, 
but that's the same article. If I haven't just stopped on this algorithm, 
it's because of the initial warning at the beginning: 


This article is obsolete. Its replacement - which will fix some errors 
and better explain the relevant issues - is being crafted as a multi-part 
series


And from Dawson's blog: "Years ago I wrote an article about how to 
do epsilon floating-point comparisons by using integer comparisons. 
That article has been quite popular (it is frequently cited, and 
the code samples have been used by a number of companies) and this 
worries me a bit, because the article has some flaws. I’m not going 
to link to the article because I want to replace it, not send people 
looking for it." 


So it appears that there are some errors (probably some edge cases 
not handled properly).

It's not imo the same article, it is supposed to be a fixed one, 
where its author does not even want to point out to what you posted 
as 2) :-) But - that does not matter, what matters is, what works 
for you/Red :-)

ah, now I can see it. It's in the frame, my intention was to post 
link "here" URL, but cutted and pasted the wrong one ...

Th new fixed version of Dawson's article is split in a new series 
of blog entries: http://randomascii.wordpress.com/category/floating-point/


But, alas, he hasn't yet reached the point where he shows the fixed 
version of `almost-equal` function.

I did not take anything else than the principle (did not copy/use 
his code) to implement the comparison.

To Geomol: the speed should not be bad, but that is not the main 
reason why this approach is good.

Variadic and typed functions now support floats datatypes (float-partial 
branch).


https://github.com/dockimbel/Red/commit/331a50e8e5535ffec710e86c7f5f2704292fc433

BTW, math operations (+, -, *, /) have been implemented for floats 
(only float! / float! or float32! / float32!).

Remainder and modulo should also work, but untested. The meaning 
of these operators has been adapted for floats, remainder (%) will 
return a standard IEEE rounded remainder, while modulo (//) will 
return a floating point remainder.

That's really great!

How do 64-bits typed values affect the other typed values?

The stack layout has changed for typed functions arguments, but it 
is backward-compatible, so all your dialects should work with the 
new version (not tested yet).

For float!, you need to do an extra type-casting on the arguments 
pointer to convert it to typed-float! instead of typed-value!. You 
have an example of that in %red-system/runtime/utils.reds in _print 
function.

Oldes, could you please test your bindings with the lastest revision 
of float-partial branch? They should run fine now.

Continuing here, briefly:


When I'm tired, I usually stick to running "monkey tests" or button 
mashing and take notes for the next day if something crashes. :-)

I made just a quick test with blur effect and it works... cool:)

Great, let me know when your ImageMagic binding will be ready, so 
I'll add it to the list of existing bindings.

I maybe need Red instead of Bash! /Reading syntax spec

DocKimber: was this possibility discussed anywhere --- re: bash vs. 
rebol or red instead of bash?

Sorry! DocKimbel

I've skimmed the BNF doc at the red-lang site. It seems it is rather 
vague and unverified. I think maybe I need to build a Coco/R EBNF 
grammar and to check the grammar formally, for unabiguity and other 
grammar errors.

maybe Doc is actually using the PARSE dialect rather than proper 
BNF.

Re: bash... i thought about this a lot, i'm not sure you can come 
up with a decent dialect, though it would be possible to create a 
bash-like language with more rebol-like syntax and semantics.

The BNF doc was written by Rudolf Meijer and it it strict EBNF. I 
did only a manual checking of the grammar and found no issues. I 
have planned to model it in ANTLR, but never found the time.

it it
 => "it is"

strict EBNF
 => "strict BNF"

The grammar rules I wrote for Red/System expressions should also 
pass through a BNF loader (except for the first five definitions): 
http://static.red-lang.org/red-system-specs.html#section-5.1

Evgeny: Red/System is a bad fit for scripting, it is too low-level, 
you'll need to wait for Red for having a good alternative to Bash.

Doc - what' coming after floating point support? Are you moving to 
work on Red itself?

I have to fix the LibC starting and ending code, then I'll move to 
Red runtime to finish it and start Red's compiler.

I'll also continue to work on the shared library support for Red/System 
at part-time, during weekends. I would like to have it working for 
the conference in Netherlands in march.

Is there something like QUIT function already?

Yes, QUIT is already defined and needs an integer return code as 
argument.

#import [
	LIBC-file cdecl [
		quit-return: "exit" [  ; Exit program, returning status code.
			status		[integer!]
		]
	]
]

yes.. just found it in C lib as well... goot to know I can use the 
quit without it - I was missing the return code

btw... is there any reason that print does not add the LF automatically?