Rounding glitch from bigint to float

[Jean-Luc-Picard-2021]

Hi,

Did some regression testing:

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Trealla Prolog 2.4.3

% ?- trealla.
% case, trealla: 0
% case2, trealla: 0
% case3, trealla: 0
% case4, trealla: 0
% case5, trealla: 0
% case6, trealla: 0

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Trealla Prolog 2.4.23

% ?- trealla.
% case, trealla: 0
% case2, trealla: 0
% case3, trealla: 0
% case4, trealla: 0
% case5, trealla: 0
% case6, trealla: 16

Is this to be expected? Test cases here:

Compliance Dogelog Spieler novacore/arithmetic
http://pages.xlog.ch/littab/doclet/docs/10_samples/04_doge_comply/novacore/arithmetic/package.jsp

Maybe this was done in favor of some performance?

[infradig]

I don't know what it means.

…

On Wed, 19 Oct 2022, 23:37 Jean-Luc-Picard-2021, ***@***.***> wrote: Hi, Did some regression testing: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Trealla Prolog 2.4.3 % ?- trealla. % case, trealla: 0 % case2, trealla: 0 % case3, trealla: 0 % case4, trealla: 0 % case5, trealla: 0 % case6, trealla: 0 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Trealla Prolog 2.4.23 % ?- trealla. % case, trealla: 0 % case2, trealla: 0 % case3, trealla: 0 % case4, trealla: 0 % case5, trealla: 0 % case6, trealla: 16 Is this to expect? Test cases here: Compliance Dogelog Spieler novacore/arithmetic http://pages.xlog.ch/littab/doclet/docs/10_samples/04_doge_comply/novacore/arithmetic/package.jsp Maybe this was done in favor of performance? — Reply to this email directly, view it on GitHub <#63>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFNKSEXXV446RPD36ZKVG53WD72TJANCNFSM6AAAAAARJEJY4M> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[Jean-Luc-Picard-2021]

Ok

[Jean-Luc-Picard-2021]

My bad, was using wrong test cases. But when I use
the most recent ones, I get these results:

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Trealla Prolog 2.4.23

case3, trealla: 5
case4, trealla: 0
case5, trealla: 0
case6, trealla: 0

The five failing test cases are among the same problems
as here for Ciao Prolog:

ciao-lang/ciao#77

For example this Trealla Prolog result is a litle too low
for HALF_EVEN. But I don't know whether it is the result

/* SWI-Prolog 8.5.20 Ok */
?- X is float(166153499473114502559719956244594689).
X = 1.6615349947311452e+35.

/* Trealla Prolog 2.4.23 Nok */
?- X is float(166153499473114502559719956244594689).
   X = 1.66153499473114e+35.

of a similar routine used by Trealla Prolog as in Ciao Prolog,
I didn't check what is used to realize float/1 in Trealla Prolog.

[infradig]

Pushed a change to devel branch, if you could try it.

…

On Fri, Oct 21, 2022 at 10:44 AM Jean-Luc-Picard-2021 < ***@***.***> wrote: My bad, was using wrong test cases. But when I use the most recent ones, I get these results: case3, trealla: 5 case4, trealla: 0 case5, trealla: 0 case6, trealla: 0 The five failing test cases are among the same problems as here for Cioa Prolog: ciao-lang/ciao#77 <ciao-lang/ciao#77> — Reply to this email directly, view it on GitHub <#63 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFNKSEVEJHWQYJ6QSIHWGYTWEHRPRANCNFSM6AAAAAARJEJY4M> . You are receiving this because you commented.Message ID: ***@***.***>

[Jean-Luc-Picard-2021]

Hi, I have streamline the test suite a little bit.
Its now only 3-sets testing float/1, (/)/2 and (**)/2.
After git pull I get for your Prolog system:

$ ./tpl -v
Trealla Prolog (c) Infradig 2020-2022, v2.4.26
$ ./tpl
?- ['trealla.p'].
   true.
?- trealla.
case3, trealla: 5
case4, trealla: 0
case6, trealla: 0
   true.

You should get the test cases and a lot of other stuff, via the following,
in the folder samples/dogelog_comply/novacore/arithmetic:

git clone http://www.antwort42.ch/dogelog/.git

And then updates of the test cases.

cd dogelog
git pull

But I hope I am now finished with updating them, except
for regular regression testing in the future, which I expect
to be in intervals of some months, and not daily as of now.

[Jean-Luc-Picard-2021]

As of now I have float/1 conversion in the test case:

?- case3(_, X, Y), Y =\= float(X).

Do you say I can remove it now? And query:

?- case3(_, X, Y), Y =\= X.

I am testing the other Prolog systems without float/1.

This would be another ticket here:

#64

[Jean-Luc-Picard-2021]

Ok, float/1 can be removed.
Closing the other ticket.

[infradig]

I don't know what the status of this is.

[Jean-Luc-Picard-2021]

The new version of Trealla Prolog returns another value:

/* Trealla Prolog 2.4.23 Nok */
?- X is float(166153499473114502559719956244594689).
   X = 1.66153499473114e+35.

/* Trealla Prolog 2.7.15 Nok */
?- X is float(166153499473114502559719956244594689).
   X = 1.6615349947311448e+35.

?- 1.66153499473114e+35 =:= 1.6615349947311448e+35.
   false.

But since this is an extremly nasty test case, the rounding
happens based on bits outside of the 53 bit mantissa, so
you need a bigint to float routine that checks the LSB,

least significant bit, of the given bigint as well.
Just check out the binary representation of the bigint:

?- atom_integer(X, 2, 166153499473114502559719956244594689).
X = '1000000000000000000000000000000000000000
000000000000010000000000000000000000000000000
000000000000000000000000000000001'.

Hence the result is not yet correct. One can compare:

/* SWI-Prolog 9.1.0 Ok */
?- X is float(166153499473114502559719956244594689).
X = 1.6615349947311452e+35.

?- 1.6615349947311452e+35 =:= 1.6615349947311448e+35.
false.

/* Trealla Prolog 2.7.15 Nok */
?- X is float(166153499473114502559719956244594689), 
   X =:= 1.6615349947311452e+35.
false

So cannot yet close this ticket.

[Jean-Luc-Picard-2021]

Scryer Prolog has the same bug:

$ target/release/scryer-prolog -v
"v0.9.1-151-g17450520"
$ target/release/scryer-prolog
?- X is float(166153499473114502559719956244594689).
   X = 1.661534994731145e35.
?- Y = 1.6615349947311452e+35.
   Y = 1.6615349947311452e35.
?- X is float(166153499473114502559719956244594689)-1.6615349947311452e+35.
   X = -3.6893488147419103e19.
?-