[WIP] Float tests

[jodavies]

Here I have added the example from the Madrid workshop slides.

[jodavies]

To do: why does F1 print with e+0 and F2 without? Which is the desired behaviour?

[tueda]

Well, if we put it into example.frm (examples in the manual), then should we put this example into the manual with the following comment?

% THIS EXAMPLE IS PART OF THE TESTSUITE. CHANGES HERE SHOULD BE APPLIED THERE AS
% WELL! (Float_1)

[jodavies]

The e+0 is a bug actually: F1 should be e+3 (easy to see from the rational form of F1/F2).

Edit: digits wise, I confirmed that there is agreement with PolyLogTools in Mathematica.

[vermaseren]

Very funny indeed. It is the term +89/18*mzv_(9,3,3) that is the culprit. If you leave it out, it prints without the e+0 and if you include it but leave the last term out, it just prints e. Gotta have a good look on the linux computers, because gdb does not work on my apple. I can leave some more terms out to just get the e, but I have not yet discovered the ‘pattern'

…

On 11 Mar 2024, at 14:26, jodavies ***@***.***> wrote: To do: why does F1 print with e+0 and F2 without? Which is the desired behaviour? — Reply to this email directly, view it on GitHub <#492 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABJPCES4BYIDJ5YJLMYDUM3YXWWKBAVCNFSM6AAAAABEQJ5KHWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSOBYGQZTSOJXHE>. You are receiving this because you are subscribed to this thread.

[jodavies]

The ouput is truncated here:
https://github.com/vermaseren/form/blob/e99543ad939ef823f5e8db0678aa18028e7de707/sources/float.c#L950

If you ask for numdigits+1 then you get e+03 (and now a bare e appears on the output of F2). I think the required buffer size for the print can be found by calling gmp_snprintf(NULL,0,"%.*Fe",numdigits,aux4) first.

[jodavies]

I was too hasty there, the issue is in the post-processing code that follows. As I understand it you would like the following:

1. truncate trailing zeros
2. don't print "e+00"

Is this correct?

[fllor]

I found the bug printing routine. One has to change

https://github.com/vermaseren/form/blob/e99543ad939ef823f5e8db0678aa18028e7de707/sources/float.c#L962-L963
to

while ( n2 < n ) { *s1++ = *s2++; n2++; }
n -= (n2-n1);

This will correctly truncate all trailing zeros, but leave the exponent intact. Removing "e+00" is not part of this routine.

[jodavies]

OK, please add the fix in your branch and lets see if the valgrind errors are gone. I can rebase this on top of your changes later.

[tueda]

By the way, why is the precision of printed numbers strongly influenced by the maximum weight?

#StartFloat 50,6
L F = 1/3;
tofloat;
P;
.end

   F =
      3.3e-01;

[jodavies]

The final digit of pi is also not rounded correctly, is it understood why this is?

[jodavies]

Updated after #532. But @tueda 's question remains: there is some interaction between the precision and weight parameters. The precision is in bits, and is modified by SetFloatPrecision called by SetupMPFTables and then again by SetupMZVTables. If AC.DefaultPrecision grows enough the printing changes...

[coveralls]

coverage: 49.801% (+0.5%) from 49.264%
when pulling 7f9388f on jodavies:float-tests
into 915a14d on vermaseren:master.