Re: [boost] Boost Test Library: general floating point comparisons

From: John Maddock (john_at_[hidden])
Date: 2007-02-19 08:37:19

Gennadiy Rozental wrote:
> How about a compromise:
>
> BOOST_CHECK_PERCENT_DIFF
> BOOST_CHECK_RELATIVE_DIFF
> BOOST_CHECK_EPSILON_DIFF
> BOOST_CHECK_ABSOLUTE_DIFF (do we need this one?)

I could live with those.  They still say what they do, but I'd like to hear 
other opinions.

> So.Would you care to provide a error message body for the non-percent
> based tools?

Well for instance:

BOOST_CHECK_PERCENT_DIFF(x, y, tol);

"Relative difference between x{%1%} and y{%2%} exceeds tol{%3%%}, the actual 
difference found was %4%%."

BOOST_CHECK_RELATIVE_DIFF(x, y, tol);

"Relative difference between x{%1%} and y{%2%} exceeds tol{%3%}}, the actual 
difference found was %4%."

BOOST_CHECK_EPSILON_DIFF(x, y, tol)

"Relative difference between x{%1%} and y{%2%} exceeds tol{%3%epsilon}, the 
actual difference found was %4%epsilon."

BOOST_CHECK_ABSOLUTE_DIFF(x, y, tol)

"Absolute difference between x{%1%} and y{%2%} exceeds tol{%3%}}, the actual 
difference found was %4%."

Probably you could figure out a way for them all to use the same code 
internally as there's not much variation between them?

Note that compared to your existing messages I've added what the actual 
*found* difference was.  This is something I've really missed in the 
existing messages: in some situations (say with 128-bit long doubles) the 
difference can be very small - if you just cut and paste the values into a 
calculator to try and work out what the error rate actually was then 
truncation of input values leads to a zero result - how many calculators do 
you know that do 35-digit precision?  At times the only way I could actually 
work out what the difference was was either to:

1) Use a pencil and paper.
2) Do it my head (not advised).
3) step through the Boost.Test code to find out.
4) Write my own code to calculate it.

None of which are exactly ideal :-(

If you're wondering why I want to know: it's because sometimes what 
constitutes an acceptable rate of error has to be a judgement call, and may 
vary from platform to platform.  I want the tolerance as low as possible - 
to catch regresions - but sometimes that means I set it just slightly too 
low for a particular compiler/platform combination.  If I could see what the 
actual relative difference was then I would know right away whether I had a 
real problem, or just a tolerance that was a bit too low.

HTH, John.

PS Thanks for persevering with this, very much appreciated!