Re: interest in math testing

["Carlos O'Donell" <carlos at redhat dot com>]

On 03/21/2014 01:26 PM, Bosz Galaxy wrote:
> sorry for using yahoomail but my gmail is having some trouble. Its
> sending both html and plain tesx From what i understand, i need to
> know the various range for each input of the functions i wish to
> test( domain of function). eg for sqrt, input parameter is >=0 For
> the test points, am not so clear.  I will like to know how to see the
> test points that are available and what are the things to take into
> consideration in comming up with new test points.
> Also, how is the testing done.
> 
> Does it entail mathematical formulara like generating the sine
> function using taylors series?
> 
> Will be waiting for reply as soon as possible. Thanks. Martin T

I assume you are talking about the glibc GSoC project to enhance
libm testing, but your email doesn't mention this at all.

Please follow general email etiquette. In general on this list we
tolerate you writing in whatever language you wish, but please use
proper sentences, grammar, and spelling. We prefer bottom posting,
but don't complain too much if you don't follow that standard.
Note that if you write in English you'll get the largest response
as the current GSoC mentors both speak English as their primary
language.

(1) Domain of the function?

It is the students responsibility to identify the values in the
domain of the function that result in 100% code coverage of the
internal implementation of the function. The only way to do this
is to read the code, determine which code patches are executed
by which domain values and then build a minimal set of input
points to get maximal code coverage.

Testing the entire domain is impossible since double and long
double are already too large for this kind of complete testing.

(2) How is testing done?

The math library testing is done by adding test points to a
database (file) of test points. The points are passed to the
function, which computes the result. The result is then compared
against the expected result. The expected result is automatically
calculated using a high-accuracy multi-precision library via a
helper program used by glibc. Therefore one need only supply the
input point to the function, regenerate the expected output,
and then run the test.

Cheers,
Carlos.