std.math unittests - accuracy of floating point results

[Kai Nacke via digitalmars-d-ldc]

Hi Kevin! Hi Johan!

On Monday, 2 March 2015 at 00:30:28 UTC, Kevin Brogan wrote:
> On Sunday, 1 March 2015 at 23:08:54 UTC, Johan Engelen wrote:
>> Hi all,
>> I am working on making more of the std.math unittests pass 
>> (I'm new to the project, and it is a nice niche thing to 
>> tinker on, learning the codebase, workflow, etc.).
>
> Guess what I'm doing. :-)
>
>> I've hit on a problem that I do not know how to handle: 
>> floating point comparison.
>>
>> There are some tests that check whether exp(x) works well, 
>> including overflow checks for different x. See 
>> phobos/std/math.d line 2083. The checks are defined for 80-bit 
>> reals, and I am converting them to 64-bit reals (Win64). The 
>> problem is that the checks are bit-precise (i.e. assert(x == 
>> y)), but the calculation results are sometimes 1 ulp off. For 
>> example, the results of two tests:
>>
>> std.math.E = 0x4005bf0a8b145769 = 2.7182818284590450
>> exp(1.0L)  = 0x4005bf0a8b145769 = 2.7182818284590450  [1]
>> Wolfram Alpha                   = 2.718281828459045235...
>>
>> E*E*E     = 0x403415e5bf6fb105 = 20.085536923187664
>> exp(3.0L) = 0x403415e5bf6fb106 = 20.085536923187668
>> Wolfram Alpha                  = 20.08553692318766774...
>>
>> I do not know how I can make the second test pass, without 
>> breaking the first one. I feel the tests are too strict and 
>> should allow an error of 1 ulp.
>
> They are too strict. floating point math is not exact between 
> different architectures, or even compilation flags. You can get 
> a different result just because the compiler reordered two 
> operations.

I agree, too. Bitwise comparison is worse if you are working with 
floating points.

BTW: These fixes should go upstream. Please create Phobos PRs for 
them. I am happy to cherry-pick these changes as soon as they 
committed.

Regards,
Kai