Re: [boost] GGL Review

Subject: Re: [boost] GGL Review
From: Simonson, Lucanus J (lucanus.j.simonson_at_[hidden])
Date: 2009-11-17 13:57:27

Barend Gehrels wrote:
> - personally I don't think that batteries of random data will cover
> *all *cases. I understand that exceptions or endless loops might
> become 
> clear. But with random it is difficult to verify the results. We can
> ask 
> Luke of course, how he verifies them, and I've no problem to add them
> then. 

Batteries of random tests won't cover all cases, but will cover more than hand designed tests and much faster.  You should use both hand designed tests for cases you think are important to test and randomly generated tests for the cases you wouldn't thing of (and didn't think of while writing the algorithm.)  I explained how to verify the correctness of the results when I suggested the random tests.  Use a known correct algorithm and perform an XOR operation on the results of your algorithm and the known good then shrink (negative offsetting) the result of that by a small value to eliminate artifacts of numerical error.  Anything left in the output is an area of discrepency that you need to investigate.

Thanks,
Luke