On Fri, 7 Jan 2011, Marc Schafer wrote:
Jeremiah,
I am looking at a particular case in the debugger and I see that is_flow returns false because of comparisons that fail by an error that is many orders of magnitude smaller than the numbers being compared. For example, x==y fails because x and y are both approximately 1. and x-y is 1.e-14. I could replace the particular comparisons that are failing for my test case, but I need to know how to make the entire algorithm generically suitable for use with floating point numbers.
is_flow just performs a consistency check after the calculation is complete. Does the calculation itself require any modifications to work correctly with floating point numbers?
I don't know, but I wouldn't imagine it would.
What are the appropriate tolerances to use?
I would guess, if is_flow() is just for debugging, that a tolerance of 1e-5 or 1e-8 would work.
Do the relational tests (<, >) also need modification to properly handle floating point?
I don't know. I had suggested that you try replacing == and != but not < and > and see if the code works correctly. There are replacements for == and != documented at http://www.boost.org/doc/libs/1_45_0/libs/test/doc/html/utf/testing-tools/fl... but I don't think there are < and > replacements there, so hopefully we won't need them. -- Jeremiah Willcock