Re: [Boost-users] [proto] generation of adaptive precision floating-point geometric predicates

Subject: Re: [Boost-users] [proto] generation of adaptive precision floating-point geometric predicates
From: Andrew Durward (andrew.durward_at_[hidden])
Date: 2010-03-22 11:00:32

Hi Kim,

Thanks very much for your response and the sample code.

> How about transform the expression a+b to sum_approx(a,b)+sum_diff(a,b)
> instead to fusion::cons<sum_approx(a,b),fusion::cons<sum_diff(a,b)> > ?
> The advantage of this approach is that you still get an ast where
> childs are held in terminal (instead in cons).

You are correct that the elements being generated will represent terms to be
summed together.  My rationale for choosing to represent them as a sequence
rather than a tree was to make it easier to modify the order in which the
summation is performed.  Namely, I plan to sort the terms according to their
expected magnitude so that I can add them up until the magnitude of the
result exceeds the maximum possible magnitude of the error (hence
guaranteeing that the sign is correct).  The value of any remaining terms
should never actually be computed.

Having said that, if it makes life easier to generate an AST with an n-ary
"sum" function rather than a sequence or a deep tree of binary "plus" nodes,
then that would be fine too.  The final tree should then consist of a single
"sum" node with a long list of terms.

> I dont see the meaning behind this step. If you use
> proto::_make_function you create a child in the ast, but you put this
> child in a fusion::cons.

Yes, exactly.  The resulting sequence should consist of AST nodes that can
be evaluated one by one until the stop condition is reached.

> Are you sure that you want to get a sequence of expression instead an
> ast. In case you really want to get a sequence you can use
> proto::tag::comma to separate your child.

Yes, I think so -- but that doesn't exclude the possibility that I'm going
about this in completely the wrong way :)

> Here is a toy project.
> Hth.

Thanks again for taking the time to respond.

andrew