dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

We are currently using "real" (almost) everywhere in the code instead
of "double". real is a typedef for double defined in dolfin/common/types.h:

  typedef double real;

The reason for this typedef has been to prepare for changing from
double to another (higher) precision representation at some point in
the future.

We (Benjamin Kehlet) have now tried to do this using GMP
(http://gmplib.org), but this fails for many reasons. For example,
standard math functions (abs, pow, exp) only work with double. Another
more obvious reason is dependencies on external libraries like PETSc,
UFC, libxml2, GTS etc where double is hard-coded.

Our conclusion is that there is little reason to keep the typedef in
DOLFIN since it is just an alias for double and it can not easily be
changed to anything else. I suggest we remove it and replace "real"
with "double" everywhere. Any objections to this?

The reason we are experimenting with GMP is that we want to use the
ODE solvers to solve ODEs with very high precision. For this reason,
we are planning to

1. Put back typedef real (after first removing it) and only use it
under dolfin/ode/. If DOLFIN is built without GMP, it will just be a
typedef for double. Otherwise, it will be a GMP type.

2. Change the ODE solver interface from

  void f(const uBLASVector& u, real t, uBLASVector& y);

to

  void f(const real* u, real t, real* y);

uBLAS will be used internally (with some additional copying of data
from real to double).

Any objections to this?

-- 
Anders

Attachment: signature.asc
Description: Digital signature