Re: Re: The support of log double

[Gong Ding]

I'd like to thank the developers of PETSC. 

I had spent years to develop a semiconductor
simulation code which based on PETSC. The code is
carefully checked by valgrind and no memory
problem should exist.

In semiconductor, the magnitude of carrier density
differs much. When the electron density is 10e18,
the hole density is only 10e2. The sum/sub of two
variables may have float precision problem and in
my code, a lot of parameters must be computed with
density arguments. In most of the situation, my
code works fine with both gcc/icc except a
heterojunction BJT example. The code generated by
icc works but not work with gcc. I guess the float
point precision in gcc is not enough. (I found a
“long double” description in the configure file of
icc. Does icc use long double internally?) 

Unfortunately, icc is not free. I have to make gcc
work. I’d like to try petsc with long double. 

Another requirement, can PETSC functions support
different “scalar type”? For example, by adding a
postfix letter as BLAS?  The semiconductor code
needs “real” value but I’d like to add optical
modular which needs “complex” value. At present, I
can only link real lib. As a result, I have to do
a lot of work on complex->real.   

Yous
G.D.
--
USTC Alumni Email System