El 23/08/2012, a las 18:13, Jed Brown escribió:

On Thu, Aug 23, 2012 at 11:02 AM, Jose E. Roman <[email protected]> wrote:

El 22/08/2012, a las 18:52, Jose E. Roman escribió:

...

f2cblaslapack is more or less prepared for this case already (it is not much tested though, we will make some tests for the w* subroutines tomorrow).

What is not ready is PETSc itself. Some definitions are missing. For instance, PetscScalar is undefined, so one should add the following to petscmath.h:106:

#elif defined(PETSC_USE_REAL___FLOAT128) typedef __complex128 PetscScalar;

#define PetscRealPart(a) crealq(a) #define PetscImaginaryPart(a) cimagq(a) #define PetscAbsScalar(a) cabsq(a) #define PetscConj(a) conjq(a) #define PetscSqrtScalar(a) csqrtq(a) #define PetscPowScalar(a,b) cpowq(a,b) #define PetscExpScalar(a) cexpq(a) #define PetscLogScalar(a) clogq(a) #define PetscSinScalar(a) csinq(a) #define PetscCosScalar(a) ccosq(a)

Similarly, MPIU_SCALAR should be defined. Maybe other as well.

Jose

I have pushed a change to petsc-dev that allows to configure --with-precision=__float128 --with-scalar-type=complex

I have tested this with src/ksp/ksp/examples/tutorials/ex12.c with 1 and 2 processes. It seems to work.

Please review the change, since I had to modify pinit.c

It looks good to me. Thanks.

A couple of comments:

1) The f2cblaslapack interface to quad precision subroutines (q* and w*) is defined only in the case of "uscore" mangling. Maybe a test in configure should assure this.

Don't we just need to copy and update the list of defines to the different mangling conventions? Seems easier to do than to test/document that it doesn't work.

Yes. But the "stdcall" case seems more tedious. Anyway, __float128 is only available in GCC which uses underscore mangling, right?

2) I could not make it work --with-clanguage=c++ - I get compilation errors like this. Any idea?

C++, in all it's template metaprogramming glory, standardized that std::complex may *only* be used with float, double, and long double. Using std::complex<anything_else> is invalid. Just one of a very long list of reasons why std::complex should not be used.