[petsc-dev] Mac OSX complex BLAS problem

[Jack Poulson]

I think I mentioned this to several PETSc folks two years ago. It is a bad
idea to return complex numbers from Fortran, as there is no portable
solution as far as I know; some compilers need you to declare it as

complex zdotc_( <regular args> )

some as

void zdotc_( <regular args>, complex* z )

and I think I ran into problems with both on OSX (but don't quote me on
it). I ended up settling on writing my own complex dotc and dotu routines.

Jack

On Sat, Feb 18, 2012 at 5:28 PM, Matthew Knepley <knepley at gmail.com> wrote:

> We have a problem with complex dot product on OSX. Here is the easiest way
> to see it (I think):
>
> knepley:/PETSc3/petsc/petsc-dev$ ./arch-complex-fftw-debug/lib/ex5-obj/ex5
> -snes_monitor
> ./arch-complex-fftw-debug/lib/ex5-obj/ex5 -snes_monitor
>   0 SNES Function norm 0.000000000000e+00
>
> With the debugger:
>
> Breakpoint 1, VecNorm_Seq (xin=0x1022fc170, type=NORM_2, z=0x7fff5fbfea10)
> at bvec2.c:238
> 238         ierr = VecGetArrayRead(xin,&xx);CHKERRQ(ierr);
> (gdb) n
> Current language:  auto; currently c++
> 239         *z = BLASdot_(&bn,xx,&one,xx,&one);
> (gdb) p bn
> $1 = 16
> (gdb) p xx[0]@16
> $2 = {{
>     _M_value = 0 + 0 * I
>   }, {
>     _M_value = 0 + 0 * I
>   }, {
>     _M_value = 0 + 0 * I
>   }, {
>     _M_value = 0 + 0 * I
>   }, {
>     _M_value = 0 + 0 * I
>   }, {
>    _M_value = -0.10378182131158753 + 0 * I
>   }, {
>     _M_value = -0.10378182131158753 + 0 * I
>   }, {
>     _M_value = 0 + 0 * I
>   }, {
>     _M_value = 0 + 0 * I
>   }, {
>     _M_value = -0.10378182131158753 + 0 * I
>   }, {
>     _M_value = -0.10378182131158753 + 0 * I
>   }, {
>     _M_value = 0 + 0 * I
>   }, {
>     _M_value = 0 + 0 * I
>   }, {
>     _M_value = 0 + 0 * I
>   }, {
>     _M_value = 0 + 0 * I
>   }, {
>     _M_value = 0 + 0 * I
>   }}
> (gdb) p one
> $3 = 1
> (gdb) n
> 240         *z = PetscSqrtReal(*z);
> (gdb) p *z
> $4 = 0
> (gdb) p zdotc
> $5 = {<text variable, no debug info>} 0x7fff863e71e2 <zdotc_>
>
> It appears that the dot product is producing nothing. Didn't we have a
> similar problem a little while ago?
>
>    Matt
>
> --
> What most experimenters take for granted before they begin their
> experiments is infinitely more interesting than any results to which their
> experiments lead.
> -- Norbert Wiener
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20120218/5281d15a/attachment.html>