Over 2 weeks, I've struggled to call one of the METIS library written in C from my fortran code. And, unfortunately, It doesn't seem to be a HAPPY END without your help. I found some posts about direct calling and using interface. I prefer the latter because I could monitor the variables for debugging. There are three codes I attached.
1. c function I'd like to use
2. fortran interface module
3. fortran program
(1) c function
int METIS_PartMeshNodal(idx_t *ne, idx_t *nn, idx_t *eptr, idx_t *eind,
idx_t *vwgt, idx_t *vsize, idx_t *nparts, real_t *tpwgts,
idx_t *options, idx_t *objval, idx_t *epart, idx_t *npart)
I removed the c funciton body. It's not necessary to understand my problem
Here, idx_t is integer and real_t is single or double precision. From ne to options are input and last three arguments are output. And vwgt, vsize, tpwgts and options can receive null as an input for default setting I wrote the interface module for using c function like this
(2) Fortran interface module
Fixed!
- Insert use iso_c_bind under use constants
- Use integer(c_int) instead of integer for ne, nn and other variables.
- Remove unused module constants
.
module Calling_METIS
!use constants, only : p2 !this is for double precision
use iso_c_bind !inserted later
implicit none
!integer :: ne, nn !modified
integer(c_int) :: ne, nn
!integer, dimension(:), allocatable :: eptr, eind !modified
integer(c_int), dimension(:), allocatable :: eptr, eind
!integer, dimension(:), allocatable :: vwgt, vsize !modified
type(c_ptr) :: vwgt, vsize
!integer :: nparts !modified
integer(c_int) :: nparts
!real(p2), dimension(:), allocatable :: tpwgts !modified
type(c_ptr) :: tpwgts
!integer, dimension(0:39) :: opts !modified
integer(c_int), dimension(0:39) :: opts
!integer :: objval !modified
integer(c_int) :: objval
!integer, dimension(:), allocatable :: epart, npart !modified
integer(c_int), dimension(:), allocatable :: epart, npart
interface
subroutine METIS_PartMeshNodal( ne, nn, eptr, eind, vwgt, vsize, nparts, tpwgt, &
opts, objval, epart, npart) bind(c)
use intrinsic :: iso_c_binding
!use constants, only : p2
implicit none
integer (c_int), intent(in) :: ne, nn
integer (c_int), dimension(*), intent(in) :: eptr, eind
!integer (c_int), dimension(*), intent(in) :: vwgt, vsize !modified
type(c_ptr), value :: vwgt, vsize
integer (c_int), intent(in) :: nparts
!real(c_double), dimension(*), intent(in) :: tpwgt !modified
type(c_ptr), value :: tpwgt
integer (c_int), dimension(0:39), intent(in) :: opts
integer (c_int), intent(out) :: objval
integer (c_int), dimension(*), intent(out) :: epart
integer (c_int), dimension(*), intent(out) :: npart
end subroutine METIS_PartMeshNodal
end interface
end module
And here is my program code calling the function
(3) Fortran program
Fixed!
- allocation size of npart is fixed. Not ne but nn
- opts(7)=1 is added to get Fortran-style array of epart, npart(no effect until now)
.
program METIS_call_test
!some 'use' statments
use Calling_METIS
use iso_c_binging !added
implicit none
! Local variable
integer :: iC
character(80) :: grid_file !grid_file
grid_file = 'test.grid'
! (1) Read grid files
call read_grid(grid_file)
! (2) Construction Input Data for calling METIS Function
! # of cells, vertices
ne = ncells
nn = nvtxs
! eptr, eind allocation
allocate(eptr(0:ne), eind(0:3*ntria + 4*nquad - 1))
! eptr and eind building
eptr(0) = 0
do iC=1, ncells
eptr(iC) = eptr(iC-1) + cell(iC)%nvtxs
eind(eptr(iC-1):eptr(iC)-1) = cell(iC)%vtx
end do
! epart, npart building
!allocate(epart(ne), npart(ne))
allocate(epart(ne), npart(nn)) ! modified
! # of partition setting
nparts = 2
vwgt = c_null_ptr !added
vsize = c_null_ptr !added
tpwgt = c_null_ptr !added
! (3) Call METIS_PartMeshNodal
call METIS_SetDefaultOptions(opts)
opts(7) = 1 !Added. For fortran style output array epart, npart.
call METIS_PartMeshNodal(ne, nn, eptr, eind, vwgt, vsize, nparts, tpwgt, &
opts, objval, epart, npart)
!call METIS_PartMeshNodal(ne, nn, eptr, eind, null(), null(), nparts, null(), &
! opts, objval, epart, npart) !wrong...
end program
But the problem is that I get an error message as below though I put null for tpwgt.
Input Error: Inorrect sum of 0.000000 for tpwgts for constraint 0.
And this message is handled in the code below.
for (i=0; i<ctrl->ncon; i++) {
sum = rsum(ctrl->nparts, ctrl->tpwgts+i, ctrl->ncon);
if (sum < 0.99 || sum > 1.01) {
IFSET(dbglvl, METIS_DBG_INFO,
printf("Input Error: Incorrect sum of %"PRREAL" for
tpwgts for constraint %"PRIDX".\n", sum, i));
return 0;
}
}
Anyway, in order to see what I would get if I put an array for tpwgts intead of null, tpwgts(:) = 1.0/nparts, which makes sum of tpwgts equal 1.0. But I got same message with 1.75 for the sum.
These are my questions
1. Did I use null() for passing arguments correctly?
2. Do I have to pass pointers for all arguments to c function? then how?
3. Is putting an integer to opts(0:39) enough for use? For example, in a post without 'interface module', simple code like options(3)=1 is used. But in the c code, options has 16 named variable like options[METIS_OPTION_NUMBERING], options[METIS_OPTION_UFACTOR]. I think some thing is necessary to set options but I have no idea.
4. Is there an example for METIS in fortran?
Any kind of hint/advice will be a great help for me. Thank you.
Conclution
The problem I had was that c function couldn't recognize null pointer from fortran code.
There were some miss declations of variables in interface module(see 'Fixed' and comments)
It looks like the code works properly. But option(7) = 1 for fortran style output didn't work and now I'm looking at it.
I think your
opts(7)
does not work because you also need an interface for the METIS functionMETIS_SetDefaultOptions
. Based on the answer from http://glaros.dtc.umn.edu/gkhome/node/877, I created a wrapper module (metisInterface.F90
) with the interfaces I needed:Then, in the main program (or wherever you make the call to the METIS functions) you need to have (for completeness, I also added the call to
METIS_PartMeshDual
):Note that
epart
andnpart
will have Fortran numbering as you want (starting at 1). However, the processors will also start numbering at 1. For example, if you are running in 4 processors, root processor is 1 and you may haveepart(n)=4
, and you will not have anyepart(n)=0
.Finally, a file
metis.c
is also needed with a single line:Compiling instructions
metis.c
with a C compilermetisInterface.F90
with a Fortran compiler linking with the compiled C objectmetisInterface.o
No, you cannot pass
null()
, that is a Fortran pointer constant. You must passC_NULL_PTR
from the moduleISO_C_BINDING
and the interface must reflect this. The dummy argument must betype(c_ptr)
, most probably withVALUE
attribute. It may actually work because of the same internal representation, but I wouldn't count on it.No, if you pass some normal variable, you can pass it directly by reference. Just like normally in Fortran. If the interface is
BIND(C)
, the compiler knows it must send a pointer.There is a new TS to update Fortran 2008, where you can define dummy arguments in the interoperable procedures as
OPTIONAL
. Then you can pass the null pointer just by omitting them. Gfortran should already support this.Note: Here I can see a much different C signature of your function, are you sure yours is OK? http://charm.cs.uiuc.edu/doxygen/charm/meshpart_8c.shtml