The preferred method for compiling a CAF program is by invoking the "caf" bash script that the OpenCoarrays CMake scripts install in the "bin" subdirectory of the installation path. This is an experimental script with limited but useful capabilities that will grow over time. Please submit bug reports and feature requests via our Issues page.
The "caf" script liberates the source code and workflow from explicit dependence on the underlying compiler and communication library in the following ways:
- With an OpenCoarrays-aware (OCA) CAF compiler, the "caf" script passes the unmodified
source code to the underlying compiler with the necessary arguments for building a
CAF program, embedding the paths to OpenCoarrays libraries (e.g., libcaf_mpi.a) installed
in the "lib" subdirectory of the OpenCoarrays installation path. The "caf" script also
embeds the path to the relevant module file in the "mod" subdirectory of the installation
path (e.g., opencoarrays.mod). This supports use association with module entities via
use opencoarrays. - With a non-CAF compiler (including gfortran 4.9), "caf" supports a subset of CAF by replacing CAF statements with calls to procedures in the opencoarrays module before passing the source code to the compiler.
When using GCC 4.9, we recommend using the use statement's "only" clause to
avoid inadvertent procedure name clashes between OpenCoarrays procedures and their
GCC counerparts. For example, use "use opencoarrays, only : co_reduce".
With a non-OCA and OCA CAF compilers, the extensions that "caf" imports include the collective subroutines proposed for Fortran 2015 in the draft Technical Specification TS 18508 Additional Parallel Features in Fortran.
The latter use case provides an opportunity to mix a compiler's CAF support with that of OpenCoarrays.
For example, a non-OCA CAF compiler, such as the Cray or Intel compilers, might support all of a
program's coarray square-bracket syntax, while OpenCoarrays supports the same program's calls to
collective subroutine such as co_sum and co_reduce.
The following program listing, compilation, and execution workflow exemplify the use of an OCA compiler (e.g., gfortran 5.1.0 or later) in a Linux bash shell with the "bin" directory of the chosen installation path in the user's PATH environment variable:
$ cat tally.f90
program main
use iso_c_binding, only : c_int
use iso_fortran_env, only : error_unit
implicit none
integer(c_int) :: tally
tally = this_image() ! this image's contribution
call co_sum(tally)
verify: block
integer(c_int) :: image
if (tally/=sum([(image,image=1,num_images())])) then
write(error_unit,'(a,i5)') "Incorrect tally on image ",this_image()
error stop
end if
end block verify
! Wait for all images to pass the test
sync all
if (this_image()==1) print *,"Test passed"
end program
$ caf tally.f90 -o tally
$ cafrun -np 4 ./tally
Test passed
where "4" is the number of images to be launched at program start-up.
## An advanced workflow ##To extend the capabilities of a non-OCA CAF compiler (e.g., the Intel or Cray compilers),
access the types and procedures of the opencoarrays module by use assocication. We
recommend using a use statement with an only clause to reduce the likelihood of a
name clash with the compiler's native CAf support. For eample, innsert the following
at line 2 of tally.f90 above:
use opencoarrays, only : co_sum
To extend the capabilities of a non-CAF compiler (e.g., GCC 4.9), use an unqualified
use statement with no only clause. The latter practrice reduces the likelihood of
name clashes with the compiler's or programs existing capabilities.
If the "caf" compiler wrapper cannot process the source code in question, invoke the underlying communication library directly:
mpif90 -fcoarray=lib -L/opt/opencoarrays/ tally.f90 \ -lcaf_mpi -o htally-I<OpenCoarrays-install-path>/mod
and also run the program with the lower-level communication library:
mpirun -np <number-of-images> ./tally