> I just wanted to notify those who are interested in a free compiler
> supporting Fortran coarrays that this is available in the OpenUH compiler.
> OpenUH is developed by the HPCTools group at the University of Houston. It
> is not a full Fortran 2008 implementation. It is based on the Open64
> compiler, which support Fortran 95 and *some* Fortran 2003 features. We
> have implemented the coarray features that were added to Fortran 2008.

> You can download a tarball of the prebuilt compiler or, if you wish,
> access the source code from here:

> http://www2.cs.uh.edu/~openuh/download/

> I am interested in getting feedback from the community on all aspects of
> the implementation -- performance issues, features that may not be
> implemented correctly, ease of installation on your particular system,
> etc.

> A few notes about the implementation:

> (1) OpenUH should support all the coarray features that are in Fortran
> 2008. This includes coarrays of basic types and derived types (including
> coarrays with pointer/allocatable components), the inquiry and
> transformational functions (this_image, image_index, num_images, ucobound,
> lcobound), all the sync statements (sync all, sync images, sync memory),
> locks, critical sections, error stop, and the atomic_define/atomic_ref
> subroutines.

> (2) The runtime can be built against GASNet (1.18.2) or ARMCI (5.0). These
> are communication libraries designed to support PGAS implementations. If
> you are using the prebuilt compiler, then you need to download the
> separate caf-runtime source package and the build/install it into
> whereever you extracted the compiler. Both GASNet and ARMCI support a
> variety of networks or may operate in an SMP environment, so they provide
> good portability.

> (3) The compiler name is uhcaf. The launcher program is cafrun. E.g.

> # compile the program and links in the necessary runtime libraries:
> $ uhcaf sample.f90 -o sample

> # run the executable with 4 images, more options available with --help
> $ cafrun -n 4 ./sample

> (4) Performance is something we're focused on improving. We leverage the
> bulk-communication capabilities available in GASNet/ARMCI for contiguous
> and strided accesses across images. Currently, for the program to benefit
> from this it must be written using the array assignment syntax. For
> example, this:

>    x(:) = y(:)[p]

> will get you better performance than this:

>    do i=lbound(x), ubound(x)
>       x(i) = y(i)[p]
>    end do

> Another performance limitation is the compiler doesn't yet generate
> non-blocking calls to help overlap your communication and computation, but
> this is something we're working on and hope to have ready by the end of
> this year.

> I hope the community finds this useful, and again I welcome any feedback
> or suggestions.

> Thanks,
> Deepak Eachempati
> HPCTools Group
> University of Houston