with MPI. This will work for one node with several devices and one MPI
process per device.

* Now messages_info has the all_nodes optional argument.

* It seems that AMD OpenCL has problems compiling just before execution.

fortcom: Error: ../../../src/basic/cl_kernel.F90, line 140: The dummy argument of a derived type containing a pointer component has the INTENT(IN) attribute therefore it shall not appear as the expr of an assignment-stmt in a PURE procedure.   [THIS]
    ref = this%kernel
----------^

and this warning

fortcom: Warning: ../../../src/basic/opencl.F90, line 411: The return value of this FUNCTION has not been defined.   [WORKGROUP_SIZE]
    integer function opencl_kernel_workgroup_size(kernel) result(workgroup_size)
-----------------------------------------------------------------^

* Made filename agree with module name for opt_control_iter.

in the place of the code where they are used and are compiled the
first time they are called.

* Fixed a bug in the calculation of the OpenCL complex dot product.

to local memory.

certain extension.

* Now the compilation flags for OpenCL are generated in the Fortran
code.

and the other builds it.

routines to messages_info, messages_fatal, etc.

Octopus now prints information about the available OpenCL platforms. The platform can be selected by OpenCLPlatform.

New form for push_sub/pop_sub. Doing it via a preprocessor macro lets the filename be inseerted automatically, avoiding any discrepancies, and also lets the "X" of the real/complex versions of functions be resolved, so that the debug trace shows which version was called.

projectors.

executed.

variable.

* cl_finish now returns the error code.