ompi_mpi_init(): we no longer *query* the components to see what they
can support -- instead, we *tell* the components what they need to
support (if they can't support it, they should disqualify themselves
from selection).
This involved a bunch of things:
- pass the configure-time-decided OMPI_ENABLE_PROGRESS_THREADS and
OMPI_ENABLE_MPI_THREADS constants to each framework selection
function
- eliminate mca_base_init_select_components()
- eliminate all "allow_multi_user_threads" and "have_hidden_threads"
kinds of arguments and aggregation from various framework selection
and component/module functions
- correctly determine the max MPI thread level that we can support and
limit the final MPI thread level as appropriate during
ompi_mpi_init()
As a side effect, while editing nearly every MPI component I also did
the following:
- found 2 places (gm and ib ptls) where ompi_set_using_threads() was
erroneously being called (I think this is left over kruft from
misunderstandings by the original authors). I removed these.
This commit was SVN r5055.
built. The issue is that these tests are trying to test specific
components, and is calling the functions directly -- and therefore
needs to have the component linked in. This is fine when the
component is statically linked as part of libmpi, but presents a
problem when the component is a DSO.
GNU compilers/linkers allow us to link in the DSO as part of the test
executable (and everything "just works"), but this is not portable. A
better solution is going to involve:
- a better unit test support library that can load a DSO on demand
- using function pointers in the unit tests (rather than direct
function invocation)
This commit was SVN r5051.
less confusing (when you have 5-6 parameters, it's easy to mis-count
and have the wrong type if you don't have the associated names).
- Use proper types for the fortran callbacks
- Simplify the predefined Fortran attribute callbacks; do the work
right there rather than calling the predefined C attribute callbacks
(which would simply require further FINT_2_INT / INT_2_FINT
translation anyway)
This commit was SVN r5038.
- Ensure to return properly if we error in the fortran callbacks
- Don't bump the refcount of attributes an extra 1 when we create
them; this just ensures that we never really delete them (oops)
This commit was SVN r5037.
technically speaking, this is a cosmetic change, but having the wrong
parameter names was extremely confusing -- it took me a while (and I
had to cross-reference MPI-1) to figure out that they were wrong, and
I was thinking about the functions incorrectly as a result.
This commit was SVN r5035.
technically speaking, this is a cosmetic change, but having the wrong
parameter names was extremely confusing -- it took me a while (and I
had to cross-reference MPI-1) to figure out that they were wrong, and
I was thinking about the functions incorrectly as a result. Our mpi.h
should have the Right names in order to not confuse users (and
developers!).
This commit was SVN r5034.
set after we call mca_pml.pml_comm_add(). This allows us, in the
destructor, to only call mca_pml.pml_comm_del() if the corresponding
pml_comm_add() was called. In this way, if an error occurs during the
construction of a communicator (e.g., attributes failing to copy, as
in the Intel test MPI_Keyval3_c), we simply propagate the error up the
stack so that an MPI exception can be generated -- not cause a failed
assert() in teg's pml_del_comm().
This commit was SVN r5032.
more obvious (and documented)
- VERSION now has lots of comments in it
- separate out whether we want the SVN r number in the version and
what the SVN r number is
- clearly describe the process for building the full version number
string
- make ompi info have a separate line item specifically for the SVN r
number, regardless of what the version number is
- update "make dist" process to substitute in the SVN r version
This commit was SVN r5022.
interface is unchanged):
- the first call to OBJ_CONSTRUCT and OBJ_NET on a given class is now
thread safe
- fix memory leaks for the vtable constructor / destructors;
orte_finalize() makes a call to free all the memory allocated for
that stuff. After ompi_class_finalize(), any calls to OBJECT
functionality will be met with a swift, painful SEGV death!
This commit was SVN r5017.
- check return value correctly when posting non-blocking recvs
- use any values that have been set in the global structs as the
defaults when registering mca parameters - this prevents any
values that have been set in the structs from the command line
parser from being overwritten
This commit was SVN r5011.
