Even if the compiler supports an "alternate" short float type (e.g.,
_Float16), check to make sure that the compiler will correctly link
applications that perform mathematical operations on that type.
Carefully choose the mathematical test in the configure check to
ensure the mathematical operation is not removed by compiler
optimization (when setting CFLAGS=-O1 or higher).
Out of the box, clang 6.0.x and 7.0.x will fail to link applications
that try to perform addition (and other mathematical operations) on
_Float16 variables (an additional CLI flag is required to enable
software emulation of _Float16). If we detect a situation where the
type is supported by a sample program fails to link and the basename
of $CC is "clang", emit a warning and point the user to a relevant
README.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Signed-off-by: KAWASHIMA Takahiro <t-kawashima@fujitsu.com>
Keep all comments in the user-facing mpi.h.in as "old style" C
comments: /* */. This gives us maximum portability, just on the off
chance that a user's C compiler does not support //-style comments.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
1. __STDC_VERSION__ isn't necessarily defined (e.g., by C++
compilers). So check to make sure it is defined before we actually
check the value.
2. If we're in C++11 (or later), use static_assert().
3. Split the static assert macro in two macros:
* THIS_SYMBOL_WAS_REMOVED_IN_MPI30(...): Insert a valid expression
(i.e., 0, because it's only used with MPI_Datatype values, and
since MPI_Datatype is a pointer, 0 is a valid RHS expression)
before invoking the static assert so that we don't get a syntax
error instead of the actual static assert error.
* THIS_FUNCTION_WAS_REMOVED_IN_MPI30(...): No need for the valid
expression; just invoke the assert functionality.
Also remove an errant "\".
Thanks to Constantine Khrulev and Martin Audet for identifying the
issue and suggesting to use C11's static_assert().
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
This is a fix based on a bugreport on github/mailing list from CGNS.
The core of the problem was that different processes entered different branches of
our aggregator selection logic, due to the fact that in some cases processes had
a matching file_view size and contiguous chunk size (thus assuming 1-D distribution),
and some processes did not (thus assuming 2-D distribution). The fix is to calculate
the avg. file view size across all processes and use this value, thus ensuring that
all processes enter the same branch.
Fixes issue #7809
Signed-off-by: Edgar Gabriel <egabriel@central.uh.edu>
As discussed, a feature is being added to libpsm2 to correctly handle
the case where the library is opened by multiple OMPI transports in the same
process. (For example, the OFI BTL and the PSM2 MTL).
* Improved error message to indicate required libpsm2 version.
* Adds a test at autogen/configure time for the existence of
PSM2_LIB_REFCOUNT_CAP.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Signed-off-by: Michael Heinz <michael.william.heinz@intel.com>
The gather and scatter operations did not use the correct message size
(Only did datatype size * com size). This did not correctly reflect the
total message size and prevents fine tuning within a com size. This
patch multiplies the value by the number of elements sent.
Signed-off-by: William Zhang <wilzhang@amazon.com>
Change ompi_mtl_ofi_get_endpoint() to call the active PML's add_procs()
rather than the OFI MTL add_procs() directly when discovering a new
process during operation.
Functionally, this has no impact in correct operation. However, the
current behavior means that the heterogenous and active PML checks
are not being executed in the dynamic discovery case.
Signed-off-by: Michael Heinz <michael.william.heinz@intel.com>
also add common verbose variable.
Note the verbosity thing is a little tricky owing to the way the MCA frameworks and components are registered and
and initialized. The BTL's are registered/initialized prior to the MTL components even getting registered.
Signed-off-by: Howard Pritchard <howardp@lanl.gov>
Also added infrastructure to have developers write man pages in
Markdown (vs. nroff). Pandoc >=v1.12 is used to convert those
Markdown files into actual nroff man pages.
Dist tarballs will contain generated nroff man pages; we don't want to
require users to have Pandoc installed. Anyone who builds Open MPI
from a git clone will need to have Pandoc installed (similar to how we
treat Flex). You can opt out of Open MPI's Pandoc-generated man pages
by configuring Open MPI with --disable-man-pages. This will also
disable "make dist" (i.e., "make dist" will error if you configured
with --disable-man-pages).
Also removed the stuff to re-generate man pages.
This commit also:
1. Includes a new man page, written in Markdown
(ompi/mpi/man/man5/MPI_T.5.md) that contains Open MPI-specific
information about MPI_T.
2. Includes a converted ompi/mpi/man/man3/MPI_T_init_thread.3.md (from
MPI_T_init_thread.3in -- i.e., nroff) just to show that Markdown
can be used throughout the Open MPI code base for man pages.
3. Made the Makefiles in ompi/mpi/man/man?/ be full-fledged
Makefile.am's (vs. Makefile.extras that are designed to be included
in ompi/Makefile.am). It is more convenient to test generation /
installation of man pages when you can "make" and "make install" in
their respective directories (vs. doing a build / install for the
entire ompi project).
4. Removed logic from ompi/Makefile.am that re-generated man pages if
opal_config.h changes.
Other man pages -- hopefully all of them! -- will be converted to
Markdown over time.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
- added detection of new API into configuration
- added tag_send call implemented using new API
- added MPI_Send/MPI_Isend/MPI_Recv/MPI_Irecv implementations
Signed-off-by: Sergey Oblomov <sergeyo@mellanox.com>
Adds the capability to select a NIC based on hardware locality.
Creates a list of NICs that share the same cpuset as the process,
then selects the NIC based on the (local rank) % (number of NICs).
If no NICs are available that share the same cpuset, the selection process
will create a list of all available NICs and make a selection based on
(local rank) % (number of NICs)
Signed-off-by: Nikola Dancejic <dancejic@amazon.com>
Deprecate the current OMPI-specific MPI_Info key definitions for
MPI_Comm_spawn and replace them with their PMIx equivalents. Issue a
deprecation/conversion warning as this is done. Also issue deprecation
warnings for options such as "ompi_non_mpi" that are no longer used.
Handle both cases where the user might pass either the PMIx attribute
name itself (e.g., "PMIX_MAPBY") or the string value of the attribute
(e.g., PMIX_MAPBY, which translates to "pmix.mapby"). This can only be
done for PMIx v4 and above, so protect that code.
Silence a couple of Coverity warnings and add a test along the way.
Signed-off-by: Ralph Castain <rhc@pmix.org>
Consolidate the ompi_process_info and opal_process_info structs to
remove duplicate storage and conversion issues. Unwind some interweaving
of include files using opal.h. Silence a couple of warnings.
For now, set the arch to local if PMIX_ARCH is not found.
Signed-off-by: Ralph Castain <rhc@pmix.org>
For direct modex, all procs publish the selected pml module
and then at add_procs pml module for each proc is checked
against every other proc in the add_proc call.
For full modex, there is no change in functionality. Only Rank0
publishes its selected pml, all other procs in the add_proc call
check their selected pml against Rank0.
If pml's do not match, throw error and exit.
Signed-off-by: Dipti Kothari <dkothar@amazon.com>
