# Copyright 1999-2024 Gentoo Authors # Distributed under the terms of the GNU General Public License v2 EAPI=8 FORTRAN_NEEDED=fortran inherit cuda flag-o-matic fortran-2 libtool MY_P=${P/-mpi} IUSE_OPENMPI_FABRICS=" openmpi_fabrics_ofed openmpi_fabrics_knem" IUSE_OPENMPI_RM=" openmpi_rm_pbs openmpi_rm_slurm" DESCRIPTION="A high-performance message passing library (MPI)" HOMEPAGE="https://www.open-mpi.org" SRC_URI="https://www.open-mpi.org/software/ompi/v$(ver_cut 1-2)/downloads/${MY_P}.tar.bz2" S="${WORKDIR}/${MY_P}" LICENSE="BSD" SLOT="0" KEYWORDS="~amd64 -arm ~arm64 -ppc ~ppc64 -x86 ~amd64-linux" IUSE="cma cuda fortran ipv6 peruse romio valgrind ${IUSE_OPENMPI_FABRICS} ${IUSE_OPENMPI_RM}" REQUIRED_USE=" openmpi_rm_slurm? ( !openmpi_rm_pbs ) openmpi_rm_pbs? ( !openmpi_rm_slurm ) " RDEPEND=" !sys-cluster/mpich !sys-cluster/mpich2 !sys-cluster/nullmpi >=dev-libs/libevent-2.0.22:=[threads(+)] >=sys-apps/hwloc-2.0.2:= sys-cluster/pmix:= sys-cluster/prrte:= >=sys-libs/zlib-1.2.8-r1 cuda? ( >=dev-util/nvidia-cuda-toolkit-6.5.19-r1:= ) openmpi_fabrics_ofed? ( sys-cluster/rdma-core ) openmpi_fabrics_knem? ( sys-cluster/knem ) openmpi_rm_pbs? ( sys-cluster/torque ) openmpi_rm_slurm? ( sys-cluster/slurm ) " DEPEND="${RDEPEND} valgrind? ( dev-debug/valgrind )" pkg_setup() { fortran-2_pkg_setup elog elog "OpenMPI has an overwhelming count of configuration options." elog "Don't forget the EXTRA_ECONF environment variable can let you" elog "specify configure options if you find them necessary." elog } src_prepare() { default elibtoolize use cuda && cuda_src_prepare # Avoid test which ends up looking at system mounts echo "int main() { return 0; }" > test/util/opal_path_nfs.c || die # Necessary for scalibility, see # http://www.open-mpi.org/community/lists/users/2008/09/6514.php echo 'oob_tcp_listen_mode = listen_thread' \ >> opal/etc/openmpi-mca-params.conf || die } src_configure() { # -Werror=lto-type-mismatch, -Werror=strict-aliasing # The former even prevents successfully running ./configure, but both appear # at `make` time as well. # https://bugs.gentoo.org/913040 # https://github.com/open-mpi/ompi/issues/12674 # https://github.com/open-mpi/ompi/issues/12675 append-flags -fno-strict-aliasing -fpermissive filter-lto local myconf=( --disable-mpi-java # configure takes a looooong time, but upstream currently force # constriants on caching: # https://github.com/open-mpi/ompi/blob/9eec56222a5c98d13790c9ee74877f1562ac27e8/config/opal_config_subdir.m4#L118 # so no --cache-dir for now. --enable-mpi-fortran=$(usex fortran all no) --enable-prte-prefix-by-default --enable-pretty-print-stacktrace --sysconfdir="${EPREFIX}/etc/${PN}" --with-hwloc=external --with-libevent=external --with-prrte=external # Oiriginally supposed to be re-enabled for 5.0! # See https://github.com/open-mpi/ompi/issues/9697#issuecomment-1003746357 # and https://bugs.gentoo.org/828123#c14 # # However as of 5.0.3 the docs still say: # # As such, supporting data heterogeneity is a feature that has fallen # into disrepair and is currently known to be broken in this release # of Open MPI. --disable-heterogeneous $(use_enable ipv6) $(use_enable peruse) $(use_enable romio io-romio) $(use_with cma) $(use_with cuda cuda "${EPREFIX}"/opt/cuda) $(use_with valgrind) $(use_with openmpi_fabrics_knem knem "${EPREFIX}"/usr) $(use_with openmpi_rm_pbs tm) $(use_with openmpi_rm_slurm slurm) ) CONFIG_SHELL="${BROOT}"/bin/bash econf "${myconf[@]}" } src_compile() { emake V=1 } src_test() { emake -C test check } src_install() { default # Remove la files, no static libs are installed and we have pkg-config find "${ED}" -name '*.la' -delete || die }