zhangyiss/spack
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[Fix for the GitLab CI] phist: prefer @1.9.5 (1.9.6 is not compatible w/ mpich%gcc:9) (#26773)

Browse Source 
* phist: Prefer 1.9.5 (1.9.6 uses mpi_f08, but not available in CI)

* phist: remove dupe of 1.9.5, missing preferred=True

Also, for 1.9.6, patch the (most, one does not work) tests to use
This commit is contained in:
Bernhard Kaindl 2021-10-16 20:18:07 +02:00 committed by GitHub
parent b2f059e547
commit 64143f970e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 70 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

73
var/spack/repos/builtin/packages/phist/package.py
View File

@ -3,6 +3,8 @@
#
# SPDX-License-Identifier: (Apache-2.0 OR MIT)
import llnl.util.tty as tty
import spack.hooks.sbang as sbang
from spack import *
@ -25,10 +27,37 @@ class Phist(CMakePackage):
maintainers = ['jthies']
tags = ['e4s']
# phist is a required part of spack GitLab CI pipelines. In them, mpich is requested
# to provide 'mpi' like this: spack install phist ^mpich %gcc@7.5.0
# Failure of this command to succeed breaks spack's gitlab CI pipelines!
version('develop', branch='devel')
version('master', branch='master')
# phist-1.9.6 updated from the older "use mpi" to the newer "use mpi_f08" (MPI3.1):
# The motivation was fixing it on Cray: https://github.com/spack/spack/issues/26002
# Documentation: https://www.mpi-forum.org/docs/mpi-3.1/mpi31-report/node408.htm
# mpich does not provide mpi_f08.mod with gfortran-[789], it needs gfortran>=10:
# https://stackoverflow.com/questions/65750862
version('1.9.6', sha256='98ed5ccb22bb98d5b6bf9de0c9960105473e5244978853070b9a3c44138db662')
version('1.9.5', sha256='24faa3373003f185c82a658c510e36cba9acc4110eb60cbfded9de370ae9ea32')
# As spack GitLab CI pipelines use ^mpich %gcc@7.5.0, @1.9.6 with it can't be used:
# A conflict would be possible, but when %gcc@10: or another compiler is available,
# clingo can select the other compiler despite a request for %gcc@7.5.0 in place,
# at least when the version is requested: spack solve phist@1.9.6 ^mpich %gcc@7.5.0
# With conflicts('mpich', when='@1.9.6:%gcc@:9'), this is the result:
# spack solve phist@1.9.6 ^mpich %gcc@7.5.0|grep -e phist -e mpich|sed 's/+.*//'
# phist@1.9.6%gcc@11.2.0
# ^mpich@3.4.2%gcc@7.5.0~argobots
# A mismatch of gfortan between gcc@7.5.0(for mpich) and @10:(phist) would not work,
# and also does not solve the build problem.
# Instead, a check with a helpful error message is added to the build (see below).
# and we use preferred=True to select 1.9.5 by default:
version(
'1.9.5',
sha256='24faa3373003f185c82a658c510e36cba9acc4110eb60cbfded9de370ae9ea32',
preferred=True,
)
version('1.9.4', sha256='9dde3ca0480358fa0877ec8424aaee4011c5defc929219a5930388a7cdb4c8a6')
version('1.9.3', sha256='3ab7157e9f535a4c8537846cb11b516271ef13f82d0f8ebb7f96626fb9ab86cf')
version('1.9.2', sha256='289678fa7172708f5d32d6bd924c8fdfe72b413bba5bbb8ce6373c85c5ec5ae5')
@ -150,7 +179,24 @@ class Phist(CMakePackage):
# the phist repo came with it's own FindMPI.cmake before, which may cause some other
# MPI installation to be used than the one spack wants.
def patch(self):
if self.spec.satisfies('@1.9.6'):
filter_file('USE mpi', 'use mpi_f08',
'src/kernels/builtin/crsmat_module.F90')
# filter_file('use mpi', 'use mpi_f08', -> Needs more fixes
# 'fortran_bindings/phist_testing.F90')
# These are not needed for the build but as a reminder to be consistent:
filter_file('use mpi', 'use mpi_f08',
'fortran_bindings/test/core.F90')
filter_file('use mpi', 'use mpi_f08',
'fortran_bindings/test/jada.F90')
filter_file('use mpi', 'use mpi_f08',
'fortran_bindings/test/kernels.F90')
if '^mpich' in self.spec and self.spec.satisfies('%gcc@:9'):
raise InstallError("PR 26773: gcc<10 can't build phist>1.9.5 w/ ^mpich")
if self.spec.satisfies('@:1.9.5'):
# Tag '1.9.5' has moved to an older commit later without fixing the version:
filter_file('VERSION_PATCH 4', 'VERSION_PATCH 5', 'CMakeLists.txt')
force_remove('cmake/FindMPI.cmake')
# mpiexec -n12 puts a lot of stress on a pod and gets stuck in a loop very often
test = FileFilter('CMakeLists.txt')
@ -205,12 +251,29 @@ def cmake_args(self):
define('MPI_HOME', spec['mpi'].prefix),
])
# Workaround for non-compliant and possibly broken code needed for gcc@10:
# MPI_Allreduce(localRes, res, 1, MPI_LOGICAL, MPI_LAND, map1%comm, ierr);
# MPI_Allreduce(my_ierr,global_ierr,1,MPI_INTEGER,MPI_MAX,map1%comm,mpi_ierr)
# Error: Type mismatch between actual argument at (1)
# and actual argument at (2) (LOGICAL(4)/INTEGER(4)).
if spec.satisfies('%gcc@10:'):
args.append(define('CMAKE_Fortran_FLAGS', '-fallow-argument-mismatch'))
return args
# removing @run_after('build') fixes it from getting called twice
def check(self):
with working_dir(self.build_directory):
make("check")
# This affects all versions of phist with ^mpich with all gcc versions:
if '^mpich' in self.spec:
hint = 'Expect tests to timeout with mpich. Should work with: ^openmpi.'
tty.warn('========================== %s =======================' % hint)
try:
make("check")
except spack.util.executable.ProcessError:
raise InstallError('run-test of phist ^mpich: Hint: ' + hint)
else:
make("check")
@run_after('install')
@on_package_attributes(run_tests=True)
@ -220,3 +283,7 @@ def test_install(self):
join_path(self.stage.path, 'exampleProjects'))
with working_dir(self.build_directory):
make("test_install")
@property
def parallel(self):
return self.spec.satisfies('@1.8:')