c214e4dd2d
* PGI compiler has trouble with avx2 SIMD support (https://github.com/FFTW/fftw3/issues/78) * Hew to the project's preferred indentation standard. * Expand '%nvhpc' logic to include '%pgi'. * Exceeded the max line-length. * Break up the long compound statement into nested if's. * Inadvertently picked up an extraneous file. * PGI compiler has trouble with avx2/avx-512 SIMD support, too.
232 lines
8.9 KiB
Python
232 lines
8.9 KiB
Python
# Copyright 2013-2020 Lawrence Livermore National Security, LLC and other
|
|
# Spack Project Developers. See the top-level COPYRIGHT file for details.
|
|
#
|
|
# SPDX-License-Identifier: (Apache-2.0 OR MIT)
|
|
|
|
import os
|
|
import os.path
|
|
import llnl.util.lang
|
|
from spack import *
|
|
|
|
|
|
class FftwBase(AutotoolsPackage):
|
|
"""Base class for building Fftw, shared with the AMD optimized version
|
|
of the library in the 'amdfftw' package.
|
|
"""
|
|
|
|
variant(
|
|
'precision', values=any_combination_of(
|
|
'float', 'double', 'long_double', 'quad'
|
|
).prohibit_empty_set().with_default('float,double'),
|
|
description='Build the selected floating-point precision libraries'
|
|
)
|
|
variant('openmp', default=False, description="Enable OpenMP support.")
|
|
variant('mpi', default=True, description='Activate MPI support')
|
|
|
|
depends_on('mpi', when='+mpi')
|
|
depends_on('llvm-openmp', when='%apple-clang +openmp')
|
|
|
|
# https://github.com/FFTW/fftw3/commit/902d0982522cdf6f0acd60f01f59203824e8e6f3
|
|
conflicts('%gcc@8:8.9999', when="@3.3.7")
|
|
conflicts('precision=long_double', when='@2.1.5',
|
|
msg='Long double precision is not supported in FFTW 2')
|
|
conflicts('precision=quad', when='@2.1.5',
|
|
msg='Quad precision is not supported in FFTW 2')
|
|
|
|
@property
|
|
def libs(self):
|
|
|
|
# Reduce repetitions of entries
|
|
query_parameters = list(llnl.util.lang.dedupe(
|
|
self.spec.last_query.extra_parameters
|
|
))
|
|
|
|
# List of all the suffixes associated with float precisions
|
|
precisions = [
|
|
('float', 'f'),
|
|
('double', ''),
|
|
('long_double', 'l'),
|
|
('quad', 'q')
|
|
]
|
|
|
|
# Retrieve the correct suffixes, or use double as a default
|
|
suffixes = [v for k, v in precisions if k in query_parameters] or ['']
|
|
|
|
# Construct the list of libraries that needs to be found
|
|
libraries = []
|
|
for sfx in suffixes:
|
|
if 'mpi' in query_parameters and '+mpi' in self.spec:
|
|
libraries.append('libfftw3' + sfx + '_mpi')
|
|
|
|
if 'openmp' in query_parameters and '+openmp' in self.spec:
|
|
libraries.append('libfftw3' + sfx + '_omp')
|
|
|
|
libraries.append('libfftw3' + sfx)
|
|
|
|
return find_libraries(libraries, root=self.prefix, recursive=True)
|
|
|
|
def patch(self):
|
|
# If fftw/config.h exists in the source tree, it will take precedence
|
|
# over the copy in build dir. As only the latter has proper config
|
|
# for our build, this is a problem. See e.g. issue #7372 on github
|
|
if os.path.isfile('fftw/config.h'):
|
|
os.rename('fftw/config.h', 'fftw/config.h.SPACK_RENAMED')
|
|
|
|
def autoreconf(self, spec, prefix):
|
|
if '+pfft_patches' in spec:
|
|
autoreconf = which('autoreconf')
|
|
autoreconf('-ifv')
|
|
|
|
@property
|
|
def selected_precisions(self):
|
|
"""Precisions that have been selected in this build"""
|
|
return self.spec.variants['precision'].value
|
|
|
|
def configure(self, spec, prefix):
|
|
# Base options
|
|
options = [
|
|
'--prefix={0}'.format(prefix),
|
|
'--enable-shared',
|
|
'--enable-threads'
|
|
]
|
|
if not self.compiler.f77 or not self.compiler.fc:
|
|
options.append("--disable-fortran")
|
|
if spec.satisfies('@:2'):
|
|
options.append('--enable-type-prefix')
|
|
|
|
# Variants that affect every precision
|
|
if '+openmp' in spec:
|
|
options.append('--enable-openmp')
|
|
if spec.satisfies('@:2'):
|
|
# TODO: libtool strips CFLAGS, so 2.x libxfftw_threads
|
|
# isn't linked to the openmp library. Patch Makefile?
|
|
options.insert(0, 'CFLAGS=' + self.compiler.openmp_flag)
|
|
if '+mpi' in spec:
|
|
options.append('--enable-mpi')
|
|
|
|
# Specific SIMD support.
|
|
# all precisions
|
|
simd_features = ['sse2', 'avx', 'avx2', 'avx512', 'avx-128-fma',
|
|
'kcvi', 'vsx', 'neon']
|
|
# float only
|
|
float_simd_features = ['altivec', 'sse']
|
|
|
|
# Workaround PGI compiler bug when avx2 is enabled
|
|
if spec.satisfies('%pgi') and 'avx2' in simd_features:
|
|
simd_features.remove('avx2')
|
|
|
|
# Workaround NVIDIA/PGI compiler bug when avx512 is enabled
|
|
if spec.satisfies('%nvhpc') or spec.satisfies('%pgi'):
|
|
if 'avx512' in simd_features:
|
|
simd_features.remove('avx512')
|
|
|
|
# NVIDIA compiler does not support Altivec intrinsics
|
|
if spec.satisfies('%nvhpc') and 'vsx' in simd_features:
|
|
simd_features.remove('vsx')
|
|
if spec.satisfies('%nvhpc') and 'altivec' in float_simd_features:
|
|
float_simd_features.remove('altivec')
|
|
|
|
# NVIDIA compiler does not support Neon intrinsics
|
|
if spec.satisfies('%nvhpc') and 'neon' in simd_features:
|
|
simd_features.remove('neon')
|
|
|
|
simd_options = []
|
|
for feature in simd_features:
|
|
msg = '--enable-{0}' if feature in spec.target else '--disable-{0}'
|
|
simd_options.append(msg.format(feature))
|
|
|
|
# If no features are found, enable the generic ones
|
|
if not any(f in spec.target for f in
|
|
simd_features + float_simd_features):
|
|
# Workaround NVIDIA compiler bug
|
|
if not spec.satisfies('%nvhpc'):
|
|
simd_options += [
|
|
'--enable-generic-simd128',
|
|
'--enable-generic-simd256'
|
|
]
|
|
|
|
simd_options += [
|
|
'--enable-fma' if 'fma' in spec.target else '--disable-fma'
|
|
]
|
|
|
|
# Double is the default precision, for all the others we need
|
|
# to enable the corresponding option.
|
|
enable_precision = {
|
|
'float': ['--enable-float'],
|
|
'double': None,
|
|
'long_double': ['--enable-long-double'],
|
|
'quad': ['--enable-quad-precision']
|
|
}
|
|
|
|
# Different precisions must be configured and compiled one at a time
|
|
configure = Executable('../configure')
|
|
for precision in self.selected_precisions:
|
|
opts = (enable_precision[precision] or []) + options[:]
|
|
|
|
# SIMD optimizations are available only for float and double
|
|
# starting from FFTW 3
|
|
if precision in ('float', 'double') and spec.satisfies('@3:'):
|
|
opts += simd_options
|
|
|
|
# float-only acceleration
|
|
if precision == 'float':
|
|
for feature in float_simd_features:
|
|
if feature in spec.target:
|
|
msg = '--enable-{0}'
|
|
else:
|
|
msg = '--disable-{0}'
|
|
opts.append(msg.format(feature))
|
|
|
|
with working_dir(precision, create=True):
|
|
configure(*opts)
|
|
|
|
def for_each_precision_make(self, *targets):
|
|
for precision in self.selected_precisions:
|
|
with working_dir(precision):
|
|
make(*targets)
|
|
|
|
def build(self, spec, prefix):
|
|
self.for_each_precision_make()
|
|
|
|
def check(self):
|
|
self.for_each_precision_make('check')
|
|
|
|
def install(self, spec, prefix):
|
|
self.for_each_precision_make('install')
|
|
|
|
|
|
class Fftw(FftwBase):
|
|
"""FFTW is a C subroutine library for computing the discrete Fourier
|
|
transform (DFT) in one or more dimensions, of arbitrary input
|
|
size, and of both real and complex data (as well as of even/odd
|
|
data, i.e. the discrete cosine/sine transforms or DCT/DST). We
|
|
believe that FFTW, which is free software, should become the FFT
|
|
library of choice for most applications."""
|
|
|
|
homepage = "http://www.fftw.org"
|
|
url = "http://www.fftw.org/fftw-3.3.4.tar.gz"
|
|
list_url = "http://www.fftw.org/download.html"
|
|
|
|
version('3.3.8', sha256='6113262f6e92c5bd474f2875fa1b01054c4ad5040f6b0da7c03c98821d9ae303')
|
|
version('3.3.7', sha256='3b609b7feba5230e8f6dd8d245ddbefac324c5a6ae4186947670d9ac2cd25573')
|
|
version('3.3.6-pl2', sha256='a5de35c5c824a78a058ca54278c706cdf3d4abba1c56b63531c2cb05f5d57da2')
|
|
version('3.3.5', sha256='8ecfe1b04732ec3f5b7d279fdb8efcad536d555f9d1e8fabd027037d45ea8bcf')
|
|
version('3.3.4', sha256='8f0cde90929bc05587c3368d2f15cd0530a60b8a9912a8e2979a72dbe5af0982')
|
|
version('2.1.5', sha256='f8057fae1c7df8b99116783ef3e94a6a44518d49c72e2e630c24b689c6022630')
|
|
|
|
variant(
|
|
'pfft_patches', default=False,
|
|
description='Add extra transpose functions for PFFT compatibility')
|
|
|
|
depends_on('automake', type='build', when='+pfft_patches')
|
|
depends_on('autoconf', type='build', when='+pfft_patches')
|
|
depends_on('libtool', type='build', when='+pfft_patches')
|
|
|
|
provides('fftw-api@2', when='@2.1.5')
|
|
provides('fftw-api@3', when='@3:')
|
|
|
|
patch('pfft-3.3.5.patch', when="@3.3.5:+pfft_patches", level=0)
|
|
patch('pfft-3.3.4.patch', when="@3.3.4+pfft_patches", level=0)
|
|
patch('pgi-3.3.6-pl2.patch', when="@3.3.6-pl2%pgi", level=0)
|
|
patch('intel-configure.patch', when="@3:3.3.8%intel", level=0)
|