1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
|
# Copyright 2013-2021 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
from spack import *
class Dihydrogen(CMakePackage, CudaPackage, ROCmPackage):
"""DiHydrogen is the second version of the Hydrogen fork of the
well-known distributed linear algebra library,
Elemental. DiHydrogen aims to be a basic distributed
multilinear algebra interface with a particular emphasis on the
needs of the distributed machine learning effort, LBANN."""
homepage = "https://github.com/LLNL/DiHydrogen.git"
url = "https://github.com/LLNL/DiHydrogen/archive/v0.1.tar.gz"
git = "https://github.com/LLNL/DiHydrogen.git"
maintainers = ['bvanessen']
version('develop', branch='develop')
version('master', branch='master')
version('0.2.1', sha256='11e2c0f8a94ffa22e816deff0357dde6f82cc8eac21b587c800a346afb5c49ac')
version('0.2.0', sha256='e1f597e80f93cf49a0cb2dbc079a1f348641178c49558b28438963bd4a0bdaa4')
version('0.1', sha256='171d4b8adda1e501c38177ec966e6f11f8980bf71345e5f6d87d0a988fef4c4e')
variant('al', default=True,
description='Builds with Aluminum communication library')
variant('developer', default=False,
description='Enable extra warnings and force tests to be enabled.')
variant('half', default=False,
description='Enable FP16 support on the CPU.')
variant('distconv', default=False,
description='Support distributed convolutions: spatial, channel, '
'filter.')
variant('nvshmem', default=False,
description='Builds with support for NVSHMEM')
variant('openmp', default=False,
description='Enable CPU acceleration with OpenMP threads.')
variant('rocm', default=False,
description='Enable ROCm/HIP language features.')
variant('shared', default=True,
description='Enables the build of shared libraries')
variant('docs', default=False,
description='Builds with support for building documentation')
# Variants related to BLAS
variant('openmp_blas', default=False,
description='Use OpenMP for threading in the BLAS library')
variant('int64_blas', default=False,
description='Use 64bit integers for BLAS.')
variant('blas', default='openblas', values=('openblas', 'mkl', 'accelerate', 'essl'),
description='Enable the use of OpenBlas/MKL/Accelerate/ESSL')
conflicts('~cuda', when='+nvshmem')
depends_on('mpi')
depends_on('catch2', type='test')
# Specify the correct version of Aluminum
depends_on('aluminum@0.4:0.4.99', when='@0.1:0.1.99 +al')
depends_on('aluminum@0.5.0:0.5.99', when='@0.2.0 +al')
depends_on('aluminum@0.7.0:0.7.99', when='@0.2.1 +al')
depends_on('aluminum@0.7.0:', when='@:0.0,0.2.1: +al')
# Add Aluminum variants
depends_on('aluminum +cuda +nccl +ht +cuda_rma', when='+al +cuda')
depends_on('aluminum +rocm +rccl +ht', when='+al +rocm')
for arch in CudaPackage.cuda_arch_values:
depends_on('aluminum cuda_arch=%s' % arch, when='+al +cuda cuda_arch=%s' % arch)
# variants +rocm and amdgpu_targets are not automatically passed to
# dependencies, so do it manually.
for val in ROCmPackage.amdgpu_targets:
depends_on('aluminum amdgpu_target=%s' % val, when='amdgpu_target=%s' % val)
for when in ['+cuda', '+distconv']:
depends_on('cuda', when=when)
depends_on('cudnn', when=when)
depends_on('cub', when='^cuda@:10.99')
# Note that #1712 forces us to enumerate the different blas variants
depends_on('openblas', when='blas=openblas')
depends_on('openblas +ilp64', when='blas=openblas +int64_blas')
depends_on('openblas threads=openmp', when='blas=openblas +openmp_blas')
depends_on('intel-mkl', when="blas=mkl")
depends_on('intel-mkl +ilp64', when="blas=mkl +int64_blas")
depends_on('intel-mkl threads=openmp', when='blas=mkl +openmp_blas')
depends_on('veclibfort', when='blas=accelerate')
conflicts('blas=accelerate +openmp_blas')
depends_on('essl', when='blas=essl')
depends_on('essl +ilp64', when='blas=essl +int64_blas')
depends_on('essl threads=openmp', when='blas=essl +openmp_blas')
depends_on('netlib-lapack +external-blas', when='blas=essl')
# Distconv builds require cuda
conflicts('~cuda', when='+distconv')
conflicts('+distconv', when='+half')
conflicts('+rocm', when='+half')
depends_on('half', when='+half')
generator = 'Ninja'
depends_on('ninja', type='build')
depends_on('cmake@3.17.0:', type='build')
depends_on('py-breathe', type='build', when='+docs')
depends_on('doxygen', type='build', when='+docs')
depends_on('llvm-openmp', when='%apple-clang +openmp')
depends_on('nvshmem', when='+nvshmem')
# Idenfity versions of cuda_arch that are too old
# from lib/spack/spack/build_systems/cuda.py
illegal_cuda_arch_values = [
'10', '11', '12', '13',
'20', '21',
]
for value in illegal_cuda_arch_values:
conflicts('cuda_arch=' + value)
@property
def libs(self):
shared = True if '+shared' in self.spec else False
return find_libraries(
'libH2Core', root=self.prefix, shared=shared, recursive=True
)
def cmake_args(self):
spec = self.spec
args = [
'-DCMAKE_CXX_STANDARD=17',
'-DCMAKE_INSTALL_MESSAGE:STRING=LAZY',
'-DBUILD_SHARED_LIBS:BOOL=%s' % ('+shared' in spec),
'-DH2_ENABLE_ALUMINUM=%s' % ('+al' in spec),
'-DH2_ENABLE_CUDA=%s' % ('+cuda' in spec),
'-DH2_ENABLE_DISTCONV_LEGACY=%s' % ('+distconv' in spec),
'-DH2_ENABLE_OPENMP=%s' % ('+openmp' in spec),
'-DH2_ENABLE_FP16=%s' % ('+half' in spec),
'-DH2_ENABLE_HIP_ROCM=%s' % ('+rocm' in spec),
'-DH2_DEVELOPER_BUILD=%s' % ('+developer' in spec),
]
if '+cuda' in spec:
if spec.satisfies('^cuda@11.0:'):
args.append('-DCMAKE_CUDA_STANDARD=17')
else:
args.append('-DCMAKE_CUDA_STANDARD=14')
archs = spec.variants['cuda_arch'].value
if archs != 'none':
arch_str = ",".join(archs)
args.append('-DCMAKE_CUDA_ARCHITECTURES=%s' % arch_str)
if '+cuda' in spec or '+distconv' in spec:
args.append('-DcuDNN_DIR={0}'.format(
spec['cudnn'].prefix))
if spec.satisfies('^cuda@:10.99'):
if '+cuda' in spec or '+distconv' in spec:
args.append('-DCUB_DIR={0}'.format(
spec['cub'].prefix))
# Add support for OpenMP with external (Brew) clang
if spec.satisfies('%clang +openmp platform=darwin'):
clang = self.compiler.cc
clang_bin = os.path.dirname(clang)
clang_root = os.path.dirname(clang_bin)
args.extend([
'-DOpenMP_CXX_FLAGS=-fopenmp=libomp',
'-DOpenMP_CXX_LIB_NAMES=libomp',
'-DOpenMP_libomp_LIBRARY={0}/lib/libomp.dylib'.format(
clang_root)])
if '+rocm' in spec:
args.extend([
'-DHIP_ROOT_DIR={0}'.format(spec['hip'].prefix),
'-DHIP_CXX_COMPILER={0}'.format(self.spec['hip'].hipcc)])
archs = self.spec.variants['amdgpu_target'].value
if archs != 'none':
arch_str = ",".join(archs)
args.append(
'-DHIP_HIPCC_FLAGS=--amdgpu-target={0}'
' -g -fsized-deallocation -fPIC'.format(arch_str)
)
return args
def setup_build_environment(self, env):
if self.spec.satisfies('%apple-clang +openmp'):
env.append_flags(
'CPPFLAGS', self.compiler.openmp_flag)
env.append_flags(
'CFLAGS', self.spec['llvm-openmp'].headers.include_flags)
env.append_flags(
'CXXFLAGS', self.spec['llvm-openmp'].headers.include_flags)
env.append_flags(
'LDFLAGS', self.spec['llvm-openmp'].libs.ld_flags)
|
