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
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
|
# Copyright 2013-2024 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.package import *
# This limits the versions of lots of things pretty severely.
#
# - Only v1.5.2 and newer are buildable.
# - CMake must be v3.22 or newer.
# - CUDA must be v11.0.0 or newer.
class Hydrogen(CachedCMakePackage, CudaPackage, ROCmPackage):
"""Hydrogen: Distributed-memory dense and sparse-direct linear algebra
and optimization library. Based on the Elemental library."""
homepage = "https://libelemental.org"
url = "https://github.com/LLNL/Elemental/archive/v1.5.1.tar.gz"
git = "https://github.com/LLNL/Elemental.git"
tags = ["ecp", "radiuss"]
maintainers("bvanessen")
license("GPL-2.0-or-later")
version("develop", branch="hydrogen")
version("1.5.3", sha256="faefbe738bd364d0e26ce9ad079a11c93a18c6f075719a365fd4fa5f1f7a989a")
version("1.5.2", sha256="a902cad3962471216cfa278ba0561c18751d415cd4d6b2417c02a43b0ab2ea33")
version("1.5.1", sha256="447da564278f98366906d561d9c8bc4d31678c56d761679c2ff3e59ee7a2895c")
# Older versions are no longer supported.
variant("shared", default=True, description="Enables the build of shared libraries.")
variant(
"build_type",
default="Release",
description="The build type to build",
values=("Debug", "Release"),
)
variant("int64", default=False, description="Use 64-bit integers")
variant("al", default=True, sticky=True, description="Use Aluminum communication library")
variant(
"cub", default=True, when="+cuda", description="Use CUB/hipCUB for GPU memory management"
)
variant(
"cub", default=True, when="+rocm", description="Use CUB/hipCUB for GPU memory management"
)
variant("half", default=False, description="Support for FP16 precision data types")
# TODO: Add netlib-lapack. For GPU-enabled builds, typical
# workflows don't touch host BLAS/LAPACK all that often, and even
# less frequently in performance-critical regions.
variant(
"blas",
default="any",
values=("any", "openblas", "mkl", "accelerate", "essl", "libsci"),
description="Specify a host BLAS library preference",
)
variant("int64_blas", default=False, description="Use 64-bit integers for (host) BLAS.")
variant("openmp", default=True, description="Make use of OpenMP within CPU kernels")
variant(
"omp_taskloops",
when="+openmp",
default=False,
description="Use OpenMP taskloops instead of parallel for loops",
)
# Users should spec this on their own on the command line, no?
# This doesn't affect Hydrogen itself at all. Not one bit.
# variant(
# "openmp_blas",
# default=False,
# description="Use OpenMP for threading in the BLAS library")
variant("test", default=False, description="Builds test suite")
conflicts("+cuda", when="+rocm", msg="CUDA and ROCm support are mutually exclusive")
conflicts("+half", when="+rocm", msg="FP16 support not implemented for ROCm.")
depends_on("cmake@3.22.0:", type="build", when="@1.5.2:")
depends_on("cmake@3.17.0:", type="build", when="@1.5.1")
depends_on("mpi")
depends_on("blas")
depends_on("lapack")
# Note that #1712 forces us to enumerate the different blas variants
# Note that this forces us to use OpenBLAS until #1712 is fixed
depends_on("openblas", when="blas=openblas")
depends_on("openblas +ilp64", when="blas=openblas +int64_blas")
depends_on("openblas@0.3.21:0.3.23", when="blas=openblas arch=ppc64le:")
depends_on("intel-mkl", when="blas=mkl")
depends_on("intel-mkl +ilp64", when="blas=mkl +int64_blas")
# I don't think this is true...
depends_on("veclibfort", when="blas=accelerate")
depends_on("essl", when="blas=essl")
depends_on("essl +ilp64", when="blas=essl +int64_blas")
depends_on("netlib-lapack +external-blas", when="blas=essl")
depends_on("cray-libsci", when="blas=libsci")
# Specify the correct version of Aluminum
depends_on("aluminum@0.7.0:", when="@1.5.2: +al")
# Add Aluminum variants
depends_on("aluminum +cuda +ht", when="+al +cuda")
depends_on("aluminum +rocm +ht", when="+al +rocm")
for arch in CudaPackage.cuda_arch_values:
depends_on("aluminum +cuda 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 +rocm amdgpu_target=%s" % val, when="+al +rocm amdgpu_target=%s" % val
)
depends_on("cuda@11.0.0:", when="+cuda")
depends_on("hipcub +rocm", when="+rocm +cub")
depends_on("half", when="+half")
depends_on("llvm-openmp", when="%apple-clang +openmp")
@property
def libs(self):
shared = True if "+shared" in self.spec else False
return find_libraries("libHydrogen", root=self.prefix, shared=shared, recursive=True)
def cmake_args(self):
args = []
return args
def get_cuda_flags(self):
spec = self.spec
args = []
if spec.satisfies("^cuda+allow-unsupported-compilers"):
args.append("-allow-unsupported-compiler")
if spec.satisfies("%clang"):
for flag in spec.compiler_flags["cxxflags"]:
if "gcc-toolchain" in flag:
args.append("-Xcompiler={0}".format(flag))
return args
def std_initconfig_entries(self):
entries = super(Hydrogen, self).std_initconfig_entries()
# CMAKE_PREFIX_PATH, in CMake types, is a "STRING", not a "PATH". :/
entries = [x for x in entries if "CMAKE_PREFIX_PATH" not in x]
cmake_prefix_path = os.environ["CMAKE_PREFIX_PATH"].replace(":", ";")
entries.append(cmake_cache_string("CMAKE_PREFIX_PATH", cmake_prefix_path))
# IDK why this is here, but it was in the original recipe. So, yeah.
entries.append(cmake_cache_string("CMAKE_INSTALL_MESSAGE", "LAZY"))
return entries
def initconfig_compiler_entries(self):
spec = self.spec
entries = super(Hydrogen, self).initconfig_compiler_entries()
# FIXME: Enforce this better in the actual CMake.
entries.append(cmake_cache_string("CMAKE_CXX_STANDARD", "17"))
entries.append(cmake_cache_option("BUILD_SHARED_LIBS", "+shared" in spec))
entries.append(cmake_cache_option("CMAKE_EXPORT_COMPILE_COMMANDS", True))
entries.append(cmake_cache_option("MPI_ASSUME_NO_BUILTIN_MPI", True))
if spec.satisfies("%clang +openmp platform=darwin") or spec.satisfies(
"%clang +omp_taskloops platform=darwin"
):
clang = self.compiler.cc
clang_bin = os.path.dirname(clang)
clang_root = os.path.dirname(clang_bin)
entries.append(cmake_cache_string("OpenMP_CXX_FLAGS", "-fopenmp=libomp"))
entries.append(cmake_cache_string("OpenMP_CXX_LIB_NAMES", "libomp"))
entries.append(
cmake_cache_string(
"OpenMP_libomp_LIBRARY", "{0}/lib/libomp.dylib".format(clang_root)
)
)
return entries
def initconfig_hardware_entries(self):
spec = self.spec
entries = super(Hydrogen, self).initconfig_hardware_entries()
entries.append(cmake_cache_option("Hydrogen_ENABLE_CUDA", "+cuda" in spec))
if spec.satisfies("+cuda"):
entries.append(cmake_cache_string("CMAKE_CUDA_STANDARD", "17"))
if not spec.satisfies("cuda_arch=none"):
archs = spec.variants["cuda_arch"].value
arch_str = ";".join(archs)
entries.append(cmake_cache_string("CMAKE_CUDA_ARCHITECTURES", arch_str))
# FIXME: Should this use the "cuda_flags" function of the
# CudaPackage class or something? There might be other
# flags in play, and we need to be sure to get them all.
cuda_flags = self.get_cuda_flags()
if len(cuda_flags) > 0:
entries.append(cmake_cache_string("CMAKE_CUDA_FLAGS", " ".join(cuda_flags)))
entries.append(cmake_cache_option("Hydrogen_ENABLE_ROCM", "+rocm" in spec))
if spec.satisfies("+rocm"):
entries.append(cmake_cache_string("CMAKE_HIP_STANDARD", "17"))
if not spec.satisfies("amdgpu_target=none"):
archs = self.spec.variants["amdgpu_target"].value
arch_str = ";".join(archs)
entries.append(cmake_cache_string("CMAKE_HIP_ARCHITECTURES", arch_str))
entries.append(cmake_cache_string("AMDGPU_TARGETS", arch_str))
entries.append(cmake_cache_string("GPU_TARGETS", arch_str))
entries.append(cmake_cache_path("HIP_ROOT_DIR", spec["hip"].prefix))
return entries
def initconfig_package_entries(self):
spec = self.spec
entries = super(Hydrogen, self).initconfig_package_entries()
# Basic Hydrogen options
entries.append(cmake_cache_option("Hydrogen_ENABLE_TESTING", "+test" in spec))
entries.append(cmake_cache_option("Hydrogen_GENERAL_LAPACK_FALLBACK", True))
entries.append(cmake_cache_option("Hydrogen_USE_64BIT_INTS", "+int64" in spec))
entries.append(cmake_cache_option("Hydrogen_USE_64BIT_BLAS_INTS", "+int64_blas" in spec))
# Advanced dependency options
entries.append(cmake_cache_option("Hydrogen_ENABLE_ALUMINUM", "+al" in spec))
entries.append(cmake_cache_option("Hydrogen_ENABLE_CUB", "+cub" in spec))
entries.append(cmake_cache_option("Hydrogen_ENABLE_GPU_FP16", "+cuda +half" in spec))
entries.append(cmake_cache_option("Hydrogen_ENABLE_HALF", "+half" in spec))
entries.append(cmake_cache_option("Hydrogen_ENABLE_OPENMP", "+openmp" in spec))
entries.append(
cmake_cache_option("Hydrogen_ENABLE_OMP_TASKLOOP", "+omp_taskloops" in spec)
)
# Note that CUDA/ROCm are handled above.
if "blas=openblas" in spec:
entries.append(cmake_cache_option("Hydrogen_USE_OpenBLAS", "blas=openblas" in spec))
# CMAKE_PREFIX_PATH should handle this
entries.append(cmake_cache_string("OpenBLAS_DIR", spec["openblas"].prefix))
elif "blas=mkl" in spec or spec.satisfies("^intel-mkl"):
entries.append(cmake_cache_option("Hydrogen_USE_MKL", True))
elif "blas=essl" in spec or spec.satisfies("^essl"):
entries.append(cmake_cache_string("BLA_VENDOR", "IBMESSL"))
# IF IBM ESSL is used it needs help finding the proper LAPACK libraries
entries.append(
cmake_cache_string(
"LAPACK_LIBRARIES",
"%s;-llapack;-lblas"
% ";".join("-l{0}".format(lib) for lib in self.spec["essl"].libs.names),
)
)
entries.append(
cmake_cache_string(
"BLAS_LIBRARIES",
"%s;-lblas"
% ";".join("-l{0}".format(lib) for lib in self.spec["essl"].libs.names),
)
)
elif "blas=accelerate" in spec:
entries.append(cmake_cache_option("Hydrogen_USE_ACCELERATE", True))
elif spec.satisfies("^netlib-lapack"):
entries.append(cmake_cache_string("BLA_VENDOR", "Generic"))
return entries
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)
|
