blob: a4ac15369af926138aa6858a1457b6838431ad53 (
plain)
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
|
# Maintainer: Jakub Klinkovský <lahwaacz at archlinux dot org>
_pkgname=ginkgo
pkgbase=ginkgo-hpc-git
pkgname=(ginkgo-hpc{,-docs,-cuda,-hip}-git)
pkgver=r8256.abe9ecf7f9
pkgrel=1
pkgdesc="Numerical linear algebra software package"
arch=(x86_64)
url=https://ginkgo-project.github.io/
license=(BSD-3-Clause)
depends=(
gcc-libs
glibc
hwloc
openmpi
)
# FIXME: system 'rapidjson' does not work due to some C++ issues
# (see the 'declared protected here' error in https://github.com/Tencent/rapidjson/issues/1338#issuecomment-449849215 )
makedepends=(
git
cmake
ninja
nlohmann-json
gflags # for benchmarks (not installed yet)
gtest
numactl
doxygen
graphviz
texlive-bin
texlive-latexextra
# -cuda
cuda
# -hip
hip-runtime-amd
hipblas
hipfft
hiprand
hipsparse
rocthrust
roctracer
)
source=(
"git+https://github.com/ginkgo-project/$_pkgname.git"
split_cuda_library.patch
)
b2sums=('SKIP'
'3dd2558f399a4e9e7176021415b53e8f06e41c328df920ed254890ffc4c2a430344162c5054bb8d8f4ef2c53ae4029084fd85f89f46a047ebf5707a70fb6d7fe')
pkgver() {
cd $_pkgname
printf "r%s.%s" "$(git rev-list --count HEAD)" "$(git rev-parse --short HEAD)"
}
prepare() {
# Ginkgo is too big to build all CUDA archs into a single library, we need to split it
# https://github.com/ginkgo-project/ginkgo/issues/1734#issuecomment-2517784197
patch --directory=$_pkgname -Np1 < split_cuda_library.patch
}
build() {
local common_cmake_flags=(
-S $_pkgname -G Ninja
-DCMAKE_BUILD_TYPE=None
-DCMAKE_INSTALL_PREFIX=/usr
-DGINKGO_BUILD_REFERENCE=ON
-DGINKGO_BUILD_OMP=ON
-DGINKGO_BUILD_MPI=ON
-DGINKGO_HAVE_GPU_AWARE_MPI=ON
-DGINKGO_BUILD_BENCHMARKS=ON
-DGINKGO_BUILD_EXAMPLES=ON
-DGINKGO_BUILD_DOC=ON
-DGINKGO_BUILD_TESTS=ON
)
# In general, we want to list all real archs (sm_XX) and the latest virtual arch (compute_XX) for future PTX compatibility.
# Valid values can be discovered from nvcc --help
local _cuda_archs="50;52;53;60;61;62;70;72;75;80;86;87;89;90;90a;90a-virtual"
# archs gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102 are not supported: https://github.com/ginkgo-project/ginkgo/issues/1429
local _amdgpu_archs="gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942"
# base package
cmake -B build "${common_cmake_flags[@]}" \
-DGINKGO_BUILD_CUDA=OFF \
-DGINKGO_BUILD_HIP=OFF \
-DGINKGO_BUILD_SYCL=OFF
cmake --build build
# -cuda package
cmake -B build-cuda "${common_cmake_flags[@]}" \
-DCMAKE_CUDA_ARCHITECTURES="$_cuda_archs" \
-DGINKGO_BUILD_CUDA=ON \
-DGINKGO_BUILD_HIP=OFF \
-DGINKGO_BUILD_SYCL=OFF
cmake --build build-cuda
# -hip package
# LTO does not work with HIP
local _hip_flags="${CXXFLAGS/-flto=auto/}"
local _cxx_flags="${CXXFLAGS/-flto=auto/}"
# HIP does not support -fcf-protection
_hip_flags="${_hip_flags/-fcf-protection/}"
# Ginkgo does not support _GLIBCXX_ASSERTIONS for device builds https://github.com/ginkgo-project/ginkgo/issues/1143#issuecomment-2036957897
_hip_flags="${_hip_flags/-Wp,-D_GLIBCXX_ASSERTIONS/}"
cmake -B build-hip "${common_cmake_flags[@]}" \
-DCMAKE_CXX_COMPILER=/opt/rocm/lib/llvm/bin/amdclang++ \
-DCMAKE_CXX_FLAGS="$_cxx_flags" \
-DCMAKE_HIP_FLAGS="$_hip_flags" \
-DCMAKE_HIP_ARCHITECTURES="$_amdgpu_archs" \
-DGINKGO_BUILD_CUDA=OFF \
-DGINKGO_BUILD_HIP=ON \
-DGINKGO_BUILD_SYCL=OFF
cmake --build build-hip
}
check() {
# some tests fail due to capturing stderr and getting different number of mpirun warnings
# see https://github.com/ginkgo-project/ginkgo/issues/1567
local excluded_tests="benchmark_.*_distributed"
# limit parallel execution of tests to 4 threads and 8 processes
# note that without a GPU we cannot run tests for -cuda and -hip
(
export OMP_NUM_THREADS=4
export CTEST_PARALLEL_LEVEL=8
export CTEST_OUTPUT_ON_FAILURE=1
ctest --test-dir build --exclude-regex "$excluded_tests"
)
}
_package() {
DESTDIR="$pkgdir" cmake --install build"$1"
# install the license
install -vDm 644 $_pkgname/LICENSE -t "$pkgdir/usr/share/licenses/$pkgname/"
# TODO: install the benchmarks and examples: https://github.com/ginkgo-project/ginkgo/issues/1144
}
package_ginkgo-hpc-git() {
optdepends=(
'ginkgo-hpc-docs: for documentation'
)
conflicts=(ginkgo-hpc)
provides=(ginkgo-hpc)
_package ""
}
package_ginkgo-hpc-docs-git() {
pkgdesc+=" - documentation"
depends=()
conflicts=(ginkgo-hpc-docs)
provides=(ginkgo-hpc-docs)
# install the documentation
install -vdm755 "$pkgdir/usr/share/doc/$pkgbase"
cp -r build/doc/usr "$pkgdir/usr/share/doc/$pkgbase/html"
# install the license
install -vDm 644 $_pkgname/LICENSE -t "$pkgdir/usr/share/licenses/$pkgname/"
}
package_ginkgo-hpc-cuda-git() {
pkgdesc+=" (with CUDA)"
depends+=(cuda)
optdepends=(
'ginkgo-hpc-docs: for documentation'
)
conflicts=(ginkgo-hpc)
provides=(ginkgo-hpc)
_package -cuda
}
package_ginkgo-hpc-hip-git() {
pkgdesc+=" (with ROCm/HIP)"
depends+=(
hip-runtime-amd
hipblas
hipfft
hiprand
hipsparse
roctracer
)
optdepends=(
'ginkgo-hpc-docs: for documentation'
)
conflicts=(ginkgo-hpc)
provides=(ginkgo-hpc)
_package -hip
}
# vim:set ts=2 sw=2 et:
|