Repository navigation
Expand file tree
/
Copy pathfobos
More file actions
147 lines (123 loc) · 5.95 KB
/
Copy pathfobos
File metadata and controls
147 lines (123 loc) · 5.95 KB
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
[project]
name = ADAM
version = VERSION
summary = Accelerated fluid Dynamics on Adaptive Mesh refinement grids
repository = https://github.com/szaghi/adam
website = https://szaghi.github.io/adam
email = stefano.zaghi@gmail.it
authors = "Andrea Di Mascio" "Federico Negro" "Giacomo Rossi" "Francesco Salvadore" "Stefano Zaghi"
[dependencies]
deps_dir = src/third_party
BeFoR64 = https://github.com/szaghi/BeFoR64
FACE = https://github.com/szaghi/FACE
FiNeR = https://github.com/szaghi/FiNeR
FUNDAL = https://github.com/szaghi/FUNDAL
FOSSIL = https://github.com/szaghi/FOSSIL
FoXy = https://github.com/szaghi/FoXy
MORTIF = https://github.com/szaghi/MORTIF
MOTIOn = https://github.com/szaghi/MOTIOn
PENF = https://github.com/szaghi/PENF
StringiFor = https://github.com/szaghi/StringiFor
VTKFortran = https://github.com/szaghi/VTKFortran
VecFor = https://github.com/szaghi/VecFor
[include]
paths = fobos.d/templates.fobos
fobos.d/rules.fobos
fobos.d/prism.fobos
?fobos.d/adam.fobos
?fobos.d/ascot.fobos
?fobos.d/chase.fobos
?fobos.d/flume.fobos
?fobos.d/ripple.fobos
?fobos.d/patch.fobos
?fobos.d/nasto.fobos
?fobos.d/tests_fdv.fobos
?fobos.d/tests_amr.fobos
?fobos.d/tests_poisson.fobos
?fobos.d/tests_rk.fobos
?fobos.d/tests_turbulence.fobos
# user-tunable variables (FoBiS fobos variables)
[common-variables]
# AMD GPU target architecture for amdflang OpenMP offload (FNL/FUNDAL backend
# built with DEV_OMP). Set to match your GPU, e.g. gfx90a (MI200), gfx942 (MI300),
# gfx1100 (RDNA3). This is the default; override it from the CLI by selecting one
# of the [varset:gfx*] sections below, e.g.
# fobis build --mode nasto-fnl-omp-amd --varset "local_amd gfx1100"
$GPU_ARCH = gfx942
# Device link-time optimisation of the optimized NVF OpenACC modes (`-gpu=$NVF_CC$NVF_LTO`): `,lto` inlines the device
# routines across modules at the device link (profile-couette3d README, "Results": the FLUME FNL loop on an A30 -35%
# uniform, -27% refined walls). Off by default; a varset turns it on once its nvfortran is known to accept it (26.9:
# local_nvf, iac_nvf); the overlay varset `nolto` turns it off again, e.g. --varset "iac_nvf nolto". The NVF templates
# keep their compile flags (cflags_heritage), so switching it recompiles everything instead of mixing objects.
$NVF_LTO =
# Parallel device LTO link (a link flag, ignored without ,lto), issue #131:
# `-Wnvlink,--split-compile=N` runs nvvm on N threads, `-Wnvlink,--split-compile-extended=N` splits at the linker.
# - CUDA 13.3 (local_nvf): --split-compile=16 links flume-fnl-nvf in 157 s instead of 491 s.
# - CUDA 12.9 (iac_nvf): --split-compile=4 or 16 never reaches its threads (one thread at 100 % for 25-50 min, killed);
# --split-compile-extended=16 links in 531 s instead of 1001 s, but the A30 runs 7-10 % slower (profile-couette3d
# medium, 1 rank: loop 18.0 -> 19.3 s refined walls, 33.4 -> 36.8 s uniform), so iac_nvf links on one thread and
# the overlay varset `fastlink` (--varset "iac_nvf fastlink") takes the faster link for development builds.
# Off by default; a varset sets it once its CUDA is measured. local_nvf (WSL, correctness only) trades the run time
# of the split link, not measured there, for the build time.
$NVF_SPLIT =
[varsets]
default = local_gnu
[varset:local_gnu]
$HDF5_PREFIX = lib/hdf5/develop/gnu/16.2.0
# GitHub Actions CI: Ubuntu 24.04's apt gfortran-14 (.github/actions/setup-build-env) cannot read the .mod files of the
# gfortran 16.2 HDF5 above, so CI builds against the gfortran 14 HDF5 kept in the repo (ci.yml passes --varset ci_gnu).
[varset:ci_gnu]
$HDF5_PREFIX = lib/hdf5/develop/gnu/14.2.0
[varset:local_nvf]
$HDF5_PREFIX = lib/hdf5/develop/nvf/26.9
$NVF_CC = cc89
$NVF_LTO = ,lto
$NVF_SPLIT = -Wnvlink,--split-compile=16
[varset:local_amd]
$HDF5_PREFIX = lib/hdf5/develop/amd/latest
[varset:thera_amd2]
# thera, system module tree (rocm/10.1.0-afar-24.3.0, openmpi/5.0.10-ucx1.22.0-ucc1.9.0, hdf5/2.2.0): HDF5 2.2.0 installs
# its libraries in lib64 and this fobos file reads $HDF5_PREFIX/lib, so the prefix is a directory of two symlinks
# (include -> .../include, lib -> .../lib64). LOCAL, not for commit.
$HDF5_PREFIX = /home/giarossi/hdf5-2.2.0-shim
[varset:thera_amd]
$HDF5_PREFIX = /mnt/thera/data/incoming/mbycklin/spack/install/linux-zen4/hdf5-1.14.6-zwo77eq6dyjllcapikwnnc7v4lrjknik
# CLI-selectable AMD GPU architecture overrides (set only $GPU_ARCH, so combine
# with a varset that provides $HDF5_PREFIX, e.g. --varset "local_amd gfx90a").
[varset:gfx906]
$GPU_ARCH = gfx906
[varset:gfx908]
$GPU_ARCH = gfx908
[varset:gfx90a]
$GPU_ARCH = gfx90a
[varset:gfx940]
$GPU_ARCH = gfx940
[varset:gfx942]
$GPU_ARCH = gfx942
[varset:gfx950]
$GPU_ARCH = gfx950
[varset:gfx1030]
$GPU_ARCH = gfx1030
[varset:gfx1100]
$GPU_ARCH = gfx1100
[varset:leonardo]
$HDF5_PREFIX = /leonardo/prod/spack/06/install/0.22/linux-rhel8-icelake/nvhpc-24.5/hdf5-1.14.3-2soj72l4cjp6mx7chrqu3h2qxzmrugnc
$NVF_CC = cc80
[varset:nolto]
# overlay (after an NVF varset): the optimized NVF modes without device link-time optimisation, the build before it
$NVF_LTO =
[varset:fastlink]
# overlay (after an NVF varset): the device LTO link split at the linker on 16 threads, any CUDA (issue #131): about
# half the link time, 7-10 % slower device code on an A30 (CUDA 12.9); for development builds, not for timings
$NVF_SPLIT = -Wnvlink,--split-compile-extended=16
[varset:iac_gnu]
$HDF5_PREFIX = lib/hdf5/develop/iac/gnu/16.2.0/openmpi-5.0.11
[varset:iac_nvf]
# CUDA pinned (issue #102): nvfortran 26.9 targets the CUDA of the build machine's driver, 13.3 where there is none (the
# login node); the A30 nodes' driver 575 runs CUDA 12.9 at most, so a login-node build failed at the first GPU call.
$HDF5_PREFIX = lib/hdf5/develop/iac/nvf/26.9
$NVF_CC = cc80,cuda12.9
$NVF_LTO = ,lto
[varset:spacehpc]
$HDF5_PREFIX = /opt/cray/pe/hdf5/1.14.3.3/nvidia/23.3
$NVF_CC = cc90