-
Notifications
You must be signed in to change notification settings - Fork 559
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Libxc] update to version 7.0.0 #9910
base: master
Are you sure you want to change the base?
Conversation
L/Libxc/Libxc_GPU/build_tarballs.jl
Outdated
|
||
mkdir libxc_build | ||
cd libxc_build | ||
cmake -DCMAKE_INSTALL_PREFIX=$prefix -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TARGET_TOOLCHAIN} \ | ||
|
||
mv ${WORKSPACE}/destdir/cuda/lib ${WORKSPACE}/destdir/cuda/lib64 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Doesn't this leave leftovers in the final tarballs? Why do you need to move it?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
CMake looks for the cuda installation in cuda/lib64
rather than cuda/lib
. Other CUDA packages either do this, or they create a symlink. I'm happy to change for the latter, if it is indeed cleaner.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The point is that the tarball has a cuda
directory full of empty subdirectories. That shouldn't exist at all.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Cleaned-up. It looks like the offending line was skip_audit=true, dont_dlopen=true
in the arguments of the build_tarball()
function. I am not sure what it did, but it does not seem necessary. I successfully tested the new build as well.
L/Libxc/Libxc_GPU/build_tarballs.jl
Outdated
-DCMAKE_BUILD_TYPE=Release -DENABLE_XHOST=OFF -DBUILD_SHARED_LIBS=ON \ | ||
-DENABLE_CUDA=ON -DCMAKE_CUDA_COMPILER=$prefix/cuda/bin/nvcc -DBUILD_TESTING=OFF \ | ||
-DENABLE_FORTRAN=OFF -DDISABLE_KXC=ON .. | ||
-DENABLE_CUDA=ON -DCMAKE_CUDA_COMPILER=$prefix/cuda/bin/nvcc \ |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
libxc is using enable_language(CUDA)
, i.e. the FindCUDAToolkit CMake module, so it should be sufficient (and a bit more proper) to specify the CUDA toolkit root, i.e. CUDAToolkit_ROOT=$prefix/cuda
- instead of specifying the path to nvcc
.
... but for unknown reasons, it seems CUDA_TOOLKIT_ROOT_DIR=$prefix/cuda
is usually what actually works - though that is the variable specified for the deprecated FindCUDA
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
None of the CUDA_TOOLKIT_ROOT_DIR
or CUDAToolkit_ROOT
option works out of the box in this build, without exporting the location of the nvcc
compiler. I left the DCMAKE_CUDA_COMPILER
as it is.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Did you try setting CUDA_PATH?
L/Libxc/Libxc_GPU/build_tarballs.jl
Outdated
-DENABLE_CUDA=ON -DCMAKE_CUDA_COMPILER=$prefix/cuda/bin/nvcc \ | ||
-DBUILD_TESTING=OFF -DENABLE_FORTRAN=OFF \ | ||
-DDISABLE_KXC=ON .. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Breaking lists into separate parts and sorting them is good practice (helps shorten future diffs etc.):
-DENABLE_CUDA=ON -DCMAKE_CUDA_COMPILER=$prefix/cuda/bin/nvcc \ | |
-DBUILD_TESTING=OFF -DENABLE_FORTRAN=OFF \ | |
-DDISABLE_KXC=ON .. | |
-DBUILD_TESTING=OFF \ | |
-DCMAKE_CUDA_COMPILER=$prefix/cuda/bin/nvcc \ | |
-DENABLE_CUDA=ON \ | |
-DENABLE_FORTRAN=OFF \ | |
-DDISABLE_KXC=ON .. |
This should of course be extended for the entire cmake "configure" call, but GitHub prevents making a suggestion for the entire cmake configure call (due to some deleted lines, apparently).
L/Libxc/Libxc_GPU/build_tarballs.jl
Outdated
|
||
build_tarballs(ARGS, name, version, sources, script, [platform], | ||
products, [dependencies; cuda_deps]; lazy_artifacts=true, | ||
julia_compat="1.7", augment_platform_block, | ||
skip_audit=true, dont_dlopen=true) | ||
julia_compat="1.8", augment_platform_block=CUDA.augment, preferred_gcc_version=v"8") |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Consider sorting keyword args (also lazy_artifacts) and putting each on a separate line.
Applied cosmetic changes suggested by @stemann |
L/Libxc/Libxc_GPU/build_tarballs.jl
Outdated
# The products that we will ensure are always built | ||
products = [ | ||
LibraryProduct("libxc", :libxc) | ||
] | ||
|
||
# Dependencies that must be installed before this package can be built | ||
dependencies = [ | ||
Dependency(PackageSpec(name="CompilerSupportLibraries_jll", uuid="e66e0078-7015-5450-92f7-15fbd957f2ae")), | ||
Dependency(PackageSpec(name="CompilerSupportLibraries_jll", uuid="e66e0078-7015-5450-92f7-15fbd957f2ae")) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Please don't make unrelated unneeded and undesired cosmetic changes such as removing trailing commas that are there for a very good reason (i.e. not to break git blame when adding new lines)
Dependency(PackageSpec(name="CompilerSupportLibraries_jll", uuid="e66e0078-7015-5450-92f7-15fbd957f2ae")) | |
Dependency(PackageSpec(name="CompilerSupportLibraries_jll", uuid="e66e0078-7015-5450-92f7-15fbd957f2ae")), |
Straight forward version update. No need for patches anymore, because of upstream fixes. Note that I gave up on a previous approach to add
aarch64
CPU + CUDA binaries (#9676).