gitlab ci: Set resource requests explicitly
This PR sets resource requests for the Kubernetes executor, which should aid in
better workload scheduling in the cluster. The specific values were derived from
profile data taken from several full "from scratch" rebuilds in a separate worker pool.
Co-authored-by: Zack Galbreath <zack.galbreath@kitware.com>
* serialbox: setup the run and dependent build environments
* Update var/spack/repos/builtin/packages/serialbox/package.py
Co-authored-by: Adam J. Stewart <ajstewart426@gmail.com>
Co-authored-by: Adam J. Stewart <ajstewart426@gmail.com>
* rocmlibs: relax rocm-cmake version requirements
The rocm-cmake modules tend to be backwards-compatible, to the extent
that most ROCm math libraries were built using rocm-cmake@master
for a long while without anybody noticing. (That was fixed in
97f0c3ccd9f0a40896998a7580150a514ec3bc37.)
Some packages, like comgr, barely use rocm-cmake for anything, and
we can easily set a very minimal version requirement. For most
packages, however, it would be a lot of effort to determine the
minimum rocm-cmake version required for each release. For those
packages, I just turned the exact version requirement into a
minimum version requirement.
Since I was looking through the CMakeLists.txt for a large number of
libraries, I also took note of the cmake_minimum_required and adjusted
the cmake minimum requirements to match.
* Add rocblas build dependency to hipblas
The rocblas library is required both for both building and linking
hipblas.
* Remove rocm-cmake from vtk-m dependency list
The rocm-cmake package provides CMake scripts that facilitate common
build configuration tasks in the ROCm libraries. It is never needed at
link-time. Also, there are no calls to find_package(ROCM) or
include(ROCM.*)in vtk-m, so this dependency will never be used.
- older versions are no longer available for download so mark them
deprecated
- set manual_download
- set url_for_version
- only install the binary that matches the cuda version
In #26630, I assumed "glu" was needed by glew because it included glu.h, but
actually, glew can be used without glu when GLEW_NO_GLU is defined and this
is documented in the announcement of glew-1.6.0:
> https://www.geeks3d.com/20110430/opengl-glew-1-6-0-available/
> * Define GLEW_NO_GLU for no glu dependency
It is therefore the duty of users of glew to decide if they use glu,
and then they need to have a depends_on("glu").
Thus, move the depends_on("glu") which I changed from "gl" in #26630
to vapor, which itself uses glu as well.
For about a decade GCC has an option `-f[no]-canonical-system-headers`
which basically runs `realpath` on all "system headers", to possibly
reduce the length of paths in diagnostics. [1]
Spack usually installs the "system headers" of GCC in very deeply nested
directories. Calling `realpath` there results in stat calls on every
level, for every header file. On some slow filesystem I have,
`-fno-canonical-system-headers` gives about 5x speedup to compile hello
world in C, meaning that ./configure scripts would be much faster when
using this flag by default.
[1] https://codereview.appspot.com/6495088
Add option to allow using OpenSSL (by default this uses the SSL
implementation that comes with Windows, since that is more likely
to have needed certificates).
