* Kokkos: add release 4.0.0
* Kokkos: updating default c++ standard requirement
Now Kokkos requires c++17 as its new minimum c++ standard library.
* Kokkos: adding support for new GPU architectures
The new updates include NVIDIA Hopper and AMD Navi
* Kokkos: fixing style...
* paraview +rocm: constrain kokkos dep to @:3.7.01
---------
Co-authored-by: eugeneswalker <38933153+eugeneswalker@users.noreply.github.com>
Co-authored-by: eugeneswalker <eugenesunsetwalker@gmail.com>
* initial commit for enabling test for rocm-smi-lib, rocm-cmake and rocm-clang-ocl
* fix styling and cleaning code
* disabling some tests for rocm-smi-lib
* fix style errors
* msmpi has no wrappers so don't set MPI_CXX_COMPILER etc. for that
MPI implementation
* hdf5 on Windows does not have h5cc etc., so do not try to filter
them on Windows
* py-uproot: new versions 5.0.4, 5.0.5
No changed in dependency versions
* py-awkward-cpp: new versions
* py-awkward: new versions
* py-awkward: new version in 1.10.* series
Previously `spack -e bla config update <section>` would treat the
environment config scope as standard config file instead of a single
file config scope. This fixes that.
When app is uninstalled, if it matches a default, then remove the
default symlink targeting its modulefile.
Until now, when a default were uninstalled, the default symlink were
left pointing to a nonexistent modulefile.
- Update default image to Ubuntu 22.04 (previously was still Ubuntu 18.04)
- Optionally use depfiles to install the environment within the container
- Allow extending Dockerfile Jinja2 template
- Allow extending Singularity definition file Jinja2 template
- Deprecate previous options to add extra instructions
Unless the amdgpu_target is overriden, the libraries will default to
being built for cuda, since amdgpu_target=none is both default and in
conflict with +rocm. This requires a custom Disjoint set to include
both the 'auto' variant used by the rocm mathlibs and the 'none'
variant used by ROCmPackage.
* Fix search for hip+cuda in hipcub@5.1 and later
This patch is not strictly necessary, but it may fix the search for HIP
in certain environments.
* Backport fix for CUDA 11.5 to hipsparse
* Reduce effort on grounding by employing cardinality constraints
If we use a cardinality constraint instead of a rule
using pair of values, we'll end up grounding 1 rule
instead of all the possible pair combinations of the
allowed values.
* Display all errors from concretization, instead of just one
If clingo produces multiple "error" facts, we now print all
of them in the error message. Before we were printing just
the one with the least priority.
Consolidate a few common patterns in concretize.lp to ensure
that certain node attributes have one and only one value
assigned.
All errors are displayed, so use a single criterion
instead of three.
* Account for weights in concretize.lp
To recover the optimization order we had before, account
for weights of errors when minimizing.
The priority is mapped to powers of 10, so to effectively
get back the same results as with priorities.
openblas likes to concurrently writes to the same archive from different
targets, and solves that through .NOTPARALLEL: all
We run `make x y z` which is not affected by `NOTPARALLEL`, running into
races.
When generating modulefile, correctly detect software installation asked
by user as explicit installation.
Explicit installation status were previously fetched from database
record of spec, which was only set after modulefile generation.
Code is updated to pass down the explicit status of software
installation to the object that generates modulefiles.
Fixes#34730.
Fixes#12105.
A value for the explicit argument has to be set when creating a new
installation, but for operations on existing installation, this value is
retrieved from database. Such operations are: module rm, module refresh,
module setdefaults or when get_module function is used.
Update on the way tests that mimics an installation, thus explicit
argument has to be set under such situation.
