* ExaGO: Handling of cuda architectures and amdgpu targets changed
to effectively handle multiple targets. See #28441.
* Add ROCm support to ExaGO and update ROCm support in HiOp
* ExaGO+rocm requires HiOp+rocm
* Newer versions of CMake may set HIP_CLANG_INCLUDE_PATH incorrectly:
add comments to the ExaGO/HiOp packages explaining how to address
this problem if it occurs.
* cmake: use CMAKE_INSTALL_RPATH_USE_LINK_PATH
Spack has a heuristic to add rpaths for packages it knows are required,
but it's really a heuristic, and it does not work when the dependencies
put their libraries in a different folder than `<prefix>/lib{64,}`.
CMake patches binaries after install with the "install rpaths", which by
default are provided by Spack and its heuristic through
`CMAKE_INSTALL_RPATH`.
CMake however knows better what libraries are effectively being linked
to, and has an option to include those in the install rpath too, through
`CMAKE_INSTALL_RPATH_USE_LINK_PATH`.
These two CMake options are complementary, repeated rpaths seem to be
filtered, and the "use link path" paths are appended to Spack's
heuristic "install rpath".
So, it seems like a good idea to enable "use link path" by default, so
that:
- `dlopen` by library name uses Spack's heuristic search paths
- linked libraries in non-standard locations within a prefix get an
rpath thanks to CMake.
* docs
- Use define/define_from_variant
- Remove unused "fortran_flags"
- Fix CUDA architectures when using multiple (needs semicolon not comma
separators)
- Add `when=` variant restrictions to simplify logic
Add output of build- and install-time tests to info command
Enable dependencies, variants, and versions by default (i.e., provide --no*
options; add gcc to test_info_fields to increase coverage for c_names->v_names
* New package: spiner
* Update dependencies for spiner package
* Update var/spack/repos/builtin/packages/spiner/package.py
Co-authored-by: Seth R. Johnson <johnsonsr@ornl.gov>
* Update var/spack/repos/builtin/packages/spiner/package.py
Co-authored-by: Seth R. Johnson <johnsonsr@ornl.gov>
* Remove versions that can't be installed and use ports-of-call@1.1.0
Co-authored-by: Seth R. Johnson <johnsonsr@ornl.gov>
* py-torch: fix build with fujitsu-ssl2
* fix to use fujitsu-ssl2 in py-torch v1.5.0 to v1.11.0
* fix to use fujitsu-ssl2 in py-torch v1.2.0 to v1.11.0
* Delete fj-ssl2.patch
* renamed the patches
* Rename fj-ssl2.1.5.patch to fj-ssl2_1.5.patch
* Delete fj-ssl2_1.5.patch
We shouldn't be using "remove_linked_tree" to remove the lock file,
since that function expects to receive a directory path as an
argument.
Also, as a further measure to avoid regression, this commit restores
the "ignore_errors=True" argument on linux and adds a unit test
checking that "remove_linked_tree" doesn't change file permissions
as a side effect of a failure to remove.
* Fix py-onnx-runtime recipe
* Add missing dependencies
* Update var/spack/repos/builtin/packages/py-cerberus/package.py
Co-authored-by: Adam J. Stewart <ajstewart426@gmail.com>
* Update package.py
* Better fix for py-onnx-runtime
Co-authored-by: Adam J. Stewart <ajstewart426@gmail.com>
* omegah: v10.1.0
this version is from the SCOREC fork of Omega_h
* prefix version with scorec
Co-authored-by: Seth R. Johnson <johnsonsr@ornl.gov>
Co-authored-by: Seth R. Johnson <johnsonsr@ornl.gov>
Reduces the number of stat calls to a bare minimum:
- Single pass over src prefixes
- Handle projection clashes in memory
Symlinked directories in the src prefixes are now conditionally
transformed into directories with symlinks in the dst dir. Notably
`intel-mkl`, `cuda` and `qt` has top-level symlinked directories that
previously resulted in empty directories in the view. We now avoid
cycles and possible exponential blowup by only expanding symlinks that:
- point to dirs deeper in the folder structure;
- are a fixed depth of 2.
* py-cffi: add compiler flags to fix build with clang
For %clang@13.0.1, this avoids the
```
clang-13: warning: optimization flag '-ffat-lto-objects' is not supported [-Wignored-optimization-argument]
```
warning being turned into an error, and fixes this link error:
```
build/temp.linux-x86_64-3.10/c/_cffi_backend.o: file not recognized: file format not recognized
```
* style
Currently `old_root` is computed by reading the symlink at `self.root`.
We should be more defensive in removing it by checking that it is in the
same directory as the new root. Otherwise, in the worst case, when
someone runs `spack env create --with-view=./view -d .` and `view`
already exists and is a symlink to `/`, Spack effectively runs `rm -rf /`.
`file` was used to detect Python scripts with shebangs, so that the interpreter could be changed from <python prefix> to <view path>. With this change, we detect shebangs using Python instead, so that `file` is no longer required.
