GitLab

- Cuda 11 dropped support for -arch=sm_30. The default was therefore updated to sm_35.
- Furthermore, it was impossible to select virtual architectures such as compute_70,
  since sm_ is always prepended to the CMake variable.
  Now, one specifies
    -DCUDA_ARCH=sm_35
  instead of
    -DCUDA_ARCH=35

Download all test data files to the test/ directory under the build
area. Only download ROOT files when VECGEOM_ROOT is enabled.

The DATA_DOWNLOAD option for testing is effectively unused because

- Testing always(*) requires the data, downloaded or locally copied
  by the client.
- Download is only triggered if the data is not present.
- Tests will fail if the data isn't present

(*) only if the full suite of tests is enabled. It may be possible
to restrict check/download for subsets of tests.

Remove the DATA_DOWNLOAD option, always checking for the data files
and downloading if not present or there is a hash mismatch. Rewrite
download/hashing to use CMake commands only.

All options related to testing Vecgeom should only be enabled if the
tests themselves are activated.

Use cmake_dependent_option to disable these options if the core
BUILD_TESTING variable is not true.

VecGeom uses CTest for testing, and hence its builtin BUILD_TESTING
option, but activates it by another custom variable, CTEST.

For clarity ans standardization, only use the CTest module and the
BUILD_TESTING variable/option. Update scripts/CI to use this.

Seth Johnson authored Dec 21, 2020 and Guilherme Lima committed Dec 21, 2020

When not validating input against the XML schema, text and comment nodes
remain in the DOM. This could result in `DOMNode` that were not
`DOMElement` instances, leading to a crash in the whitespace check.
Since only the whitespace/comment nodes are *not* elements, we can
eliminate the whitespace check more effectively.

Benjamin Morgan authored Dec 14, 2020 and Sandro Christian Wenzel committed Dec 16, 2020

CMake 3.8 (3.9 for Windows) supports CUDA as a first-class language.

To allow VecGeom to use this feature, hence simplifying CUDA support,
bump minimum required CMake to 3.8. Add version range support using
current 3.19 as upper tested limit.

3.8 is a conservative bump to begin the migration process. CUDA support
improves further in CMake 3.17 (FindCUDAToolkit) and 3.18 (CUDA compile
features).

Temporary as a proper project/site should be set up for VecGeom.

Seth Johnson authored Dec 14, 2020 and Benjamin Morgan committed Dec 14, 2020

This improves the logic so that:
- wget/curl is not called when DATA_DOWNLOAD is set to false
- downloaded files are checksummed and a warning is issued if they're
  incorrect

Guilherme Lima authored Nov 25, 2020 and Andrei Gheata committed Nov 30, 2020

This is required because globaldevicegeomdata::gMaxDepth is not initialized
properly for USE_NAVINDEX=OFF.

The static on-device variable globaldevicegeomdata::gMaxDepth is initialized
by the CUDA kernel named InitDeviceNavIndexPtrCudaKernel(), and this kernel
does not get called unless USE_NAVINDEX=ON.

This reverts commit e295ae2c. The method is used to fill a Transformation3D from data cached in the navigation index table, which is stored in unaligned memory, so may produce crashes.

Part of the algorithm was enabled only #ifndef VECCORE_CUDA, and
it has now been enabled everytime.

Thanks to Giulherme Lima