GitLab

Benjamin Morgan authored May 26, 2020 and Gabriele Cosmo committed May 26, 2020

Use install(... EXPORT ...), export() to create/install files defining
VecGeom's targets for import by other projects.

Remove specialization of lib install "by file" on macOS to ensure
behaviour is same on all platforms.

If validation was requested but the schema file (gdml.xsd) was not found, the DocumentElement
is a nullptr.

For backward compatibility, allow BACKEND argument to be case insensitive.
Internally convert value to mixed-case required by call to find VecCore.
Make find of VecCore QUIET to avoid warnings about no FindVecCore module
or VecCoreConfig.

Benjamin Morgan authored Apr 06, 2020 and Gabriele Cosmo committed Apr 06, 2020

Organize setup of configuration/build settings related to a CMake
option as close to the definition of that option as possible.
When the option requires later setup, such as adding code to the build,
group setup for that option in the relevant section.

Benjamin Morgan authored Mar 10, 2020 and Gabriele Cosmo committed Mar 10, 2020

Replace custom install location variables with standard and FHS
compilant set provided by CMake's builtin GNUInstallDirs module.

Provide a basic set of CMake settings to check for absolute install
locations and to reduce verbosity of install messages when files
have not changed.

Replace all occurences (except Jenkins and RecordCmdLine.cmake) of
CMAKE_{SOURCE,BINARY}_DIR variables with PROJECT_{SOURCE,BINARY}_DIR.
The PROJECT versions are more robust under the use case that VecGeom
is imported/built as a submodule of a downstream project. Jenkins
scripts are left as-is as they are project specific. RecordCmdLine.cmake
is unchanged as it must run before project() is called and thus the
PROJECT variables do not exist.

RecordCmdLine is called before Project

Benjamin Morgan authored Mar 05, 2020 and Gabriele Cosmo committed Mar 05, 2020

VecGeom enables/disables features in its API and ABI via preprocessor
conditionals, e.g. "#ifdef VECGEOM_SCALAR". The definition of these
symbols is done by appending -D flags to a CMake variable, and then to
CMAKE_CXX_FLAGS. Additionally, this VECGEOM_DEFINITIONS variable is
exported to VecGeomConfig.cmake, and applied globally to any including
project with add_definitions.

This is suboptimal as the build time flags fix the API and ABI of the
compiled library, requiring any client to use exactly the same set of -D
flags. Whilst VecGeomConfig uses add_definitions, client projects are
at liberty to change the build settings, and clients may not use CMake at
all.

Provide a Config.h header, configured at CMake time with #define/undef
directives for VecGeom's symbols as configured by the client. For initial
clarity and checking, retain -D flag setting and do not #include generated header
nor install it.

Mihaela Gheata authored Feb 12, 2020 and Gabriele Cosmo committed Feb 24, 2020

[VECGEOM-540] Added executable to run shape tester from GDML file. Fixed errors reported by ShapeTester in BooleanUnionImplementation.

Benjamin Morgan authored Feb 18, 2020 and Gabriele Cosmo committed Feb 18, 2020

As GDML is an optional component and Xerces is a common and easy
to build external, remove ability to build it as an external.

Make find of external Xerces required. Link vgdml to imported target
supplied by CMake's FindXerces module.

Remove un-needed GLOBing of sources as only have a few sources, so clearer
to list explicitly.

At present, there is no export of the vgdml target, as neither are the
core VecGeom targets.

Benjamin Morgan authored Feb 17, 2020 and Gabriele Cosmo committed Feb 17, 2020

To reduce congestion in main CMakeLists.txt, move all code building an
"internal" copy of VecCore into a dedicated module.

This addresses Sandro's comment about dynamic heap allocation, but eliminating
dynamic bead allocation requires new functions to support NavStatePool instead
of NavigationState**.

- Keep warm-up call to GPU kernels, otherwise there is a 20% effect
  on GPU timing -- 0.195ms vs. 0.155ms with warm-up call beforehand.

- If CUDA enabled at compile time, require GPU device at run time.
  This avoids unexpected "illegal instructions" due to missing GPUs.
  NavigationBenchmark only changed, maybe shape benchmarks could also
  be modified likewise.