These are distinct from the other build instructions (see also #499), as there is additional effort required on the different platforms to get everything running correctly.
need to describe how to use/set CUDA_TOOLKIT_ROOT_DIR from commandline or in CMakeLists
need to describe when and how to use USER_COMPILER_C, USER_COMPILER_CXX and USER_PATH_COMPILER
Issues to keep in mind on Windows:
so far, we have never successfully built the CUDA backend of Spirit on Windows
the CUDA installation may need to be added to the system PATH
for some reason, the conversion of 1.0 etc. to float does not happen automatically, which may appear as a build bug here and there, as we do not test this in CI (compare to https://github.com/facebookarchive/caffe2/issues/2183)
These are distinct from the other build instructions (see also #499), as there is additional effort required on the different platforms to get everything running correctly.
The requirements (CUDA version, compute capability) need to be stated clearly as well, see e.g. https://devblogs.nvidia.com/unified-memory-in-cuda-6/
This should therefore be in a separate file.
Issues to keep in mind:
CUDA_TOOLKIT_ROOT_DIR
from commandline or in CMakeListsUSER_COMPILER_C
,USER_COMPILER_CXX
andUSER_PATH_COMPILER
Issues to keep in mind on Windows:
PATH
1.0
etc. to float does not happen automatically, which may appear as a build bug here and there, as we do not test this in CI (compare to https://github.com/facebookarchive/caffe2/issues/2183)