Building with cmake doesn't set CUDA_ARCHITECTURES correctly
Issue #68
new
… at least on my Windows machine.
Setting GPU_TARGET to any value doesn’t work because it ends up setting the variable CUDA_ARCHITECTURES which is not used anywhere. Therefore, the default value of CMAKE_CUDA_ARCHTECTURES remains 52 and nvcc compiles for sm52 only no matter what.
Fix is simple:
replace set(CUDA_ARCHITECTURES "${__cuda_architectures}")
to set(CMAKE_CUDA_ARCHITECTURES "${__cuda_architectures}" CACHE STRING "CUDA architectures" FORCE)
https://bitbucket.org/icl/magma/pull-requests/32/bug-fix-some-misnamed-cmake-cache