39e7d213cf
[Reason] To support maximum CUDA features in offline tests + Add defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 600 restriction for atomicAdd on doubles in atomics.cu. So if LLVM < 7 and --cuda-gpu-arch doesn't work, __CUDA_ARCH__ is unset too (350 by default in clang); if LLVM >= 7 --cuda-gpu-arch is used and __CUDA_ARCH__ is set based on it.