This replies on the fact that clang allows using device-side features
from __host__/__device__ functions from __host__ ones as long as we
don't have to generate code for that. Wrapping thread/blockIdx in
__host__ __device__ function allows using CUDA_LOG everywhere during
host and device compilation.
Clang is less agressive than nvccnvcc, so number of functions did not getn
inlined into the kernel by default. That prevented SROA from eliminating
loads/stores to temporary buffers and resulted in abysmal performance.
Replaced inline with __forceinline__ to ensure that we do inline the
functions necessary for optimal performance.
Unlike nvcc, clang always sees both host and device-side code during
compilation. CUDA_LOG macro is used in both host and device code, so when it
expanded to contain device-only code, that resulted in errors when it was used
from the host-side functions.
In order to make CUDA_LOG work with clang it was split into two parts -- a pair
of target-attribute-based overloaded functions that perform host or device
specific parts of logging, and a printf which works on both sides.
