I'm not sure why they imposed these limitations for inlineable
functions. It seems that unless there is an ambiguity (you need a
__host__ and a __device__ function), and the function is inlineable,
nvcc should easily be able to consume these functions. In fact, in
early editions, this often worked (though perhaps was a bug).
Brian
On Tue, Nov 2, 2010 at 11:34 PM, joel falcou
On 02/11/10 23:39, Manjunath Kudlur wrote:
Seriously? I'm afraid I have no suggestions for you. How does anybody use any third-party library on CUDA?
The short answer probably is, nobody does. The only libraries you can use from the device side of CUDA are those that you wrote yourself (so you would have the __device__ attributes on functions) or those that you have the source code for and painstakingly put __device__ attributes everywhere.
Yup and this is a MAJOR PITA :/ I tried to come up with a PP trick but nothing ends up working. I wish it was the other way around: default to be callable on device and specify the host only functions. _______________________________________________ Boost-users mailing list Boost-users@lists.boost.org http://lists.boost.org/mailman/listinfo.cgi/boost-users