cuda in direct show filters


I have a pretty large library, used inside a DS filter.

The library should use CUDA for acceleration. The problem is that as soon as I add to the *.cu files a memcpyToSymbol, I get the following error: invalid device symbol. There is nothing else (I kept it simple for testing purposes) in the *.cu file, except for some cudaMalloc s.

The device constant array is:

device constant float d_kernel_8[8][8];

and inside the same cu file I have:

cudaMemcpyToSymbol(d_kernel_8, h_kernel, 64*sizeof(float);

Anyway, not very relevant, since the same library works fine with a console application.

Has anyone encountered a similar behaviour?