parallel.gpu.CUDAKernel chokes on simple prototypes

I just noticed that the GPU module will recognize the following prototype

     __global__ void gaussianBlur( float* out,const float* image,
                  int w, int h)
{

but not this one:

     void __global__  gaussianBlur( float* out,const float* image,
                  int w, int h)
{

When using Matlab

k = parallel.gpu.CUDAKernel('matlab_kernel.ptx',...
'matlab_kernel.cu');

Despite both compiling... Maybe Mathworks should get a grab a copy of EDG? (I wanted to make a post somewhere because I couldn't find a solution on Google for a while.)