In the CUDA SDK there is an example called convolutionSeparable which demonstrates how to implement a separable convolution filter. On p.4 it says that “The convolutionSeparable code sample in the NVIDIA CUDA SDK uses a separable Gaussian  blur filter.”
When I looked at the source code (main.cpp, line 93), the filter kernel is being set as follows:
for(unsigned int i = 0; i < KERNEL_LENGTH; i++) h_Kernel[i] = (float)(rand() % 16);
Where is the kernel being set to the Gaussian? Am I missing something here?