Calling built-in gpu functions from Cuda Kernel.

Hi,

I want to write my own Cuda kernel function. What I'd like to do is be able to call a matlab-gpu function from the device and not the host. Is this possible? I'd like to use svd because I have thousands of 3x3 matrices that I'd like to use svd on, but currently, its faster on the cpu as I can only do one small matrix at a time on the gpu.