Dynamic parallelisim and cusparse

Hi,
Is this possible or all cusparse\cublas function can be called just from host?
I have an algorithm that includes tridiagonal solver and i want to use cusparse instead without the overhead of going back to the host can i do that?

cublas functions can be called from the device. There is cublas documentation and there is a sample code that demonstrate how:

http://docs.nvidia.com/cuda/cublas/index.html#device-api

http://docs.nvidia.com/cuda/cuda-samples/index.html#simpledevlibcublas-gpu-device-api-library-functions–cuda-dynamic-parallelism-

As far as I know, cusparse functions cannot be called from device code.