Executing CUBLAS / CULA methods from within a Kernel

[shay_m]

Hi all,

I was wondering if there is any way I can invoke CUBLAS / CULA methods from within a kernel.
I was trying to do so, and I got a compilation error, saying that I am trying to run a host function from the device.
However, for an application I want to write, I need to execute the same CULA method N times in parallel.
Must I do a loop and perform these CULA methods in a sequential order?

Thanks,
Shay M.

[avidday]

Cublas and cula are host side apis. They cannot be used in kernel code.

Cublas exposes the streams interface, so if you use streams and run on a Fermi card, there might be some possibility of running simultaneous kernels. But I don’t remember cula using streams, although it was some time ago the last time I tried it.