Using older version of cuda to call cuBlas and cuSparse from device code


Thanks in advance for your help. I’m trying to figure out the best way to call cuBLAS and cuSparse library functions from within a kernel. I found other posts stating that this is no longer possible as of Cuda 10 and so I would like to know if I can run my code using an older version of cuda that still supports this feature. I’m using a GeForce RTX 3060 Ti card on Windows using Visual Studio 2019.


AFAIK cusparse never had any in-kernel call capability.

For in-kernel cublas, you would need to install a CUDA 9.x toolkit on your machine and see what happens. You would use that toolkit to build for an earlier target (e.g. cc 7.0) but also specify PTX during the build process, and see if it works. When you install the CUDA 9.x toolkit, don’t install or update the driver. You want to keep your current driver that works with RTX 3060 TI.

NVIDIA recommends CUDA 11.1 or newer for RTX 30 cards, so you’re in an area that isn’t really supported, at all, by NVIDIA.