I have some code written using the CUDA runtime API. I would like to be able to load and launch kernels from cubin files in the way possible using the driver API. However, my code depends on the CUDPP library which will make it hard to port to the driver API. Even if I could get of CUDPP, I’d still like to keep using the runtime API .
I think loadable module support can be a very useful feature, especially for runtime generated code. Any good reason this isn’t supported in the runtime API? Are there any plans to support it in the runtime API in an upcoming release?
If not, any suggestions how to get around this? Is this a case where mixing runtime and driver API’s might be an ugly but possible solution?