Question: any way I can pass a `cl_mem` to your library, and get the convolutions to run against that?

[hughperkins]

Question: any way I can pass a cl_mem to your library, and get the winograd convolutions to run against that?

Obviously, you'd probably need to do some driver-level hacking, possibly with the help of nvidia, to do this, in order to retrieve the appropriate float * from it somehow. I cannot do this myself, because I dont want to expose myself to nvidia's IP, or vioalte any license agreements etc, but ... maybe.... you are able to do this???

Then I can make it that my variuos opencl libraries can use winograd on cuda gpus, and other libraries on non-cuda gpus.

(for context, here are my opencl libraries:

• https://github.com/hughperkins/DeepCL
• https://github.com/hughperkins/cltorch https://github.com/hughperkins/clnn )