Use differing technology APIs for differing portions of algorithm / transfers

[trumpetlicks]

I have been doing a ton of work with halide recently, and one of the things I have noticed as I have been performance testing against all available APIs for a given OS is that in general, Direct3D on windows gives the best buffer transfer times to and from the GPU, while OpenCL gives the best algorithm runtimes (in general). I think that it would be great (and Im not sure if this is possible, but I cant see why not) if I could setup an algorithm/scheduling to use one technology for transfers, and a completely different one for "acting" on said buffer (i.e. performing the algorithm).

this could be furthered by allowing differing technologies to act on differing portions of the algorithm as well (this would probably be much more difficult).

[trumpetlicks]

Just as a note, currently my use model is 100% JIT compile.