Automatic Differentiation for OpenCL. Requires a GPU with fp64 extension

Status: Patiently waiting for for OpenCL 2.0 so we can use shared virtual memory. This will significantly reduce the time to transfer the gradient buffers from the host to the device.