By bind, do you mean mapping the resources to the parameters before you call the kernel, or do you mean allocating the resource?
If the contexts are on multiple devices, it in theory would depend on how they're allocated. Obviously, a local resource allocated on GPU1 cannot be used by GPU2. However, if you declare the resource as remote, you allocate the same resource on multiple devices, according to the API. From what I've read in the forums however, this is broken and numDevices must be 1.