It seems that AF cannot run operations transparantely across multiple gpus. For instance, I'm trying to do the following:

std::vector d = {1};
af::setDevice(0);
const af::array l1(1, d.data());
af::setDevice(1);
const af::array l2(1, d.data());
af::array r = l1 + l2;
std::vector o(1);
r.host(&o[0]);

This yields the error:
AfError: Input Array not created on current device

It would be really cool feature if arrayfire could handle this by transparently.

Related:

• given the af::array l1, is there a method or function to know to retrieve that it was created in device 0, or do i have to do the book keeping myself?