At a high level, calling a callable or kernel should be thread safe and should not normally incur any allocation overhead beyond, perhaps, allocating space for the output. It should also be possible to use a custom allocator for the output array. I have no idea on what the interface should look like for this, at least as of right now.