I need ability to copy back to host memory only one array from 2D-table. AFAIK it should be possible by using stream domain. But this would involve additional data copy inside GPU memory. Is any way exist to write part of 2D stream back to host memory w/o any additional data copying?
There is no way other than domain operator. But, assuming data copies within GPUs is about 100 GBps, it should still be a good idea to copy a part of stream to host than the complete stream copy.