this is essential for the streaming weights to work correctly
due to the depth of the mem addr gen pipeline

turns out HLS does not let us specify stream depths on top level
AXI streams, need another solution