High-performance AXI-MM in HLS

My personal experience: the native AXI-MM in HLS is horrible. It fails to generate efficient code. The best practice I found is the use an external Datamover. In HLS, all memory access is made via AXI-Stream. Using AXI-Stream means we can wait the result asynchronously, hence we can deal with long memory access in a more informed manner.

Usually using AXI-Stream and Datamover delivers code with II=1.