Backends and Data Handling

Backends

hls4ml supports the concept of a backend that determines the target HLS language. Currently, Vivado HLS, Intel HLS, and Vitis HLS (experimental) are supported.

Data Handling

hls4ml supports multiple styles for handling data between layers. Data is passed in parallel between the layers when using the the io_parallel type. In contrast, data is passed one “pixel” at a time when using io_stream. What is actually streamed is an array of channels, which are always sent in parallel. This definition is most pertinent when performing CNNs. For Dense layers, all the inputs are streamed in parallel as a single array.