AI Engine kernels in the data flow graph operate on data streams that are infinitely long sequences of typed values. These data streams can be broken into separate blocks called windows and processed by a kernel. Kernels consume input blocks of data and produce output blocks of data. An initialization function can be specified to run before the kernel starts processing input data. The kernel can read scalars or vectors from the memory, however, the valid vector length for each read and write operation must be either 128 or 256 bits. Windows of input data and output buffer are locked for kernels before they are executed. Because the input data window needs to be filled with input data before kernel start, it increases latency compared to stream interface. The kernel can perform random access within a window of data and there is the ability to specify a window margin for algorithms that require some number of bytes from the previous window.
Kernels can also access the data streams in a window-by-window fashion. Streams are used for continuous data and using blocking or non-blocking calls to read and write. Cascade stream only supports blocking access. The AI Engine supports two 32-bit stream input ports and two 32-bit stream output ports. Valid vector length for reading or writing data streams must be either 32 or 128 bits. Packet streams are useful when the number of independent data streams in the program exceeds the number of hardware stream channels or ports available.
A PLIO port attribute is used to make external stream connections that cross the AI Engine to PL boundary. The PLIO port can be connected to the AI Engine window buffer via DMA S2MM or MM2S channels, or directly to AI Engine stream interfaces. Both of these connections (PL from/to window or stream) are limited by the stream interface of AI Engine tiles, that has a limit of 32-bit per cycle. However, for the window interface, the ping or pong buffer of the window needs to be filled up before the kernel can start. Therefore, window interfaces from / to PL usually have a larger latency than a stream interface when both fit into design architecture.
The following table summarizes the differences in window and stream connections between kernels.
|Connection||Margin||Packet Switching||Back Pressure||Lock||Max throughput by VLIW (per cycle)||Multicast as a Source|
|Window||Yes||Yes||Yes 1||Yes||2*256-bit load + 1*256-bit store||Yes|
2*32-bit read + 1*32-bit write, or
1*32-bit read + 2*32-bit write
Graph code is C++ and available in a separate file from kernel source files. The compiler places the AI Engine kernels into the AI Engine array, taking care of the memory requirements and making all the necessary connections for data flow. Multiple kernels with low core usage can be placed into a single tile.