22.214.171.124. Streaming System Buffer Management
Before machine learning inference operations can occur, the system requires some initial configuration.
As in the M2M variant, the S2M application allocates sections of system memory to handle the various Intel® FPGA AI Suite IP buffers at startup. These include the graph buffer, which contains the weights, biases and configuration, and the input and output buffers for individual inference requests.
Instead of fully managing these buffers, the input-data buffer management is offloaded to the Nios® processor. The Nios® processor owns the Avalon® streaming to memory-mapped mSGDMA, and the processor programs this DMA to push the formatted data into system memory.
As the buffers are allocated at startup, the input buffer locations are written into the mailbox. The Nios® V processor then holds onto these buffers until a new set is received. All stream data is now constantly pushed into these buffers in a circular ring-buffer concept.