Direct memory access operation for neural network accelerator

A neural network accelerator, comprising:

• a local memory;
• a programmable address translation table;
• an address mapping direct memory access (DMA) queue to store address mapping direct memory access (DMA) descriptors, each of the address mapping DMA descriptors including, as a source address, a physical address of address mapping data between physical addresses of a host memory and indirect addresses, each of the address mapping DMA descriptors further including, as a destination address, a physical address of the programmable address translation table;
• an input data DMA descriptor queue configured to store input data DMA descriptors, each input data DMA descriptor including an indirect address as a source address;
• an output data DMA descriptor queue configured to store output data DMA descriptors, each output data DMA descriptor including an indirect address as a destination address;
• a DMA engine;
• and
• a hardware processor configured to execute instructions associated with a neural network operation to:
• control the DMA engine to fetch a first address mapping DMA descriptor from the address mapping DMA queue, and to fetch a first address mapping to the programmable address translation table based on executing the first address mapping DMA descriptor;
• program the programmable address translation table based on the first address mapping;
• control the DMA engine to fetch a subset of the input data DMA descriptors from the input data DMA descriptor queue, and to translate first indirect addresses of the subset of the input data DMA descriptors to first physical addresses of the host memory via the programmable address translation table and based on the first address mapping;
• control the DMA engine to transfer input data from the first physical addresses of the host memory to the local memory based on executing the subset of the input data DMA descriptors;
• perform the neural network operation based on the input data from the local memory to generate output data, and to store the output data at the local memory;
• control the DMA engine to fetch a subset of the output data DMA descriptors from the output data DMA descriptor queue, and to translate second indirect addresses of the subset of the output data DMA descriptors to second physical addresses of the host memory via the programmable address translation table and based on the first address mapping;
• and
• control the DMA engine to transfer the output data from the local memory to the second physical addresses of the host memory based on executing the subset of the output data DMA descriptors.