Memory access optimisation using per-layer computational mapping and memory allocation for CNN application

Abstract MEMORY ACCESS OPTIMISATION USING PER-LAYER COMPUTATIONAL MAPPING AND MEMORY ALLOCATION FOR CNN APPLICATION A method (100) of configuring a System on Chip (110, 1314) to execute a CNN process (101, 1303) comprising CNN layers (1304, 1306, ... , 1306), the method comprising, for each 5 schedule: determining (202) memory access amount information (207) describing how many memory accesses are required; expressing (203) the memory access amount information (207) as relationships (204) describing reusability of data; combining (305) the relationships (204) with a cost (304) of writing and reading from external memory (1309), to form memory access information (306); determining (400) a memory allocation (413) for on-chip memory 10 (1310, 1320) of the SoC (110) for the input FMs and the output FMs; and determining (500), dependent upon the memory access information (306) and the memory allocation (413) for each schedule; a schedule (508) which minimises the memory access information (306) of external memory access for the CNN layer of the CNN process; and a memory allocation (413) associated with the determined schedule (508). P274796 / 14051477_2 (See Figs. 8, 9) 101 102 103 Convolutional Set of Memory Neural Network Scheduling Cntan 7306f (CNN) process Schm es z osta External Memory (See Figs. 2, 3) access 104 estimation model Minimum Buffer Build estimation models for (See Fig. 4) requirement A External Mem Access 400 (See Figs. Bestmemory 2 3,4,5) allocation 3 413 Determine memory allocation for combination of scheduling schemes and Layer Best scheduling scheme Determine Preferred 508 Scheduling Scheme per 5) Layer Target accelerator Program the controller z hardware ExecutionofCNNBest scheduling application scheme and memory allocation erlaer P274796 / 13569828_1 Fig.1 107