A dual-objective NPU scheduling optimization method constrained by data handling volume

Abstract

This invention relates to the field of artificial intelligence technology and discloses a dual-objective NPU scheduling optimization method constrained by data transport volume. The steps include: loading basic scheduling results containing optimized cache allocation, SPILL operations, scheduling sequences, and baseline transport volume; setting a tolerable growth range for transport volume; analyzing buffer time and physical address intervals to identify overlapping conflict buffer pairs; sequentially executing three types of incremental operations: adding positive-reward SPILL operations, node rearrangement under topological constraints, and timing optimization within the feasible time window of SPILL operations; verifying each round of optimization through hardware simulation, retaining schemes that satisfy multiple constraints and have better execution time; iterating according to a preset convergence criterion until execution time stabilizes, and outputting the optimized scheduling sequence, cache allocation scheme, and SPILL operation set. This invention maximizes the reduction of total NPU execution time while controlling the growth of data transport volume, thus balancing resource overhead and execution efficiency.

Images