The application discloses a task scheduling method for general applications of a heterogeneous integrated
chip system. The method aims at the problem that the computing capacity, on-
chip storage capacity and cross-grain communication overhead of different grains in the heterogeneous integrated
chip are quite different, constructs a
system architecture graph and an operator
dependency graph, and establishes unified computing
delay, communication
delay, storage constraint, cost and
power consumption models. On this basis, the
task mapping process is modeled as a Markov
decision process, a local subgraph sampling, a communication-aware graph
attention network and a cross-attention mechanism are used to
encode the scheduling state, and a proximal policy optimization
algorithm is used to search the mapping scheme of the operator to the processor core. The method can consider the computing load balancing, on-chip storage feasibility and cross-grain communication cost under the premise of meeting the resource and constraint conditions, and is suitable for
application scheduling of
large model inference, graph computing,
deep learning training and
inference and other applications with complex dependency relationships.