Configuration of a processing module to process a workload

Abstract

Processing a workload according to parameters, performance dependent on parameters; training agent to determine parameters, receiving workload, determining workload and hardware parameters, manipulating workload with workload parameters and hardware with hardware parameters. Hardware parameter may relate to parallel processing, arithmetic logic, streaming. Agent may be trained with reinforcement learning with performance as reward. May store agent representation where agent is implemented using retrieved representation. Agent may be actor or critic neural network (NN). Training may comprise iterations where actor NN generates probability distribution over parameters, action selection logic selects parameter set based on probabilities, critic NN predicts reward, workload processed using selected parameters, actual reward determined and compared. Actual reward proportional or inversely proportional to reward. May comprise representation learning NN generating embeddings representing agent input state. Workload parameter may comprise kernel parameter defining how workload is split into chunks for respective hardware passes in processing. Workload may be NN. Layers of NN may process tensor, parameters specifyin: width / height split per pass, tensor planes per pass, filters processed per pass. Workload may comprise matrix multiplication, parameters specifying numbers included per chunk and parallelisation. Workload may be module of code to be compiled, parameters specifying compiler modes, options, flags. Figure 5