Method for improving GPU efficiency via programmatic tests of synchronization primitive progress

A method, comprising:

• executing instructions of a wave by each of a plurality of vector processing circuits;
• monitoring, by circuitry, hardware behavior of a plurality of threads of a given wave;
• responsive to executing a synchronization instruction by a first thread of the plurality of threads, generating, by circuitry, a prediction of an amount of time for the synchronization instruction to complete based on the hardware behavior of the plurality of threads;
• and
• updating, by circuitry, control flow of the first thread based on the prediction.

A computing system comprising:

• a memory;
• and
• a processing circuit comprising:
• a plurality of compute circuits, each comprising:
• a plurality of vector processing circuits, each configured to execute instructions of a wave;
• and
• circuitry configured to:
• monitor hardware behavior of a plurality of threads of a given wave;
• responsive to executing a synchronization instruction by a first thread of the plurality of threads, generate a prediction of an amount of time for the synchronization instruction to complete based on the hardware behavior of the plurality of threads;
• and
• update control flow of the first thread based on the prediction.