Graphics processing apparatus and method for performance metric sampling

Abstract

Graphics processor performing: workload execution on circuitry including performance counting circuitry to count instances of performance metrics; responding to a performance counter sampling command configuring a performance counter sampling context that defines a sampling interval and a performance metric to be sampled; monitoring the sampling interval and when it elapses writing out a sample value for the performance metric. Also disclosed – controlling a graphics processor, comprising: generating and providing to the graphics processor the performance counter sampling command; and allocating a memory region to store the sampled values. Workloads may comprise tile-based rendering, texturing, fragment shading, geometry or vertex processing, draw calls, machine learning, neural processing. Performance metrics may comprise numbers of: cache misses, processed primitives, rendered fragments, particular operation types, machine learning acceleration tasks, computing tasks, cycles a workload execution circuitry element or core has been active. Performance counters may be integrated into the processing cores of the GPU. Sample values may be aggregated or stored with metadata identifying the workload and timestamp. Execution circuits may comprise endpoints for different job types, eg. rendering, neural processing, general computing. Triggering performance counting from the GPU rather than CPU may improve efficiency and sampling speed and reduce latency. [Fig. 5]