GPU Long Instruction Tracing for Low-Overhead Kernel Profiling

The patent creates a copy of the GPU execution pipeline through software simulation. The GLIT engine captures actual GPU execution traces and replays them in software, allowing detailed instruction-level analysis without impacting actual GPU performance. This copying approach enables precise measurement of each GPU instruction while maintaining productivity by performing analysis on the replicated execution data rather than interfering with the original execution.

The patent introduces an intermediary GLIT engine that sits between the GPU execution and the analysis tools. This intermediary component captures GPU states and instruction traces in a standardized format, then provides this data to various analysis tools. The intermediary approach improves measurement precision by systematically capturing detailed GPU states while managing device complexity by centralizing the instrumentation logic in the GLIT engine rather than distributing it throughout the entire system.

The patent performs preliminary action by capturing complete GPU execution traces before any detailed analysis is performed. The GLIT engine records instruction-level data and GPU states during a single execution pass, storing this information for subsequent offline analysis. This preliminary capture approach improves measurement precision by having all execution data available for analysis while minimizing time loss, since the tracing overhead is incurred only once during the initial execution rather than during each analysis operation.