OpenVINO Profiling using VTune & GTPin

The existing GPU profiling workflow in OpenVino, through benchmark_app connects kernels to the model layers (primitives) and reports their execution times. However, this level of abstraction is too coarse to identify kernel instruction level insights. Intel GTPin, on the other hand, provides instruction level GPU kernel metrics like clock cycle counts, but is unable to directly map such metrics to specific inference or layers. This project aims to build a GTPin-centered profiling capability within OpenVINO to identify hotspot kernels and analyze their execution characteristics. The approach involves integrating controlled GTPin profiling into the inference flow, developing a custom GTPin tool for collecting kernel-level metrics, and progressively interpreting these metrics using signals from OpenVINO performance counters, and kernel metadata. The expected deliverable is a GTPin-based profiling workflow integrated into OpenVINO, capable of identifying bottleneck kernels, providing insight into their behavior, and offering indicative optimization directions, with validation using tools such as VTune.