Profiling Halide DSL with CPU Performance Events for Schedule Optimization

Halide is a domain-specific language (DSL) for image processing that enforces a separation of the algorithm and the execution schedule, allowing the generation of specialized code for distinct computer architectures by rewriting only the execution schedule, instead of the whole algorithm. In order to support the creation of good Halide schedules, our work extends the Halide DSL by adding a profiling API that uses the CPU Performance Events to measure events supported by the target processor during the application runtime. The proposed extension offers profiling of the application loop levels and functions' producer and consumer relations, embedding calls to a profiling library in the loop nests of the generated code. It also supports individualized profiling by threads on parallel regions. As a case study we use the PAPI library in order to count events such as L1 cache misses, number of float operations (FLOP) and L3 data volume on an Intel Core i5-7500 CPU, and discuss how the reported results can be used to manually or automatically generate better schedules for an image processing pipeline.