Schedule visualization and analysis for halide image processing language

[摘要] Image processing applications require high performance software implementations in order to satisfy large input data and run on smaller mobile devices that require high efficiency. Halide is a language and compiler for optimizing image processing pipelines. Halide introduces a separation between algorithm, the logics behind the program, and a schedule, the order of execution. This thesis focuses on providing interactive GUI for visual analysis of Halide schedules. It creates a visualization of the order of execution and provides tools for analyzing three important aspects of image processing schedules: redundancy, locality and parallelism. Tool is designed for Halide programers who want to gain better understanding of scheduling in Halide and receive guidance for schedule optimizations.

[发布机构] Massachusetts Institute of Technology