Halide is a high performance tensor processing DSL that separates algorithm from execution schedule. Currently, Halide’s type system is limited by a lack of a Q number format, as well as optimizations for variable bit calculations. This project will add the Q number format (which expresses the number of integer and fractional bits) to Halide’s type infrastructure, as well as develop compiler optimizations that analyze the degree of numeric calculations in Halide to achieve better performance in image processing pipelines. We hope to extend the Halide language to allow for a number of numeric type optimizations. We will assess these optimizations based on speed of execution and scalability measures.

“I am excited to apply the knowledge I’ve gained in Professor Durand’s class, as well as others, towards this research project. I want to gain high-level research experience, and this project ties together a number of my interests: high performance computing, compilers, graphics, and mathematics. I am excited to learn more about parallel programming techniques and how to produce better optimization techniques.”