ROLLER - Fast and Efficient Tensor Compilation for Deep Learning

This course aims to teach learners how to efficiently compile tensors for deep learning tasks, specifically focusing on reducing the time it takes to generate kernels for operators in deep neural networks on various accelerators. The course introduces ROLLER, a construction-based approach that utilizes a new tile abstraction called rTile to limit shape choices and improve execution efficiency. By adopting a recursive rTile-based construction algorithm, ROLLER can generate efficient kernels quickly, offering comparable performance to existing solutions on popular accelerators like GPUs and superior performance on less mature accelerators like IPUs. The course covers topics such as excessive compilation time, the black-box compiler approach, improving pipeline throughput, handling small and irregular shapes, and performance evaluations on GPUs. The intended audience for this course includes developers, researchers, and engineers working on deep learning projects who are interested in optimizing tensor compilation processes for improved performance on various hardware accelerators.