Efficient memory and GPU operations for Tiramisu compiler

Abstract

The increasing complexity of computer architectures with different memory hierarchies and parallelism characteristics makes generating efficient code a difficult task. Achieving high performance requires complex schedules and data layout transformations which might not be easy to express in a low level language. TIRAMISU [3] is an optimization framework for generating efficient code for different platforms including CPU, GPU, and distributed systems. It combines the polyhedral intermediate representation with rich scheduling and data layout commands, creating a high level interface to generate high performance code. In this thesis, we present new memory interfaces and GPU operators implemented to extend TIRAMISU compiler. We demonstrate that these features enable users to generate high performance GPU code with concise TIRAMISU programs. We also evaluate TIRAMISU's GPU backend with two benchmarks, matrix multiplication and a recurrent neural network architecture, showing that TIRAMISU outperforms other polyhedral compilers and popular library implementations.