Program Sparsification with DaCe

Abstract

The optimization of sparse computations is critical for a broad spectrum of applications, from material sciences to machine learning. These computations, characterized by their irregular data structures, present unique challenges in optimizing performance. This thesis investigates the limitations of existing frameworks like DaCe, which are primarily designed for dense data applications and struggle with the irregularity inherent in sparse data. Building on the framework’s strengths, we propose enhancements that enable it to handle the optimization of sparse data applications better. By integrating a modular approach into DaCe, we facilitate the transition between dense and sparse implementations, allowing developers to interchange storage formats and compute kernels easily. This approach not only improves productivity by simplifying the optimization process but also has the potential to enhance performance and portability through the use of various underlying libraries suited to different hardware architectures. We evaluate our proposed workflow by applying it to synthetic and real-world applications, illustrating significant productivity improvements in optimizing sparse computations.