Improving the Programmability of A Distributed Hardware Accelerator

Abstract

Sparse iterative matrix algorithms are critical to many scientific and engineering workloads, yet they perform poorly on conventional hardware. (Ōmeteōtl, a new hardware accelerator with a distributed-memory and task-based execution model, aims to address these performance bottlenecks. However, programming for (Ōmeteōtl is low-level, error-prone, and far removed from the simplicity of typical iterative formulations. This thesis presents Lapis, a domain-specific language and compiler that allows users to express sparse matrix algorithms in high-level Python code and automatically generates efficient C++ code for (Ōmeteōtl. Lapis abstracts away data partitioning and task orchestration, reducing implementation complexity: for example, it lowers lines of code by 30× for conjugate gradients and 46× for power iteration. Despite this abstraction, generated code achieves 75.7% to 92.6% of the performance of manually written implementations across several benchmarks.