Sang-Woo Jun
Esmerald Aliaj
Department of Computer Science, University of California, Irvine

Abstract
Cosmology has historically been a computationally heavy discipline, which is often used in justification for new supercomputer development. Modernly, application-specific hardware acceleration of computation-intensive scientific kernels has provided significant performance and power efficiency improvements over general-purpose software. However, it is difficult and costly to incorporate them into existing software systems.
So far, designing such hardware accelerators using FPGAs or ASICS, as the venerable GRAPE special-purpose computers for n-body simulations, and modifying legacy software to incorporate them, are highly complex tasks. Furthermore, identifying a suitable kernel for acceleration is complicated by the PCIe-attached architecture of commodity FPGA cards, meaning hardware details such as bandwidth and latency overhead must be considered for software kernel selection.
As a remedy to these challenges, we present the current state of our FarSlayer project, a high-level source-to-source compiler for end-to-end acceleration of legacy software.FarSlayer analyzes previously existing legacy software code and emits an accelerated version of it, where the kernel is automatically selected considering data movement over PCIe. The entire process is automatic, meaning the domain scientist does not necessarily need to understand the existing code, or reason about hardware development. We demonstrate FarSlayer on using physically relevant software, such as the n-body codes Gizmo and Nemo, and the General Relativity ray tracer Gyoto, and demonstrate it can automatically achieve significant performance improvements.

2023 July 18, 11:00

IA/U.Lisboa
Faculdade de Ciências da Universidade de Lisboa (C8.1.69)
Campo Grande, 1749-016 Lisboa