Virtual Conference: 2022 SIAM Conference on Parallel Processing for Scientific Computing

Part of MS54 Co-Design of Data Flow Accelerators for Scientific Simulations and Machine Learning - Part I of III
COMET: A Compiler Framework for Next-Generation Heterogenous Systems

Abstract. As Moore’s Law is coming to an end, simple technology scaling cannot be relied on for performance gain, and new technologies and computing paradigms must be developed to achieve application performance, which is leading to an explosion of accelerator designs from both industry and academia. Leveraging the new types of compute resources will require new methods and tools to not only achieve high-performance on current heterogeneous systems but also interface with hardware performance modeling and simulation tools to estimate application performance on future specialized accelerators. Compiler technologies have considerably evolved during the last decades and now support sophisticated code analysis and translation methodologies as well as efficient code generation for target heterogeneous systems. In the context of co-design, a compiler-based tool is desirable because it offers the possibility of automatically generate code that leverages new hardware concepts without or with little code modifications in the applications. In this talk, I present COMET tensor algebra compiler targeting quantum chemistry and graph analytics applications. COMET provides an opportunity to perform hardware/software co-design and design space exploration efficiently and to assess the performance of the entire application, instead of only the innermost kernel.

Authors  
 
 
PP22 Home 2022 Program Speaker Index
Powered by MathJax