This online webinar, presented by Thomas Applencourt and Abhishek Bagusetty from ALCF, is part of the Argonne Leadership Computing Facility (ALCF) Developer Sessions and is open to NERSC users.
Abstract
This webinar will cover the process of porting CUDA code to SYCL, focusing on high-performance math libraries like cuBLAS and cuFFT. ALCF’s Thomas Applencourt and Abhishek Bagusetty will discuss key challenges, such as differences in API, memory management, and execution models, and provide strategies for achieving portability and performance.
Special attention will be given to common pitfalls in porting, including synchronization issues and device memory handling. Additionally, we'll explore how to optimize SYCL code for Aurora’s advanced architecture, highlighting techniques to avoid bottlenecks, map kernels efficiently, and leverage multi-level parallelism. Through case studies and practical examples, this talk will guide developers transitioning CUDA applications to SYCL without sacrificing performance, particularly on heterogeneous platforms like Aurora.
