Extend SYCL Support to Multivendor GPUs and Accelerators

This talk introduces how cross-architecture Data Parallel C++ (DPC++, a oneAPI implementation of SYCL*) is extended to an AMD* GPU platform. The basic architecture and constituent modules of the AMD GPU computing platforms (ROCM and DPC++) are briefly analyzed. Based on that analysis, learn how the platforms integrate an AMD device-compilation tool chain and DPC++ (SYCL) compilation tool chain. To complete the support of DPC++ for an AMD GPU, identify the mapping from the SYCL runtime to AMD device runtime. Sample programs are provided at the end of the talk.

Speaker

Leping Wang has a bachelor's degree of engineering in automation from Nankai University and a master's degree in computer science and intelligent control from University of Nebraska-Lincoln and Nankai University. He has over 10 years of experience in full stack software development and system architecture design. Leping is an engineer in the High-Performance Computer Research Center of The Institute of Computing Technology at the Chinese Academy of Sciences. His work focuses on cloud computing and unified parallel programming frameworks.