Title
DPC++: A Direct Programming Model for Diverse Architectures
Abstract
Computation offloading is one of the fast growing trends in the data-centric computing and AI application domains for diverse architecture. OpenMP, CUDA and OpenCL programming languages have delivered excellent results nowadays for diverse architectures such as CPU, GPU and FPGA. However, there are tremendous opportunities and demands to push the edge of programming technology envelop further. This talk will present the DPC++ programming model, which is the core of the new oneAPI initiative and Intel’s top choice for GPU programming. DPC++ is based on Khronos SYCL language specification and additionally provides a rich set of Intel extensions for GPU offload computing and achieving optimal performance. Furthermore, we share our LLVM-based compiler technologies that have been developed in Intel’s DPC++ compiler for diverse architectures.
Bio
Xinmin Tian holds a Ph.D. in computer science, who is a Senior Principal Engineer and Compiler Architect in Intel Corporation, driving compiler parallelization, vectorization and offloading technology development in supporting Intel’s software strategy for Intel’s diverse architectures. Xinmin holds 25 US patents, published 52 technical papers in refereed international journals and conferences, co-authored two books on compiler optimizations and performance tuning.