Heterogeneous Programming in C++ with SYCL 2020
summary
SYCL 2020 standardizes unified shared memory (USM) for pointer-based data, adds a CUDA backend via DPC++ so SYCL can target NVIDIA GPUs, and raises the minimum to C++17 with CTAD to cut template noise. It ships a host_task API, explicit memcopy/usm malloc/free, and optional in-order queues; four implementations exist (ComputeCpp, DPC++, hipSYCL, triSYCL). DPC++’s CUDA plugin runs BabelStream on a GTX 980 with throughput close to native CUDA, and the provisional spec targets finalization in late 2020.
AI-written summary · about
chapters (6)
- 0:00Introduction to SYCL 2020
- 2:55SYCL past and lessons learned
- 6:27SYCL present and roadmap
- 14:30Deep dive into SYCL 2020 features
- 40:59SYCL future and ecosystem growth
- 50:52Q&A session
related talks
01h 07mHeterogeneous Modern C++ with SYCL
41mSYCL building blocks for C++ libraries
18moneAPI with SYCL gives Software Portability including Nvidia
01h 06mEfficient GPU Programming with Modern C++
48mBridging the Gap: Writing Portable C++ Programs for CPU and GPU
53mC++ Horizons
19mTask-Based Tensor Computations on Modern GPUs
59mUsing Template Magic to Automatically Generate Hybrid CPU/GPU-Code