C++ for multi-accelerator, multi-vendor systems
summary
Claims that C++ code can run unchanged across CPUs, GPUs, FPGAs and vendors using oneAPI / SYCL and shows a laptop demo where switching the execution policy offloads a sort to the GPU and raises GPU utilization. Tools shown: oneDPL for parallel STL, DPC++/cyclomatic for CUDA→SYCL migration (reported ~90–95% automated), and Intel Advisor which flags three loops covering 75% of runtime and predicts offloading them could yield ~10% overall speedup.
AI-written summary · about
chapters (7)
- 0:00Start
- 4:32Higher abstraction level = Less performance!
- 5:07Exponential Compute Demand Growth
- 7:33Specialized workloads need specialized HW
- 21:16Device discovery and information
- 22:37Device control
- 23:12Cross devices memory management
related talks
31mGPU Performance Portability Using Standard C++ with SYCL
01h 01mTake Advantage of All the MIPS - SYCL & C++
01h 32mHeterogeneous Programming in C++ today
56mA Smooth Introduction to SYCL for C++20 afficionados
56m“Towards Heterogeneous Programming in C++"
01h 02mC++ Performance Portability - A Decade of Lessons Learned
01h 26mKhronos Sycl Language Framework for C++ Accelerators: Take Advantage of All the MIPS
01h 07mHeterogeneous Modern C++ with SYCL