|
| TOPIC | Presenter |
|---|
| 09:00 | 09:10 | Welcome and Introduction to Day 2 (presentation recording) | Edmund Preiss (Intel),
Thomas Steinke (ZIB) |
| 09:10 | 10:10 | A oneAPI Case Study: easyWave - A Tsunami Simulations Application
with demos ZIB's experiences with Intel oneAPI
- Porting a tsunami application from CUDA to DPC++
- Running DPC++ code on GPU
- From CUDA to DPC++ back to Nvidia GPUs... and FPGAs
- A oneAPI case study with the tsunami simulation easyWave
Live Demos: Christgau_playbook.tar.gz
(1) overview on CUDA code,
(2) CUDA code execution
(3) guided transformation with Compatibility Tool,
(4) presentation/analysis of generated DPC++ code,
(5) execution on different architectures (Intel CPU + GPU, Nvidia GPU) (presentation recording) | Steffen Christgau (ZIB)
(Marius Knaust, ZIB) |
| 10:10 | 10:50 | Intel OpenMP – with Demos
- Parallelizing heterogenous applications with Intel OpenMP and OpenMP offloading (presentation recording) | Michael Steyer (Intel) |
| 10:50 | 11:00 | Bio Break |
|
| 11:00 | 11:45 | Intel HW (public) roadmap/XPU and architecture specifics
- Server CPUs
- Client CPUs (i.e., Tigerlake; iCore with integrated graphics)
- Intel hardware accelerators GPUs (DG1, SG1, ATS) and Intel FPGAs
(presentation recording) | Jean-Laurent Philippe (Intel) |
| 11:45 | 12:15 | A oneAPI Case Study: Ginkgo – a sparse linear algebra library for OneAPI Hardware
Experience and potential of using oneAPI for the Ginkgo sparse linear algebra library
- Porting of numerical linear algebra kernels from Cuda to DPC++
- Pitfalls and solutions in the OneAPI Compatibility Tool
- Performance of Ginkgo’s DPC++ backend on Intel GPU
- Live demo running Ginkgo in the Intel DevCloud (presentation recording) | Hartwig Anzt (KIT)
(Y. Tsai, T. Cojean, KIT) |
| 12:15 | 13:15 | Lunch Break |
|
| 13:15 | 14:30 | Application profiling for heterogenous hardware - Demos
- Profile DPC++ and GPU Workload VTune Profiler and VTune Offload
- Share experiences/key findings with Gromacs related porting and optimization efforts efforts
(presentation recording)
| Heinrich Bockhorst (Intel) |
| 14:30 | 14:35 | Bio Break |
|
| 14:35 | 15:35 | Application profiling for heterogenous hardware - Demos
- Profile DPC++ and GPU Workload with Advisor include and Roofline analyser
- Estimate performance potential gains with Offload Advisor ( CPU -> HW Accelerator) Playbook_Advisor_GPU_NBody.txt (presentation recording) | Klaus-Dieter Oertel (Intel) |
| 15:35 | 16:00 | - Questions and Answers
- Wrap up | Intel / ZIB |