🎓 GPU Porting for HPC - Free Online Webinar at HLRS

This two-day online webinar introduces GPU porting for high-performance computing (HPC), focusing on opportunities, programming approaches and portable tools. It will help participants assess where GPU acceleration may benefit their applications and understand the first steps in adapting scientific and engineering codes.

The programme begins with an introduction to CPU and GPU architectures, their implications for performance, and the main GPU programming models. The second session introduces alpaka, an open-source C++ library for portable parallel programming across supported CPU and GPU backends. Topics include project setup, core concepts, a first kernel, and higher-level algorithms. Friday's session will start with an introduction to the multipysics framework m-AIA. The final session will introduce RedSim, a GPU-accelerated CFD simulation framework designed to model urban air pollution.

This course is organised by HLRS in cooperation with EXCELLERAT P2, HiDALGO2 and Plasma-PEPSC, under the umbrella of High-Performance GPU Alliance.

Preliminary Agenda:

Thursday, 5. November

8:45 Log in to the videoconference
9:00 Welcome
9:05 - 12:00 (including breaks)
Introduction to GPU Programming - Opportunities and Challenges (HLRS)

• Differences between CPU and GPU
• GPU Architecture and Consequences
• Programming models
• Summary
• Discussion

12:00 Lunch break

13:00 - 16:00 (including breaks)
The alpaka library: task-parallel kernel execution on heterogenous hardware (PLASMA-PEPSC)

• Introduction
• Getting alpaka & project setup
• Basic concepts
• First kernel
• Higher-level algorithms & helpers
• Summary
• Discussion / Q&A
   

Friday, 6. November

9:00 - 12:00 (including breaks)
The multiphysics framework m-AIA (EXCELLERAT)

• Welcome and Introduction to the multiphysics framework m-AIA
• GPU porting in C++ with parallel STL in m-AIA
• Discussion and Coffee break
• Adaptation and dynamic load balancing on GPUs

12:00 Lunch break

13:00 - 16:00 (including breaks)
Efficient GPU Implementation of the Finite Volume Method on Unstructured Meshes (HIDALGO)

• Building the Mesh
  • Introduction
  • Sorting as a mechanism for adjacency, RADIX Sort on the GPU
  • Sub-Meshes, Halo cells
  • Reordering Meshes using Space Filling Curves
  • Grouping interior and boundary touching cells
• Simulation, Runtime Communication and Kernels
  • Halo communication patterns
  • Maximizing work while communicationg
  • GPU topology: inter-node vs extra-node GPUs
  • Load balancing
  • Custom standard lobrary overview
• Physical Method Overview, demonstration
  • Solving scalar transport, kernel overviews
  • Demonstration of RedSIMon a cube mesh
    • SOD problem: defining the IC, BC via CUDA kernel programming
    • Taylor-Green problem: defining the IC, BC, and Source term via custom CUDA kernel programming