GPU computing

Parallel programming of image processing algorithms on GPU (Graphics Processing Unit) using the CUDA (Nvidia GPUs) and/or HIP (AMD GPUs) languages that are multi-threaded extensions of C/C++, and smartly exploiting the GPU’s memory architecture in an embedded and/or High-Performance Computing (HPC) system.

By the end of the course, the student will have implemented 2D image processing algorithms in CUDA, targeting (near) real-time applications. 

Required and Corequired knowledge and skills

C/C++ programming skills (medium level; no need to master Object-Oriented concepts).

The course follows a “learn by example” approach.
The exercises will prepare the students (in groups of 2) to the parallel implementation of a 2D image processing algorithm described in a scientific report.

References, bibliography, and recommended reading

• John Cheng, Max Grossman, Ty McKercher, "Professional CUDA C Programming," John Wiley & Sons, 2014.
• Yifan Sun, Trinayan Baruah, David Kaeli, "Accelerated computing with HIP," Advanced Micro Devices, Inc, December 2022.

Course notes

• Université virtuelle

Contacts

Prof. Gauthier Lafruit, LISA-VR, ULB
Prof. Jan Lemeire, ETRO, VUB

Campus

Solbosch

Method(s) of evaluation

• Project
• Oral presentation

Project

Oral presentation

The evaluation is mainly based on an oral exam, presenting an imaging project with CUDA acceleration (report and source code to be submitted, one week prior to the exam). Complementary questions on the theory might also be punctually asked during the CUDA project presentation.

Mark calculation method (including weighting of intermediary marks)

A mini-project (typically measuring basic GPU performance figures in the roofline model) with written report has to be delivered mid-term, enabling us to check whether the students are ready for their final CUDA project. It counts for 20% of the marks; the remaining 80% come from the oral exam itself.

Language(s) of evaluation

• english
• (if applicable french, Dutch )