Parallel Computations on GPU
PCG Acad. year 2021/2022 Winter semester 5 credits
The course covers the architecture and programming of graphics processing units by the NVidia and partially AMD. First, the architecture of GPUs is studied in detail. Then, the model of the program execution using hierarchical thread organisation and the SIMT model is discussed. Next, the memory hierarchy and synchronization techniques are described. After that, the course explains novel techniques of dynamic parallelism and data-flow processing concluded by practical usage of multi-GPU systems in environments with shared (NVLink) and distributed (MPI) memory. The second part of the course is devoted to high level programming techniques and libraries based on the OpenACC technology.
Language of instruction
Subject specific learning outcomes and competences
Knowledge of the parallel programming on GPUs in the area of general purpose computing, orientation in the area of accelerated systems, libraries and tools.
Generic learning outcomes and competences
Understanding of hardware limitations having impact on the efficiency of software solutions.
To familiarize yourself with the architecture and programming of graphics processing unit in the area of general purpose computuing using the NVidia libraries and OpenACC standard. To learn how to design and implement accelerated programs exploiting the potential of GPUs. To gain knowledge about the available libraries for programming on GPUs.
Why is the course taught
The future of computation systems ranging from ordinary PC up to top supercomputers is seen in heterogeneous systems where the sequential parts and the logic is processed by CPUs while the computational parts are offload to accelerators, in this case GPUs. This course will teach you the architecture and software libraries for programming graphics processing units in the area of general purpose computations.
Prerequisite kwnowledge and skills
Knowledge gained in courses AVS and partially in PRL and PPP.
- Kirk, D., and Hwu, W.: Programming Massively Parallel Processors: A Hands-on Approach, Elsevier, 2010, s. 256, ISBN: 978-0-12-381472-2
- Sanders, J., & Kandrot, E: CUDA by Example: An Introduction to General-Purpose GPU Programming. Review Literature And Arts Of The Americas. Addison-Wesley, 2010.
- Storti,D., and Yurtoglu, M.: CUDA for Engineers: An Introduction to High-Performance Parallel Computing, Addison-Wesley Professional; 1 edition, 2015. ISBN 978-0134177410.
- Chandrasekaran, S., and Juckeland, G.: OpenACC for Programmers: Concepts and Strategies, Addison-Wesley Professional, 2017, ISBN 978-0134694283
Syllabus of lectures
- Architecture of graphics processing units.
- CUDA programming model, tread execution.
- CUDA memory hierarchy.
- Synchronization and reduction.
- Dynamic parallelism and unified memory.
- Design and optimization of GPU algorithms.
- Stream processing, computation-communication overlapping.
- Multi-GPU systems.
- Nvidia Thrust library.
- OpenACC basics.
- OpenACC memory management.
- Code optimization with OpenACC.
- Libraries and tools for GPU programming.
Syllabus of computer exercises
- CUDA: Memory transfers, simple kernels
- CUDA: Shared memory
- CUDA: Texture and constant memory
- CUDA: Dynamic parallelism and unified memory.
- OpenACC: basic techniques.
- OpenACC: advanced techniques.
Syllabus - others, projects and individual work of students
- Development of an application in Nvidia CUDA
- Development of an application in OpenACC
Assessment of two projects, 14 hours in total and, computer laboratories and a midterm examination.
- Missed labs can be substituted in alternative dates.
- There will be a place for missed labs in the last week of the semester.
To get 20 out of 40 points for projects and midterm examination.
Course inclusion in study plans