Course on CUDA Programming on NVIDIA GPUs, July 21-25, 2025

The course will be taught by Prof. Mike Giles and Prof. Wes Armour. They have both used CUDA in their research for many years, and set up and manage JADE, the first national GPU supercomputer for Machine Learning.

The 2024 course is now finished. We will run it again on July 21-25, 2025. Online registration should be set up by the end of April 2025, with a link from this webpage.

This is a one-week hands-on course for students, postdocs, academics and others who want to learn how to develop applications to run on NVIDIA GPUs using the CUDA programming environment. All that will be assumed is some proficiency with C and basic C++ programming. No prior experience with parallel computing will be assumed.

The course consists of approximately 3 hours of lectures and 4 hours of practicals each day. The aim is that by the end of the course you will be able to write relatively simple programs and will be confident and able to continue learning through studying the examples provided by NVIDIA on GitHub.

All attendees should bring a laptop to access the GPUs servers which will be used for the practicals.

The costs for the course are: Anyone with a status which does not fit into one of the categories above, including those outside the UK who are not from a university, company or government lab, should contact me (mike.giles@maths.ox.ac.uk) to discuss the appropriate fee category.

The intention is that these costs should not deter anyone from attending the course. The higher costs for certain participants correspond to the fact that they will be paying more for their travel and accommodation, and/or their organisations will be paying more for their time spent attending the course. It also reflects the UK funding for the facilities being used.

Venue

The lectures and practicals will all take place in the Mathematical Institute. Attendees should bring laptops for accessing the remote servers to carry out the practicals. It would be good to use fully-charged laptops, but we will try to provide adequate charging points as far as possible.

Travel to Oxford

For those coming to Oxford, especially from abroad, there is travel advice here.

Accommodation and food

Those attending the course must arrange their own accommodation. These are within a few minutes walk (or bus ride), and are arranged roughly in order of increasing cost: Alternatively, you might consider using Airbnb.

For coffee, breakfast and lunch, there is a good cafe in the basement of the Mathematical Institute. Little Clarendon Street, which is nearby, has several restaurants for dinner (and an excellent ice cream shop), and there are two sandwich shops for lunch on either side of its junction with Woodstock Road (A4144 on Google Maps).

Timetable

For the first three days we will follow this timetable: On the last two days we will switch to having both lectures in the morning, and then have practicals all afternoon. This provides more time for longer practicals, and will also allow those coming to Oxford from far away to leave when they wish on Friday afternoon.

Preliminary Reading

Please read chapters 1 and 2 of the NVIDIA CUDA C Programming Guide which is available both as PDF and online HTML.

CUDA is an extension of C/C++, so if you are a little rusty with C/C++ you should refresh your memory of it. Here are links to a couple of introductory lectures on C and an online resource.

Additional References


Lectures

Practicals

Most attendees will be provided with accounts on the ARC/HTC system which has a number of NVIDIA GPU nodes. Before starting the practicals, please read these ARC notes. Some details on the Slurm batch queueing system are available here.

The practicals all use these header files (helper_cuda.h, helper_string.h) which came originally from the CUDA SDK. They provide routines for error-checking and initialisation.

Tar file for all practicals

practicals.tar.gz contains all of the files needed for the practicals. Follow the instructions in the ARC notes to copy it to your ARC account and untar it.

Practical 1

Application: a trivial "hello world" example

CUDA aspects: launching a kernel, copying data to/from the graphics card, error checking and printing from kernel code Note: the instructions above explain how a tar file of all files can be copied from this webpage, so there's no need to download individual files from here

Practical 2

Application: Monte Carlo simulation using NVIDIA's cuRAND library for random number generation

CUDA aspects: constant memory, random number generation, kernel timing, minimising device memory bandwidth requirements

Practical 3

Application: 3D Laplace finite difference solver

CUDA aspects: thread block size optimisation, multi-dimensional memory layout, performance profiling

Practical 4

Application: reduction

CUDA aspects: dynamic shared memory, thread synchronisation

Practical 5

Application: using Tensor Cores and cuBLAS and other libraries

Practical 6

Application: revisiting the simple "hello world" example

CUDA aspects: using g++ for the main code, building libraries, using templates

Practical 7

Application: tri-diagonal equations -- see Lecture 7, slide 8, and also this research talk

Practical 8

Application: scan operation and recurrence equations

Practical 9

Application: pattern matching

Practical 10

Application: auto-tuning

Practical 11

Application: streams and OpenMP multithreading

Practical 12

Application: more on streams and overlapping computation and communication

Acknowledgements

Many thanks to:
webpage link checker