Open Compute Language (OpenCL) provides a programming framework for harnessing the compute capabilities of multicore processors. In this teaching series we cover a majority of the concepts for working with OpenCL in the context of supercomputing. We use the OpenCL 3.0 standard.

• course_material - contains the course material.
• deployment - contains tools for deploying course material to Github and managing users.
• resources - helpful tools and information for use within the course.

In this course we cover the following topics. Each topic is a subfolder in course_material.

• Lesson 1 - Introduction to OpenCL and high level features
• Lesson 2 - How to build and run OpenCL applications on Cray AMD systems like Frontier and Setonix
• Lesson 3 - A complete example of matrix multiplication, explained line by line
• Lesson 4 - Debugging OpenCL applications
• Lesson 5 - Measuring the performance of OpenCL applications with profiling and tracing tools
• Lesson 6 - Memory management
• Lesson 7 - Shared virtual memory
• Lesson 8 - Strategies for improving the performance of OpenCL kernels
• Lesson 9 - Strategies for optimising application performance with concurrent IO.

Lessons are in the form of Jupyter notebooks which can be viewed on the student's machine with JupyterLab or html files that can be viewed with a web browser. All exercises may be performed on the command line using an SSH connection to a remote server that has one or more implementations of OpenCL installed.

A recent version of Anaconda Python is helpful for viewing the notebook files. Once Anaconda is installed create a new environment

conda create --name opencl_course python=3.10 nodejs=18.15.0
conda activate opencl_course

A list of helpful packages for viewing the material may then be installed with this command when run from the OpenCL_Course folder.

pip install -r deployment/requirements.txt

then run

jupyter-lab

from the command line to start the Jupyter Lab environment. The command

conda activate opencl_course

is to enter the created environment, and the command

conda deactivate

will leave the environment.

A C++ compiler must be installed and available through the CC command.

Access to at least one OpenCL implementation is necessary. For teaching purposes it is preferred to have at least one CPU implementation and at least one GPU implementation installed, each with their icd files placed in /etc/OpenCL/vendors. The path to the OpenCL ICD loader libOpenCL.so must be in both LIBRARY_PATH and LD_LIBRARY_PATH. The location of the OpenCL headers, the CL directory (i.e where the file opencl.h lives) must be in CPATH. Khronos has a recent ICD loader and headers that can be used to compile with and link against. Alternatively a fairly recent package for the OpenCL headers and ICD loader may be installed through a package manager.

Here are some computing frameworks that contain OpenCL implementations. The trademark for each company is the property of their respective owners.

• AMD ROCM
• NVIDIA CUDA
• Intel oneAPI
• Portable CL

The debugging sections of this course use the excellent open source tool Oclgrind. This can be installed through a package manager.

Tau stands for Tuning and Analysis Utilities. In the profiling section we use Tau to trace OpenCL calls.

CLBlast is a linear algebra library we use to compare the performance of OpenCL kernels. This may be installed through a package manager.