Notes on Programming Apple Silicon
Preface
Programming Massively Parallel Processors is widely considered essential reading for GPU programmers. The book, a relatively thin paperback with an orange cover, contains over a dozen chapters which cumulatively present fundamental concepts in GPU hardware and software, an overview of parallel execution and memory models, performance factors, and intermediate to advanced parallel programming patterns; the 4th edition includes new content in four additional chapters: scheduling, the stencil pattern, reduction optimizations, and sorting algorithms adapted to parallel processors. The textbook focuses on CUDA, but explicitly states that the presented concepts are relevant to parallel programming in general [1].
Although the chapters throughout this book are based on CUDA, they help the readers to build up the foundation for parallel programming in general. We believe that humans understand best when we learn from concrete examples. That is, we must first learn the concepts in the concext of a particular programming model, which provides us with solid footing when we generalize our knowledge to other programing models. As we do so, we can draw on our concrete experience from the CUDA examples. In-depth experience with CUDA also enables us to gain maturity, which will help us to learn concepts that may not even be pertinent to the CUDA model.
The final line in the quote above is particularly relevant for Apple Silicon programmers. Apple Silicon differs from Nvidia hardware, particularly in its use of unified memory; the CPU and GPU both share access to the same dynamic random-access memory (DRAM); this allows developers to skip memory copy steps which are common to naive CUDA kernels.
This note-set walks through select chapters of Programming Massively Parallel Processors, with examples modified to apply directly to Apple Silicon. To do so, Apple’s open-source array framework — MLX — is used in illustrative code examples. All examples are executable, thanks to the Quarto literate programming framework in which this note-set was written. To run each example, you can install MLX directly using the Git source control and uv package manager commands below to open an IPython REPL with similar1 package dependency versions which were used to execute examples in the rendering of this note-set. Alternatively, you can try executing each example in your own Python environment with mlx installed; versions should be stable enough that all code examples in this note-set are executable for any recent Python release.
git clone -b stable https://github.com/cadojo/metalogue
cd metalogue
uv sync
uv run ipythonWe must say similar package dependency versions because package versions may change slightly for different operating systems and Python versions.↩︎