Introduction

The core of katsdpsigproc is a framework to simplify programming for CUDA and OpenCL, including writing code that will work on both APIs. OpenCL allows for programming a range of devices, including CPUs, GPUs and FPGAs, but where it simplifies the flow of text I will talk about GPUs as they’re the most commonly available accelerator.

GPU programming is a massive topic and attempting to provide background in it is far beyond the scope of this user guide. I will thus assume that you are familiar with either CUDA or OpenCL (but not necessarily PyCUDA or PyOpenCL — that is not necessary to be able to use katsdpsigproc). The API naming tends towards using OpenCL terminology, as it is more generic. The table below may be handy for users more familiar with CUDA. If these terms are unfamiliar to you, I recommend learning about them before trying to use katsdpsigproc.

CUDA	OpenCL
Thread	Work item
Thread block	Work group
Stream	Command queue
Local memory	Private memory
Shared memory	Local memory
Device memory	Global memory

The sections in the guide are organized from basics towards more advanced features, and later sections often build on knowledge from earlier ones, so new users are advised to read the guide in order rather than jumping ahead.