Chapter 6. GPU Programming with Accelerate
The most powerful processor in your computer may not be the CPU. Modern graphics processing units (GPUs) usually have something on the order of 10 to 100 times more raw compute power than the general-purpose CPU. However, the GPU is a very different beast from the CPU, and we can’t just run ordinary Haskell programs on it. A GPU consists of a large number of parallel processing units, each of which is much less powerful than one core of your CPU, so to unlock the power of a GPU we need a highly parallel workload. Furthermore, the processors of a GPU all run exactly the same code in lockstep, so they are suitable only for data-parallel tasks where the operations to perform on each data item are identical.
In recent years GPUs have become less graphics-specific and more suitable for performing general-purpose parallel processing tasks. However, GPUs are still programmed in a different way from the CPU because they have a different instruction set architecture. A special-purpose compiler is needed to compile code for the GPU, and the source code is normally written in a language that resembles a restricted subset of C. Two such languages are in widespread use: NVidia’s CUDA and OpenCL. These languages are very low-level and expose lots of details about the workings of the GPU, such as how and when to move data between the CPU’s memory and the GPU’s memory.
Clearly, we would like to be able to make use of the vast computing power of the GPU from ...