Julia High Performance - Second Edition

Writing your own kernels

The Julia CUDA packages allow the programmer to create their own low-level kernels for the GPU, where you write code to manually handle the parallelism. This is something unique in Julia—the ability to use a high-level language to write these kernels. Previously, CUDA C++ was pretty much the only option to implement your own kernel.

As an example, consider a simple function that adds two vectors in place. A kernel can be written as follows: 

function add_gpu!(y, x)        index = threadIdx().x    stride = blockDim().x           for i = index:stride:length(y)                 @inbounds y[i] += x[i]           end         return nothing end

This is fairly low-level code, but is still recognizably Julia, while the semantics would be familiar to anyone who has written ...