https://hgpu.org/?p=2377
Data transformations enabling loop vectorization on multithreaded data parallel architectures