Writing CUDA Kernels For Interpolation

In real-life applications, we want big images: when we watch a video clip on a PC, we like to see it in the full-screen mode. We want high-quality images: if a block of pixels gets damaged during the transmission, we want to repair it. We want cool images: by digital image manipulation, fancy artistic effects as seen in movies can be rendered. We want fast processing, especially when the images are big and many. To process even faster, we want that the various image pixels are processed in parallel.

CUDA (Compute Unified Device Architecture) is a hardware architecture and programming model introduced by NVIDIA for the parallel processing of Graphics Processing Units (GPUs). It represents by now an assessed tool for parallel programming and permits low-level programming capable of achieving very high performance by directly and properly managing the thread work.

In this course, the direct use of CUDA for a simple yet common problem like image interpolation is illustrated. This will enable the attendee to get familiar with the functions running on the GPU, namely, the kernel functions. Being interpolation very common in technical and scientific applications, the development of parallel interpolation codes permits having a tool that can be reused when needed.