Domain Specific Description in Halide for Randomized Image Convolution

We propose a domain-specific language for finite impulse response (FIR) filters using a randomized algorithm. FIR filters such as Gaussian and bilateral filters are fundamental tools for signal processing and image processing. The computational time of these convolutions increases as the convolution kernel size becomes more extensive because the computational time depends on the kernel size. Approximating the kernel by random sub-sampling is one of the methods to reduce the computational time. The advantage of this approximation is easily controllable for the balance between its approximation accuracy and computational time. However, it is tricky to optimize this approximation to suit the multi-core and SIMD instructions provided in recent CPUs. Halide is a domain-specific language for image processing and can powerfully accelerate image processing with a concise description. One of the remarkable points of Halide is that it can be described as a separate description of the algorithm, which is the essence of image processing, and scheduling, which optimizes the processing such as parallelization and vectorization. Using Halide, we optimize the approximation of the FIR filter by the randomized algorithm. Also, we design a new domain-specific language that outputs the optimized Halide code in a concise description.