CUDA Block Size Optimization for Gaussian and Sobel Filters: Benchmarking Against CPU Implementations
Аннотация
This article evaluates the efficiency of central processing unit (CPU) and graphics processing unit (GPU)-based computations in image filtering. Gaussian and Sobel filters were selected as the most effective methods for image filtering; these filtering processes are performed sequentially on the CPU and in parallel on the GPU. In order to organize parallel computations in image filtering, the most effective CUDA (Compute Unified Device Architecture) technology is used to implement parallel programming on the GPU. The CUDA programming model provides parallel processing of images by performing all computations in the kernel function of thousands of threads in blocks, which are grouped together. Therefore, the correct selection of GPU blocks in the program also affects the speed of the computation process. During the study, using the CUDA programming model, 8×8, 16×16, and 32×32 block sizes were selected appropriately for image filtering and allowed optimal use of GPU resources. As a result, large blocks (16x16 and 32x32) of images were filtered much faster than sequential image filtering on the CPU and small blocks (8x8) on the GPU due to the improved use of shared memory. Parallel image filtering on 8, 16 and 32 block sizes was accelerated by an average of 96.26% for the Gaussian filter and 91.41% for the Sobel filter compared to a conventional processor (CPU).