Optimizing Vision Benchmark for GPU

Abstract

The San Diego Vision Benchmark Suite (SD-VBS), a suite of diverse vision appli- cations drawn from the vision domain. From the assembly line to home entertain- ment systems, the need for e cient real-time computer vision systems is growing rapidly.NVIDIA has developed the CUDA (Compute Uni ed Device Architecture) which can be used to speedup computer vision application. CUDA enables soft- ware developers to access the GPU through standard programming languages such as 'C'. It also gives developers access to the GPU's virtual instruction set, onboard memory and the parallel computational elements. Taking advantage of parallel computation will signi cantly speedup applications. Here we explore the potential power of using CUDA and NVIDIA GPUs to speedup common computer vision algorithms along with algorithmic optimizations. Approaches to optimize few ap- plications of SD-VBS on GPU are part of this thesis. To analyze simulation time of these applications inputs of di erent size are feeded.