high performance computing on graphics processing units: hgpu.org

Computer vision algorithms are becoming increasingly important in mobile, embedded, and wearable devices and applications. These compute-intensive workloads are challenging to implement with good performance and power-efficiency. In many applications, implementing critical portions of computer vision workloads on a general-purpose graphics processing unit (GPU) is an attractive solution. Qualcomm enables programming of the Adreno GPU in its Snapdragon application processors via the open standard OpenCL language and API. OpenCL support enables programmers to offload computer vision algorithm kernels to the GPU, which in turn provides speed and power-consumption advantages over a CPU implementation. BDTI developed an Android application demonstrating computer vision functionality utilizing the Adreno 420 GPU in Qualcomm’s Snapdragon 805. The BDTI application can run compute-intensive computer vision functions on the GPU or the CPU, enabling comparisons of GPU and CPU performance and power-efficiency. This paper discusses the BDTI application, implementation and optimization techniques used in its development, and the substantial benefits observed when offloading compute-intensive kernels to the GPU.