Optimizing Performance of Image Processing Algorithms on GPUs
Honghui Zhou, Ruyi Qin, Zihan Liu, Ying Qian, Xiaoming Ju · Lecture notes in electrical engineering · 2022
Abstract The application of machine learning algorithms in the field of power grid improves the service level of power enterprises and promotes the development of power grid. NVIDIA Volta and Turing GPUs powered by Tensor Cores can accelerate training and learning performance for these algorithms. With Tensor Cores enabled, FP32 and FP16 mixed precision matrix multiplication dramatically accelerates the throughput and reduces AI training times. In order to explore the cause of this phenomenon, we choose a convolutional neural network (CNN), which is widely used in computer vision, as an example and show the performance characteristics with tensor core on general matrix multiplications and convolution calculations as benchmark. Building a CNN based on cuDNN and TensorFlow, we analyze the performance of CNN from various aspects and optimize performance of it by changing the shape of convolution kernel and using texture memory, etc. The experimental results prove the effectiveness of our methods.