Optimizing Deep Learning Workloads on ARM GPU with TVM
Lanmin Zheng, Tianqi Chen · 2018
With the great success of deep learning, the demand for deploying deep neural networks to mobile devices is growing rapidly. However, current popular deep learning frameworks are often poorly optimized for mobile devices, especially mobile GPU. In this paper, we follow the pipeline proposed by TVM/NNVM, and optimize both kernel implementations and dataflow graph for ARM Mali GPU. Compared with vendor-provided ARM Compute Library, our kernel implementations and end-to-end pipeline are 1.7x faster on VGG16 and 2.2x faster on mobilenet.