UHCL-Darknet
Longlong Liao, Kenli Li, Keqin Li, Canqun Yang, Qi Tian · 2018
As the majority of popular deep neural network (DNN) frameworks focus on a closed format CUDA implementations based on one or more NVIDIA GPUs, they cannot efficiently leverage other devices in cluster mode to accelerate the training and inference of DNNs except NVIDIA GPUs. To accelerate DNNs using heterogeneous multi-/many-core clusters, we propose an OpenCL-based DNN framework called UHCL-Darknet. First, we design a unified OpenCL platform model for the heterogeneous cluster called UHCL, and an adaptive runtime system with the affinity-based dynamic scheduler for UHCL, enabling transparent utilization of a wide variety of vendor-specific OpenCL devices in the heterogeneous cluster. Then, we extend Darknet to UHCL by introducing the parallel optimization of DNNs, such as paralleling Winogrand-based convolutions and auto-tuning parameterized OpenCL kernels. The training and inference of art-of-data DNN models (e.g., YOLOv2, ResNet-50, and DenseNet-201) are executed on an experimental heterogeneous cluster. Results show that UHCL-Darknet is a scalable and portable DNN framework for heterogeneous clusters, and achieves 1.9x and 2.2x speedups on average respectively for the image throughput of data-parallel training and inference on the experimental heterogeneous cluster.