high performance computing on graphics processing units: hgpu.org

The main contribution of this thesis is the design and development of an optimized framework to realize the deep neural classifiers on the embedded platforms. Deep convolutional networks exhibit unmatched performance in image classification. However, these deep classifiers demand huge computational power and memory storage. That is an issue on embedded devices due to limited onboard resources. The computational demand of neural networks mainly stems from the convolutional layers. A significant improvement in performance can be obtained by reducing the computational complexity of these convolutional layers, making them realizable on embedded platforms. In this thesis, we proposed a CUDA (Compute Unified Device Architecture)- based accelerated scheme to realize the deep architectures on the embedded platforms by exploiting the already trained networks. All required functions and layers to replicate the trained neural networks were implemented and accelerated using concurrent resources of embedded GPU. Performance of our CUDA-based proposed scheme was significantly improved by performing convolutions in the transform domain. This matrix multiplication based convolution was also compared with the traditional approach to analyze the improvement in inference performance. The second part of this thesis focused on the optimization of the proposed framework. The flow of our CUDA-based framework was optimized using unified memory scheme and hardware-dependent utilization of computational resources. The proposed flow was evaluated over three different image classification networks on Jetson TX1 embedded board and Nvidia Shield K1 tablet. The performance of proposed GPU-only flow was compared with its sequential and heterogeneous versions. The results showed that the proposed scheme brought the higher performance and enabled the real-time image classification on the embedded platforms with lesser storage requirements. These results motivated us towards the realization of useful real-time classification and recognition problems on the embedded platforms. Finally, we utilized the proposed framework to realize the neural network-based automatic license plate recognition (ALPR) system on a mobile platform. This highly-precise and computationally demanding system was deployed by simplifying the flow of trained deep architecture developed for powerful desktop and server environments. A comparative analysis of computational complexity, recognition accuracy and inference performance was performed.