Optimizing convolutional neural network on DSP

Abstract

Deep learning techniques like Convolutional Neural Networks (CNN) are getting traction for classification of objects (e.g. traffic signs, pedestrian, vehicles etc.) in Advanced Driver Assistance Systems (ADAS). Typical CNN based trained networks poses huge computational complexity in feed forward path during operation due to multiple layers and within layer operations like 2D convolution, spatial pooling and non-linear mapping. The paper proposes optimization techniques to efficiently map such networks on Digital Signal processors (DSP). These techniques consist of fixed point conversion, data re-organization, weight placement and LUT usage resulting in optimal utilization of resources on C66x™ DSP. The proposed kernels are developed and simulated on Texas Instruments (TI)'s Driver Assist TDA3X platform with optimal utilization of compute and data resources inside DSP. These optimization techniques are applicable for multiple network topologies published in the literature.