A Convolutional Neural Network Accelerator Architecture with Fine-Granular Mixed Precision Configurability