Quantization of constrained processor data paths applied to convolutional neural networks