Exploring Perceptual Limitation of Multimodal Large Language Models