This paper systematically clarifies granular rough theory from three aspects for its motivation, theory and implementation. Three expectations that motivate granular rough theory are analyzed as follows: 1) to emphasize representative semantics of roughness, with explicitly encoded semantic contexts in underlying representation model; 2) to extend applicability of roughness to a wider range of information sources, with representation model designed to accommodate semi-structured data; 3) to describe a variety of application contexts of information structure, to adapt roughness methodology to disciplines driven by mereology, and to exhibit potentials of combining mereology and computer science in the sense of developing innovative interdisciplinary methodologies, with a pure mereological approach to roughness. From theoretic perspective, granular representation calculus is defined, which plays the role of common representation model for both ordinary information sources and roughness methodology. In terms of this model, corresponding to the notion of lower approximation, border region and upper approximation for roughness, Kernel granule, hull granule and corpus granule are constructed respectively. From pragmatic perspective, upon open source implementation of "Entity-Attribute-Value" model, a rapid prototyping method for granular rough theory is described to provide a test-bed for verification purpose and to apply the roughness methodology for analyzing clinical data more naturally. Significance of granular rough theory, some open problems and further research are summarized.