Automated Application-Specific Instruction Generation For TTA-Based ASIP Design

Computing performances and flexibility are critical and conflicting design goals in embedded system development. In the recent years, embedded applications, especially mobile devices, have been increasing requirement for these microprocessors with high performances, small sizes, and low power dissipation. The applications in one device are always distinct from each other on computing characteristics. Since the general-purpose microprocessor cannot satisfy these applications, ASIC are always used to solve these challenging problems. However, ASIC are costly to design and always short of flexibility to meet the computing characteristics of different applications. It is known that application-specific instruction set processor (ASIP) provides a good tradeoff between efficiency and flexibility.Recent ASIP researches are always based on a flexible and programmable architecture template. In this paper we propose the usage of Transport Triggered Architectures (TTA) for such an ASIP researches architecture template. We are particularly interested in automated application-specific instruction generation which is one field of ASIP important research fields. Designing application specific instructions is critical to ASIP design. However, application specific instructions are difficult to design and maybe iterate design for a lot of times. There is a great need for tools for supporting the designers.The impacts of hardware costs are not fully considered in current researches when selecting instructions. Moreover, the instruction patterns are fixed on the single output or chain-shaped or tree-shaped instruction patterns. In this paper, we design and implement a fully automatic extending application specific instructions system for specific applications based on TTA. The characteristics of our systems are as follows: firstly it can automatically detects most frequently occurring operation template, and uses some optimization techniques to reduces the complexity of searching the data flow graphs; secondly it can support the non-tree-shaped, multi-output operation pattern; thirdly it can automatically modify software toolsets.The system is tested with several benchmarks from security and telecomm domains. The results demonstrate our method gains a significant performance enhancement.