Gerald George Pechanek

Any-size instruction abbreviation technique for embedded DSPs

There exist a whole class of systems, which presents critical requirements for code density, efficient memory usage, low power and performance. A representative of this class are embedded DSP systems for SOC. This work presents a method for entropy-bounded encoding of an original ISA and decoupling it from a DSP core. The encoding allows the instruction storage to be used with high efficiency, which is only bounded by the information contents of an application, and relaxes any restrictions imposed on the ISA by the physical memory and branching mechanism. The concept is illustrated with an exemplary commercial DSP processor showing a reduction of the required instruction memory space of greater than 40% without significant impact on the instruction fetch stage of the DSP pipeline.