General Overview of an Adaptive Dynamic Extensible Processor

This paper describes an approach for adaptive dynamic instruction set extension, tuning embedded processors to specific applications which are going to be executed in their life-time. These new instructions are generated after production. The processor has two modes: training mode and normal mode. Training mode can be done offline or online. The applicationspecific instructions are extracted from the critical portions of the code detected by a binary-level profiler at training mode. At normal mode custom instructions are executed on a reconfigurable coarse grain accelerator. The sequencer (an augmented hardware to the base processor) decides when, which custom instruction should be executed and switches between functional unit and accelerator outputs. It also takes the responsibility for switching between different accelerator configurations. In this methodology there is no need to a new compiler, extra opcodes or recompiling the source code. This paper provides a general overview of the system.