A novel approach for the creation of a multimedia processor is presented in this paper. Three steps to customize a general-purpose processor (GPP) based on the detailed analysis of a target application are described, and these steps are demonstrated on some multimedia applications. The approach consists of analysis of an application in terms of memory requirements, potential parallelism, and control operations. These metrics are further used to customize a GPP with additional dedicated hardware modules.

The first step is the analysis of the targeted application using three metrics. The first is gama, an indicator of the amount of parallelism present in a subroutine. Functions with a high gama value should be implemented in hardware because they have a higher potential for acceleration with dedicated circuits. Functions with a small gama value do not result in any speedup with dedicated hardware, hence they may be mapped onto a GPP. The memory-oriented metric (MOM) is an indicator of the frequency of the global memory accesses (accesses to data not present in registers) in a function, normalized to one. Applications with higher MOMs require good data bandwidth from the lower levels of the memory, via big local memories or parallel data accesses. The third metric, control-oriented metric (COM), is a measure of the control operations in a function (such as greater than, less than, and so on) that cannot be resolved at compile time. Functions with higher COM values should be implemented in a GPP rather than in a specialized hardware processor because such functions require the implementation of rather large and complicated state machines. The second step is the customization of a GPP, and the third step is the architecture generation.

The above three steps constitute a framework for a detailed analysis of an application from its description in a higher level language. The paper argues that this framework is well suited to improve application-architecture matching. It is a first step in the development of embedded devices with energy, time, and quality of service tradeoffs.