If one considers an embedded processor connected to a ROM, it may be viewed as one large FSM. Since for any given piece of software, the ROM is unlikely to be full and there are likely to be resources in the processor that are not used by that software: the application of a good quality logic minimiser to the system, while it is in the design database, could trim it greatly. In most real designs, this will not be helpful: for instance, the advantages of full-custom applied to the processor core will be lost. In fact, the minimisation function may be too complex for most algorithms to tackle on today's computers.
On the other hand, algorithms to create a good static scheduling of a fixed number of hardware resources work quite well. A processing algorithm typically consists of multiple processing stages (e.g. called pre-emphasis, equalisation, coefficient adaptation, FFT, deconvolution, reconstruction and so on). Each of these steps normally has to be done within tight real-time bounds and so parallelism through multiple instances of ALU and register hardware is needed. The Cathedral DSP compiler was an early tool for helping design such circuits. Such tools can perform time/space folding/unfolding of the algorithm to generate the static shedule that maps operations and variables in a high-level description to actual resources in the hardware. Cache misses, contention for resources and operations that have data-dependent runtime will cause time-domain deviations from a static schedule, so a potentially a dynamic schedule could then make better use of resources but the overhead of dynamic scheduling can outweigh the cost of the resources saved if the data dependant variations are rare.
Custom hardware is generally much more energy efficient that general-purpose processors. Reasons include (also see DJG 9-points in `accelerators' section)
| 3: (C) 2012-18, DJ Greaves, University of Cambridge, Computer Laboratory. |