[Next] [Previous] [Top]
Software Synthesis for DSP Using Ptolemy
Although code generation is beginning to mature in Ptolemy, it is by no means finished. We have only begun to explore buffer management techniques to use memory more efficiently, Currently, in the assembly language domains, all stars must communicate through memory, not registers. Hence, the more fine-grained a star is, the more penalty it suffers. For example, a simple add star must first read in its two inputs from memory and then write its output to memory. Even though a simple operation like add might take one cycle on a DSP, the add could potentially take four or more cycles. Future versions of Ptolemy will use registers to exchange data, as done in [1]. Because there are no data-dependent decisions in the SDF domain, it is possible in principle to do more efficient register allocation than can be done for more conventional high-level languages (although since the problem of optimal register allocation, like so many others in this area, has combinatorial complexity, heuristics still must be used).
Work is in progress to extend our code generation techniques to support more general models of computation, such as the token flow model [19] and dynamic constructs [14].
Software Synthesis for DSP Using Ptolemy - 04 SEP 94
[Next] [Previous] [Top]