High-level Synthesis Tools for High-throughput Digital Signal Processing Applications (High2) (TES.5226)

Important classes of embedded systems are systems for digital signal processing, DSP, in which algorithms are applied to streams of signals. The variety of these algorithms is large and encompasses algorithms for filtering audio and video signals, algorithms for error-correction, compression, picture-in-picture, etcetera. Modern multi-media systems could not exist without this kind of embedded systems. The algorithms are all characterized by a repeated application of some functions on a stream of input values, resulting in a stream of output values.

Important parameters from the application domain which strongly influence the synthesis method are the throughput and the synchronicity. This proposal is focussed on high-throughput applications. With high-throughput only less than 100 time steps are available for executing a function once.

Real-time constraints play a crucial role in the design of high-throughput DSP applications. The available hardware must be able to perform the specified function at any time and with a sufficient small latency. In case the input is provided as a synchronous stream of equally sized blocks and the control flow of the algorithm is not data dependent, currently, synthesis algorithms for high throughput applications are available which give (near) optimal schedules. These synthesis algorithms are based on integer linear programming. Unfortunately, if the input streams are not fully synchronous, or in case of data dependent control flow, no optimal synthesis algorithms are available. Scheduling and resource allocation in that case rely on heuristic algorithms.

A second problem of this kind of synthesis tools is that they are often only suitable for a parts of a design. So they have to cooperate with other synthesis tools. This poses additional requirement to the synthesis tool.

This proposal aims at diminishing these disadvantages of not being able to handle irregular input streams or data dependent control flow, without losing the advantage of being able to determine optimal schedules (by means of integer linear programming). Moreover we will investigate how such a synthesis tool needs to be adapted in order to fit into existing design flows.

For results, see http://wwwhome.cs.utwente.nl/~krol/projects-current/High2.htm.

Four companies are involved in this project.