ETD

• data stream processing
• DSP
• FPGA
• framework
• parallel computing

The Data Stream Processing (DSP) paradigm studies novel algorithms and parallel processing techniques to gain valuable information from data streams. Accelerating DSP with FPGAs is a promising idea since such devices provide low latency, high energy efficiency, and low hardware engineering cost. Unfortunately, state-of-art DSP systems do not provide support for those devices. In this thesis, our aim is to determine if FPGAs can be considered a viable solution for accelerating DSP applications, and we propose a code generation approach which, starting from a high-level API in Python, allows programmer to easily instantiate streaming pipelines. We demonstrate the effectiveness of this approach, in terms of programmability and performance. We give a careful analysis of different ways to parallelize streaming operators, and we show the performance of the different techniques on a set of synthetic benchmarks.