High-level synthesis of functional patterns with lift

Abstract

High-level languages are commonly seen as a good fit to tackle the problem of performance portability across parallel architectures. The Lift framework is a recent approach which combines high-level, array-based programming abstractions, with a system of rewrite-rules that express algorithmic as well as low-level hardware optimizations. Lift has successfully demonstrated its ability to address the challenge of performance portability across multiple types of CPU and GPU devices by automatically generating code that is on-par with highly optimized hand-written code. This paper demonstrates the potential of Lift for targeting FPGA-based platforms. It presents the design of new Lift parallel patterns operating on data streams, and describes the implementation of a Lift VHDL backend. This approach is evaluated on a Xilinx XC7Z010 FPGA using matrix multiplication, leading to a 10x speed-up over highly optimized CPU code and a commercial HLS tool. Furthermore, by considering the potential of design space exploration enabled by Lift, this work is a stepping stone towards automatically generated competitive code for FPGAs.