Abstract:
The reconfigurable computing (RC) aims to combine the flexibility of General-Purpose Processor (GPP) with performance of Application Specific Integrated Circuits (ASIC). There are several architectures proposed since RC’s inception in 1960s, but all have failed to become mainstream. The main factor preventing RC to become common practice is its requirement for implementers of algorithms (programmers) to be familiar with hardware design. In RC, a hardened processor cooperates with a dynamic reconfigurable Hardware Accelerator (HA) which is implemented on Field-Programmable Gate Array (FPGA). The HA implements crucial software kernel on hardware to increase performance and its design demands digital circuit expertise. In this paper a novel RC architecture is proposed that keeps the decades old programming practices intact while harnessing the power of HA. The architecture uses LLVM compiler infrastructure to receive an algorithm and then outputs the equivalent machine language, it then finds the most frequent instruction pairs and generates equivalent RC circuit called “Miniature Accelerator (MA)”. The instruction pairs are dynamically removed from pipeline and MA computed result replaces them in parallel. To demonstrate the concept the Fast Fourier Transform (FFT) algorithm which is core Digital signal processing (DSP) kernel is written in C and then executed on an ARM Cortex-M0. The execution of FFT function is improved by 14.12%. The proposed adaptive processor is fully backward compatible, compilation is automated, and no modification of exiting software or established programming paradigms is required.
