A SystemC profiling framework to improve fixed-point hardware utilization
Resumo
Common hardware design specifications usually describe the worst case scenario to improve fault tolerance and the quality of the final product. When taken to extreme, this approach can also lead to hardware overestimation, resulting on extra logic which, in real scenarios, may never be exercised. In this paper, we propose a SystemC profiling framework to help developers improve fixed-point hardware utilization by reporting unused bits and providing insights on data usage through the source code. We tested our framework against a set of publicly available synthesizable benchmarks (S2CBench), obtaining improvements in latency and area, demonstrating the potential of this technology. We also show a case study based on a proprietary Digital Signal Processor block that had their configurations tuned, for a year, by a group of specialist designers. According to our analysis, the methodology presented in this paper reduced bit requirements on some intermediate computations as much as 55.6%, reducing the energy consumption by 2.46%.
Palavras-chave:
Hardware, Tools, Observers, Optimization, Numerical models, Runtime, Digital signal processors, Fixed-point, SystemC, Profiling, Hardware Modeling, Mixed-Precision
Publicado
24/08/2020
Como Citar
LINHARES, Alisson; RUSA, Henrique; FORMIGA, Daniel; AZEVEDO, Rodolfo.
A SystemC profiling framework to improve fixed-point hardware utilization. In: SYMPOSIUM ON INTEGRATED CIRCUITS AND SYSTEMS DESIGN (SBCCI), 33. , 2020, Evento Online.
Anais [...].
Porto Alegre: Sociedade Brasileira de Computação,
2020
.
p. 163-168.