基于粗粒度可重构处理器的浮点乘加算法

Floating-Point Multiplication and Addition Algorithm Based on Coarse-Grained Reconfigurable Processor

在粗粒度可重构处理器中,往往采用以定点代替浮点或嵌入专用的浮点计算单元的方式来处理应用中的浮点运算,从而导致在面对大动态范围运算时精度不够,造成系统面积与功耗的增加.本文提出了一种在通用粗粒度可重构处理器上用定点运算单元实现浮点乘加运算的方法,采用8个可重构处理单元实现1次乘或加的浮点运算,该方法不仅兼容了IEEE-754的单精度浮点标准而且没有增加任何浮点运算硬件.在模拟器上对系统性能进行测试,使用本文的方法,在通用粗粒度可重构处理器上实现浮点乘法运算性能提升2.09倍,浮点加法运算性能提升1.68倍.

For coarse-grained reconfigurable processors,there are two approaches to dealing with floating-point operations.The first one is using fixed-point operations instead of floating-point,and the second one is embedding a floating-point calculation unit.But both the two methods will cause low accuracy in large dynamic range calculation,and increase the system area and power consumption.This paper presents a method of floating-point multiplication and addition implemented on a general coarse-grained reconfigurable processor with fixed-point arithmetic units by using eight reconfigurable processing units to implement a floating-point multiplication or addition operation.The method is compatible with the IEEE-754single precision floating point standard but does not add any floating-point arithmetic hardware.The performance of the system is evaluated by SOC designer.The results show that the performance of floating-point multiplication operation improves2.09times,and addition operation improves1.68times compared with the conventional one on the general coarse-grained reconfigurable processor architecture.