Multiply-add fused float point unit with on-fly denormalized number processing
Hu He, Zheng Li, Yihe Sun · 2005
Denormalized numbers are the most difficult type of numbers to implement in float-point units. They are so complex that many designs avoid handling them in hardware. The denormalized number processing cost two much extra clock cycle in software implementation. In this paper, an on-fly floating point denormalized number processing implemented in a multiply-add-fused (MAF) with little extra latency is presented. The denormalized number processing is embedded in a popular MAF data path and fused with the MAF smoothly by representing the denormalized number. The extra latency introduced by the denormalized number processing is cost by the denormalized number detection