Suggestions for Implementing a Fast Ieee Multiply-add-fused Instruction

We studied three possible strategies to overlap the operations in a floating-point add (FADD) and a floating-point multiply (FMPY) for implementing a multiply-add-fused (MAF) instruction, whose result would be compatible with the IEEE floating-point standard. The operations in FMPY and FADD are: (a) non-overlapped, (b) fully-overlapped, and (c) partially-overlapped. The first strategy corresponds to multiply-add-chained (MAC) widely used in vector processors. The second (Greedy) strategy uses a greedy algorithm, yielding an implementation similar to the IBM RS/SOOO one. The third and final (SNAP) strategy uses a less aggressive starting configuration and corresponds to the Stanford Nanosecond Arithmetic Processor (SNAP) implementation. Two observations have prompted this study. First, in the IBM RS/GOOO implementation, the design tradeoffs have been made for high internal data precision, which facilitates the execution of elementary functions. These tradeoff decisions, however, may not be valid for an IEEE-compatible MAF. Second, the RS/6000 implementation assumed a different critical path for FADD and FMPY, which does not reflect the current state-of-the-art in floating-point technology. Using latency and hardware costs as the performance metrics we show that: (1) MAC has the lowest FADD latency and consumes the least hardware. But its MAF latency is the highest. (2) Greedy has an intermediate MAF latency but the highest FADD latency. And finally (3) SNAP provides the lowest MAF latency at the expense of a small increase in FADD latency over MAC and in area over Greedy. Both Greedy and SNAP have higher design complexity arising from rounding for the IEEE standard. SNAP has an additional wire complexity, which Greedy does not have because of its simpler datapath. If rounding for the IEEE standard is not a requirement, the Greedy strategy and therefore the RS/6000 seems a reasonable middle ground for applications with a high MAF to FADD ratio.