Reaearch Of The Key SIMD Function Unit In DSP

The application in embedded system is developping towards large-scale and realtime in recent years. High performance function unit is the key element of thedevelopment of processor. Based on research of high performance float point DSP, thispaper studies subword SIMD architecture and proposed several Subword calculationarithmetic. The main work of the paper is as follow.1) Based on special characteristic of function unit, the performance model offunction unit in various applications is founded. The speed up of multiplication andaddition is studied. The speed up/enegy is being described.2) This article proposed a subword SIMD multiplication arithmetic which is highbandwidth, high performance and low hardware spending. A novel Booth code methodis been used. The ES code and CS code has been segmented rationally. The arithmeticsupports three work modes. In the paper, a multiplier which has capable of supportingone64, four32or sixteen16bits signed or unsigned multiply is described. A carrycompensate arithmetic is proposed to support the arithmetic used in the dot-multiply.3) The Galois multiplier is described in this paper. The multiplier support subwordSIMD. The Galois multiplier has an advantage of both bandwidth and field ischangeable.4) Based on describe to various addition arithmetics, a novel subword additionarthmeitc is describe in this article. The arthmetics has the characteristic of highperformance and easy to be expanded.5) All arthmeitcs aboved is being used in FT-X DSP. All function units are beingcarefully designed. Function verification and synthesize result is shown at the end of thearticle.The multiplication arthmeitc has a small delay and powerful function. Thesynthesize result shown that the arithmetic decrease the delay about4%. The additionarithmetic shown in the article enhanced the performance of subword SIMD about11%.The M fuction unit based on the mulitiplication arithmetic satisfied the requirement ofinstruction set. And the function unit’s peak performance reaches1GHz in theenvironment of TSMC40nm with area spending of142275(um2) and powerconsumption28.6863(mW).