高性能处理器.ppt

《高性能处理器.ppt》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、ControlHazardSolution（3）:软件循环展开减少分支指令例：向量的每个元素加常数For(j=1;j<=1000;j++)x[j]=x[j]+s;编译：整数寄存器R1：循环计数器，初值为向量中最高端地址元素的地址浮点寄存器F2：保存常数s1Loop:LDF0,0(R1);F0=vectorelement2ADDF4,F0,F2;addscalarinF23SD0(R1),F4;storeresult4SUBIR1,R1,8;decrementpointer8B(DW)5BNEZR1,Loop;branchR1!=zeroScalarvs.Superscala

2、rvs.VLIWsingle-threaded:(a)scalar(b)superscalar(c)VLIW(a)Time(processcycles)(a)Time(processorcycles)Issueslots(c)NNNNNNSingleissuevs.multipleissueprocessorvonNeumannprocessor(Scalarprocessor)asingleprogramcounterreally:IPC≤1ILPprocessor(Superscalar,VILW)asingleprogramcounterreally:IPC≤2(wh

3、en4~8issue)IMTvs.BMT(c)ContextswitchNNNNNNNNNNN(b)Contextswitch(a)multi-threaded:(a)cycle-by-cycleinterleavingsuperscalar（Fine-GrainedMultithreading）(b)blockinterleavingsuperscalar（Course-GrainedMultithreading）(c)cycle-by-cycleinterleavingVLIW（Fine-GrainedMultithreading）ContextswitchGlobal

4、BranchPredictionScheme:设计思想例：考虑下面的代码段：if(aa==2)//branchb1aa=0;if(bb==2)//branchb2bb=0;if(aa!=bb){//branchb3……}这段代码由编译器转换为以下汇编程序（假定aa和bb分别分配为R1和R2）：SUBIR3，R1，#2BNEZR3，L1；branchb1(aa!=2),takenADDR1,R0,R0；aa==0,nottakenL1:SUBIR3，R2，#2BNEZR3，L2；branchb2(bb!=2),takenADDR2,R0,R0；bb==0,nottakenL2

5、:SUBR3，R1，R2；R3=aa-bbBEQZR3，L3；branchb3(aa==bb),taken通向b3的路径A：0-0B：0-1C：1-0D：1-1在b3执行之前可以推测aa=0bb=0aa=0bb2aa2bb=0aa2bb2每次循环需要9clocksRewritecodetominimizestalls?InstructionInstructionLatencyinproducingresultusingresultclockcyclesFPALUopAnotherFPALUop3FPALUopStoredouble2LoaddoubleFPALUo

6、p11Loop:LDF0,0(R1);F0=vectorelement2stall3ADDDF4,F0,F2;addscalarinF24stall5stall6SD0(R1),F4;storeresult7SUBIR1,R1,8;decrementpointer8B(DW)8BNEZR1,Loop;branchR1!=zero9stall;delayedbranchslot每次循环都有数据相关/控制相关引起的停顿每次循环需要6clocksUnrollloop(循环展开)4timescodetomakefaster?1Loop:LDF0,0(R1)2stall3ADDDF4

7、,F0,F24SUBIR1,R1,85BNEZR1,Loop;delayedbranch6SD8(R1),F4;alteredwhenmovepastSUBIPutSUBItoADDD’sfirststallslotAddjustthedisplacementofSDfrom0(R1)to8(R1)(3)SwapBNEZandSDbychangingaddressofSD静态调度的例子循环展开+寄存器重命名（RegisterRenaming）的例子1Loop:LDF0,0(R1)2ADDDF4,F0,F23SD0(