CS 152: Computer Architecture and Engineering Lecture 14 Advanced Pipelining/Compiler Scheduling Randy H. Katz, Instructor Satrajit Chatterjee, Teaching Assistant George Porter, Teaching Assistant CS 152 Lec 14.1 Review: Pipelining • Key to pipelining: smooth flow – Making all instructions the same length can increase performance! • Hazards limit performance – Structural: need more HW resources – Data: need forwarding, compiler scheduling – Control: early evaluation & PC, delayed branch, prediction • Data hazards must be handled carefully: – RAW (Read-After-Write) data hazards handled by forwarding – WAW (Write-After-Write) and WAR (Write-After-Read) hazards don’t exist in 5-stage pipeline • MIPS I instruction set architecture made pipeline visible (delayed branch, delayed load) – Change in programmer semantics to make hardware simpler CS 152 Lec 14.2 Recap: Data Hazards I-Fet ch DCD MemOpFetch OpFetch IFetch Structural Hazard I-Fet ch DCD DCD OpFetch Jump IFetch IF DCD EX IF Mem WB DCD EX IF DCD IF °°° Control Hazard °°° WB Mem WB WAW Data Hazard (write after write) DCD IF Store RAW (read after write) Data Hazard Mem DCD EX Exec DCD OF OF Ex RS Ex Mem WAR Data Hazard (write after read) CS 152 Lec 14.3 Recap: Data Stationary Control • Main Control generates control signals during Reg/Dec – Control signals for Exec (ExtOp, ALUSrc, ...) are used 1 cycle later – Control signals for Mem (MemWr Branch) are used 2 cycles later – Control signals for Wr (MemtoReg MemWr) are used 3 cycles later Reg/Dec ALUSrc ALUSrc ALUOp ALUOp RegDst MemWr Branch MemtoReg RegWr RegDst MemWr Branch MemtoReg RegWr MemWr Branch MemtoReg RegWr Wr Mem/Wr Register ExtOp Mem Ex/Mem Register ExtOp ID/Ex Register IF/ID Register Main Control Exec MemtoReg RegWr CS 152 Lec 14.4 Review: Resolve RAW by “Forwarding” (or Bypassing) IAU npc I mem Regs op rw rs rt Forward mux B A im n op rw alu S n op rw D mem m Regs PC • Detect nearest valid write op operand register and forward into op latches, bypassing remainder of the pipe • Increase muxes to add paths from pipeline registers • Data Forwarding = Data Bypassing n op rw CS 152 Lec 14.5 Question: Critical Path??? • Bypass path is invariably trouble • Options? – Make logic really fast – Move forwarding after muxes Regs PC Sel Forward mux B A alu im Equal » Problem: screws up branches that require forwarding! » Use same tricks as “carry-skip” adder to fix this? » This option may just push delay around….! S – Insert an extra cycle for branches that need forwarding? D mem » Or: hit common case of forwarding from EX stage and stall for forward from memory? m Regs CS 152 Lec 14.6 What About Interrupts, Traps, Faults? • External Interrupts: – Allow pipeline to drain, Fill with NOPs – Load PC with interrupt address • Faults (within instruction, restartable) – Force trap instruction into IF – Disable writes till trap hits WB – Must save multiple PCs or PC + state • Recall: Precise Exceptions State of the machine is preserved as if program executed up to the offending instruction – All previous instructions completed – Offending instruction and all following instructions act as if they have not even started – Same system code will work on different implementations CS 152 Lec 14.7 Exception/Interrupts: Implementation questions 5 instructions, executing in 5 different pipeline stages! • Who caused the interrupt? Stage Problem interrupts occurring IF Page fault on instruction fetch; misaligned memory access; memory-protection violation Undefined or illegal opcode Arithmetic exception Page fault on data fetch; misaligned memory access; memory-protection violation; memory error ID EX MEM • How do we stop the pipeline? How do we restart it? • Do we interrupt immediately or wait? • How do we sort all of this out to maintain preciseness? CS 152 Lec 14.8 Exception Handling IAU npc I mem Regs B lw $2,20($5) A S D mem Regs n op rw PC Excp detect bad instruction Excp detect overflow alu m im detect bad instruction address Excp detect bad data address Excp Allow exception to take effect CS 152 Lec 14.9 Another Look at the Exception Problem Time Bad Inst Inst TLB fault Overflow IFetch Dcd Exec IFetch Dcd Program Flow Data TLB Mem WB Exec Mem WB Exec Mem WB Exec Mem IFetch Dcd IFetch Dcd WB • Use pipeline to sort this out! – Pass exception status along with instruction. – Keep track of PCs for every instruction in pipeline. – Don’t act on exception until it reache WB stage • Handle interrupts through “faulting noop” in IF stage • When instruction reaches end of MEM stage: – Save PC EPC, Interrupt vector addr PC – Turn all instructions in earlier stages into noops! CS 152 Lec 14.10 Resolution: Freeze Above & Bubble Below IAU npc I mem freeze Regs op rw rs rt PC bubble B A im n op rw alu S n op rw D mem m Regs • Flush accomplished by setting “invalid” bit in pipeline n op rw CS 152 Lec 14.11 FYI: MIPS R3000 clocking discipline phi1 phi2 • 2-phase non-overlapping clocks • Pipeline stage is two (level sensitive) latches Edge-triggered phi1 phi2 phi1 CS 152 Lec 14.12 MIPS R3000 Instruction Pipeline Decode Reg. Read Inst Fetch TLB I-Cache RF ALU / E.A Memory Operation E.A. TLB Write Reg WB D-Cache Resource Usage TLB TLB I-cache RF WB ALUALU D-Cache Write in phase 1, read in phase 2 => eliminates bypass from WB CS 152 Lec 14.13 Recall: Data Hazard on r1 xor r10,r1,r11 Reg Dm Im Reg Dm Im Reg Dm Im Reg ALU or r8,r1,r9 W B Reg ALU and r6,r1,r7 Im ME M Dm ALU sub r4,r1,r3 E X ALU O r d e r add r1,r2,r3 ID/R FReg ALU I n s t r. Time (clock cycles) I F Im Reg Reg Reg Dm Reg With MIPS R3000 pipeline, no need to forward from WB stage CS 152 Lec 14.14 MIPS R3000 Multicycle Operations Use control word of local stage to step through multicycle operation op Rd Ra Rb Stall all stages above multicycle operation in the pipeline mul Rd Ra Rb Rd A B R Rd T to reg file Drain (bubble) stages below it Alternatively, launch multiply/divide to autonomous unit, only stall pipe if attempt to get result before ready - This means stall mflo/mfhi in decode stage if multiply/divide still executing - Extra credit in Lab 5 does this Ex: Multiply, Divide, Cache Miss CS 152 Lec 14.15 Is CPI = 1 for our pipeline? • Remember that CPI is an “Average # cycles/inst IFetch Dcd Exec IFetch Dcd Mem WB Exec Mem WB Exec Mem WB Exec Mem IFetch Dcd IFetch Dcd WB • CPI here is 1, since the average throughput is 1 instruction every cycle. • What if there are stalls or multi-cycle execution? • Usually CPI > 1. How close can we get to 1?? CS 152 Lec 14.16 Recall: Compute CPI? • Start with Base CPI • Add stalls CPI CPIbase CPI stall CPI stall STALLtype1 freq type1 STALLtype 2 freq type 2 • Suppose: – – – – CPIbase=1 Freqbranch=20%, freqload=30% Suppose branches always cause 1 cycle stall Loads cause a 100 cycle stall 1% of time • Then: CPI = 1 + (10.20)+(100 0.300.01)=1.5 • Multicycle? Could treat as: CPIstall=(CYCLES-CPIbase) freqinst CS 152 Lec 14.17 Administrivia • Get moving on Lab 5! – This lab is even harder than Lab 4. – Much trickier to debug … ! – Start with unpipelined version? » Interesting thought… May or may not help • Dynamic scheduling techniques discussed in the Other Hennessy & Patterson book: – “Computer Architecture: A Quantitative Approach” – Chapter 3, handout of relevant section for HW that is due Friday CS 152 Lec 14.18 Administrivia: Be Careful About Clock Edges in Lab5! D B Reg. File M Mem Acces s Data Mem PC Next PC Equal IRmem WB Ctrl Ex Ctrl Exec S IRwb IRex A Mem Ctrl Dcd Ctrl Reg File IR Inst. Mem Valid • Since Register file has edge-triggered write: – Must have everything set up at end of memory stage – This means that “M” register here is not necessary! CS 152 Lec 14.19 Case Study: MIPS R4000 (200 MHz) • 8 Stage Pipeline: – IF–first half of fetching of instruction; PC selection happens here as well as initiation of instruction cache access. – IS–second half of access to instruction cache. – RF–instruction decode and register fetch, hazard checking and also instruction cache hit detection. – EX–execution, which includes effective address calculation, ALU operation, and branch target computation and condition evaluation. – DF–data fetch, first half of access to data cache. – DS–second half of access to data cache. – TC–tag check, determine whether the data cache access hit. – WB–write back for loads and register-register operations. • 8 Stages: What is impact on Load delay? Branch delay? Why? CS 152 Lec 14.20 Case Study: MIPS R4000 TWO Cycle IF IS IF RF IS IF EX RF IS IF DF EX RF IS IF DS DF EX RF IS IF TC DS DF EX RF IS IF WB TC DS DF EX RF IS IF THREE Cycle IF IS IF RF IS IF EX RF IS IF DF EX RF IS IF DS DF EX RF IS IF TC DS DF EX RF IS IF WB TC DS DF EX RF IS IF Load Latency Branch Latency (conditions evaluated during EX phase) Delay slot plus two stalls Branch likely cancels delay slot if not taken CS 152 Lec 14.21 MIPS R4000 Floating Point • FP Adder, FP Multiplier, FP Divider • Last step of FP Multiplier/Divider uses FP Adder HW • 8 kinds of stages in FP units: Stage Functional unit Description A D E M N R S U Mantissa ADD stage Divide pipeline stage Exception test stage First stage of multiplier Second stage of multiplier Rounding stage Operand shift stage Unpack FP numbers FP adder FP divider FP multiplier FP multiplier FP multiplier FP adder FP adder CS 152 Lec 14.22 MIPS FP Pipe Stages FP Instr Add, Subtract Multiply Divide Square root Negate Absolute value FP compare Stages: M N R S U 1 U U U U U U U 2 S+A E+M A E S S A 3 4 5 A+R R+S M M M R D28 … (A+R)108 … 6 7 8 … N N+A R D+A D+R, D+R, D+A, D+R, A, R A R R First stage of multiplier Second stage of multiplier Rounding stage Operand shift stage Unpack FP numbers A D E Mantissa ADD stage Divide pipeline stage Exception test stage CS 152 Lec 14.23 R4000 Performance Not ideal CPI of 1: Base Load stalls Branch stalls FP result st alls tomcatv su2cor spice2g6 ora li gcc espresso 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 nasa7 FP structural stalls: Not enough FP hardware (parallelism) FP result stalls: RAW data hazard (latency) Branch stalls (2 cycles + unfilled slots) Load stalls (1 or 2 clock cycles) doduc – – – – eqntott • FP structural stalls CS 152 Lec 14.24 Can We Somehow Make CPI Closer to 1? • Let’s assume full pipelining: – If we have a 4-cycle instruction, then we need 3 instructions between a producing instruction and its use: multf $F0,$F2,$F4 delay-1 delay-2 delay-3 addf $F6,$F10,$F0 Earliest forwarding for 4-cycle instructions Earliest forwarding for 1-cycle instructions Fetch Decode addf Ex1 Ex2 Ex3 delay3 delay2 delay1 Ex4 WB multf CS 152 Lec 14.25 FP Loop: Where are the Hazards? Loop: • LD ADDD SD SUBI BNEZ NOP F0,0(R1) F4,F0,F2 0(R1),F4 R1,R1,8 R1,Loop ;F0=vector element ;add scalar from F2 ;store result ;decrement pointer 8B (DW) ;branch R1!=zero ;delayed branch slot Instruction producing result Instruction using result Latency in clock cycles FP ALU op FP ALU op Load double Load double Integer op Another FP ALU op Store double FP ALU op Store double Integer op 3 2 1 0 0 Where are the stalls? CS 152 Lec 14.26 FP Loop Showing Stalls 1 Loop: LD F0,0(R1) 2 stall 3 ADDD F4,F0,F2 4 stall 5 stall 6 SD 0(R1),F4 7 SUBI R1,R1,8 8 BNEZ R1,Loop 9 stall ;F0=vector element ;add scalar in F2 ;store result ;decrement pointer 8B (DW) ;branch R1!=zero ;delayed branch slot Instruction producing result Instruction using result Latency in clock cycles FP ALU op FP ALU op Load double Another FP ALU op Store double FP ALU op 3 2 1 • 9 clocks: Rewrite code to minimize stalls? CS 152 Lec 14.27 Revised FP Loop Minimizing Stalls 1 Loop: LD F0,0(R1) 2 stall 3 ADDD F4,F0,F2 4 SUBI R1,R1,8 5 BNEZ R1,Loop 6 SD 8(R1),F4 ;delayed branch ;altered when move past SUBI Swap BNEZ and SD by changing address of SD Instruction producing result Instruction using result Latency in clock cycles FP ALU op FP ALU op Load double Another FP ALU op Store double FP ALU op 3 2 1 6 clocks: Unroll loop 4 times code to make faster? CS 152 Lec 14.28 Unroll Loop Four Times (Straightforward Way) 1 Loop:LD 2 ADDD 3 SD 4 LD 5 ADDD 6 SD 7 LD 8 ADDD 9 SD 10 LD 11 ADDD 12 SD 13 SUBI 14 BNEZ 15 NOP F0,0(R1) F4,F0,F2 0(R1),F4 F6,-8(R1) F8,F6,F2 -8(R1),F8 F10,-16(R1) F12,F10,F2 -16(R1),F12 F14,-24(R1) F16,F14,F2 -24(R1),F16 R1,R1,#32 R1,LOOP 1 cycle stall 2 cycles stall ;drop SUBI & BNEZ Rewrite loop to minimize stalls? ;drop SUBI & BNEZ ;drop SUBI & BNEZ ;alter to 4*8 15 + 4 x (1+2) = 27 clock cycles, or 6.8 per iteration Assumes R1 is multiple of 4 CS 152 Lec 14.29 Unrolled Loop That Minimizes Stalls 1 Loop:LD 2 LD 3 LD 4 LD 5 ADDD 6 ADDD 7 ADDD 8 ADDD 9 SD 10 SD 11 SD 12 SUBI 13 BNEZ 14 SD F0,0(R1) F6,-8(R1) F10,-16(R1) F14,-24(R1) F4,F0,F2 F8,F6,F2 F12,F10,F2 F16,F14,F2 0(R1),F4 -8(R1),F8 -16(R1),F12 R1,R1,#32 R1,LOOP 8(R1),F16 • What assumptions made when moved code? – OK to move store past SUBI even though changes register – OK to move loads before stores: get right data? – When is it safe for compiler to do such changes? ; 8-32 = -24 14 clock cycles, or 3.5 per iteration When safe to move instructions? CS 152 Lec 14.30 Getting CPI < 1: Issuing Multiple Instructions/Cycle • Two main variations: Superscalar and VLIW • Superscalar: varying no. instructions/cycle (1 to 6) – Parallelism and dependencies determined/resolved by HW – IBM PowerPC 604, Sun UltraSparc, DEC Alpha 21164, HP 7100 • Very Long Instruction Words (VLIW): fixed number of instructions (16) parallelism determined by compiler – Pipeline is exposed; compiler must schedule delays to get right result • Explicit Parallel Instruction Computer (EPIC)/ Intel – 128 bit packets containing 3 instructions (can execute sequentially) – Can link 128 bit packets together to allow more parallelism – Compiler determines parallelism, HW checks dependencies and fowards/stalls CS 152 Lec 14.31 Getting CPI < 1: Issuing Multiple Instructions/Cycle • Superscalar DLX: 2 instructions, 1 FP & 1 anything else – Fetch 64-bits/clock cycle; Int on left, FP on right – Can only issue 2nd instruction if 1st instruction issues – More ports for FP registers to do FP load & FP op in a pair Type Int. instruction FP instruction Int. instruction FP instruction Int. instruction FP instruction Pipe Stages IF IF ID IF ID EX MEM EX MEM WB IF ID EX ID EX MEM IF ID IF ID EX WB MEM WB WB EX MEM WB MEM WB • 1 cycle load delay expands to 3 instructions in SS – instruction in right half can’t use it, nor instructions in next slot CS 152 Lec 14.32 Loop Unrolling in Superscalar Integer instruction Loop: LD F0,0(R1) LD F6,-8(R1) LD F10,-16(R1) LD F14,-24(R1) LD F18,-32(R1) SD 0(R1),F4 SD -8(R1),F8 SD -16(R1),F12 SD -24(R1),F16 SUBI R1,R1,#40 BNEZ R1,LOOP SD -32(R1),F20 FP instruction ADDD F4,F0,F2 ADDD F8,F6,F2 ADDD F12,F10,F2 ADDD F16,F14,F2 ADDD F20,F18,F2 Clock cycle 1 2 3 4 5 6 7 8 9 10 11 12 • Unrolled 5 times to avoid delays (+1 due to SS) • 12 clocks, or 2.4 clocks per iteration CS 152 Lec 14.33 Limits of Superscalar • While Integer/FP split is simple for the HW, get CPI of 0.5 only for programs with: – Exactly 50% FP operations – No hazards • If more instructions issue at same time, greater difficulty of decode and issue – Even 2-scalar => examine 2 opcodes, 6 register specifiers, & decide if 1 or 2 instructions can issue • VLIW: tradeoff instruction space for simple decoding – The long instruction word has room for many operations – By definition, all the operations the compiler puts in the long instruction word can execute in parallel – E.g., 2 integer operations, 2 FP ops, 2 Memory refs, 1 branch » 16 to 24 bits per field => 7*16 or 112 bits to 7*24 or 168 bits wide – Need compiling technique that schedules across several branches CS 152 Lec 14.34 Loop Unrolling in VLIW Memory reference 1 Memory FP reference 2 operation 1 FP op. 2 Int. op/ branch Clock LD F0,0(R1) LD F6,-8(R1) LD F10,-16(R1) LD F14,-24(R1) LD F18,-32(R1) LD F22,-40(R1) ADDD F4, F0, F2 ADDD F8,F6,F2 LD F26,-48(R1) ADDD F12, F10, F2 ADDD F16,F14,F2 ADDD F20, F18, F2 ADDD F24,F22,F2 SD 0(R1),F4 SD -8(R1),F8 ADDD F28, F26, F2 SD -16(R1),F12 SD -24(R1),F16 SD -32(R1),F20 SD -40(R1),F24 SUBI R1,R1,#48 SD -0(R1),F28 BNEZ R1,LOOP Unrolled 7 times to avoid delays 7 results in 9 clocks, or 1.3 clocks per iteration Need more registers in VLIW(EPIC => 128int + 128FP) CS 152 Lec 14.35 1 2 3 4 5 6 7 8 9 Software Pipelining • Observation: if iterations from loops are independent, then can get more ILP by taking instructions from different iterations • Software pipelining: reorganizes loops so that each iteration is made from instructions chosen from different iterations of the original loop ( Tomasulo in SW) Iteration 0 Iteration Iteration 1 2 Iteration 3 Iteration 4 Soft warepipelined it eration CS 152 Lec 14.36 Software Pipelining Example After: Software Pipelined 1 2 3 4 5 • Symbolic Loop Unrolling SD ADDD LD SUBI BNEZ 0(R1),F4 ; Stores M[i] F4,F0,F2 ; Adds to M[i-1] F0,-16(R1); Loads M[i-2] R1,R1,#8 R1,LOOP overlapped ops Before: Unrolled 3 times 1 LD F0,0(R1) 2 ADDD F4,F0,F2 3 SD 0(R1),F4 4 LD F6,-8(R1) 5 ADDD F8,F6,F2 6 SD -8(R1),F8 7 LD F10,-16(R1) 8 ADDD F12,F10,F2 9 SD -16(R1),F12 10 SUBI R1,R1,#24 11 BNEZ R1,LOOP SW Pipeline Time Loop Unrolled – Maximize result-use distance – Less code space than unrolling Time – Fill & drain pipe only once per loop vs. once per each unrolled iteration in loop unrolling CS 152 Lec 14.37 Software Pipelining with Loop Unrolling in VLIW Memory reference 1 Memory reference 2 FP operation 1 FP op. 2 Int. op/ branch Clock LD F0,-48(R1) ST 0(R1),F4 ADDD F4,F0,F2 LD F6,-56(R1) ST -8(R1),F8 ADDD F8,F6,F2 SUBI R1,R1,#24 2 LD F10,-40(R1) ST 8(R1),F12 ADDD F12,F10,F2 BNEZ R1,LOOP 3 1 • Software pipelined across 9 iterations of original loop – In each iteration of above loop, we: » Store to m,m-8,m-16 » Compute for m-24,m-32,m-40 » Load from m-48,m-56,m-64 (iterations I-3,I-2,I-1) (iterations I,I+1,I+2) (iterations I+3,I+4,I+5) • 9 results in 9 cycles, or 1 clock per iteration • Average: 3.3 ops per clock, 66% efficiency Note: Need less registers for software pipelining (only using 7 registers here, was using 15) CS 152 Lec 14.38 Can we use HW to get CPI closer to 1? • Why in HW at run time? – Works when can’t know real dependence at compile time – Compiler simpler – Code for one machine runs well on another • Key idea: Allow instructions behind stall to proceed: DIVD ADDD SUBD F0,F2,F4 F10,F0,F8 F12,F8,F14 • Out-of-order execution => out-of-order completion. • Disadvantages? – Complexity – Precise interrupts harder! (Talk about this next time) CS 152 Lec 14.39 Problems? • How do we prevent WAR and WAW hazards? • How do we deal with variable latency? – Forwarding for RAW hazards harder. Clock Cycle Number Instruction LD F6,34(R2) LD F2,45(R3) MULTD F0,F2,F4 SUBD F8,F6,F2 DIVD F10,F0,F6 ADDD F6,F8,F2 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 IF ID EX MEM WB RAW IF ID EX MEM WB IF ID stall M1 M2 M3 M4 M5 M6 M7 M8 M9 M10 MEM WB IF ID A1 A2 MEM WB IF ID stall stall stall stall stall stall stall stall stall D1 IF ID A1 A2 MEM WB WAR CS 152 Lec 14.40 D2 Summary • Precise interrupts are easy to implement in a 5-stage pipeline: – Mark exception on pipeline state rather than acting on it – Handle exceptions at one place in pipeline (memory-stage/beginning of writeback) • Loop unrolling Multiple iterations of loop in software: – Amortizes loop overhead over several iterations – Gives more opportunity for scheduling around stalls • Software Pipelining take one instruction from each of several iterations of the loop – Software overlapping of loop iterations • Very Long Instruction Word machines (VLIW) Multiple operations coded in single, long instruction – Requires compiler to decide which ops can be done in parallel – Trace scheduling find common path and schedule code as if branches didn’t exist (+ add “fixup code”) CS 152 Lec 14.41