1. Single Cycle Datapath Lecture notes from MKP, H. H. Lee and S. Yalamanchili

2. (2) Reading Section 4.1-4.4 Appendices B.7, B.8, B.11, D.2 Practice Problems: 1, 4, 6, 9

3. (3) Introduction We will examine two MIPS implementations  A simplified version  this module  A more realistic pipelined version Simple subset, shows most aspects  Memory reference: lw, sw  Arithmetic/logical: add, sub, and, or, slt  Control transfer: beq, j

4. (4) Instruction Execution PC  instruction memory, fetch instruction Register numbers  register file, read registers Depending on instruction class 1.Use ALU to calculate oArithmetic result oMemory address for load/store oBranch target address 2.Access data memory for load/store 3.PC  An address or PC + 4 8d0b0000 014b5020 21080004 2129ffff 1520fffc 000a082a ….. An Encoded Program Address

5. (5) Basic Ingredients Include the functional units we need for each instruction – combinational and sequential

6. (6) Sequential Elements (4.2, B.7, B.11) Register: stores data in a circuit  Uses a clock signal to determine when to update the stored value  Edge-triggered: update when Clk changes from 0 to 1 D Clk Q D Q falling edgerising edge

7. (7) Sequential Elements Register with write control  Only updates on clock edge when write control input is 1  Used when stored value is required later D Clk Q Write D Q Clk cycle time

8. (8) Clocking Methodology Combinational logic transforms data during clock cycles  Between clock edges  Input from state elements, output to state element  Longest delay determines clock period Synchronous vs. Asynchronous operation Recall: Critical Path Delay

9. (9) Built using D flip-flops (remember ECE 2030!) Register File (B.8)

10. (10) Register File Note: we still use the real clock to determine when to write

11. (11) Building a Datapath (4.3) Datapath  Elements that process data and addresses in the CPU oRegisters, ALUs, mux’s, memories, … We will build a MIPS datapath incrementally  Refining the overview design

12. (12) High Level Description Single instruction single data stream model of execution (Remember Flynn’s Taxonomy)  Serial execution model Commonly known as the von Neumann execution model  Stored program model  Instructions and data share memory Fetch Instructions Execute Instructions Memory Operations Control SISDSIMD MISDMIMD Data Streams Instruction Streams

13. (13) Instruction Fetch Increment by 4 for next instruction 32-bit register clk cycle time Start instruction fetchComplete instruction fetch clk

14. (14) R-Format Instructions Read two register operands Perform arithmetic/logical operation Write register result op rs rt rdshamtfunct

15. (15) Executing R-Format Instructions ALU control RegWrite Write register Read data 1 Read data 2 Read register 1 Read register 2 Write data ALU result ALU Zero 5 5 5 3 op rs rt rdshamtfunct

16. (16) Load/Store Instructions Read register operands Calculate address using 16-bit offset  Use ALU, but sign-extend offset Load: Read memory and update register Store: Write register value to memory oprsrt16-bit constant

17. (17) Executing I-Format Instructions 1632 Sign extend MemRead MemWrite Data memory Write data Read data Address RegWrite Read register 1 Read register 2 Write register oprsrt16-bit constant

18. (18) Branch Instructions Read register operands Compare operands  Use ALU, subtract and check Zero output Calculate target address  Sign-extend displacement  Shift left 2 places (word displacement)  Add to PC + 4 oAlready calculated by instruction fetch oprsrt16-bit constant

19. (19) Branch Instructions Just re-routes wires Sign-bit wire replicated oprsrt16-bit constant

20. (20) Updating the Program Counter PC Instruction memory Read address Instruction [31–0] Instruction [20–16] Instruction [25–21] Add 4 16 32 Instruction [15–0] Sign extend 1 M u x 0 Instruction [15–11 Shift Branch Add ALU result Computation of the branch address loop: beq $t0, $0, exit addi $t0, $t0, -1 lw $a0, arg1($t1) lw $a1, arg2($t2) jal func add $t3, $t3, $v0 addi $t1, $t1, 4 addi $t2, $t2, 4 j loop

21. (21) Composing the Elements First-cut data path does an instruction in one clock cycle  Each datapath element can only do one function at a time  Hence, we need separate instruction and data memories Use multiplexers where alternate data sources are used for different instructions 014b5020 21080004 2129ffff 1520fffc 000a082a ….. An Encoded Program Address PC

22. (22) Full Single Cycle Datapath Destination register is “instruction- specific” lw$t0, 0($t4) vs. add $t0m $t1, $t2

23. (23) The Main Control Unit Control signals derived from instruction 0rsrtrdshamtfunct 31:265:025:2120:1615:1110:6 35 or 43rsrtaddress 31:2625:2120:1615:0 4rsrtaddress 31:2625:2120:1615:0 R-type Load/ Store Branch opcodealways read read, except for load write for R-type and load sign-extend and add

24. (24) ALU Control (4.4, D.2) ALU used for  Load/Store: Function = add  Branch: Function = subtract  R-type: Function depends on funct field ALU controlFunction 0000AND 0001OR 0010add 0110subtract 0111set-on-less-than 1100NOR

25. (25) ALU Control Assume 2-bit ALUOp derived from opcode  Combinational logic derives ALU control opcodeALUOpOperationfunctALU functionALU control lw00load wordXXXXXXadd0010 sw00store wordXXXXXXadd0010 beq01branch equalXXXXXXsubtract0110 R-type10add100000add0010 subtract100010subtract0110 AND100100AND0000 OR100101OR0001 set-on-less-than101010set-on-less-than0111 How do we turn this description into gates?

26. (26) ALU Controller inst[5:0] Generated from Decoding inst[31:26] ALU control ALU result ALU Zero 3 add sub add sub and or slt lw/sw beq arith ALU control ALUOp funct = inst[5:0]

27. (27) ALU Control Simple combinational logic (truth tables)

28. (28) Datapath With Control Use rt not rd InstructionRegDstALUSrc Memto- Reg Write Mem Read Mem WriteBranchALUOp1ALUp0 R-format100100010 lw 011110000 sw X1X001000 beq X0X000101

29. (29) Commodity Processors ARM 7 Single Cycle Datapath

30. (30) Control Unit Signals To harness the datapath Inst[31:26] InstructionRegDstALUSrc Memto- Reg Write Mem Read Mem WriteBranchALUOp1ALUp0 R-format100100010 lw 011110000 sw X1X001000 beq X0X000101

31. (31) Controller Implementation LIBRARY IEEE; USE IEEE.STD_LOGIC_1164.ALL; USE IEEE.STD_LOGIC_ARITH.ALL; USE IEEE.STD_LOGIC_SIGNED.ALL; ENTITY control IS PORT( SIGNAL Opcode : IN STD_LOGIC_VECTOR( 5 DOWNTO 0 ); SIGNAL RegDst : OUT STD_LOGIC; SIGNAL ALUSrc : OUT STD_LOGIC; SIGNAL MemtoReg : OUT STD_LOGIC; SIGNAL RegWrite : OUT STD_LOGIC; SIGNAL MemRead : OUT STD_LOGIC; SIGNAL MemWrite : OUT STD_LOGIC; SIGNAL Branch : OUT STD_LOGIC; SIGNAL ALUop : OUT STD_LOGIC_VECTOR( 1 DOWNTO 0 ); SIGNAL clock, reset: IN STD_LOGIC ); END control;

32. (32) Controller Implementation (cont.) ARCHITECTURE behavior OF control IS SIGNAL R_format, Lw, Sw, Beq : STD_LOGIC; BEGIN -- Code to generate control signals using opcode bits R_format <= '1' WHEN Opcode = "000000" ELSE '0'; Lw <= '1' WHEN Opcode = "100011" ELSE '0'; Sw <= '1' WHEN Opcode = "101011" ELSE '0'; Beq <= '1' WHEN Opcode = "000100" ELSE '0'; RegDst <= R_format; ALUSrc <= Lw OR Sw; MemtoReg <= Lw; RegWrite <= R_format OR Lw; MemRead <= Lw; MemWrite <= Sw; Branch <= Beq; ALUOp( 1 ) <= R_format; ALUOp( 0 ) <= Beq; END behavior; Implementation of each table column InstructionRegDstALUSrc Memto- Reg Write Mem Read Mem WriteBranchALUOp1ALUp0 R-format100100010 lw 011110000 sw X1X001000 beq X0X000101

33. (33) R-Type Instruction

34. (34) Load Instruction

35. (35) Branch-on-Equal Instruction

36. (36) Implementing Jumps Jump uses word address Update PC with concatenation of  Top 4 bits of old PC  26-bit jump address  00 Need an extra control signal decoded from opcode 2address 31:2625: 0 Jump

37. (37) Datapath With Jumps Added

38. (38) Energy Behavior combinational activity storage read/write access

39. (39) Recall Hierarchy of Energy Models V in V out V dd PMOS Ground NMOS a b c x y ALU Switch level activity (dynamic) and leakage (static) energy costs Aggregate energy expenditure into gate level estimates Aggregate energy expenditure into higher level modules

40. (40) A Simple Architecture Energy Model To a first order, we can use the per-access energy of each major component  Obtain this for a technology generation Use this per-access energy to compute the energy of each instruction Note:  This is a high level approximation. The actual physics is more complicated.  However, this useful for several purposes What components do each instruction exercise?

41. (41) Example: Updating the PC MemtoReg MemRead MemWrite ALUOp ALUSrc RegDst PC Instruction memory Read address Instruction [31–0] Instruction [20–16] Instruction [25–21] Add Instruction [5–0] RegWrite 4 16 32 Instruction [15–0] 0 Registers Write register Write data Write data Read data 1 Read data 2 Read register 1 Read register 2 Sign extend ALU result Zero Data memory Address Read data M u x 1 1 M u x 0 1 M u x 0 1 M u x 0 Instruction [15–11] ALU control Shift left 2 ALU Add ALU result Branch What is the energy cost of this operation?

42. (42) Example: Register Instructions MemtoReg MemRead MemWrite ALUOp ALUSrc RegDst PC Instruction memory Read address Instruction [31–0] Instruction [20–16] Instruction [25–21] Add Instruction [5–0] RegWrite 4 16 32 Instruction [15–0] 0 Registers Write register Write data Write data Read data 1 Read data 2 Read register 1 Read register 2 Sign extend ALU result Zero Data memory Address Read data M u x 1 1 M u x 0 1 M u x 0 1 M u x 0 Instruction [15–11] ALU control Shift left 2 ALU Add ALU result Branch What is the energy cost of this operation?

43. (43) Example: I-type Instructions MemtoReg MemRead MemWrite ALUOp ALUSrc RegDst PC Instruction memory Read address Instruction [31–0] Instruction [20–16] Instruction [25–21] Add Instruction [5–0] RegWrite 4 16 32 Instruction [15–0] 0 Registers Write register Write data Write data Read data 1 Read data 2 Read register 1 Read register 2 Sign extend ALU result Zero Data memory Address Read data M u x 1 1 M u x 0 1 M u x 0 1 M u x 0 Instruction [15–11] ALU control Shift left 2 ALU Add ALU result Branch What is the energy cost of this operation?

44. (44) Example: I-Type for Branches What is the energy cost of this operation?

45. (45) Converting Energy to Power For this data path, except for data memory, all components are active every cycle, and dissipating energy on every cycle  Later we will see how data paths can be made more energy efficient Computing power  Compute the total energy consumed over all cycles (instructions)  Divide energy by time to get power in watts Example:

46. (46) Example: A Simple Energy Model We can use a simple model of per-access energy for the architecture components Common ComponentsAccess Energy (10 -12 joules) Inst. DecodeLogic Switching 16.78 Inst. RegistersRead 2.74Write 4.38 FP. RegistersRead 1.26Write 1.98 Other BuffersRead 9.74Write 11.18 ALU + Result Bus (interconnect)Logic Switching 123.2 FPU + Result Bus (interconnect)Logic Switching 241.02 Each unit can be accessed multiple times depending on instruction type An Intel/AMD x86 instruction consume 600pJ ~ 4nJ dynamic energy. @16nm

47. (47) ITRS Roadmap for Logic Devices From: “ExaScale Computing Study: Technology Challenges in Achieving Exascale Systems,” P. Kogge, et.al, 2008

48. (48) All of the logic is combinational We wait for everything to settle down, and the right thing to be done  ALU might not produce “right answer” right away  we use write signals along with clock to determine when to write Cycle time determined by length of the longest path Our Simple Control Structure We are ignoring some details like setup and hold times

49. (49) Performance Issues Longest delay determines clock period  Critical path: load instruction  Instruction memory  register file  ALU  data memory  register file Not feasible to vary period for different instructions Violates design principle  Making the common case fast We will improve performance by pipelining

50. (50) Summary Single cycle datapath  All instructions execute in one clock cycle  Not all instructions take the same amount of time  Software sees a simple interface  Can memory operations really take one cycle? Improve performance via pipelining, multi- cycle operation, parallelism or customization We will address these next

51. (51) Study Guide Given an instruction, be able to specify the values of all control signals required to execute that instruction Add new instructions: modify the datapath and control to affect its execution  E.g., jal, jr, shift, etc.  Modify the VHDL controller Given delays of various components, determine the cycle time of the datapath Distinguish between those parts of the datapath that are unique to each instruction and those components that are shared across all instructions

52. (52) Study Guide (cont.) Given a set of control signal values determine what operation the datapath performs Given the per access energies of each component:  Compute the energy required of any instruction  Given a program and clock rate compute the power dissipation of the datapath

53. (53) Glossary Asynchronous Clock Controller Critical path Flip Flop ITRS Roadmap Per-access energy Program counter Register Synchronous