CSC 4250 Computer Architectures September 26, 2006 Appendix A. Pipelining
Checks before Instruction Issue Check for structural hazards ─ FP divide, register write port Check for RAW hazard ─ check if source registers of instruction in ID is listed as a destination in ID/A1, A1/A2, A2/A3, ID/M1, M1/M2, M2/M3, …, M5/M6, D Check for WAW hazard ─ check if instruction in ID has the same register destination as any instruction in A1, …, A4, M1, …, M7, D Have we forgotten about WAR hazard?
Example on why no WAR hazards Modify Figure A.33: 1. L.DF4,O(R2) 2. MUL.DF0,F4,F6 3. ADD.DF4,F0,F8 4. S.DF4,O(R2) Clock cycle Number In IF ID EX ME WB 2. IF 3. IF 4. IF Fill in the blanks
How Imprecise Exception May Arise Consider code: DIV.DF0,F2,F4 ADD.DF10,F10,F8 SUB.DF12,F12,F14 Will get out-of-order completion Suppose SUB.D causes a FP exception after ADD.D completes (but before DIV.D finishes) Next, DIV.D causes a FP exception Cannot restore the state to before DIV.D, as ADD.D has destroyed one of its operands
Performance of MIPS FP Pipeline The MIPS FP pipeline generates both structural stalls for the divide unit and stalls for RAW hazards (it can also have WAW hazards, but this rarely occurs in practice). Figure A.35 shows the number of stall cycles for each type of FP operation. The stall cycles per operation track the latency of the FP operations, varying from 46% to 59% of the latency of the FU.
Fig. A.35. Stalls per FP operation for FP SPEC89
Stalls for SPEC89 FP Benchmarks Consider data in Figure A.35 Except for the divide structural hazards, these data do not depend on the frequency of an operation, only on its latency and the # of cycles before the result is used The number of stalls from RAW hazards roughly tracks the latency of the FP unit. For example, the average number of stalls per FP add, subtract, or convert is 1.7 cycles, or 56% of the latency (3 cycles). Likewise, the average number of stalls for multiplies and divides are 2.8 and 14.2, resp., or 46% and 59% of the corr. latency Structural hazards for divides are rare, since the divide frequency is low
Fig. A.36. Stalls per instr. for FP SPEC89
MIPS R4000 Pipeline Implements MIPS64 but uses a deeper pipeline Achieve higher clock rates by decomposing five- stage integer pipeline into eight stages Extra pipeline stages from decomposing memory access Sometimes called superpipelining
Eight-stage Pipeline Structure of R4000 It uses pipelined instruction and data caches
Functions of the Eight Stages (1) 1. IF─ First half of instruction fetch PC selection Initiation of instruction cache access 2. IS─ Second half of instruction fetch Complete instruction cache access 3. RF─ Instruction decode and register fetch Hazard check Instruction cache hit detection 4. EX─ Execution (including effective address calculation, ALU operation, branch-target calc. and condition evaluation.)
Functions of the Eight Stages (2) 5. DF─Data fetch, 1st half of data cache access 6. DS─2nd half of data fetch, Complete data cache access 7. TC─Tag check, Determine if data cache access is hit 8. WB─Write back for loads and RR operations
Two-cycle Load Delay for R4000
Example of Load 2-cycle Stall Clock number Instruction no LDR1,…IF IS RF EX DF DS TC WB DADDR2,R1,… IF IS RF st st EX DF DS DSUBR3,R1,… IF IS st st RF EX DF ORR4,R1IF st st IS RF EX
Three-cycle Branch Delay for R4000 Evaluate branch condition during EX
Example of Taken Branch Clock number Instruction no Branch instructionIF IS RF EX DF DS TC WB Delay slot IF IS RF EX DF DS TC WB Stall st st st st st st st Stall st st st st st st Branch instruction IF IS RF EX DF
Example of Untaken Branch Clock number Instruction no Branch instructionIF IS RF EX DF DS TC WB Delay slot IF IS RF EX DF DS TC WB Branch instruction+2 IF IS RF EX DF DS TC Branch instruction+3 IF IS RF EX DF DS R4000 uses a predicted-not-taken strategy for the remaining two cycles of the branch delay. Advantage over predicted-taken strategy?
R4000 FP Pipeline FP instructionLatencyInitiation interval Add, subtract43 Multiply84 Divide3635 Square root Negate21 Absolute value21 FP compare32
Major Causes of R4000 Pipeline Stalls Load stalls ─ use of a load result 1 or 2 cycles after the load Branch stalls ─ 2-cycle stall on taken branch plus unfilled or cancelled branch delay slots FP result stalls ─ RAW hazards for operand FP structural stalls ─ conflicts for func. units WAW stalls are not common
Pipeline CPI for 10 SPEC92 Benchmarks The pipeline CPI varies from 1.2 to 2.8. The left five programs are integer programs, and branch delays are the major CPI contributor. The right five programs are FP, and FP result stalls are the major contributors.
Pipeline CPI and Major Sources of Stalls BenchmarkCPILoadBranchResultStructure Gcc Int. average Hydro2d FP average Overall average
MIPS R4300 Pipeline Manufactured by NEC 64 bit processor implements MIPS64 IS Embedded applications Nintendo-64 game processor High-end color laser printer Multiple EX stages for FP operations Instructions complete out of order FP instruction generates exception after a following integer instruction has completed, leading to an imprecise exception