A Problem

• Problem 4.26 gives the following table
  • EX to 1st only: 5%
  • Mem to 1st only: 20%
  • EX to 2nd only: 5%
  • Mem to 2nd only: 10%
  • Ex to 1st and EX to 2nd 10%
  • no hazard 100 - (5+20+5+10+10) = 50%
• I added the last line. We will need it.
• The unit (EX, MEM) is what is producing the data
• The 1st or 2nd is the delay in the instruction consuming the data.
• EX to 1st only would be:

  •  
    add r1, r5, r6
    add r7, r1, r2
            

• They ask us to create examples of each, but I think we already did that.
  • If not, do it yourself.
• Without forwarding hardware how many stalls would each type of instruction cause?:
  • EX to 1st only: 5%, 2 bubbles
  • Mem to 1st only: 20%, 2 bubbles
  • EX to 2nd only: 5%, 1 bubbles
  • Mem to 2nd only: 10%, 1 bubbles
  • Ex to 1st and EX to 2nd 10%, 2 bubbles
  • no hazard 50%, 0 bubbles
• So without forwarding what would the new CPI be?

  2 Bubble instruction percent 35%
  1 Bubble instruction percent 15% 
  0 Bubble instruction percent 50%
  
  Average bubbles   $2 \times .35 + 1 \times .15 + 0 \times .5$
  Average Cycles    $3 \times .35 + 2 \times .15 + 1 \times .5$
  In any case, the CPI would come out to be 1.85
       

• What happens if we do full forwarding.
  • As far as I can tell, Mem to 1st only: 20%, 2 bubbles is the only one that we have trouble with.
  • This will lead to $ 2 \times .2 + 1 \times .8 = 1.20$.