Home: Contents

Oh crap, more numbers...

Sorry, you're not getting away without a few calculations. (The complexity of this part is debatable.) Two things we need to keep track of: the CPIs of the loop initiation interval and the loop latency. These terms are NOT DEFINED IN THE BOOK: I wanted a way to easily show the CPI of the loop by itself and the all of the unrolled code. In order to take this into account, I have to include a NOP in the branch delay slot for ``rolled'' code, and include it in the CPI. This is only fair, because when the bds is used, it changes the numbers slightly. I borrowed these terms from the multicycle DLX FP pipelines.

When you are asked on an exam to calculate the CPI of a loop, you are to calculate the CPI of the contents of the loop up until the branch instruction.

The Loop Initiation Interval

This is the time between the start of one loop iteration and the start of the next loop iteration, including the NOP in the bds. As we change the amount of work done in each loop, this gives a measure of how efficient the loop is, and makes it easy to count the number of cycles in the unrolled loop.

The Loop Latency

This is the time that each loop will take to execute AS IF IT STOOD ON ITS OWN. Namely, the number of cycles between the first instruction of the loop to the last instruction of the branch. This gives us an idea of how moving instructions into the bds helps the code, and reminds us of a possible stall cycle before the branch.