We do the latter by defining an "optimized machine," which is the set of timings we obtain for the AbOps for optimized code.
First, a set of vector AbOps which characterizes the performance of individual vector operations is added to the original set.
When a large fraction of the execution time of a benchmark is accounted for by a small amount of code, the relative running time of that benchmark may vary widely between machines, depending on the execution time of the relevant AbOps on each machine, i.
Figure 11 shows the cumulative distribution of AbOps for the different benchmark suites.
We can also compute the skewness of the ordered distribution of AbOps in the same way as we did with basic blocks, i.
This lack of correlation seems to be due to two factors: (a) those programs with the most highly skewed distributions emphasize AbOps such as floating point, for which measurement errors are small, and (b) prediction errors are mostly due to other factors (e.