In the Tool Option Exploration stage, the goal is to get the maximum QoR return from tool options. To achieve this, three implementation runs using tool options predicted by ML strategies are run. This is shown in the following diagram:
QoR suggestions are carried forward from the best run in stage 1. If later steps in stage 1 showed a QoR degradation, their suggestions are dropped. By running three ML strategy runs (where three or more are available), QoR fluctuations from any single run are smoothed out.
The ML strategy phase uses standard implementation runs that can be run in parallel if compute resources allow. As such, when used in parallel, the compile time of this phase is approximately one implementation run. When the implementation runs are finished, the best run selector determines which run to take forward. The flow behavior is as follows:
- Exit if timing is met.
- Proceed to stage 3 if Last Mile Timing Closure criteria are met.
- Exit if Last Mile Timing Closure criteria are not met.