search

efabless / openlane2

The next generation of OpenLane, rewritten from scratch with a modular architecture
https://openlane2.readthedocs.io/
Apache License 2.0
189 stars 35 forks source link

Optimize Global Routing and Resizer PostGRT steps #538

Open kareefardi opened 2 weeks ago

kareefardi commented 2 weeks ago

Description

As it turns out estimating routing parasitics in OR needs a call for GRT in the same script (it cannot load all the info from the db). Some steps in the Classic flow use this. At the moment there are a lot of redundant GRT calls that slows down the flow in large designs.

Proposal

Access and revisit the situation. Additionally, GRT_RESIZER_RUN_GRT should only disable the final GRT step in the script. Finally, we should access incremental routing as well.

donn commented 2 weeks ago

Wasn't this addressed in https://github.com/efabless/openlane2/commit/a8a6e84d4970c94641f2be62c8bc709ca353ed59

kareefardi commented 2 weeks ago

Wasn't this addressed in a8a6e84

No. The solution is not complete and wrong given the current state of OR db.

donn commented 2 weeks ago

Elaborate.

kareefardi commented 2 weeks ago

The solution doesn't consider other places in the flow that uses GRT parasitics estimation. It doesn't consider that there is a redundant GRT run before OpenROAD.RepairDesignPostGRT for example. It is wrong because it turns out that GRT is needed in the same script as the ones modified in the PR. A similar PR should only disable GRT ran after calling resizer.

donn commented 2 weeks ago

If this is a Classic flow issue, not a step issue, that's a separate discussion. For the giant designs in question our PnR engineers tend to define custom sequential flows in the JSON…