search

limbo018 / DREAMPlace

Deep learning toolkit-enabled VLSI placement
BSD 3-Clause "New" or "Revised" License
714 stars 205 forks source link

Compared to dc/invc/openroad upon my cases, dreaplace has the shock congestion Morphologically #177

Open tigereatsheep opened 7 months ago

tigereatsheep commented 7 months ago

dcg: 0.+% H + 0.+% V innovus: 1.+% H + 1.+% V OR: 22.+% H + 33.+% V dreamplace: 58+% H + 38.+% V

and dreaplace' routing length is 4x dcg's result.

I think here maybe hpwl is not a good cost function?

limbo018 commented 7 months ago

Hi, by default, dreamplace does not turn on routability optimization. It looks like your case needs heavy optimization on routability.

tigereatsheep commented 7 months ago

Hi, by default, dreamplace does not turn on routability optimization. It looks like your case needs heavy optimization on routability.

You are right, I'm focusing on solving congestion problem in EPlace framework. Do you know better WL estimation method (better than LSE/WA)？ By the way, DreamPlace is really fast.

limbo018 commented 7 months ago

You may check section III.F in the TCAD version of DREAMPlace paper (https://yibolin.com/publications/papers/PLACE_TCAD2020_Lin.pdf). It explains a typical way to do routability optimization. The default flow of DREAMPlace does not turn on routability optimization though.

tigereatsheep commented 7 months ago

You may check section III.F in the TCAD version of DREAMPlace paper (https://yibolin.com/publications/papers/PLACE_TCAD2020_Lin.pdf). It explains a typical way to do routability optimization. The default flow of DREAMPlace does not turn on routability optimization though.

Thank you for your help. I would working on this problem for several weeks