limbo018 / DREAMPlace

Deep learning toolkit-enabled VLSI placement

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

question about the macros

Icamd opened this issue · comments

commented

Dear professor Lin:
Thank you for your code! When I was using the DreamPlace to solve the placement of the ispd2005 benchmarks, I find that all the macros are already fixed and cannot be moved. But there are a lot of papers that optimize the position of macros using reinforcement learning and claim to have good results. Does the DreamPlace algorithm optimize the location of macros? Or is it just optimizing standard cells? And are those fixed macros in the ispd2005 benchmark "expert data" that have eliminated overlap? And do you how to generate the placement pictures/ is there any program that visualizes the Dreamplace algorithm? Thank you professor!

You can modify the files in the ispd2005 benchmarks to make FIXED nodes movable. DREAMPlace supports mixed-size placement.

@limbo018 Thank you! Could you please tell me how to set FIXED nodes movable in DREAMPlace? And how to generate the pictures of DREAMPlace result ( is there any open-source drawing program that visualizes DREAMPlace )? Thank you!

Please check dreamplace/params.json for all the options. You can set plot_flag in JSON to 1 to enable plotting the results.

If you don't want to change the input files, you may need to modify the code in PlaceDB.py to make fixed nodes movable.

@limbo018 Thank you! But how to change the input files ( just delete the "FIXED" sign?) to make FIXED nodes movable? I tried to remove the FIXED flag, which resulted in a large increase in HPWL. Does making fixed node movable cause HPWL to increase? Thank you!

Hi, it is not recommended to change all fixed nodes to movable, which can make the optimization unstable. If you don't know how to choose movable macros, I recommend to you the mms benchmarks. Previous papers used it to study macro placement.

 J. Z. Yan, N. Viswanathan, and C. Chu. Handling Complexities in Modern Large-Scale Mixed-Size Placement. In DAC, pages 436–441, 2009.

J. Lu et al., "ePlace-MS: Electrostatics-Based Placement for Mixed-Size Circuits," in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 34, no. 5, pp. 685-698, May 2015, doi: 10.1109/TCAD.2015.2391263.

https://disk.pku.edu.cn:443/link/2F11552C2A8AA27B24CC28EB8ECE8CE1
Valid Until: 2023-12-31 23:59

@limbo018 Thank you professor! I wish you all the best in your scientific research!

@limbo018 Hi Professor, sorry to bother you again. I have encountered new problems in generating pictures of the result. I set plot_flag to 1, but it appears that the program only visualizes part of macros (it seems to be 1/10 of the total macros). How can I visualize all macros and standard cells? Thank you!

Are your macros out of the diearea or are they actually IO pins outside the diearea? The plot function only draws stuff within the diearea. You may modify dreamplace/ops/draw_place/draw_place.py a little bit to enlarge the region defined by (xl, yl, xh, yh).