[AP] Global Placer #2806
[AP] Global Placer #2806
Conversation
github-actions
bot
added
VPR
|
@vaughnbetz This has passed CI. Please review. This is a much smaller PR than the last one. It adds the Global Placer class into the AP flow which handles how the Analytical Solver interacts with the Partial Legalizer. |
AlexandreSinger
force-pushed
the
feature-ap-global-placement
branch
from
c45d0b4
to
dea8847
Compare
|
@vaughnbetz Thank you for the review. I have resolved your comments. I agree that using an absolute gap was not a sustainable idea. I changed it to a percent gap, as you suggested, where the percent gap is equal to (UB - LB) / UB [since UB should be larger than LB). I set the target percent gap to 10%, which seems to work relatively well for some of the smaller circuits I test on. In the future I will tune this to a better value. Let me know if you have any further comments or if we can merge this. |
|
Great, thanks. |
Created the Global Placer base class which will create a partial placement based on the netlist and the architecture. This attempts to find a "globally" good placement, without considering all of the complex constraints of the FPGA architecture (without clustering). Implemented a SimPL-based Global Placer which maintains an upper and lower bound solution which slowly approach each other over several iterations. The lower-bound is the analytically solved solution which tries to optimize the placement (hinted by the upper bound solution). The upper-bound solution is the lower-bound solution which has been partially legalized.
AlexandreSinger
force-pushed
the
feature-ap-global-placement
branch
from
dea8847
to
fe92b3e
Compare
|
@vaughnbetz Completely agree. I like that way better than percent! I have updated the code and documentation. Shall we merge? |
|
Done! |
Created the Global Placer base class which will create a partial placement based on the netlist and the architecture. This attempts to find a "globally" good placement, without considering all of the complex constraints of the FPGA architecture (without clustering).
Implemented a SimPL-based Global Placer which maintains an upper and lower bound solution which slowly approach each other over several iterations. The lower-bound is the analytically solved solution which tries to optimize the placement (hinted by the upper bound solution). The upper-bound solution is the lower-bound solution which has been partially legalized.