You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
@amin1377 has created a place delay matrix loader that uses the router lookahead to populate it. Since the router lookahead uses a Djikstra flood algorithm (all dests path search) it is faster than many (dx,dy) routings, which is what the placer currently uses to load the place delay matrix.
Proposed Behaviour
We should change to building the place delay matrix from the router lookahead, for speed.
Current Behaviour
Slower profiling of the architecture using repeated invocations of the router for different dx,dy
Possible Solution
Make the new method the default, after gathering QoR data on vtr, titan and koios to show that works well. Test on both 2D and 3D architectures.
Context
place delay matrix calculation can get slow (600 s for a big circuit) and slows more for 3D architectures (can be 24 hours for a big design).
The text was updated successfully, but these errors were encountered:
@amin1377 has created a place delay matrix loader that uses the router lookahead to populate it. Since the router lookahead uses a Djikstra flood algorithm (all dests path search) it is faster than many (dx,dy) routings, which is what the placer currently uses to load the place delay matrix.
Proposed Behaviour
We should change to building the place delay matrix from the router lookahead, for speed.
Current Behaviour
Slower profiling of the architecture using repeated invocations of the router for different dx,dy
Possible Solution
Make the new method the default, after gathering QoR data on vtr, titan and koios to show that works well. Test on both 2D and 3D architectures.
Context
place delay matrix calculation can get slow (600 s for a big circuit) and slows more for 3D architectures (can be 24 hours for a big design).
The text was updated successfully, but these errors were encountered: