added caching mechanism #492
Conversation
|
The caching messes up error injection because normally error injection will catch flipped bits during the initialization of the possible matches. |
bfasst/utils/structural.py
Outdated
| @@ -22,11 +30,12 @@ class StructuralCompareError(Exception): | |||
| class StructuralCompare: | |||
| """Structural compare and map""" | |||
|
|
|||
| def __init__(self, named_netlist_path, reversed_netlist_path, log_path) -> None: | |||
| def __init__(self, named_netlist_path, reversed_netlist_path, log_path, debug=False) -> None: | |||
There was a problem hiding this comment.
Did you change other files to use this debug parameter? Or when you use the debug feature how are you setting it?
There was a problem hiding this comment.
I've been using a script that runs comparison directly without using ninja to allow me more flexibility and customization. Should I add it in as a flow argument?
There was a problem hiding this comment.
I think what we should do is have it dump the debug file by default, but only look for if the file exists if it is passed in as a command line argument. Maybe --debug should tell it to look for the file, and it dumps the file regardless. How does that sound?
There was a problem hiding this comment.
That sounds good to me.
|
I decided to do it as a flow arg, by default it's false but if you add --flow_arguments "{'debug': True}" to the end of your "python scripts/run.py ... " it will use the cache if there is one available, and use the RapidWright functionality as well. If debugging mode is False, and/or if there is no cache available, it will create a cache during that run. |
For the larger designs like jpegencode, a lot of the runtime is spent initializing the list of possible matches for each primitive based on cell type and properties. This pr creates a caching mechanism which rebuilds the dict based on the instance names. For jpegencode, this brought the runtime down by 29 minutes.