Create lazy_dyndeps to avoid caffe2 import costs. (#41343)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/41343
Currently caffe2.InitOpLibrary does the dll import uniliaterally. Instead if we make a lazy version and use it, then many pieces of code which do not need the caffe2urrenoperators get a lot faster.
One a real test, the import time went from 140s to 68s. 8s.
This also cleans up the algorithm slightly (although it makes a very minimal
difference), by parsing the list of operators once, rather than every time a
new operator is added, since we defer the RefreshCall until after we've
imported all the operators.
The key way we maintain safety, is that as soon as someone does an operation
which requires a operator (or could), we force importing of all available
operators.
Future work could include trying to identify which code is needed for which
operator and only import the needed ones. There may also be wins available by
playing with dlmopen (which opens within a namespace), or seeing if the dl
flags have an impact (I tried this and didn't see an impact, but dlmopen may
make it better).
Note that this was previously landed and reverted. The issue was that if a import failed and raised an exception, the specific library would not be removed from the lazy imports. This caused our tests which had libraries that failed to poison all other tests that ran after it. This has been fixed and a unit test has been added for this case (to help make it obvious what failed).
Test Plan:
I added a new test a lazy_dyndep_test.py (copied from all_compare_test.py).
I'm a little concerned that I don't see any explicit tests for dyndep, but this
should provide decent coverage.
I've added a specific test to handle the poisoning issues mentioned above, which caused the previous version to get reverted.
Differential Revision: D22506369
fbshipit-source-id: 7395df4778e8eb0220630c570360b99a7d60eb83