pytorch
66ccc9f7 - Augment inlined call stack with module type/instance info.

Commit
4 years ago
Augment inlined call stack with module type/instance info. Summary: Motivation: During inlinng we lose information about where the node/op is coming from in module hierarchy. The SourceDebug info is useful in providing where in the source code the definition of the node exist. One can follow the callstack to find the relevant module hierarchy. However, this is less accessible, particularly in the context of aggregating node level op latencies to produce module level latencies. Facilitiating this enables identifying which module might be expensive and can be targeted either for optimization or redesign. This also aids network architecture search approaches. Approach: This PR adds module type/instance information to inlined call stack during inlining. During inlining where the subsequent CallMethods are being replaced with their corresponding graph, we find the module instance information corresponding to the CallMethod following the GetAttr node that produces the module type input to CallMethod. This approach has limitations in that such information will not be avaiable to CallFunctions and hence we will only be able to annotate type information and not instance information. But the intuition is that, it should cover majority of the cases. And we should be able to address this in another way. Test Plan: python test/test_jit.py TestScopeWithInlinedCallSTack Reviewers: Subscribers: Tasks: Tags:
Author
Committer
taivu
Parents
Loading