Discussion: 2019/06/14

[ArmageddonKnight]

Discussion Questions:

• AttrScope vs. NameManager
  • What is the difference? Should we just benefit directly from one of these scoping methods rather than introducing the extra with mx.profiler.scope(...) approach?
• std::string profiler_scope; as member of the NodeAttr class?
• std::string vs. std::vector < std::string > as the type for the profiler_scope variable.

[ArmageddonKnight]

• Those are all obsolete classes that are no longer used anymore. We could, however, benefit from the AttrScope class as it is very similar to what we are going to implement.
• We could directly make use of the std::unordered_map < std::string, std::string > dict inside the NodeAttr class to store the profiler scope information. In this way, we could preserve the backward compatibility as the name member itself is not affected. Note that, however, it is important we skip the check on profiler_scope attribute when a symbol is loaded from the JSON file.
• As NodeAttr::dict is a dictionary of {std::string : std::string}, we should probably store the profiler scope information as a single string, with hierachy of profiler scopes separated by ":". With this in mind, we should also check the profiler scope string at the Python frontend to make sure that the users are not providing ":" as part of the input.