In tensorflow\compiler\xla\service\Flatten_call_graph.h
, it says:
// Flattening associates each call site with a unique computation (for
// sequential calling contexts) This simplifies buffer assignment and
// points-to analysis (see b/36865746 for details).
What is b/36865746
?
They're internal bug references. If you work at Google and enter b/<number>
into Chrome on your corporate machine, it will take you to that bug. If you don't work at Google, you can't access the internal bug tracker.
You'll also see some comments in the code that are in the form TODO(b/<number>): ...
. Again, b/<number>
in these TODO comments is a reference to the relevant internal bug.
Obligatory disclaimer: I'm posting this in a personal capacity on my personal computer in my personal time. I'm not representing my employer here.