Hi folks
I'm trying to understand some implementation details of xla.
// NOTE(dimvar): this is quite ugly and should be changed. It's unclear
// why we care about transitive sizes; when scheduling a node, its input
// and output buffers should be all that matters, not its "history".
My confusion is the same as dimvar noted in the code:
Why is extra_users and total_sizes accumulated? "why we care about transitive sizes"?
For each node(hlo_instruction), when look at extra_users of that node, it indicates there were this many branches before we visit current node.
why is "extra_users" used as the major parameter for DFS memory scheduler?
Thanks a lot for your time
Kevin