[GSoC 2015 Week 4]

This week I'm working with an optimization of inplace_elemwise_optimizer. The idea is described here. In the current version, when inplace_elemwise_optimizer trying to replace outputs of a node, the graph can become invalid, therefore validate() is called frequently to make sure the graph unbroken. But validate() is very time consuming, and the goal of this optimization is to make this optimizer more efficient by applying new validating strategy.

However, this optimization did not work as expected. The total optimization time become 10 times larger:

370.541s for fgraph.validate()
...
1497.519540s - ('inplace_elemwise_optimizer', 'FromFunctionOptimizer', 33) - 315.055s

The origin version:

72.644s for fgraph.validate()
...
143.351832s - ('inplace_elemwise_optimizer', 'FromFunctionOptimizer', 34) - 30.064s

After several small optimization pointed out by my mentor, the time become 1178s.

Why is it slower? I think it is because we are trying to apply the optimizer successfully on all nodes. It is a trade-off between the time took by validate() and the number of nodes optimized. In the past, all failed nodes are ignored directly, so it was fast. Now we are trying to apply on them again and again. validate() is called for much more times than before.

Here is a figure I just plotted to display the nodes number to optimize in each iteration.

In this figure, we know that although it is slower now, there is more nodes are optimized. A better balance should be taken in the trade-off, maybe to make the iteration stop earlier is a good choice? Or maybe the validate() can be optimized?

I'm still working on this. Please tell me if you have any idea.

Thank you.