Grappler and XLA Optimizations

[ja...@nervanasys.com]

We just noticed the new grappler graph optimization code in the TF tree. From a cursory investigation it looks like there are overlaps between the two in terms of functionality (cost based optimization, GPU memory layout, constant propagation, etc). Can anyone provide guidelines for how these two graph optimizers are supposed to work together?

Thanks,

Jason

[Peter Hawkins]

I think we are still figuring this out ourselves :-)

Grappler performs optimizations at the TF graph level. XLA has its own intermediate language at a level under the TF graph representation. I think both complement each other. Some optimizations will be easier on the TF graph, some easier at the HLO level. I don't see any major downside to using both, although you probably won't get as much benefit from Grappler if you also are using XLA.

Peter

--
You received this message because you are subscribed to the Google Groups "XLA development" group.
To unsubscribe from this group and stop receiving emails from it, send an email to xla-dev+u...@googlegroups.com.
To post to this group, send email to xla...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/xla-dev/db98f094-3b5d-4b85-8cf7-69766edd6d97%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.