Adrian Kuegel

Thanks for the reproducer. I have played around with it, and tried several layer sizes. With your workaround you are making sure that several XlaModules are created. The size of...

I am not familiar with tf.function, assigning to @cheshire who is familiar with this and might know who can help.

So we finally figured out where the difference between XLA:GPU and eager GPU is coming from. In XLA:GPU, we set a flag so that div.full ptx instructions are emitted (which...

1) I am not familiar with this, but would it also act as a fusion blocker? 2) The custom-call representation will be a fusion blocker, so it is not the...

FYI: https://github.com/openxla/xla/commit/7bb072f7ee224488a0c743f68d5082e988b2addf added the possibility to specify an option to preserve the layout as part of metadata

@burmako It seems we have a duplicate UnfuseBatchNorm pass. Do you think that one of them can be removed? And which one is better to keep?

@jreiffers given that you worked on the reduction emitter quite a bit and know the latest state, can you please review this?

@ekuznetsov139 @draganmladjenovic @pemeliya maybe one of you can help?

> @ddunl This fails to import internally because copybara isn't sure where to put the new spirv files. It seems like it requires some extra configuration. Could you please advise...

> > > @ddunl This fails to import internally because copybara isn't sure where to put the new spirv files. It seems like it requires some extra configuration. Could you...