module mlir supported on ROCm

[weihanmines]

Thanks for your contribution! Unfortunately, tensorflow/runtime is currently not
accepting contributions. Please see the
Contribution Guidelines for
more information.
@deven-amd @rsanthanam-amd

[weihanmines]

@chsigg @hanbinyoon

[weihanmines]

@chsigg The print error enum is added in this PR.

[chsigg]

Hi Wei, could you please explain why this change is necessary?

Also, please remove the unrelated changes.

Thanks!

[weihanmines]

Hi Wei, could you please explain why this change is necessary?

Also, please remove the unrelated changes.

Thanks!

Hi Christian, the main purpose of this PR is to add jit compilation option for ROCm and get module.mlir working on ROCm. I will create two or three PRs later this week, so these PRs make reviewing the changes easier. Thank you.

[chsigg]

I think we came to the conclusion that HSACO is the appropriate serialization for a ROCm binary, even if it's architecture specific (PTX is forward compatible, but we could also use a FATBIN and update it for new archs).

The binary blob is only persistently stored in a few tests and does not need to be human-readable.

[weihanmines]

I think we came to the conclusion that HSACO is the appropriate serialization for a ROCm binary, even if it's architecture specific (PTX is forward compatible, but we could also use a FATBIN and update it for new archs).

The binary blob is only persistently stored in a few tests and does not need to be human-readable.

Yes, we did. Thank you for the update.