[TVMC][VitisAI] Enable Vitis AI target through TVMC

[jtuyls]

This adds Vitis AI as a composite target to TVMC. As we need access to the CLI options in the partitioning function for the Vitis AI target, this PR updates the TVMC compiler and tuner interface to pass those options to the partitioning functions. The Arm Compute Lib and Ethos-N partitioning functions are adjusted accordingly.

Documentation
I think we should document the use of these targets but not sure what the best place is. For example, we can add documentation to https://tvm.apache.org/docs/tutorials/get_started/tvmc_command_line_driver.html or in the target deploy and integration documentation (for example here: https://tvm.apache.org/docs/deploy/vitis_ai.html). Or we can do both?

CI
The following PR should be merged first for tests to pass: #7575

@leandron @comaniac @mbaret

[leandron]

In general LGTM, my only concern is with the already overused term target, that if we can find some other term to make it clear, it would be good. BTW, @jtuyls thanks for the effort to implement TVMC support.

Apart from that, some test nits.

[leandron]

In general LGTM, two minor comments.

[jtuyls]

@leandron Do you have any thoughts on documenting TVMC target usage?

Documentation
I think we should document the use of these targets but not sure what the best place is. For example, we can add documentation to https://tvm.apache.org/docs/tutorials/get_started/tvmc_command_line_driver.html or in the target deploy and integration documentation (for example here: https://tvm.apache.org/docs/deploy/vitis_ai.html). Or we can do both?

[leandron]

@leandron Do you have any thoughts on documenting TVMC target usage?

Documentation
I think we should document the use of these targets but not sure what the best place is. For example, we can add documentation to https://tvm.apache.org/docs/tutorials/get_started/tvmc_command_line_driver.html or in the target deploy and integration documentation (for example here: https://tvm.apache.org/docs/deploy/vitis_ai.html). Or we can do both?
I agree this is

Yes, I agree this is something we need to improve. Actually in general for targets in TVM it is not easy to obtain meta-information without inspecting the code.

Ideally, I'd like to make the targets self-documenting, so that we don't need to update a page and keep it in sync with changes we do in the code, and users can find useful help on the --help option - it is fine if you have your own documentation page, to write about the options, but I'm not sure this should be our official documentation. So that would be my recommendation for now.

I was planning to submit a separate PR, making --help list the available TVM targets plus our composite targets here. This one is very easy (yet not 100% complete as a feature), but I think should be done in a separate PR.

For the inner options, such as -mcpu for llvm, and -dpu for vitis-ai, I think it would require a bit of refactoring, to make that an exposed API in TVM level, just consuming that on TVMC.

Concretely, I don't think this should block your PR here. But certainly is a very important topic to start a broader discussion. (cc @masahi @comaniac @junrushao1994)

[jtuyls]

For the inner options, such as -mcpu for llvm, and -dpu for vitis-ai, I think it would require a bit of refactoring, to make that an exposed API in TVM level, just consuming that on TVMC.

Concretely, I don't think this should block your PR here. But certainly is a very important topic to start a broader discussion.

I think it would be very useful to solve this target documentation issue in general and proceeding in that manner is better than relying on hard to maintain documentation. So I definitely agree this shouldn't block this PR but I think it would be good to start/continue looking at this so we can make these options transparent to the user.

[leandron]

@jtuyls with the merge of #7299 there might be a come minor conflicts with the repo now.

[leandron]

LGTM

[comaniac]

LGTM. Please fix the CI and let's merge this PR.

[jtuyls]

@comaniac The ci docker will have be updated first for the changes in PR #7575 to take effect.

[jtuyls]

@comaniac With the CI docker update and lazy loading the pyxir package also inside the the Vitis AI partitioning pass, the tests are now passing. Could you have another look?

[comaniac]

Thanks @jtuyls @leandron

[leandron]

Thanks @jtuyls