[tvmc] Introduce 'tune' subcommand (part 3/4)

[leandron]

Introduce a sub-command to drive auto-tuning via command line, using AutoTVM, to produce the tuning records that can be used at compile time, or to further improve other tuning sessions. It follow good practices present on the auto-tuning tutorials.

In case you want to try it out (supposing you have your tvm built and working with this PR), you can have a look on the example below:

$ wget https://github.com/onnx/models/raw/master/vision/classification/resnet/model/resnet50-v2-7.onnx
$ python -m tvm.driver.tvmc --verbose tune --trials 1 --output log1.txt --target="llvm" resnet50-v2-7.onnx
(...messages...)
WARNING:root:Attribute spatial is ignored in relay.sym.batch_norm
[Task  1/23]  Current/Best:    0.00/   0.00 GFLOPS | Progress: (1/1) | 1.32 s Done.
[Task  2/23]  Current/Best:    0.00/   0.00 GFLOPS | Progress: (1/1) | 1.23 s Done.
[Task  3/23]  Current/Best:    0.00/   0.00 GFLOPS | Progress: (1/1) | 1.07 s Done.
...

cc @comaniac @masahi @FrozenGene for reviews if possible

[comaniac]

Also cc @FrozenGene @kevinthesun

[FrozenGene]

Thanks @leandron , great job. Besides this, I think we could also add some features I think it is important too:

1. We could add --num-threads to control how many cores to tune. This is important for remote devices like ARM
2. We could add --cpu-mode to control BIG core or LITTLE core to tune. Imagine remote devices like have A72 + A35. This feature maybe should require us to expose one interface in our AutoTVM to use API runtime.config_threadpool.

[leandron]

We could add --num-threads to control how many cores to tune. This is important for remote devices like ARM

@FrozenGene that looks interesting. Just to clarify, is this not related to the parameter n_parallel, from autotvm.RPCRunner, already supported via --parallel?

We could add --cpu-mode to control BIG core or LITTLE core to tune. Imagine remote devices like have A72 + A35. This feature maybe should require us to expose one interface in our AutoTVM to use API runtime.config_threadpool.

Yes, that looks useful. Can we add it in a separate PR, considering it requires some infra to be exposed to AutoTVM?

[FrozenGene]

We could add --num-threads to control how many cores to tune. This is important for remote devices like ARM

@FrozenGene that looks interesting. Just to clarify, is this not related to the parameter n_parallel, from autotvm.RPCRunner, already supported via --parallel?

Yes, it is not related with n_parallel, which is to support we have multi devices to tune.

We could add --cpu-mode to control BIG core or LITTLE core to tune. Imagine remote devices like have A72 + A35. This feature maybe should require us to expose one interface in our AutoTVM to use API runtime.config_threadpool.

Yes, that looks useful. Can we add it in a separate PR, considering it requires some infra to be exposed to AutoTVM?

Fine to me.

[leandron]

@FrozenGene that looks interesting. Just to clarify, is this not related to the parameter n_parallel, from autotvm.RPCRunner, already supported via --parallel?

Yes, it is not related with n_parallel, which is to support we have multi devices to tune.

Understood. I'm happy to look into it. Can you expand a bit on how that would work, so that we can decide whether adding it here or on a future PR?

[FrozenGene]

@FrozenGene that looks interesting. Just to clarify, is this not related to the parameter n_parallel, from autotvm.RPCRunner, already supported via --parallel?

Yes, it is not related with n_parallel, which is to support we have multi devices to tune.

Understood. I'm happy to look into it. Can you expand a bit on how that would work, so that we can decide whether adding it here or on a future PR?

You need to do the similar thing like we control BIG / LITTLE core via runtime.config_threadpool, this api could let us control how many cores and which cpu mode we like. So I am fine leave this in the future pr together with control BIG / LITTLE cpu mode.

[leandron]

Re-triggering CI

[leandron]

This PR is passing CI again.

@comaniac @FrozenGene do you mind having another look?

[comaniac]

Thanks @leandron @FrozenGene