Add build_create_shared_func to tvm/contrib/cc.py

[apivovarov]

This builder functions allows to build create_shared function with particular default options and compile_cmd.
Returned function can be used in autotune.LocalBuilder

Example:

# Create tuning_option for ARMv7 hf device (which does not have g++ installed)
from tvm.contrib import cc
build_func = cc.build_create_shared_func(compile_cmd="arm-linux-gnueabihf-g++")

tuning_option = {
    'tuner': 'xgb',
    'n_trial': 1500,
    'early_stopping': 800,

    'measure_option': autotvm.measure_option(
        builder=autotvm.LocalBuilder(
            build_func=build_func
        ),
        runner=autotvm.RPCRunner(
            device_key, host=tr_addr, port=tr_port,
            number=5,
            timeout=10,
        ),
    ),
}

The tuning_option above can be used to tune models for ARMv7 hf devices which does not have g++ installed. Auto-tuning a convolutional network for ARM CPU

Discussion - https://discuss.tvm.ai/t/lib-export-library-cc-param-was-renamed-to-compile-cmd/3644/3

[apivovarov]

@weberlo Can you have a look?

[zhiics]

LGTM. @merrymercy Can you take a look? Thanks.

[weberlo]

Is this argument meant to specify a base of compiler options or a default set of compiler options. That meaning, if you pass in options=['a'], and the resulting create_shared_wrapper gets called with options=['b'], is the intent that options == ['a', 'b'] or options == ['b']? Because I think the code currently does the latter.

[apivovarov]

My intent was to havecreate_shared function builder which allows me to specify returned function default options and compile_cmd.

[merrymercy]

Thanks @apivovarov! It is merged