[Relay] Invoke tvm::build from relay compile_engine and interpreter

[hlu1]

relay.backend.build is implemented in python and is used in C++ in https://github.com/apache/incubator-tvm/blob/c69092ae0d39f9a5161f098d933d0a2ec570a2c5/src/relay/backend/interpreter.cc#L422 and https://github.com/apache/incubator-tvm/blob/ce807fe83825bce666ed1834ab24b5d6ddfa6bca/src/relay/backend/compile_engine.cc#L602

Similar to relay.backend.lower https://github.com/apache/incubator-tvm/blob/ce807fe83825bce666ed1834ab24b5d6ddfa6bca/src/relay/backend/compile_engine.cc#L731-L738

invoke tvm::build instead of LOG(FATAL) when relay.backend.build cannot be found, for example, in pure C++ environment.

@jroesch, @MarisaKirisame please review :)

[MarisaKirisame]

Why isnt this a line by line translation?

[hlu1]

Sorry I missed a few things at the beginning and just added them.

Do you mean translating this part line by line, https://github.com/apache/incubator-tvm/blob/master/python/tvm/build_module.py#L509-L650?

Some of the type checks are not needed in C++, and the signature of https://github.com/apache/incubator-tvm/blob/bfb4884e47f7d2c759b2ee707aa18acf35116380/python/tvm/relay/backend/_backend.py#L65-L88 is a lot simpler.

Also,
https://github.com/apache/incubator-tvm/blob/master/python/tvm/build_module.py#L635-L650 has been implemented in here https://github.com/apache/incubator-tvm/blob/master/src/codegen/build_module.cc#L278-L319. All I need to do is call this function.

[MarisaKirisame]

I know very little about tvm Module and relay build, so I might get something wrong.
But it seems like you replace 27 line with 47 line, but they are implemented completely differently - one call another function while another does lots of work. Any reason for this?

[tqchen]

cc @zhiics @jroesch please also help to take a look

[zhiics]

@hlu1 Can we simply remove this packed function and directly invoke tvm::build from compile_engine and interpreter? We have two separate pipelines from Python and C++. I think we can probably directly use tvm.build from python and tvm::build from C++. So we don't really need to interoperate between them.

[hlu1]

@zhiics, makes sense. Will do.

[zhiics]

according to the syntax we had in Python, should we do the following?

m = build(value->cached_func->funcs, key->target, Target(nullptr), BuildConfig::Current());

Same to the one in intepreter.cc

[hlu1]

Fixed

[zhiics]

LGTM

[zhiics]

Thanks @hlu1 @MarisaKirisame