[REFACTOR] Polish ffi convention. (#4912)

* [REFACTOR] Polish ffi convention.

- Remove the src/api, keep registration local to the c++ function.
- Remove the api_internal as it is no longer needed.

* Update the codebase walk through

CMakeLists.txt

@@ -133,7 +133,7 @@ file(GLOB_RECURSE COMPILER_SRCS
     src/tir/*.cc
     src/driver/*.cc
     src/printer/*.cc
-    src/api/*.cc
+    src/support/*.cc
     )
 
 file(GLOB CODEGEN_SRCS

docs/dev/codebase_walkthrough.rst

@@ -55,7 +55,7 @@ We use a simple example that uses the low level TVM API directly. The example is
    B = tvm.placeholder((n,), name='B')
    C = tvm.compute(A.shape, lambda i: A[i] + B[i], name="C")
 
-Here, types of ``A``, ``B``, ``C`` are ``tvm.tensor.Tensor``, defined in ``python/tvm/tensor.py``. The Python ``Tensor`` is backed by C++ ``Tensor``, implemented in ``include/tvm/tensor.h`` and ``src/lang/tensor.cc``. All Python types in TVM can be thought of as a handle to the underlying C++ type with the same name. If you look at the definition of Python ``Tensor`` type below, you can see it is a subclass of ``Object``.
+Here, types of ``A``, ``B``, ``C`` are ``tvm.tensor.Tensor``, defined in ``python/tvm/te/tensor.py``. The Python ``Tensor`` is backed by C++ ``Tensor``, implemented in ``include/tvm/te/tensor.h`` and ``src/te/tensor.cc``. All Python types in TVM can be thought of as a handle to the underlying C++ type with the same name. If you look at the definition of Python ``Tensor`` type below, you can see it is a subclass of ``Object``.
 
 ::
 
@@ -68,24 +68,12 @@ Here, types of ``A``, ``B``, ``C`` are ``tvm.tensor.Tensor``, defined in ``pytho
 
 The object protocol is the basis of exposing C++ types to frontend languages, including Python. The way TVM implements Python wrapping is not straightforward. It is briefly covered in `this document <https://docs.tvm.ai/dev/runtime.html#tvm-node-and-compiler-stack>`_, and details are in ``python/tvm/_ffi/`` if you are interested.
 
-``Tensor`` is created by functions in ``python/tvm/api.py``, which in turn calls into C++ functions exposed in ``src/api/api_lang.cc``. All C++ functions that are callable from Python are exposed in the ``src/api`` subdirectory. For example, the ``tvm.compute()`` function above calls into ``_ComputeOp`` API exposed in ``src/api/api_lang.cc``:
-
-::
-
-   TVM_REGISTER_GLOBAL("_ComputeOp")
-   .set_body([](TVMArgs args,  TVMRetValue* ret) {
-       *ret = ComputeOpNode::make(args[0],
-                                  args[1],
-                                  args[2],
-                                  args[3],
-                                  args[4]);
-     });
-
 We use the ``TVM_REGISTER_*`` macro to expose C++ functions to frontend languages, in the form of a `PackedFunc <https://docs.tvm.ai/dev/runtime.html#packedfunc>`_. A ``PackedFunc`` is another mechanism by which TVM implements interoperability between C++ and Python. In particular, this is what makes calling Python functions from the C++ codebase very easy.
+You can also checkout `FFI Navigator <https://github.com/tqchen/ffi-navigator>`_ which allows you to navigate between python and c++ FFI calls.
 
-A ``Tensor`` object has an ``Operation`` object associated with it, defined in ``python/tvm/tensor.py``, ``include/tvm/operation.h``, and ``src/tvm/op`` subdirectory. A ``Tensor`` is an output of its ``Operation`` object. Each ``Operation`` object has in turn ``input_tensors()`` method, which returns a list of input ``Tensor`` to it. This way we can keep track of dependencies between ``Operation``.
+A ``Tensor`` object has an ``Operation`` object associated with it, defined in ``python/tvm/te/tensor.py``, ``include/tvm/te/operation.h``, and ``src/tvm/te/operation`` subdirectory. A ``Tensor`` is an output of its ``Operation`` object. Each ``Operation`` object has in turn ``input_tensors()`` method, which returns a list of input ``Tensor`` to it. This way we can keep track of dependencies between ``Operation``.
 
-We pass the operation corresponding to the output tensor ``C`` to ``tvm.create_schedule()`` function in ``python/tvm/schedule.py``.
+We pass the operation corresponding to the output tensor ``C`` to ``tvm.create_schedule()`` function in ``python/tvm/te/schedule.py``.
 
 ::
 
@@ -103,7 +91,7 @@ This function is mapped to the C++ function in ``include/tvm/schedule.h``.
 
 ``Stage`` corresponds to one ``Operation``. In the vector add example above, there are two placeholder ops and one compute op, so the schedule ``s`` contains three stages. Each ``Stage`` holds information about a loop nest structure, types of each loop (``Parallel``, ``Vectorized``, ``Unrolled``), and where to execute its computation in the loop nest of the next ``Stage``, if any.
 
-``Schedule`` and ``Stage`` are defined in ``tvm/python/schedule.py``, ``include/tvm/schedule.h``, and ``src/schedule/schedule_ops.cc``.
+``Schedule`` and ``Stage`` are defined in ``tvm/python/te/schedule.py``, ``include/tvm/te/schedule.h``, and ``src/te/schedule/schedule_ops.cc``.
 
 To keep it simple, we call ``tvm.build(...)`` on the default schedule created by ``create_schedule()`` function above.
 
@@ -112,7 +100,7 @@ To keep it simple, we call ``tvm.build(...)`` on the default schedule created by
    target = "cuda"
    fadd = tvm.build(s, [A, B, C], target)
 
-``tvm.build()``, defined in ``python/tvm/build_module.py``, takes a schedule, input and output ``Tensor``, and a target, and returns a ``tvm.Module`` object, defined in ``python/tvm/module.py``. A ``Module`` object contains a compiled function which can be invoked with function call syntax.
+``tvm.build()``, defined in ``python/tvm/driver/build_module.py``, takes a schedule, input and output ``Tensor``, and a target, and returns a :py:class:`tvm.runtime.Module` object. A :py:class:`tvm.runtime.Module` object contains a compiled function which can be invoked with function call syntax.
 
 The process of ``tvm.build()`` can be divided into two steps:
 
@@ -133,14 +121,14 @@ Lowering is done by ``tvm.lower()`` function, defined in ``python/tvm/build_modu
       stmt = schedule.ScheduleOps(sch, bounds)
       ...
 
-Bound inference is the process where all loop bounds and sizes of intermediate buffers are inferred. If you target the CUDA backend and you use shared memory, its required minimum size is automatically determined here. Bound inference is implemented in ``src/schedule/bound.cc``, ``src/schedule/graph.cc`` and ``src/schedule/message_passing.cc``. For more information on how bound inference works, see `InferBound Pass`_.
+Bound inference is the process where all loop bounds and sizes of intermediate buffers are inferred. If you target the CUDA backend and you use shared memory, its required minimum size is automatically determined here. Bound inference is implemented in ``src/te/schedule/bound.cc``, ``src/te/schedule/graph.cc`` and ``src/te/schedule/message_passing.cc``. For more information on how bound inference works, see `InferBound Pass`_.
 
 .. _InferBound Pass: http://docs.tvm.ai/dev/inferbound.html
 
 
-``stmt``, which is the output of ``ScheduleOps()``, represents an initial loop nest structure. If you have applied ``reorder`` or ``split`` primitives to your schedule, then the initial loop nest already reflects those changes. ``ScheduleOps()`` is defined in ``src/schedule/schedule_ops.cc``.
+``stmt``, which is the output of ``ScheduleOps()``, represents an initial loop nest structure. If you have applied ``reorder`` or ``split`` primitives to your schedule, then the initial loop nest already reflects those changes. ``ScheduleOps()`` is defined in ``src/te/schedule/schedule_ops.cc``.
 
-Next, we apply a number of lowering passes to ``stmt``. These passes are implemented in ``src/pass`` subdirectory. For example, if you have applied ``vectorize`` or ``unroll`` primitives to your schedule, they are applied in loop vectorization and unrolling passes below.
+Next, we apply a number of lowering passes to ``stmt``. These passes are implemented in ``src/tir/pass`` subdirectory. For example, if you have applied ``vectorize`` or ``unroll`` primitives to your schedule, they are applied in loop vectorization and unrolling passes below.
 
 ::
 
@@ -157,7 +145,7 @@ Next, we apply a number of lowering passes to ``stmt``. These passes are impleme
 
 After lowering is done, ``build()`` function generates target machine code from the lowered function. This code can contain SSE or AVX instructions if you target x86, or PTX instructions for CUDA target. In addition to target specific machine code, TVM also generates host side code that is responsible for memory management, kernel launch etc.
 
-Code generation is done by ``build_module()`` function, defined in ``python/tvm/codegen.py``. On the C++ side, code generation is implemented in ``src/codegen`` subdirectory. ``build_module()`` Python function will reach ``Build()`` function below in ``src/codegen/codegen.cc``:
+Code generation is done by ``build_module()`` function, defined in ``python/tvm/target/codegen.py``. On the C++ side, code generation is implemented in ``src/target/codegen`` subdirectory. ``build_module()`` Python function will reach ``Build()`` function below in ``src/target/codegen/codegen.cc``:
 
 ::
 

python/tvm/_ffi/registry.py

@@ -19,7 +19,6 @@
 """FFI registry to register function and objects."""
 import sys
 import ctypes
-from .. import _api_internal
 
 from .base import _LIB, check_call, py_str, c_str, string_types, _FFI_MODE, _RUNTIME_ONLY
 
@@ -288,17 +287,11 @@ def _init_api_prefix(module_name, prefix):
     module = sys.modules[module_name]
 
     for name in list_global_func_names():
-        if prefix == "api":
-            fname = name
-            if name.startswith("_"):
-                target_module = sys.modules["tvm._api_internal"]
-            else:
-                target_module = module
-        else:
-            if not name.startswith(prefix):
-                continue
-            fname = name[len(prefix)+1:]
-            target_module = module
+        if not name.startswith(prefix):
+            continue
+
+        fname = name[len(prefix)+1:]
+        target_module = module
 
         if fname.find(".") != -1:
             continue