[DOCS] Integration guideline (#1135)

docs/how_to/deploy.md

@@ -34,3 +34,8 @@ on how to generate the library and [cpp_deploy.cc](https://github.com/dmlc/tvm/t
 - Bundle the compiled library into your project in system module mode.
 
 Dynamic loading is more flexible and can load new modules on the fly. System module is a more ```static``` approach.  We can use system module in places where dynamic library loading is banned.
+
+
+How to Deploy on Android
+------------------------
+- [How to deploy on android](deploy_android.md)
\ No newline at end of file

docs/how_to/deploy_android.md

 # How to deploy and use compiled model on Android
 
-
 This tutorial explain below aspects (Unlike the android_rpc approach we already have)
 
   * Build a model for android target
@@ -36,5 +35,3 @@ Refer [here](https://github.com/dmlc/tvm/blob/master/apps/android_deploy/README.
 ### Android Native API Reference
 
 From android java TVM API to load model & execute can be refered at this [java](https://github.com/dmlc/tvm/blob/master/apps/android_deploy/app/src/main/java/ml/dmlc/tvm/android/demo/MainActivity.java) sample source.
-
-

docs/how_to/integrate.md

+Integrate TVM into Your Project
+===============================
+
+TVM's runtime is designed to be lightweight and portable.
+There are several ways you can integrate TVM into your project.
+If you are looking for minimum deployment of a compiled module, take a look at [deployment guide](deploy.md)
+
+This article introduces possible ways to integrate TVM
+as a JIT compiler to generate functions on your system.
+
+
+## DLPack Support
+
+TVM's generated function follows the PackedFunc convention.
+It is a function that can take positional arguments including
+standard types such as float, integer, string.
+The PackedFunc takes DLTensor pointer in [dlpack](https://github.com/dmlc/dlpack) convention.
+So the only thing you need to solve is to create a corresponding DLTensor object.
+
+
+
+## Integrate User Defined C++ Array
+
+The only thing we have to do in C++ is to convert your array to DLTensor and pass in its address as
+```DLTensor*``` to the generated function.
+
+
+## Integrate User Defined Python Array
+
+Assume you have a python object ```MyArray```. There are three things that you need to do
+
+- Add ```_tvm_tcode``` field to your array which returns ```tvm.TypeCode.ARRAY_HANDLE```
+- Support ```_tvm_handle``` property in your object, which returns the address of DLTensor in python integer
+- Register this class by ```tvm.register_extension```
+
+```python
+# Example code
+import tvm
+
+class MyArray(object):
+    _tvm_tcode = tvm.TypeCode.ARRAY_HANDLE
+
+    @property
+    def _tvm_handle(self):
+        dltensor_addr = self.get_dltensor_addr()
+        return dltensor_addr
+
+# You can put registration step in a separate file mypkg.tvm.py
+# and only optionally import that if you only want optional dependency.
+tvm.register_extension(MyArray)
+```

docs/index.rst

@@ -15,6 +15,7 @@ Contents
    tutorials/index
    faq
    how_to/deploy
+   how_to/integrate
    how_to/contribute
    api/python/index
    dev/index