[DOCS] How to deploy TVM Modules (#499)

* [DOCS] How to deploy TVM Modules

* More comments

apps/README.md

@@ -8,5 +8,4 @@ If you are interested in writing optimized kernels with TVM, checkout [TOPI: TVM
 - [graph_executor](graph_executor) Build nnvm graph executor with TVM.
 - [ios_rpc](ios_rpc) iOS RPC server.
 - [android_rpc](android_rpc) Android RPC server.
-- [cpp_deploy](cpp_deploy) Example to deploy with C++ runtime.
-
+- [howto_deploy](howto_depploy) Tutorial on how to deploy TVM with minimum code dependency.

apps/cpp_deploy/README.md

-C++ Deployment API Example
-==========================
-This folder contains an example to demonstrate how to use TVM C++ Runtime
-API to load and run modules built by TVM.
-
-Type the following command to run the example under current folder(need to build TVM first).
-```bash
-./run_example.sh
-```
-

apps/cpp_deploy/run_example.sh

-#!/bin/bash
-echo "Build the libraries.."
-mkdir -p lib
-make
-echo "Run the example"
-export LD_LIBRARY_PATH=../../lib
-export DYLD_LIBRARY_PATH=../../lib
-lib/cpp_deploy

apps/cpp_deploy/Makefile → apps/howto_deploy/Makefile

-# Minimum Makefile for the extension package
+# Makefile Example to deploy TVM modules.
 TVM_ROOT=$(shell cd ../..; pwd)
 NNVM_PATH=nnvm
 DMLC_CORE=${TVM_ROOT}/dmlc-core
@@ -8,12 +8,27 @@ PKG_CFLAGS = -std=c++11 -O2 -fPIC\
 	-I${DMLC_CORE}/include\
 	-I${TVM_ROOT}/dlpack/include\
 
-PKG_LDFLAGS = -L${TVM_ROOT}/lib -ltvm_runtime
+PKG_LDFLAGS = -L${TVM_ROOT}/lib -ldl -lpthread
 
+.PHONY: clean all
 
-lib/cpp_deploy: cpp_deploy.cc lib/test_addone_sys.o
+all: lib/cpp_deploy_pack lib/cpp_deploy_normal
+
+# Build rule for all in one TVM package library
+lib/libtvm_runtime_pack.o: tvm_runtime_pack.cc
 	@mkdir -p $(@D)
-	$(CXX) $(PKG_CFLAGS) -o $@  $^ $(PKG_LDFLAGS)
+	$(CXX) -c $(PKG_CFLAGS) -o $@  $^
 
+# The code library built by TVM
 lib/test_addone_sys.o: prepare_test_libs.py
 	python prepare_test_libs.py
+
+# Deploy using the all in one TVM package library
+lib/cpp_deploy_pack: cpp_deploy.cc lib/test_addone_sys.o lib/libtvm_runtime_pack.o
+	@mkdir -p $(@D)
+	$(CXX) $(PKG_CFLAGS) -o $@  $^ $(PKG_LDFLAGS)
+
+# Deploy using pre-built libtvm_runtime.so
+lib/cpp_deploy_normal: cpp_deploy.cc lib/test_addone_sys.o
+	@mkdir -p $(@D)
+	$(CXX) $(PKG_CFLAGS) -o $@  $^ $(PKG_LDFLAGS) -ltvm_runtime

apps/howto_deploy/README.md

+How to Deploy TVM Modules
+=========================
+This folder contains an example on how to deploy TVM modules.
+It also contains an example code to deploy with C++.
+
+Type the following command to run the sample code under the current folder(need to build TVM first).
+```bash
+./run_example.sh
+```
+
+Checkout [How to Deploy TVM Modules](http://docs.tvmlang.org/how_to/deploy.html) for more information.

apps/cpp_deploy/prepare_test_libs.py → apps/howto_deploy/prepare_test_libs.py

@@ -7,11 +7,14 @@ def prepare_test_libs(base_path):
     A = tvm.placeholder((n,), name='A')
     B = tvm.compute(A.shape, lambda *i: A(*i) + 1.0, name='B')
     s = tvm.create_schedule(B.op)
+    # Compile library as dynamic library
     fadd_dylib = tvm.build(s, [A, B], "llvm", name="addone")
-    fadd_syslib = tvm.build(s, [A, B], "llvm --system-lib", name="addonesys")
     dylib_path = os.path.join(base_path, "test_addone_dll.so")
-    syslib_path = os.path.join(base_path, "test_addone_sys.o")
     fadd_dylib.export_library(dylib_path)
+
+    # Compile library in system library mode
+    fadd_syslib = tvm.build(s, [A, B], "llvm --system-lib", name="addonesys")
+    syslib_path = os.path.join(base_path, "test_addone_sys.o")
     fadd_syslib.save(syslib_path)
 
 if __name__ == "__main__":

apps/howto_deploy/run_example.sh

+#!/bin/bash
+echo "Build the libraries.."
+mkdir -p lib
+make
+echo "Run the example"
+export LD_LIBRARY_PATH=../../lib:${LD_LIBRARY_PATH}
+export DYLD_LIBRARY_PATH=../../lib:${DYLD_LIBRARY_PATH}
+
+echo "Run the deployment with all in one packed library..."
+lib/cpp_deploy_pack
+
+echo "Run the deployment with all in normal library..."
+lib/cpp_deploy_normal

apps/howto_deploy/tvm_runtime_pack.cc

+/*!
+ * \brief This is an all in one TVM runtime file.
+ *
+ *   You only have to use this file to compile libtvm_runtime to
+ *   include in your project.
+ *
+ *  - Copy this file into your project which depends on tvm runtime.
+ *  - Compile with -std=c++11
+ *  - Add the following include path
+ *     - /path/to/tvm/include/
+ *     - /path/to/tvm/dmlc-core/include/
+ *     - /path/to/tvm/dlpack/include/
+ *   - Add -lpthread -ldl to the linked library.
+ *   - You are good to go.
+ *   - See the Makefile in the same folder for example.
+ *
+ *  The include files here are presented with relative path
+ *  You need to remember to change it to point to the right file.
+ *
+ */
+#include "../../src/runtime/c_runtime_api.cc"
+#include "../../src/runtime/cpu_device_api.cc"
+#include "../../src/runtime/workspace_pool.cc"
+#include "../../src/runtime/module_util.cc"
+#include "../../src/runtime/module.cc"
+#include "../../src/runtime/registry.cc"
+#include "../../src/runtime/file_util.cc"
+#include "../../src/runtime/thread_pool.cc"
+
+// NOTE: all the files after this are optional modules
+// that you can include remove, depending on how much feature you use.
+
+// Likely we only need to enable one of the following
+// If you use Module::Load, use dso_module
+// For system packed library, use system_lib_module
+#include "../../src/runtime/dso_module.cc"
+#include "../../src/runtime/system_lib_module.cc"
+
+// Graph runtime
+#include "../../src/runtime/graph/graph_runtime.cc"
+
+// Uncomment the following lines to enable RPC
+// #include "../../src/runtime/rpc/rpc_session.cc"
+// #include "../../src/runtime/rpc/rpc_event_impl.cc"
+// #include "../../src/runtime/rpc/rpc_server_env.cc"
+
+// Uncomment the following lines to enable Metal
+// #include "../../src/runtime/metal/metal_device_api.mm"
+// #include "../../src/runtime/metal/metal_module.mm"
+
+// Uncomment the following lines to enable OpenCL
+// #include "../../src/runtime/opencl/opencl_device_api.cc"
+// #include "../../src/runtime/opencl/opencl_module.cc"

docs/how_to/deploy.md

+How to Deploy TVM Modules
+=========================
+We provide an example on how to deploy TVM modules in [apps/howto_deploy](https://github.com/dmlc/tvm/tree/master/apps/howto_deploy)
+
+To run the example, you can use the following command
+
+```bash
+cd apps/howto_deploy
+./run_example.sh
+```
+
+Get TVM Runtime Library
+-----------------------
+
+![](http://www.tvmlang.org/images/release/tvm_flexible.png)
+
+The only thing we need is to link to a TVM runtime in your target platform.
+TVM provides a minimum runtime, which costs around 300K to 600K depending on how much modules we use.
+In most cases, we can use ```libtvm_runtime.so``` that comes with the build.
+
+If somehow you find it is hard to build ```libtvm_runtime```, checkout [tvm_runtime_pack.cc](https://github.com/dmlc/tvm/tree/master/apps/howto_deploy/tvm_runtime_pack.cc).
+It is an example all in one file that gives you TVM runtime.
+You can compile this file using your build system and include this into your project.
+
+You can also checkout [apps](https://github.com/dmlc/tvm/tree/master/apps/) for example applications build with TVM on iOS, Android and others.
+
+Dynamic Library vs. System Module
+---------------------------------
+TVM provides two ways to use the compiled library.
+You can checkout [prepare_test_libs.py](https://github.com/dmlc/tvm/tree/master/apps/howto_deploy/prepare_test_libs.py)
+on how to generate the library and [cpp_deploy.cc](https://github.com/dmlc/tvm/tree/master/apps/howto_deploy/cpp_deploy.cc) on how to use them.
+
+- Store library as a shared library and dynamically load the library into your project.
+- 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.

docs/how_to/install.md

@@ -77,7 +77,7 @@ There are several ways to install the package:
     The changes will be immediately reflected once you pulled the code and rebuild the project (no need to call ```setup``` again)
 
     ```bash
-    export PYTHONPATH=/path/to/tvm/python:${PYTHONPATH}
+    export PYTHONPATH=/path/to/tvm/python:/path/to/tvm/topi/python:${PYTHONPATH}
     ```
 
 2. Install tvm python bindings by `setup.py`:
@@ -87,7 +87,6 @@ There are several ways to install the package:
     # NOTE: if you installed python via homebrew, --user is not needed during installaiton
     #       it will be automatically installed to your user directory.
     #       providing --user flag may trigger error during installation in such case.
-    cd python; python setup.py install --user
-    # or install tvm package system wide
-    cd python; sudo python setup.py install
+    cd python; python setup.py install --user; cd ..
+    cd topi/python; python setup.py install --user; cd ../..
     ```

docs/index.rst

@@ -14,6 +14,7 @@ Contents
    how_to/install
    tutorials/index
    faq
+   how_to/deploy
    how_to/contribute
    api/python/index
    dev/index

topi/python/setup.py

+# pylint: disable=invalid-name, exec-used
+"""Setup TOPI package."""
+from __future__ import absolute_import
+import sys
+
+from setuptools import find_packages
+from setuptools.dist import Distribution
+
+if "--inplace" in sys.argv:
+    from distutils.core import setup
+    from distutils.extension import Extension
+else:
+    from setuptools import setup
+    from setuptools.extension import Extension
+
+__version__ = "0.1.0"
+
+setup(name='topi',
+      version=__version__,
+      description="TOPI: TVM operator index",
+      install_requires=[
+        "numpy",
+        "decorator",
+        ],
+      packages=find_packages(),
+      url='https://github.com/dmlc/tvm')