runtime.md

TVM export the runtime module

python: lib initialized
python: -> self.module = tvm.graph_executor_factory.create(libmod, libmod_name, *args);
cpp:      -> call `make_object<GraphExecutorFactory>@src/runtime/graph_executor/graph_executor_factory.cc:214` make a `GraphExecutorFactory`
          -> call `exec->Import(args[1]);`; import from the graph_json_str

python: lib.export_library
python:  -> self.module.export_library
python:    -> export_library@python/tvm/runtime/module.py:535
python:       -> save@python/tvm/runtime/module.py:316
cpp:            -> LLVMModuleNode::SaveToFile@src/target/llvm/llvm_module.cc:242 [This is for lib0.o]
cpp:            -> PackImportsToLLVM@src/target/codegen.cc:362 [This is for devc.o]
cpp:              -> call `CodeGenBlob@src/target/llvm/codegen_blob.cc:64` [will initialize the `LLVMJITEngine`]




## TVM runtime source workflow

```text
python: load_module
  -> call `_ffi_api.ModuleLoadFromFile@python/tvm/runtime/module.py:708`
    -> call `Module::LoadFromFile@src/runtime/module.cc:79`
      -> call `runtime.module.loadfile_so@src/runtime/dso_library.cc:152`
        -> call `CreateDSOLibraryObject`
           This function will call the `dlopen` and `dlsym` to analysis the callee function.
        -> call `CreateModuleFromLibrary@src/runtime/library_module.cc:204`
          -> call `ProcessModuleBlob@src/runtime/library_module.cc:137`
            -> There are 3 tkey can be load
            -> `tkey == "_lib"` call `make_object<LibraryModuleNode>`
            -> `tkey == "_import_tree"` read `import_tree_row_ptr` and `import_tree_child_indices`
            -> `tkey == <others>` call `LoadModuleFromBinary@src/runtime/library_module.cc:107`
              -> call `runtime.module.loadbinary_<tkey>` for target
                -> call `LoadFromBinary@src/runtime/contrib/target/target_runtime.cc:140`
            -> push back into the Module List
        -> return the root module which is the first module in the list.

During the CreateModuleFromLibrary function, the net_onnx_rdl.tar will be untar and with g++ -shared -fPIC -o <output_name>.so devc.o lib0.o to create the shared library.

python: tvm.graph_executor.create
-> call the `GraphExecutorCreate@src/runtime/graph_executor/graph_executor.cc:788`
  -> call the `exec->Init`
    -> call the `this->Load(&reader)` # This is the `Load` function.
    -> call `GraphExecutor::SetupStorage()`
      -> calculate the tensor space and create a storage list and allocate the memory.
    -> call `GraphExecutor::SetupOpExecs()`
      -> create the input and output tensor list and create the `CreateTVMOp` list.
      -> `CreateTVMOp@src/runtime/graph_executor/graph_executor.cc:603`
        -> scan the input and output and create the TVMArgs
        -> `module_.GetFunction(param.func_name, true);`
          # Here the param.func_name is the function name in the json.
          # So, This will call the `TargetRuntime::GetFunction` since the name is for target.
          # -> `ModuleNode::GetFunction@src/runtime/module.cc:64`
        -> return the function ptr and args ptr.

python:
  -> call `GraphExecutor.run`
    -> call the `GraphExecutor::GetFunction@src/runtime/graph_executor/graph_executor.cc:653`
      -> call `GraphExecutor::Run()@src/runtime/graph_executor/graph_executor.cc@69`
        -> you can find source code that the run the iteration of the `CreateTVMOp` list.

Note: call the GetFuntion(default) function, This will convert the tensor into device tensor.