Improved MLF to contain workspace info

Added functionality to calculate workspace, io and constant
memory required by each primfunc and main function. Moreover,
the workspace information required by each primfunc and main
is reported in metadata.json in the Model Library Format(MLF).
- added functionality to record tir and relay primfuncs
- added tests for model_library_format changes

Change-Id: Ib4a8b787345aa35f8a1645e8a648fad84de37bce

python/tvm/micro/model_library_format.py

@@ -27,6 +27,8 @@
 from ..relay.backend import executor_factory
 from ..relay import param_dict
 
+MAIN_FUNC_NAME_STR = "run"
+
 
 class UnsupportedInModelLibraryFormatError(Exception):
     """Raised when export_model_library_format does not support the given Module tree."""
@@ -73,8 +75,16 @@ def _populate_codegen_dir(mod, codegen_dir: str):
         dso_mod.save(file_name)
 
 
-def _build_memory_map(graph_json):
-    """Build a simpler memory map from graph JSON.
+def _build_memory_map(mod):
+    ret = dict()
+    if isinstance(mod, executor_factory.GraphExecutorFactoryModule):
+        ret["sids"] = _build_sid_map(mod.graph_json)
+    ret["functions"] = _build_function_memory_map(mod.function_metadata)
+    return ret
+
+
+def _build_sid_map(graph_json):
+    """Build a simpler storage id info map from graph JSON.
 
     Parameters
     ----------
@@ -117,6 +127,81 @@ def _build_memory_map(graph_json):
     return memory_map
 
 
+def _build_function_memory_map(function_metadata):
+    """Build a simple map that shows how much workspace is required to execute
+    each primitive function. The main_func describes how much memory is required
+    to execute the main control code.
+
+    Parameters
+    ----------
+    function_metadata : Map<String, FunctionInfo>
+        This contains all the compiled metadata on a function basis
+
+    Returns
+    -------
+    dict :
+        This will have two entries:
+        1.) A list with one entry per function describing local memory it is using.
+        2.) A global memory requirement if all functions are executed sequentially
+    """
+    device_max_workspace = dict()
+    num_targets = len(function_metadata[MAIN_FUNC_NAME_STR].workspace_sizes.items())
+    func_entries = []
+    target_local_entries = dict()
+    for i in range(num_targets):
+        for func_name, finfo in function_metadata.items():
+            if func_name == MAIN_FUNC_NAME_STR:
+                continue
+            target = finfo.workspace_sizes.items()[i][0]
+            device_max_workspace[target] = 0
+            target_local_entries[func_name] = list()
+
+        for func_name, finfo in function_metadata.items():
+            if func_name == MAIN_FUNC_NAME_STR:
+                continue
+            assert len(finfo.constant_sizes.items()) == num_targets
+            assert len(finfo.io_sizes.items()) == num_targets
+            target = finfo.workspace_sizes.items()[i][0]
+            workspace_size = finfo.workspace_sizes.items()[i][1]
+            target_entry = {
+                "device": int(target.kind.device_type),
+                "workspace_size_bytes": int(workspace_size),
+            }
+            target_local_entries[func_name].append(target_entry)
+            if workspace_size > device_max_workspace[target]:
+                device_max_workspace[target] = workspace_size
+
+    for func_name, target_entries_ in target_local_entries.items():
+        func_entry = {
+            "function_name": str(func_name),
+            "workspace": target_entries_,
+        }
+        func_entries.append(func_entry)
+
+    target_main_entries = list()
+    main_func_metadata = function_metadata[MAIN_FUNC_NAME_STR]
+    for i in range(num_targets):
+        target = main_func_metadata.workspace_sizes.items()[i][0]
+        main_func_local_workspace = main_func_metadata.workspace_sizes.items()[i][1]
+        main_func_constants = main_func_metadata.constant_sizes.items()[i][1]
+        main_func_io = main_func_metadata.io_sizes.items()[i][1]
+        target_main_entries.append(
+            {
+                "device": int(target.kind.device_type),
+                "workspace_size_bytes": int(device_max_workspace[target])
+                + int(main_func_local_workspace),
+                "constants_size_bytes": int(main_func_constants),
+                "io_size_bytes": int(main_func_io),
+            }
+        )
+
+    ret = {
+        "operator_functions": func_entries,
+        "main_function": target_main_entries,
+    }
+    return ret
+
+
 def export_model_library_format(mod: executor_factory.ExecutorFactoryModule, file_name):
     """Export the build artifact in Model Library Format.
 
@@ -133,14 +218,13 @@ def export_model_library_format(mod: executor_factory.ExecutorFactoryModule, fil
     """
     tempdir = utils.tempdir()
     is_aot = isinstance(mod, executor_factory.AOTExecutorFactoryModule)
-    memory_map = [] if is_aot else _build_memory_map(mod.get_executor_config())
     runtime = ["aot"] if is_aot else ["graph"]
 
     metadata = {
-        "version": 1,
+        "version": 2,
         "model_name": mod.libmod_name,
         "export_datetime": datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%SZ"),
-        "memory": memory_map,
+        "memory": _build_memory_map(mod),
         "target": {int(k): str(v) for k, v in mod.target.items()},
         "runtimes": runtime,
     }

python/tvm/relay/backend/__init__.py

@@ -16,3 +16,4 @@
 # under the License.
 """Backend codegen modules for relay."""
 from . import compile_engine
+from . import utils

python/tvm/relay/backend/_ffi_api.py

@@ -0,0 +1,21 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""FFI APIs for tvm.relay.backend"""
+import tvm._ffi
+
+
+tvm._ffi._init_api("relay.backend", __name__)

python/tvm/relay/backend/executor_factory.py

@@ -81,15 +81,18 @@ class AOTExecutorFactoryModule(ExecutorFactoryModule):
         The name of module
     params : dict of str to NDArray
         The parameters of module
+    function_metadata : Map of String to FunctionInfo
+        This holds a map function names to their information
     """
 
-    def __init__(self, ir_mod, target, libmod, libmod_name, params):
+    def __init__(self, ir_mod, target, libmod, libmod_name, params, function_metadata):
         self.ir_mod = ir_mod
         self.target = target
         self.lib = libmod
         self.libmod_name = libmod_name
         self.params = params
         self.iter_cnt = 0
+        self.function_metadata = function_metadata
 
     def get_params(self):
         return self.params
@@ -118,9 +121,13 @@ class GraphExecutorFactoryModule(ExecutorFactoryModule):
         The name of module
     params : dict of str to NDArray
         The parameters of module
+    function_metadata : Map of String to FunctionInfo
+        This holds a map function names to their information
     """
 
-    def __init__(self, ir_mod, target, graph_json_str, libmod, libmod_name, params):
+    def __init__(
+        self, ir_mod, target, graph_json_str, libmod, libmod_name, params, function_metadata
+    ):
         assert isinstance(graph_json_str, string_types)
         fcreate = get_global_func("tvm.graph_executor_factory.create")
         args = []
@@ -136,6 +143,7 @@ def __init__(self, ir_mod, target, graph_json_str, libmod, libmod_name, params):
         self.libmod_name = libmod_name
         self.params = params
         self.iter_cnt = 0
+        self.function_metadata = function_metadata
 
     def export_library(self, file_name, fcompile=None, addons=None, **kwargs):
         return self.module.export_library(file_name, fcompile, addons, **kwargs)

python/tvm/relay/backend/utils.py

@@ -0,0 +1,29 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""The utility functions and classes for relay backend compilation"""
+from tvm.runtime import Object
+from . import _ffi_api
+
+
+class FunctionInfo(Object):
+    """A data structure to hold metadata of relay primitive functions"""
+
+    def __init__(self, dummy):
+        self.__init_handle_by_constructor__(_ffi_api.FunctionInfo, dummy)
+
+    def set_workspace_size(self, target, size):
+        _ffi_api._FunctionInfo_SetWorkspaceSize(self, target, size)

python/tvm/relay/build_module.py

@@ -83,6 +83,7 @@ def __init__(self):
         self._optimize = self.mod["optimize"]
         self._set_params_func = self.mod["set_params"]
         self._get_params_func = self.mod["get_params"]
+        self._get_function_metadata = self.mod["get_function_metadata"]
 
     def build(self, mod, target=None, target_host=None, params=None, executor="graph"):
         """
@@ -200,6 +201,12 @@ def get_module(self):
         """Return the built module."""
         return self._get_module()
 
+    def get_function_metadata(self):
+        """Return the compiled function metadata.
+        Currently, the metadata contains workspace size required by
+        each PrimFunc"""
+        return self._get_function_metadata()
+
     def get_params(self):
         """Return the updated weights."""
         params = self._get_params_func()
@@ -325,14 +332,15 @@ def build(ir_mod, target=None, target_host=None, params=None, mod_name="default"
         executor_config, runtime_mod, params = bld_mod.build(
             mod=ir_mod, target=target, params=params, executor=executor
         )
+        func_metadata = bld_mod.get_function_metadata()
 
         if executor == "aot":
             executor_factory = _executor_factory.AOTExecutorFactoryModule(
-                ir_mod, target, runtime_mod, mod_name, params
+                ir_mod, target, runtime_mod, mod_name, params, func_metadata
             )
         elif executor == "graph":
             executor_factory = _executor_factory.GraphExecutorFactoryModule(
-                ir_mod, target, executor_config, runtime_mod, mod_name, params
+                ir_mod, target, executor_config, runtime_mod, mod_name, params, func_metadata
             )
         else:
             assert False, "Executor " + executor + " not supported"

src/relay/backend/build_module.cc

@@ -62,6 +62,10 @@ struct ExecutorCodegen {
 
   virtual void UpdateOutput(BuildOutput* ret) = 0;
 
+  Map<String, FunctionInfo> GetFunctionMetadata() {
+    return CallFunc<Map<String, FunctionInfo>>("get_function_metadata", nullptr);
+  }
+
   std::unordered_map<std::string, tvm::runtime::NDArray> GetParams() {
     std::unordered_map<std::string, tvm::runtime::NDArray> ret;
     auto names = CallFunc<Array<runtime::String>>("list_params_name", nullptr);
@@ -197,6 +201,10 @@ class RelayBuildModule : public runtime::ModuleNode {
       return PackedFunc([sptr_to_self, this](TVMArgs args, TVMRetValue* rv) {
         *rv = this->executor_codegen_->GetExternalModules();
       });
+    } else if (name == "get_function_metadata") {
+      return PackedFunc([sptr_to_self, this](TVMArgs args, TVMRetValue* rv) {
+        *rv = this->executor_codegen_->GetFunctionMetadata();
+      });
     } else if (name == "optimize") {
       return PackedFunc([sptr_to_self, this](TVMArgs args, TVMRetValue* rv) {
         ICHECK_EQ(args.num_args, 2);