comaniac commented on a change in pull request #8702: URL: https://github.com/apache/tvm/pull/8702#discussion_r704643441
########## File path: python/tvm/contrib/pipeline_executor.py ########## @@ -0,0 +1,538 @@ +# 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. +"""Pipeline executor that executes a series of modules in a pipeline fashion.""" +import json +import tvm._ffi +from tvm import relay +from tvm.relay.transform import InferType +from tvm.contrib import graph_executor + + +def pipeline_executor_enabled(): + """check if pipeline executor is enabled. + + Return + ------- + enable: bool + Return pipeline executor is enabled or not. + """ + return tvm._ffi.get_global_func("tvm.pipeline_executor.create", allow_missing=True) is not None + + +def build(pipe_configs): + """build module list that can use for pipeline execution. + + Parameters + ---------- + pipe_configs: PipelineConfig + build configuration informaton. + + Returns + ------- + ret: PipeModuleConfig Review comment: PipelineExecutorFactoryModule ########## File path: python/tvm/contrib/pipeline_executor.py ########## @@ -0,0 +1,538 @@ +# 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. +"""Pipeline executor that executes a series of modules in a pipeline fashion.""" +import json +import tvm._ffi +from tvm import relay +from tvm.relay.transform import InferType +from tvm.contrib import graph_executor + + +def pipeline_executor_enabled(): + """check if pipeline executor is enabled. + + Return + ------- + enable: bool + Return pipeline executor is enabled or not. + """ + return tvm._ffi.get_global_func("tvm.pipeline_executor.create", allow_missing=True) is not None + + +def build(pipe_configs): + """build module list that can use for pipeline execution. + + Parameters + ---------- + pipe_configs: PipelineConfig + build configuration informaton. + + Returns + ------- + ret: PipeModuleConfig + the class that wrap module list and configuration. + """ + mods = {} + mod_n_configs = pipe_configs.get_config() + config_len = len(mod_n_configs) + string_config = [{} for _ in range(config_len)] + for ir_mod, mod_config in mod_n_configs.items(): + mconf = mod_config["pipeline"].copy() + mod_idx = mconf["mod_idx"] - 1 + # Get mod device config + dev = mod_config["dev"] + target = mod_config["target"] + build_func = relay.build + # if there is a self defined build function then use it. + if "build" in mod_config and mod_config["build"]: + build_func = mod_config["build"] + + # build IRModule + mod = build_func( + ir_mod, + target, + params=mod_config["params"], + target_host=mod_config["target_host"], + mod_name=mod_config["mod_name"], + ) + + mconf["dev"] = "{},{}".format(dev.device_type, dev.device_id) + # Create pipeline configuration + string_config[mod_idx] = mconf + # associate mod with device + mods[mod] = {"dev": dev} + + # return PipelineExecutorFactoryModule + return PipelineExecutorFactoryModule(mods, string_config) + + +def create(pipe_mod_config): + """Create a pipeline runtime executor. + + Parameters + ---------- + pipe_mod_config : PipeModuleConfig + class to storage IRModule list and pipeline configuration. + + Returns + submodule : PipelineModule + Runtime pipeline module. + """ + + return PipelineModule(pipe_mod_config) + + +class PipelineModule(object): + """Wrapper runtime module. This is a thin wrapper of the underlying TVM module. + + Parameters + ---------- + pipeline_mods : List[GraphModule] + The internal tvm module that holds the actual graph functions. + pipeline_config : Dict[IRModule, Dict[str, Any]] + modules and modules dependency configuration informaiton. + """ + + def __init__(self, pipe_mod_config): + self.pipeline_mods_ = pipe_mod_config.pipeline_mods_ + self.mod_config_ = pipe_mod_config.mods_config_ + mods, config = self.graph_executor_create(self.pipeline_mods_, self.mod_config_) + assert ( + pipeline_executor_enabled() + ), "Pipeline executor is not enabled. Please \ + re-build TVM with USE_PIPELINE_EXECUTOR=ON" + pipelinecreate = tvm._ffi.get_global_func( + "tvm.pipeline_executor.create", allow_missing=False + ) + assert pipelinecreate + module = pipelinecreate(mods, config) + + self.module_ = module + + def graph_executor_create(self, pipeline_mods, mod_config): + """Create graph_executor list and return string format config. + + Parameters + ---------- + pipeline_mods : List[IRModule] + list of IRModule + + mod_config : Dict[int, Dict[str, Any]] + modules and modules dependency configuration informaiton. + + Returns + ------- + mods : List[GraphModule] + Runtime graph module. + + mod_config : str + mods configuration + """ + + mods = [] + for pipeline_mod in pipeline_mods: + mod = graph_executor.GraphModule( + pipeline_mod["default"](pipeline_mods[pipeline_mod]["dev"]) + ) + mods.append(mod.module) + + return mods, json.dumps(mod_config) + + +class PipelineConfig(object): + """The wrapper of each module to be pipelined. The wrapper mainly includes the + module itself as well as the binding that represents the connections of this + module's inputs and outputs to other modules. + """ + + class ModuleWrapper: + """The class use use to represent Module and storage module idx and + Binding information. + """ + + class Binding: + """The class that use to storage module connection information. + The binding can be either "input" or "output". + + Parameters + ---------- + owner : ModuleWrapper + The class that own this interface, in such class there are + Module information like idx, module name + + io_type : str + The type of this binding. It can be either "input" or "output". + + name : str/integer + Binding name, for input it is string such as "data0"; + for output it is the idx integer such as 0. + """ + + def __init__(self, owner, stype, name, data_type=None): + self.io_owner = owner + self.io_type = stype + self.name = str(name) + # These item that have dependency relation with self + self.bindings = [] + # The item that self depend + self.parents = [] + + self.data_type = data_type + + def get_name(self): + """get owner name and self name""" + owner_name = "" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + owner_name = self.io_owner.name + + return owner_name, self.name + + def get_owner_idx(self): + """return idx if owner is ModuleWrapper, if not return 0""" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + return self.io_owner.idx + + # if not ModuleWrapper then owner is PipelineConfig, return 0 + # to identify this is global interface + return 0 + + def is_global_interface(self): + """check if this interface is global""" + return not isinstance(self.io_owner, PipelineConfig.ModuleWrapper) + + def __repr__(self): + """Get all binding(input data), exepect like |data_0: mod1:data_0""" + ret = " |{}: ".format(self.name) + for binding in self.bindings: + mname, dname = binding.get_name() + ret += "{0}:{1} ".format(mname, dname) + return ret + + def check_dag_acyclic(self, start, inputs): + """check if the DAG that current binding stay is acircle""" + for binding in inputs.values(): + if start == binding.io_owner: + return False + for p in binding.parents: + if not self.check_dag_acyclic(start, p.io_owner.input_bindings.bindings): + return False + + return True + + def connect(self, binding): + """ + # check if the bindendency setting correct. + # correct connection are following + # 1. global input to module input + # 2. module output to global output + # 3. module output to moudle input + """ + if self.io_owner == binding.io_owner: + raise RuntimeError(f"can not set self as binding.") + + if not self.is_global_interface() and self.io_type == "input": + raise RuntimeError(f"Module only can start binding from output!") + + if ( + not self.is_global_interface() + and not binding.is_global_interface() + and binding.io_type == "output" + ): + raise RuntimeError(f"Module output can not binding with module output!") + + if ( + not self.is_global_interface() + and binding.is_global_interface() + and binding.io_type == "input" + ): + raise RuntimeError(f"Module output can not binding with global input!") + + if self.is_global_interface() and self.io_type != "input": + raise RuntimeError(f"Global only can start binding from input!") + + if self.is_global_interface() and binding.io_type != "input": + raise RuntimeError(f"Global input only can set binding with module input.") + + self.bindings.append(binding) + if not self.is_global_interface(): + # check if the source and target data_type same + if ( + isinstance(binding.io_owner, PipelineConfig.ModuleWrapper) + and self.data_type != binding.data_type + ): + raise RuntimeError( + f"Illegal type (%s vs. %s): binding type is not same!" + % (self.data_type, binding.data_type) + ) + + binding.parents.append(self) + # Do acyclic check after increase the in-degree. + if not self.check_dag_acyclic( + binding.io_owner, self.io_owner.input_bindings.bindings + ): + raise RuntimeError(f"Illegal connection: cause a circle!") + + def __init__(self, mod=None): + """init class""" + self.target_host = None + self.build_func = None + self.params = None + self.target = None + self.name = None + self.dev = None + self.idx = None + self.mod = mod + self.input_params = InferType()(mod)["main"].params + self.output_values = InferType()(mod)["main"].checked_type.ret_type + self.input_bindings = PipelineConfig.BindingList(self, "input") + self.output_bindings = PipelineConfig.BindingList(self, "output") + + def __eq__(self, other): + """check if self equl other""" + if isinstance(other, PipelineConfig.ModuleWrapper): + return self.mod == other.mod + + return False + + def __getitem__(self, key): + """get item by key""" + if isinstance(key, str): + if key == "input": + return self.input_bindings + + if key == "output": + return self.output_bindings + + raise RuntimeError(f"{key} not found!") + + def get_data_type(self, key, stype): + """get module input/output data type.""" + if stype == "input": + for param in self.input_params: + if param.name_hint == key: + return param._checked_type_ + + if stype == "output": + if isinstance(self.output_values, tvm.ir.type.TupleType): + if int(key) < len(self.output_values.fields): + return self.output_values.fields[int(key)] + elif int(key) == 0: + return self.output_values + + return None + + def set_idx_name(self, idx): + """generate name by idx and storage idx value""" + self.idx = idx + self.name = "mod{}".format(str(idx)) + + def is_root_mod(self): + """use by dag topology sort, identify if this item is root item""" + for _, binding in self.input_bindings.bindings.items(): + if binding.parents: + return False + + return True Review comment: nit: ```suggestion return all([not b.parents for b in self.input_bindings.bindings.values()]) ``` ########## File path: python/tvm/contrib/pipeline_executor.py ########## @@ -0,0 +1,538 @@ +# 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. +"""Pipeline executor that executes a series of modules in a pipeline fashion.""" +import json +import tvm._ffi +from tvm import relay +from tvm.relay.transform import InferType +from tvm.contrib import graph_executor + + +def pipeline_executor_enabled(): + """check if pipeline executor is enabled. + + Return + ------- + enable: bool + Return pipeline executor is enabled or not. + """ + return tvm._ffi.get_global_func("tvm.pipeline_executor.create", allow_missing=True) is not None + + +def build(pipe_configs): + """build module list that can use for pipeline execution. + + Parameters + ---------- + pipe_configs: PipelineConfig + build configuration informaton. + + Returns + ------- + ret: PipeModuleConfig + the class that wrap module list and configuration. + """ + mods = {} + mod_n_configs = pipe_configs.get_config() + config_len = len(mod_n_configs) + string_config = [{} for _ in range(config_len)] + for ir_mod, mod_config in mod_n_configs.items(): + mconf = mod_config["pipeline"].copy() + mod_idx = mconf["mod_idx"] - 1 + # Get mod device config + dev = mod_config["dev"] + target = mod_config["target"] + build_func = relay.build + # if there is a self defined build function then use it. + if "build" in mod_config and mod_config["build"]: + build_func = mod_config["build"] + + # build IRModule + mod = build_func( + ir_mod, + target, + params=mod_config["params"], + target_host=mod_config["target_host"], + mod_name=mod_config["mod_name"], + ) + + mconf["dev"] = "{},{}".format(dev.device_type, dev.device_id) + # Create pipeline configuration + string_config[mod_idx] = mconf + # associate mod with device + mods[mod] = {"dev": dev} + + # return PipelineExecutorFactoryModule + return PipelineExecutorFactoryModule(mods, string_config) + + +def create(pipe_mod_config): + """Create a pipeline runtime executor. + + Parameters + ---------- + pipe_mod_config : PipeModuleConfig + class to storage IRModule list and pipeline configuration. + + Returns + submodule : PipelineModule + Runtime pipeline module. + """ + + return PipelineModule(pipe_mod_config) + + +class PipelineModule(object): + """Wrapper runtime module. This is a thin wrapper of the underlying TVM module. + + Parameters + ---------- + pipeline_mods : List[GraphModule] + The internal tvm module that holds the actual graph functions. + pipeline_config : Dict[IRModule, Dict[str, Any]] + modules and modules dependency configuration informaiton. + """ + + def __init__(self, pipe_mod_config): + self.pipeline_mods_ = pipe_mod_config.pipeline_mods_ + self.mod_config_ = pipe_mod_config.mods_config_ + mods, config = self.graph_executor_create(self.pipeline_mods_, self.mod_config_) + assert ( + pipeline_executor_enabled() + ), "Pipeline executor is not enabled. Please \ + re-build TVM with USE_PIPELINE_EXECUTOR=ON" + pipelinecreate = tvm._ffi.get_global_func( + "tvm.pipeline_executor.create", allow_missing=False + ) + assert pipelinecreate + module = pipelinecreate(mods, config) + + self.module_ = module + + def graph_executor_create(self, pipeline_mods, mod_config): + """Create graph_executor list and return string format config. + + Parameters + ---------- + pipeline_mods : List[IRModule] + list of IRModule + + mod_config : Dict[int, Dict[str, Any]] + modules and modules dependency configuration informaiton. + + Returns + ------- + mods : List[GraphModule] + Runtime graph module. + + mod_config : str + mods configuration + """ + + mods = [] + for pipeline_mod in pipeline_mods: + mod = graph_executor.GraphModule( + pipeline_mod["default"](pipeline_mods[pipeline_mod]["dev"]) + ) + mods.append(mod.module) + + return mods, json.dumps(mod_config) + + +class PipelineConfig(object): + """The wrapper of each module to be pipelined. The wrapper mainly includes the + module itself as well as the binding that represents the connections of this + module's inputs and outputs to other modules. + """ + + class ModuleWrapper: + """The class use use to represent Module and storage module idx and + Binding information. + """ + + class Binding: + """The class that use to storage module connection information. + The binding can be either "input" or "output". + + Parameters + ---------- + owner : ModuleWrapper + The class that own this interface, in such class there are + Module information like idx, module name + + io_type : str + The type of this binding. It can be either "input" or "output". + + name : str/integer + Binding name, for input it is string such as "data0"; + for output it is the idx integer such as 0. + """ + + def __init__(self, owner, stype, name, data_type=None): + self.io_owner = owner + self.io_type = stype + self.name = str(name) + # These item that have dependency relation with self + self.bindings = [] + # The item that self depend + self.parents = [] + + self.data_type = data_type + + def get_name(self): + """get owner name and self name""" + owner_name = "" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + owner_name = self.io_owner.name + + return owner_name, self.name + + def get_owner_idx(self): + """return idx if owner is ModuleWrapper, if not return 0""" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + return self.io_owner.idx + + # if not ModuleWrapper then owner is PipelineConfig, return 0 + # to identify this is global interface + return 0 + + def is_global_interface(self): + """check if this interface is global""" + return not isinstance(self.io_owner, PipelineConfig.ModuleWrapper) + + def __repr__(self): + """Get all binding(input data), exepect like |data_0: mod1:data_0""" + ret = " |{}: ".format(self.name) + for binding in self.bindings: + mname, dname = binding.get_name() + ret += "{0}:{1} ".format(mname, dname) + return ret + + def check_dag_acyclic(self, start, inputs): + """check if the DAG that current binding stay is acircle""" + for binding in inputs.values(): + if start == binding.io_owner: + return False + for p in binding.parents: + if not self.check_dag_acyclic(start, p.io_owner.input_bindings.bindings): + return False + + return True + + def connect(self, binding): + """ + # check if the bindendency setting correct. + # correct connection are following + # 1. global input to module input + # 2. module output to global output + # 3. module output to moudle input + """ + if self.io_owner == binding.io_owner: + raise RuntimeError(f"can not set self as binding.") + + if not self.is_global_interface() and self.io_type == "input": + raise RuntimeError(f"Module only can start binding from output!") + + if ( + not self.is_global_interface() + and not binding.is_global_interface() + and binding.io_type == "output" + ): + raise RuntimeError(f"Module output can not binding with module output!") + + if ( + not self.is_global_interface() + and binding.is_global_interface() + and binding.io_type == "input" + ): + raise RuntimeError(f"Module output can not binding with global input!") + + if self.is_global_interface() and self.io_type != "input": + raise RuntimeError(f"Global only can start binding from input!") + + if self.is_global_interface() and binding.io_type != "input": + raise RuntimeError(f"Global input only can set binding with module input.") + + self.bindings.append(binding) + if not self.is_global_interface(): + # check if the source and target data_type same + if ( + isinstance(binding.io_owner, PipelineConfig.ModuleWrapper) + and self.data_type != binding.data_type + ): + raise RuntimeError( + f"Illegal type (%s vs. %s): binding type is not same!" + % (self.data_type, binding.data_type) + ) + + binding.parents.append(self) + # Do acyclic check after increase the in-degree. + if not self.check_dag_acyclic( + binding.io_owner, self.io_owner.input_bindings.bindings + ): + raise RuntimeError(f"Illegal connection: cause a circle!") + + def __init__(self, mod=None): + """init class""" + self.target_host = None + self.build_func = None + self.params = None + self.target = None + self.name = None + self.dev = None + self.idx = None + self.mod = mod + self.input_params = InferType()(mod)["main"].params + self.output_values = InferType()(mod)["main"].checked_type.ret_type + self.input_bindings = PipelineConfig.BindingList(self, "input") + self.output_bindings = PipelineConfig.BindingList(self, "output") + + def __eq__(self, other): + """check if self equl other""" + if isinstance(other, PipelineConfig.ModuleWrapper): + return self.mod == other.mod + + return False + + def __getitem__(self, key): + """get item by key""" + if isinstance(key, str): + if key == "input": + return self.input_bindings + + if key == "output": + return self.output_bindings + + raise RuntimeError(f"{key} not found!") + + def get_data_type(self, key, stype): + """get module input/output data type.""" + if stype == "input": + for param in self.input_params: + if param.name_hint == key: + return param._checked_type_ + + if stype == "output": + if isinstance(self.output_values, tvm.ir.type.TupleType): + if int(key) < len(self.output_values.fields): + return self.output_values.fields[int(key)] + elif int(key) == 0: + return self.output_values + + return None + + def set_idx_name(self, idx): + """generate name by idx and storage idx value""" + self.idx = idx + self.name = "mod{}".format(str(idx)) + + def is_root_mod(self): + """use by dag topology sort, identify if this item is root item""" + for _, binding in self.input_bindings.bindings.items(): + if binding.parents: + return False + + return True + + def remove_self_from_bindings(self): + """use by DAG topology sort, by remove self from binding to reduce child in-degree""" + for _, binding in self.output_bindings.bindings.items(): + for child in binding.bindings: + if binding in child.parents: + child.parents.remove(binding) + + class BindingList: + """Use to storage Binding list. + + Parameters + ---------- + owner : ModuleWrapper/PipelineConfig + who own this list, it can be ModuleWrapper or PipelineConfig + + type_name : str + The type of this binding list. It can be either "input" or "output". + """ + + def __init__(self, owner, type_name): + self.bindings = {} + self.io_owner = owner + self.binding_type = type_name + + def get_binding_data_type(self, key): + """return binding data type""" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + return self.io_owner.get_data_type(key, self.binding_type) + return None + + def __getitem__(self, key): + """return item by key""" + if key not in self.bindings: + data_type = self.get_binding_data_type(key) + if not data_type and isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + raise RuntimeError(f"Illegal name: {self.binding_type}:{key} cannot find") + + self.bindings[key] = PipelineConfig.ModuleWrapper.Binding( + self.io_owner, self.binding_type, key, data_type + ) + + return self.bindings[key] + + def __init__(self): + self.mod_wrapper = {} + self.input_bindings = self.BindingList(self, "input") + self.output_bindings = self.BindingList(self, "output") + + def __str__(self): + """ Get configuration in string type""" + # topology sort to get correct module order in list. + self.dag_topology_sort() + # get input + input_dump = "Inputs\n" + for input_name in self.input_bindings.bindings: + inf = self.input_bindings.bindings[input_name] + input_dump += inf.__repr__() + "\n" + + # get connections + output = {} + connections_dump = "\nconnections\n" + for mod in self.mod_wrapper: + for _, interface in self.mod_wrapper[mod].output_bindings.bindings.items(): Review comment: ```suggestion for interface in self.mod_wrapper[mod].output_bindings.bindings.values(): ``` ########## File path: python/tvm/contrib/pipeline_executor.py ########## @@ -0,0 +1,538 @@ +# 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. +"""Pipeline executor that executes a series of modules in a pipeline fashion.""" +import json +import tvm._ffi +from tvm import relay +from tvm.relay.transform import InferType +from tvm.contrib import graph_executor + + +def pipeline_executor_enabled(): + """check if pipeline executor is enabled. + + Return + ------- + enable: bool + Return pipeline executor is enabled or not. + """ + return tvm._ffi.get_global_func("tvm.pipeline_executor.create", allow_missing=True) is not None + + +def build(pipe_configs): + """build module list that can use for pipeline execution. + + Parameters + ---------- + pipe_configs: PipelineConfig + build configuration informaton. + + Returns + ------- + ret: PipeModuleConfig + the class that wrap module list and configuration. + """ + mods = {} + mod_n_configs = pipe_configs.get_config() + config_len = len(mod_n_configs) + string_config = [{} for _ in range(config_len)] + for ir_mod, mod_config in mod_n_configs.items(): + mconf = mod_config["pipeline"].copy() + mod_idx = mconf["mod_idx"] - 1 + # Get mod device config + dev = mod_config["dev"] + target = mod_config["target"] + build_func = relay.build + # if there is a self defined build function then use it. + if "build" in mod_config and mod_config["build"]: + build_func = mod_config["build"] + + # build IRModule + mod = build_func( + ir_mod, + target, + params=mod_config["params"], + target_host=mod_config["target_host"], + mod_name=mod_config["mod_name"], + ) + + mconf["dev"] = "{},{}".format(dev.device_type, dev.device_id) + # Create pipeline configuration + string_config[mod_idx] = mconf + # associate mod with device + mods[mod] = {"dev": dev} + + # return PipelineExecutorFactoryModule + return PipelineExecutorFactoryModule(mods, string_config) + + +def create(pipe_mod_config): + """Create a pipeline runtime executor. + + Parameters + ---------- + pipe_mod_config : PipeModuleConfig + class to storage IRModule list and pipeline configuration. + + Returns + submodule : PipelineModule + Runtime pipeline module. + """ + + return PipelineModule(pipe_mod_config) + + +class PipelineModule(object): + """Wrapper runtime module. This is a thin wrapper of the underlying TVM module. + + Parameters + ---------- + pipeline_mods : List[GraphModule] + The internal tvm module that holds the actual graph functions. + pipeline_config : Dict[IRModule, Dict[str, Any]] + modules and modules dependency configuration informaiton. + """ + + def __init__(self, pipe_mod_config): + self.pipeline_mods_ = pipe_mod_config.pipeline_mods_ + self.mod_config_ = pipe_mod_config.mods_config_ + mods, config = self.graph_executor_create(self.pipeline_mods_, self.mod_config_) + assert ( + pipeline_executor_enabled() + ), "Pipeline executor is not enabled. Please \ + re-build TVM with USE_PIPELINE_EXECUTOR=ON" + pipelinecreate = tvm._ffi.get_global_func( + "tvm.pipeline_executor.create", allow_missing=False + ) + assert pipelinecreate + module = pipelinecreate(mods, config) + + self.module_ = module + + def graph_executor_create(self, pipeline_mods, mod_config): + """Create graph_executor list and return string format config. + + Parameters + ---------- + pipeline_mods : List[IRModule] + list of IRModule + + mod_config : Dict[int, Dict[str, Any]] + modules and modules dependency configuration informaiton. + + Returns + ------- + mods : List[GraphModule] + Runtime graph module. + + mod_config : str + mods configuration + """ + + mods = [] + for pipeline_mod in pipeline_mods: + mod = graph_executor.GraphModule( + pipeline_mod["default"](pipeline_mods[pipeline_mod]["dev"]) + ) + mods.append(mod.module) + + return mods, json.dumps(mod_config) + + +class PipelineConfig(object): + """The wrapper of each module to be pipelined. The wrapper mainly includes the + module itself as well as the binding that represents the connections of this + module's inputs and outputs to other modules. + """ + + class ModuleWrapper: + """The class use use to represent Module and storage module idx and + Binding information. + """ + + class Binding: + """The class that use to storage module connection information. + The binding can be either "input" or "output". + + Parameters + ---------- + owner : ModuleWrapper + The class that own this interface, in such class there are + Module information like idx, module name + + io_type : str + The type of this binding. It can be either "input" or "output". + + name : str/integer + Binding name, for input it is string such as "data0"; + for output it is the idx integer such as 0. + """ + + def __init__(self, owner, stype, name, data_type=None): + self.io_owner = owner + self.io_type = stype + self.name = str(name) + # These item that have dependency relation with self + self.bindings = [] + # The item that self depend + self.parents = [] + + self.data_type = data_type + + def get_name(self): + """get owner name and self name""" + owner_name = "" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + owner_name = self.io_owner.name + + return owner_name, self.name + + def get_owner_idx(self): + """return idx if owner is ModuleWrapper, if not return 0""" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + return self.io_owner.idx + + # if not ModuleWrapper then owner is PipelineConfig, return 0 + # to identify this is global interface + return 0 + + def is_global_interface(self): + """check if this interface is global""" + return not isinstance(self.io_owner, PipelineConfig.ModuleWrapper) + + def __repr__(self): + """Get all binding(input data), exepect like |data_0: mod1:data_0""" + ret = " |{}: ".format(self.name) + for binding in self.bindings: + mname, dname = binding.get_name() + ret += "{0}:{1} ".format(mname, dname) + return ret + + def check_dag_acyclic(self, start, inputs): + """check if the DAG that current binding stay is acircle""" + for binding in inputs.values(): + if start == binding.io_owner: + return False + for p in binding.parents: + if not self.check_dag_acyclic(start, p.io_owner.input_bindings.bindings): + return False + + return True + + def connect(self, binding): + """ + # check if the bindendency setting correct. + # correct connection are following + # 1. global input to module input + # 2. module output to global output + # 3. module output to moudle input + """ + if self.io_owner == binding.io_owner: + raise RuntimeError(f"can not set self as binding.") + + if not self.is_global_interface() and self.io_type == "input": + raise RuntimeError(f"Module only can start binding from output!") + + if ( + not self.is_global_interface() + and not binding.is_global_interface() + and binding.io_type == "output" + ): + raise RuntimeError(f"Module output can not binding with module output!") + + if ( + not self.is_global_interface() + and binding.is_global_interface() + and binding.io_type == "input" + ): + raise RuntimeError(f"Module output can not binding with global input!") + + if self.is_global_interface() and self.io_type != "input": + raise RuntimeError(f"Global only can start binding from input!") + + if self.is_global_interface() and binding.io_type != "input": + raise RuntimeError(f"Global input only can set binding with module input.") + + self.bindings.append(binding) + if not self.is_global_interface(): + # check if the source and target data_type same + if ( + isinstance(binding.io_owner, PipelineConfig.ModuleWrapper) + and self.data_type != binding.data_type + ): + raise RuntimeError( + f"Illegal type (%s vs. %s): binding type is not same!" + % (self.data_type, binding.data_type) + ) + + binding.parents.append(self) + # Do acyclic check after increase the in-degree. + if not self.check_dag_acyclic( + binding.io_owner, self.io_owner.input_bindings.bindings + ): + raise RuntimeError(f"Illegal connection: cause a circle!") + + def __init__(self, mod=None): + """init class""" + self.target_host = None + self.build_func = None + self.params = None + self.target = None + self.name = None + self.dev = None + self.idx = None + self.mod = mod + self.input_params = InferType()(mod)["main"].params + self.output_values = InferType()(mod)["main"].checked_type.ret_type + self.input_bindings = PipelineConfig.BindingList(self, "input") + self.output_bindings = PipelineConfig.BindingList(self, "output") + + def __eq__(self, other): + """check if self equl other""" + if isinstance(other, PipelineConfig.ModuleWrapper): + return self.mod == other.mod + + return False + + def __getitem__(self, key): + """get item by key""" + if isinstance(key, str): + if key == "input": + return self.input_bindings + + if key == "output": + return self.output_bindings + + raise RuntimeError(f"{key} not found!") + + def get_data_type(self, key, stype): + """get module input/output data type.""" + if stype == "input": + for param in self.input_params: + if param.name_hint == key: + return param._checked_type_ + + if stype == "output": + if isinstance(self.output_values, tvm.ir.type.TupleType): + if int(key) < len(self.output_values.fields): + return self.output_values.fields[int(key)] + elif int(key) == 0: + return self.output_values + + return None + + def set_idx_name(self, idx): + """generate name by idx and storage idx value""" + self.idx = idx + self.name = "mod{}".format(str(idx)) + + def is_root_mod(self): + """use by dag topology sort, identify if this item is root item""" + for _, binding in self.input_bindings.bindings.items(): + if binding.parents: + return False + + return True + + def remove_self_from_bindings(self): + """use by DAG topology sort, by remove self from binding to reduce child in-degree""" + for _, binding in self.output_bindings.bindings.items(): + for child in binding.bindings: + if binding in child.parents: + child.parents.remove(binding) + + class BindingList: + """Use to storage Binding list. + + Parameters + ---------- + owner : ModuleWrapper/PipelineConfig + who own this list, it can be ModuleWrapper or PipelineConfig + + type_name : str + The type of this binding list. It can be either "input" or "output". + """ + + def __init__(self, owner, type_name): + self.bindings = {} + self.io_owner = owner + self.binding_type = type_name + + def get_binding_data_type(self, key): + """return binding data type""" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + return self.io_owner.get_data_type(key, self.binding_type) + return None + + def __getitem__(self, key): + """return item by key""" + if key not in self.bindings: + data_type = self.get_binding_data_type(key) + if not data_type and isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + raise RuntimeError(f"Illegal name: {self.binding_type}:{key} cannot find") + + self.bindings[key] = PipelineConfig.ModuleWrapper.Binding( + self.io_owner, self.binding_type, key, data_type + ) + + return self.bindings[key] + + def __init__(self): + self.mod_wrapper = {} + self.input_bindings = self.BindingList(self, "input") + self.output_bindings = self.BindingList(self, "output") + + def __str__(self): + """ Get configuration in string type""" + # topology sort to get correct module order in list. + self.dag_topology_sort() + # get input + input_dump = "Inputs\n" + for input_name in self.input_bindings.bindings: + inf = self.input_bindings.bindings[input_name] + input_dump += inf.__repr__() + "\n" + + # get connections + output = {} + connections_dump = "\nconnections\n" + for mod in self.mod_wrapper: + for _, interface in self.mod_wrapper[mod].output_bindings.bindings.items(): + if interface.bindings: + mname, dname = interface.get_name() + iname = mname + ".output(" + dname + ")->" + for dep in interface.bindings: + dep_mname, dep_dname = dep.get_name() + if isinstance(dep.io_owner, PipelineConfig.ModuleWrapper): + iname += " " + dep_mname + "." + dep_dname + connections_dump += " |" + iname + "\n" + else: + output[dep_dname] = mname + ".output(" + dname + ")" Review comment: ```suggestion if isinstance(dep.io_owner, PipelineConfig.ModuleWrapper): iname += f" {dep_mname}.{dep_dname}" connections_dump += f" |{iname\n}" else: output[dep_dname] = f"{mname}.output({dname})" ``` ########## File path: python/tvm/contrib/pipeline_executor.py ########## @@ -0,0 +1,538 @@ +# 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. +"""Pipeline executor that executes a series of modules in a pipeline fashion.""" +import json +import tvm._ffi +from tvm import relay +from tvm.relay.transform import InferType +from tvm.contrib import graph_executor + + +def pipeline_executor_enabled(): + """check if pipeline executor is enabled. + + Return + ------- + enable: bool + Return pipeline executor is enabled or not. + """ + return tvm._ffi.get_global_func("tvm.pipeline_executor.create", allow_missing=True) is not None + + +def build(pipe_configs): + """build module list that can use for pipeline execution. + + Parameters + ---------- + pipe_configs: PipelineConfig + build configuration informaton. + + Returns + ------- + ret: PipeModuleConfig + the class that wrap module list and configuration. + """ + mods = {} + mod_n_configs = pipe_configs.get_config() + config_len = len(mod_n_configs) + string_config = [{} for _ in range(config_len)] + for ir_mod, mod_config in mod_n_configs.items(): + mconf = mod_config["pipeline"].copy() + mod_idx = mconf["mod_idx"] - 1 + # Get mod device config + dev = mod_config["dev"] + target = mod_config["target"] + build_func = relay.build + # if there is a self defined build function then use it. + if "build" in mod_config and mod_config["build"]: + build_func = mod_config["build"] + + # build IRModule + mod = build_func( + ir_mod, + target, + params=mod_config["params"], + target_host=mod_config["target_host"], + mod_name=mod_config["mod_name"], + ) + + mconf["dev"] = "{},{}".format(dev.device_type, dev.device_id) + # Create pipeline configuration + string_config[mod_idx] = mconf + # associate mod with device + mods[mod] = {"dev": dev} + + # return PipelineExecutorFactoryModule + return PipelineExecutorFactoryModule(mods, string_config) + + +def create(pipe_mod_config): + """Create a pipeline runtime executor. + + Parameters + ---------- + pipe_mod_config : PipeModuleConfig + class to storage IRModule list and pipeline configuration. + + Returns + submodule : PipelineModule + Runtime pipeline module. + """ + + return PipelineModule(pipe_mod_config) + + +class PipelineModule(object): + """Wrapper runtime module. This is a thin wrapper of the underlying TVM module. + + Parameters + ---------- + pipeline_mods : List[GraphModule] + The internal tvm module that holds the actual graph functions. + pipeline_config : Dict[IRModule, Dict[str, Any]] + modules and modules dependency configuration informaiton. + """ + + def __init__(self, pipe_mod_config): + self.pipeline_mods_ = pipe_mod_config.pipeline_mods_ + self.mod_config_ = pipe_mod_config.mods_config_ + mods, config = self.graph_executor_create(self.pipeline_mods_, self.mod_config_) + assert ( + pipeline_executor_enabled() + ), "Pipeline executor is not enabled. Please \ + re-build TVM with USE_PIPELINE_EXECUTOR=ON" + pipelinecreate = tvm._ffi.get_global_func( + "tvm.pipeline_executor.create", allow_missing=False + ) + assert pipelinecreate + module = pipelinecreate(mods, config) + + self.module_ = module + + def graph_executor_create(self, pipeline_mods, mod_config): + """Create graph_executor list and return string format config. + + Parameters + ---------- + pipeline_mods : List[IRModule] + list of IRModule + + mod_config : Dict[int, Dict[str, Any]] + modules and modules dependency configuration informaiton. + + Returns + ------- + mods : List[GraphModule] + Runtime graph module. + + mod_config : str + mods configuration + """ + + mods = [] + for pipeline_mod in pipeline_mods: + mod = graph_executor.GraphModule( + pipeline_mod["default"](pipeline_mods[pipeline_mod]["dev"]) + ) + mods.append(mod.module) + + return mods, json.dumps(mod_config) + + +class PipelineConfig(object): + """The wrapper of each module to be pipelined. The wrapper mainly includes the + module itself as well as the binding that represents the connections of this + module's inputs and outputs to other modules. + """ + + class ModuleWrapper: + """The class use use to represent Module and storage module idx and + Binding information. + """ + + class Binding: + """The class that use to storage module connection information. + The binding can be either "input" or "output". + + Parameters + ---------- + owner : ModuleWrapper + The class that own this interface, in such class there are + Module information like idx, module name + + io_type : str + The type of this binding. It can be either "input" or "output". + + name : str/integer + Binding name, for input it is string such as "data0"; + for output it is the idx integer such as 0. + """ + + def __init__(self, owner, stype, name, data_type=None): + self.io_owner = owner + self.io_type = stype + self.name = str(name) + # These item that have dependency relation with self + self.bindings = [] + # The item that self depend + self.parents = [] + + self.data_type = data_type + + def get_name(self): + """get owner name and self name""" + owner_name = "" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + owner_name = self.io_owner.name + + return owner_name, self.name + + def get_owner_idx(self): + """return idx if owner is ModuleWrapper, if not return 0""" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + return self.io_owner.idx + + # if not ModuleWrapper then owner is PipelineConfig, return 0 + # to identify this is global interface + return 0 + + def is_global_interface(self): + """check if this interface is global""" + return not isinstance(self.io_owner, PipelineConfig.ModuleWrapper) + + def __repr__(self): + """Get all binding(input data), exepect like |data_0: mod1:data_0""" + ret = " |{}: ".format(self.name) + for binding in self.bindings: + mname, dname = binding.get_name() + ret += "{0}:{1} ".format(mname, dname) + return ret + + def check_dag_acyclic(self, start, inputs): + """check if the DAG that current binding stay is acircle""" + for binding in inputs.values(): + if start == binding.io_owner: + return False + for p in binding.parents: + if not self.check_dag_acyclic(start, p.io_owner.input_bindings.bindings): + return False + + return True + + def connect(self, binding): + """ + # check if the bindendency setting correct. + # correct connection are following + # 1. global input to module input + # 2. module output to global output + # 3. module output to moudle input + """ + if self.io_owner == binding.io_owner: + raise RuntimeError(f"can not set self as binding.") + + if not self.is_global_interface() and self.io_type == "input": + raise RuntimeError(f"Module only can start binding from output!") + + if ( + not self.is_global_interface() + and not binding.is_global_interface() + and binding.io_type == "output" + ): + raise RuntimeError(f"Module output can not binding with module output!") + + if ( + not self.is_global_interface() + and binding.is_global_interface() + and binding.io_type == "input" + ): + raise RuntimeError(f"Module output can not binding with global input!") + + if self.is_global_interface() and self.io_type != "input": + raise RuntimeError(f"Global only can start binding from input!") + + if self.is_global_interface() and binding.io_type != "input": + raise RuntimeError(f"Global input only can set binding with module input.") + + self.bindings.append(binding) + if not self.is_global_interface(): + # check if the source and target data_type same + if ( + isinstance(binding.io_owner, PipelineConfig.ModuleWrapper) + and self.data_type != binding.data_type + ): + raise RuntimeError( + f"Illegal type (%s vs. %s): binding type is not same!" + % (self.data_type, binding.data_type) + ) + + binding.parents.append(self) + # Do acyclic check after increase the in-degree. + if not self.check_dag_acyclic( + binding.io_owner, self.io_owner.input_bindings.bindings + ): + raise RuntimeError(f"Illegal connection: cause a circle!") + + def __init__(self, mod=None): + """init class""" + self.target_host = None + self.build_func = None + self.params = None + self.target = None + self.name = None + self.dev = None + self.idx = None + self.mod = mod + self.input_params = InferType()(mod)["main"].params + self.output_values = InferType()(mod)["main"].checked_type.ret_type + self.input_bindings = PipelineConfig.BindingList(self, "input") + self.output_bindings = PipelineConfig.BindingList(self, "output") + + def __eq__(self, other): + """check if self equl other""" + if isinstance(other, PipelineConfig.ModuleWrapper): + return self.mod == other.mod + + return False + + def __getitem__(self, key): + """get item by key""" + if isinstance(key, str): + if key == "input": + return self.input_bindings + + if key == "output": + return self.output_bindings + + raise RuntimeError(f"{key} not found!") + + def get_data_type(self, key, stype): + """get module input/output data type.""" + if stype == "input": + for param in self.input_params: + if param.name_hint == key: + return param._checked_type_ + + if stype == "output": + if isinstance(self.output_values, tvm.ir.type.TupleType): + if int(key) < len(self.output_values.fields): + return self.output_values.fields[int(key)] + elif int(key) == 0: + return self.output_values + + return None + + def set_idx_name(self, idx): + """generate name by idx and storage idx value""" + self.idx = idx + self.name = "mod{}".format(str(idx)) + + def is_root_mod(self): + """use by dag topology sort, identify if this item is root item""" + for _, binding in self.input_bindings.bindings.items(): + if binding.parents: + return False + + return True + + def remove_self_from_bindings(self): + """use by DAG topology sort, by remove self from binding to reduce child in-degree""" + for _, binding in self.output_bindings.bindings.items(): + for child in binding.bindings: + if binding in child.parents: + child.parents.remove(binding) + + class BindingList: + """Use to storage Binding list. + + Parameters + ---------- + owner : ModuleWrapper/PipelineConfig + who own this list, it can be ModuleWrapper or PipelineConfig + + type_name : str + The type of this binding list. It can be either "input" or "output". + """ + + def __init__(self, owner, type_name): + self.bindings = {} + self.io_owner = owner + self.binding_type = type_name + + def get_binding_data_type(self, key): + """return binding data type""" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + return self.io_owner.get_data_type(key, self.binding_type) + return None + + def __getitem__(self, key): + """return item by key""" + if key not in self.bindings: + data_type = self.get_binding_data_type(key) + if not data_type and isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + raise RuntimeError(f"Illegal name: {self.binding_type}:{key} cannot find") Review comment: ```suggestion raise RuntimeError(f"Cannot find {key} in binding list {self.binding_type}") ``` ########## File path: python/tvm/contrib/pipeline_executor.py ########## @@ -0,0 +1,538 @@ +# 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. +"""Pipeline executor that executes a series of modules in a pipeline fashion.""" +import json +import tvm._ffi +from tvm import relay +from tvm.relay.transform import InferType +from tvm.contrib import graph_executor + + +def pipeline_executor_enabled(): + """check if pipeline executor is enabled. + + Return + ------- + enable: bool + Return pipeline executor is enabled or not. + """ + return tvm._ffi.get_global_func("tvm.pipeline_executor.create", allow_missing=True) is not None + + +def build(pipe_configs): + """build module list that can use for pipeline execution. + + Parameters + ---------- + pipe_configs: PipelineConfig + build configuration informaton. + + Returns + ------- + ret: PipeModuleConfig + the class that wrap module list and configuration. + """ + mods = {} + mod_n_configs = pipe_configs.get_config() + config_len = len(mod_n_configs) + string_config = [{} for _ in range(config_len)] + for ir_mod, mod_config in mod_n_configs.items(): + mconf = mod_config["pipeline"].copy() + mod_idx = mconf["mod_idx"] - 1 + # Get mod device config + dev = mod_config["dev"] + target = mod_config["target"] + build_func = relay.build + # if there is a self defined build function then use it. + if "build" in mod_config and mod_config["build"]: + build_func = mod_config["build"] + + # build IRModule + mod = build_func( + ir_mod, + target, + params=mod_config["params"], + target_host=mod_config["target_host"], + mod_name=mod_config["mod_name"], + ) + + mconf["dev"] = "{},{}".format(dev.device_type, dev.device_id) + # Create pipeline configuration + string_config[mod_idx] = mconf + # associate mod with device + mods[mod] = {"dev": dev} + + # return PipelineExecutorFactoryModule + return PipelineExecutorFactoryModule(mods, string_config) + + +def create(pipe_mod_config): + """Create a pipeline runtime executor. + + Parameters + ---------- + pipe_mod_config : PipeModuleConfig + class to storage IRModule list and pipeline configuration. + + Returns + submodule : PipelineModule + Runtime pipeline module. + """ + + return PipelineModule(pipe_mod_config) + + +class PipelineModule(object): + """Wrapper runtime module. This is a thin wrapper of the underlying TVM module. + + Parameters + ---------- + pipeline_mods : List[GraphModule] + The internal tvm module that holds the actual graph functions. + pipeline_config : Dict[IRModule, Dict[str, Any]] + modules and modules dependency configuration informaiton. + """ + + def __init__(self, pipe_mod_config): + self.pipeline_mods_ = pipe_mod_config.pipeline_mods_ + self.mod_config_ = pipe_mod_config.mods_config_ + mods, config = self.graph_executor_create(self.pipeline_mods_, self.mod_config_) + assert ( + pipeline_executor_enabled() + ), "Pipeline executor is not enabled. Please \ + re-build TVM with USE_PIPELINE_EXECUTOR=ON" + pipelinecreate = tvm._ffi.get_global_func( + "tvm.pipeline_executor.create", allow_missing=False + ) + assert pipelinecreate + module = pipelinecreate(mods, config) + + self.module_ = module + + def graph_executor_create(self, pipeline_mods, mod_config): + """Create graph_executor list and return string format config. + + Parameters + ---------- + pipeline_mods : List[IRModule] + list of IRModule + + mod_config : Dict[int, Dict[str, Any]] + modules and modules dependency configuration informaiton. + + Returns + ------- + mods : List[GraphModule] + Runtime graph module. + + mod_config : str + mods configuration + """ + + mods = [] + for pipeline_mod in pipeline_mods: + mod = graph_executor.GraphModule( + pipeline_mod["default"](pipeline_mods[pipeline_mod]["dev"]) + ) + mods.append(mod.module) + + return mods, json.dumps(mod_config) + + +class PipelineConfig(object): + """The wrapper of each module to be pipelined. The wrapper mainly includes the + module itself as well as the binding that represents the connections of this + module's inputs and outputs to other modules. + """ + + class ModuleWrapper: + """The class use use to represent Module and storage module idx and + Binding information. + """ + + class Binding: + """The class that use to storage module connection information. + The binding can be either "input" or "output". + + Parameters + ---------- + owner : ModuleWrapper + The class that own this interface, in such class there are + Module information like idx, module name + + io_type : str + The type of this binding. It can be either "input" or "output". + + name : str/integer + Binding name, for input it is string such as "data0"; + for output it is the idx integer such as 0. + """ + + def __init__(self, owner, stype, name, data_type=None): + self.io_owner = owner + self.io_type = stype + self.name = str(name) + # These item that have dependency relation with self + self.bindings = [] + # The item that self depend + self.parents = [] + + self.data_type = data_type + + def get_name(self): + """get owner name and self name""" + owner_name = "" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + owner_name = self.io_owner.name + + return owner_name, self.name + + def get_owner_idx(self): + """return idx if owner is ModuleWrapper, if not return 0""" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + return self.io_owner.idx + + # if not ModuleWrapper then owner is PipelineConfig, return 0 + # to identify this is global interface + return 0 + + def is_global_interface(self): + """check if this interface is global""" + return not isinstance(self.io_owner, PipelineConfig.ModuleWrapper) + + def __repr__(self): + """Get all binding(input data), exepect like |data_0: mod1:data_0""" + ret = " |{}: ".format(self.name) + for binding in self.bindings: + mname, dname = binding.get_name() + ret += "{0}:{1} ".format(mname, dname) + return ret + + def check_dag_acyclic(self, start, inputs): + """check if the DAG that current binding stay is acircle""" + for binding in inputs.values(): + if start == binding.io_owner: + return False + for p in binding.parents: + if not self.check_dag_acyclic(start, p.io_owner.input_bindings.bindings): + return False + + return True + + def connect(self, binding): + """ + # check if the bindendency setting correct. + # correct connection are following + # 1. global input to module input + # 2. module output to global output + # 3. module output to moudle input + """ + if self.io_owner == binding.io_owner: + raise RuntimeError(f"can not set self as binding.") + + if not self.is_global_interface() and self.io_type == "input": + raise RuntimeError(f"Module only can start binding from output!") + + if ( + not self.is_global_interface() + and not binding.is_global_interface() + and binding.io_type == "output" + ): + raise RuntimeError(f"Module output can not binding with module output!") + + if ( + not self.is_global_interface() + and binding.is_global_interface() + and binding.io_type == "input" + ): + raise RuntimeError(f"Module output can not binding with global input!") + + if self.is_global_interface() and self.io_type != "input": + raise RuntimeError(f"Global only can start binding from input!") + + if self.is_global_interface() and binding.io_type != "input": + raise RuntimeError(f"Global input only can set binding with module input.") + + self.bindings.append(binding) + if not self.is_global_interface(): + # check if the source and target data_type same + if ( + isinstance(binding.io_owner, PipelineConfig.ModuleWrapper) + and self.data_type != binding.data_type + ): + raise RuntimeError( + f"Illegal type (%s vs. %s): binding type is not same!" + % (self.data_type, binding.data_type) + ) + + binding.parents.append(self) + # Do acyclic check after increase the in-degree. + if not self.check_dag_acyclic( + binding.io_owner, self.io_owner.input_bindings.bindings + ): + raise RuntimeError(f"Illegal connection: cause a circle!") + + def __init__(self, mod=None): + """init class""" + self.target_host = None + self.build_func = None + self.params = None + self.target = None + self.name = None + self.dev = None + self.idx = None + self.mod = mod + self.input_params = InferType()(mod)["main"].params + self.output_values = InferType()(mod)["main"].checked_type.ret_type + self.input_bindings = PipelineConfig.BindingList(self, "input") + self.output_bindings = PipelineConfig.BindingList(self, "output") + + def __eq__(self, other): + """check if self equl other""" + if isinstance(other, PipelineConfig.ModuleWrapper): + return self.mod == other.mod + + return False + + def __getitem__(self, key): + """get item by key""" + if isinstance(key, str): + if key == "input": + return self.input_bindings + + if key == "output": + return self.output_bindings + + raise RuntimeError(f"{key} not found!") + + def get_data_type(self, key, stype): + """get module input/output data type.""" + if stype == "input": + for param in self.input_params: + if param.name_hint == key: + return param._checked_type_ + + if stype == "output": + if isinstance(self.output_values, tvm.ir.type.TupleType): + if int(key) < len(self.output_values.fields): + return self.output_values.fields[int(key)] + elif int(key) == 0: + return self.output_values + + return None + + def set_idx_name(self, idx): + """generate name by idx and storage idx value""" + self.idx = idx + self.name = "mod{}".format(str(idx)) + + def is_root_mod(self): + """use by dag topology sort, identify if this item is root item""" + for _, binding in self.input_bindings.bindings.items(): + if binding.parents: + return False + + return True + + def remove_self_from_bindings(self): + """use by DAG topology sort, by remove self from binding to reduce child in-degree""" + for _, binding in self.output_bindings.bindings.items(): Review comment: ```suggestion for binding in self.output_bindings.bindings.values(): ``` ########## File path: python/tvm/contrib/pipeline_executor.py ########## @@ -0,0 +1,538 @@ +# 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. +"""Pipeline executor that executes a series of modules in a pipeline fashion.""" +import json +import tvm._ffi +from tvm import relay +from tvm.relay.transform import InferType +from tvm.contrib import graph_executor + + +def pipeline_executor_enabled(): + """check if pipeline executor is enabled. + + Return + ------- + enable: bool + Return pipeline executor is enabled or not. + """ + return tvm._ffi.get_global_func("tvm.pipeline_executor.create", allow_missing=True) is not None + + +def build(pipe_configs): + """build module list that can use for pipeline execution. + + Parameters + ---------- + pipe_configs: PipelineConfig + build configuration informaton. + + Returns + ------- + ret: PipeModuleConfig + the class that wrap module list and configuration. + """ + mods = {} + mod_n_configs = pipe_configs.get_config() + config_len = len(mod_n_configs) + string_config = [{} for _ in range(config_len)] + for ir_mod, mod_config in mod_n_configs.items(): + mconf = mod_config["pipeline"].copy() + mod_idx = mconf["mod_idx"] - 1 + # Get mod device config + dev = mod_config["dev"] + target = mod_config["target"] + build_func = relay.build + # if there is a self defined build function then use it. + if "build" in mod_config and mod_config["build"]: + build_func = mod_config["build"] + + # build IRModule + mod = build_func( + ir_mod, + target, + params=mod_config["params"], + target_host=mod_config["target_host"], + mod_name=mod_config["mod_name"], + ) + + mconf["dev"] = "{},{}".format(dev.device_type, dev.device_id) + # Create pipeline configuration + string_config[mod_idx] = mconf + # associate mod with device + mods[mod] = {"dev": dev} + + # return PipelineExecutorFactoryModule + return PipelineExecutorFactoryModule(mods, string_config) + + +def create(pipe_mod_config): + """Create a pipeline runtime executor. + + Parameters + ---------- + pipe_mod_config : PipeModuleConfig + class to storage IRModule list and pipeline configuration. + + Returns + submodule : PipelineModule + Runtime pipeline module. + """ + + return PipelineModule(pipe_mod_config) + + +class PipelineModule(object): + """Wrapper runtime module. This is a thin wrapper of the underlying TVM module. + + Parameters + ---------- + pipeline_mods : List[GraphModule] + The internal tvm module that holds the actual graph functions. + pipeline_config : Dict[IRModule, Dict[str, Any]] + modules and modules dependency configuration informaiton. + """ + + def __init__(self, pipe_mod_config): + self.pipeline_mods_ = pipe_mod_config.pipeline_mods_ + self.mod_config_ = pipe_mod_config.mods_config_ + mods, config = self.graph_executor_create(self.pipeline_mods_, self.mod_config_) + assert ( + pipeline_executor_enabled() + ), "Pipeline executor is not enabled. Please \ + re-build TVM with USE_PIPELINE_EXECUTOR=ON" + pipelinecreate = tvm._ffi.get_global_func( + "tvm.pipeline_executor.create", allow_missing=False + ) + assert pipelinecreate + module = pipelinecreate(mods, config) + + self.module_ = module + + def graph_executor_create(self, pipeline_mods, mod_config): + """Create graph_executor list and return string format config. + + Parameters + ---------- + pipeline_mods : List[IRModule] + list of IRModule + + mod_config : Dict[int, Dict[str, Any]] + modules and modules dependency configuration informaiton. + + Returns + ------- + mods : List[GraphModule] + Runtime graph module. + + mod_config : str + mods configuration + """ + + mods = [] + for pipeline_mod in pipeline_mods: + mod = graph_executor.GraphModule( + pipeline_mod["default"](pipeline_mods[pipeline_mod]["dev"]) + ) + mods.append(mod.module) + + return mods, json.dumps(mod_config) + + +class PipelineConfig(object): + """The wrapper of each module to be pipelined. The wrapper mainly includes the + module itself as well as the binding that represents the connections of this + module's inputs and outputs to other modules. + """ + + class ModuleWrapper: + """The class use use to represent Module and storage module idx and + Binding information. + """ + + class Binding: + """The class that use to storage module connection information. + The binding can be either "input" or "output". + + Parameters + ---------- + owner : ModuleWrapper + The class that own this interface, in such class there are + Module information like idx, module name + + io_type : str + The type of this binding. It can be either "input" or "output". + + name : str/integer + Binding name, for input it is string such as "data0"; + for output it is the idx integer such as 0. + """ + + def __init__(self, owner, stype, name, data_type=None): + self.io_owner = owner + self.io_type = stype + self.name = str(name) + # These item that have dependency relation with self + self.bindings = [] + # The item that self depend + self.parents = [] + + self.data_type = data_type + + def get_name(self): + """get owner name and self name""" + owner_name = "" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + owner_name = self.io_owner.name + + return owner_name, self.name + + def get_owner_idx(self): + """return idx if owner is ModuleWrapper, if not return 0""" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + return self.io_owner.idx + + # if not ModuleWrapper then owner is PipelineConfig, return 0 + # to identify this is global interface + return 0 + + def is_global_interface(self): + """check if this interface is global""" + return not isinstance(self.io_owner, PipelineConfig.ModuleWrapper) + + def __repr__(self): + """Get all binding(input data), exepect like |data_0: mod1:data_0""" + ret = " |{}: ".format(self.name) + for binding in self.bindings: + mname, dname = binding.get_name() + ret += "{0}:{1} ".format(mname, dname) + return ret + + def check_dag_acyclic(self, start, inputs): + """check if the DAG that current binding stay is acircle""" + for binding in inputs.values(): + if start == binding.io_owner: + return False + for p in binding.parents: + if not self.check_dag_acyclic(start, p.io_owner.input_bindings.bindings): + return False + + return True + + def connect(self, binding): + """ + # check if the bindendency setting correct. + # correct connection are following + # 1. global input to module input + # 2. module output to global output + # 3. module output to moudle input + """ + if self.io_owner == binding.io_owner: + raise RuntimeError(f"can not set self as binding.") + + if not self.is_global_interface() and self.io_type == "input": + raise RuntimeError(f"Module only can start binding from output!") + + if ( + not self.is_global_interface() + and not binding.is_global_interface() + and binding.io_type == "output" + ): + raise RuntimeError(f"Module output can not binding with module output!") + + if ( + not self.is_global_interface() + and binding.is_global_interface() + and binding.io_type == "input" + ): + raise RuntimeError(f"Module output can not binding with global input!") + + if self.is_global_interface() and self.io_type != "input": + raise RuntimeError(f"Global only can start binding from input!") + + if self.is_global_interface() and binding.io_type != "input": + raise RuntimeError(f"Global input only can set binding with module input.") + + self.bindings.append(binding) + if not self.is_global_interface(): + # check if the source and target data_type same + if ( + isinstance(binding.io_owner, PipelineConfig.ModuleWrapper) + and self.data_type != binding.data_type + ): + raise RuntimeError( + f"Illegal type (%s vs. %s): binding type is not same!" + % (self.data_type, binding.data_type) + ) + + binding.parents.append(self) + # Do acyclic check after increase the in-degree. + if not self.check_dag_acyclic( + binding.io_owner, self.io_owner.input_bindings.bindings + ): + raise RuntimeError(f"Illegal connection: cause a circle!") + + def __init__(self, mod=None): + """init class""" + self.target_host = None + self.build_func = None + self.params = None + self.target = None + self.name = None + self.dev = None + self.idx = None + self.mod = mod + self.input_params = InferType()(mod)["main"].params + self.output_values = InferType()(mod)["main"].checked_type.ret_type + self.input_bindings = PipelineConfig.BindingList(self, "input") + self.output_bindings = PipelineConfig.BindingList(self, "output") + + def __eq__(self, other): + """check if self equl other""" + if isinstance(other, PipelineConfig.ModuleWrapper): + return self.mod == other.mod + + return False + + def __getitem__(self, key): + """get item by key""" + if isinstance(key, str): + if key == "input": + return self.input_bindings + + if key == "output": + return self.output_bindings + + raise RuntimeError(f"{key} not found!") + + def get_data_type(self, key, stype): + """get module input/output data type.""" + if stype == "input": + for param in self.input_params: + if param.name_hint == key: + return param._checked_type_ + + if stype == "output": + if isinstance(self.output_values, tvm.ir.type.TupleType): + if int(key) < len(self.output_values.fields): + return self.output_values.fields[int(key)] + elif int(key) == 0: + return self.output_values + + return None + + def set_idx_name(self, idx): + """generate name by idx and storage idx value""" + self.idx = idx + self.name = "mod{}".format(str(idx)) + + def is_root_mod(self): + """use by dag topology sort, identify if this item is root item""" + for _, binding in self.input_bindings.bindings.items(): + if binding.parents: + return False + + return True + + def remove_self_from_bindings(self): + """use by DAG topology sort, by remove self from binding to reduce child in-degree""" + for _, binding in self.output_bindings.bindings.items(): + for child in binding.bindings: + if binding in child.parents: + child.parents.remove(binding) + + class BindingList: + """Use to storage Binding list. + + Parameters + ---------- + owner : ModuleWrapper/PipelineConfig + who own this list, it can be ModuleWrapper or PipelineConfig + + type_name : str + The type of this binding list. It can be either "input" or "output". + """ + + def __init__(self, owner, type_name): + self.bindings = {} + self.io_owner = owner + self.binding_type = type_name + + def get_binding_data_type(self, key): + """return binding data type""" + if isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + return self.io_owner.get_data_type(key, self.binding_type) + return None + + def __getitem__(self, key): + """return item by key""" + if key not in self.bindings: + data_type = self.get_binding_data_type(key) + if not data_type and isinstance(self.io_owner, PipelineConfig.ModuleWrapper): + raise RuntimeError(f"Illegal name: {self.binding_type}:{key} cannot find") + + self.bindings[key] = PipelineConfig.ModuleWrapper.Binding( + self.io_owner, self.binding_type, key, data_type + ) + + return self.bindings[key] + + def __init__(self): + self.mod_wrapper = {} + self.input_bindings = self.BindingList(self, "input") + self.output_bindings = self.BindingList(self, "output") + + def __str__(self): + """ Get configuration in string type""" + # topology sort to get correct module order in list. + self.dag_topology_sort() + # get input + input_dump = "Inputs\n" + for input_name in self.input_bindings.bindings: + inf = self.input_bindings.bindings[input_name] + input_dump += inf.__repr__() + "\n" + + # get connections + output = {} + connections_dump = "\nconnections\n" + for mod in self.mod_wrapper: + for _, interface in self.mod_wrapper[mod].output_bindings.bindings.items(): + if interface.bindings: + mname, dname = interface.get_name() + iname = mname + ".output(" + dname + ")->" + for dep in interface.bindings: + dep_mname, dep_dname = dep.get_name() + if isinstance(dep.io_owner, PipelineConfig.ModuleWrapper): + iname += " " + dep_mname + "." + dep_dname + connections_dump += " |" + iname + "\n" + else: + output[dep_dname] = mname + ".output(" + dname + ")" + + # get output + output_dump = "\noutput\n" + for name in sorted(output.keys()): + output_dump += " |output(" + name + ") : " + output[name] + "\n" Review comment: ```suggestion output_dump += f" |output("{name}) : {output[name]}\n" ``` -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
