merrymercy commented on a change in pull request #5962: URL: https://github.com/apache/incubator-tvm/pull/5962#discussion_r453195528
########## File path: python/tvm/auto_schedule/measure.py ########## @@ -0,0 +1,480 @@ +# 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. + +""" +Distributed measurement infrastructure to measure the runtime costs of tensor programs. + +These functions are responsible for building the tvm module, uploading it to +remote devices, recording the running time costs, and checking the correctness of the output. + +We separate the measurement into two steps: build and run. +A builder builds the executable binary files and a runner runs the binary files to +get the measurement results. The flow of data structures is + + `ProgramBuilder` `ProgramRunner` +`MeasureInput` -----------------> `BuildResult` ----------------> `MeasureResult` + +We implement these in python to utilize python's multiprocessing and error handling. +""" + +import os +import time +import shutil +import traceback +import tempfile +import multiprocessing + +import tvm._ffi +from tvm.runtime import Object, module, ndarray +from tvm.driver import build_module +from tvm.ir import transform +from tvm.contrib import tar, ndk + +from . import _ffi_api +from .utils import get_const_tuple, NoDaemonPool, call_func_with_timeout + +# The maximum length of error message +MAX_ERROR_MSG_LEN = 512 + +# We use fork and a global variable to copy arguments between processings. +# This can avoid expensive serialization of TVM IR when using multiprocessing.Pool +GLOBAL_BUILD_ARGUMENTS = None + +@tvm._ffi.register_object("auto_schedule.MeasureCallback") +class MeasureCallback(Object): + """ The base class of measurement callback functions. """ + + +@tvm._ffi.register_object("auto_schedule.MeasureInput") +class MeasureInput(Object): + """ Store the input of a measurement. + + Parameters + ---------- + task : SearchTask + The SearchTask of this measure. + state : State + The State to be measured. + """ + def __init__(self, task, state): + self.__init_handle_by_constructor__(_ffi_api.MeasureInput, task, state.state_object) + + +@tvm._ffi.register_object("auto_schedule.BuildResult") +class BuildResult(Object): + """ Store the result of a build. + + Parameters + ---------- + filename : Optional[str] + The filename of built binary file. + args : List[Tensor] + The arguments. + error_no : int + The error code. + error_msg : Optional[str] + The error message if there is any error. + time_cost : float + The time cost of build. + """ + def __init__(self, filename, args, error_no, error_msg, time_cost): + filename = filename if filename else "" + error_msg = error_msg if error_msg else "" + + self.__init_handle_by_constructor__( + _ffi_api.BuildResult, filename, args, error_no, error_msg, time_cost) + + +@tvm._ffi.register_object("auto_schedule.MeasureResult") +class MeasureResult(Object): + """ Store the results of a measurement. + + Parameters + ---------- + costs : List[float] + The time costs of execution. + error_no : int + The error code. + error_msg : Optional[str] + The error message if there is any error. + all_cost : float + The time cost of build and run. + timestamp : float + The time stamps of this measurement. + """ + def __init__(self, costs, error_no, error_msg, all_cost, timestamp): + error_msg = error_msg if error_msg else "" + + self.__init_handle_by_constructor__( + _ffi_api.MeasureResult, costs, error_no, + error_msg, all_cost, timestamp) + + +@tvm._ffi.register_object("auto_schedule.ProgramBuilder") +class ProgramBuilder(Object): + """ The base class of ProgramBuilders. """ + + def build(self, measure_inputs, verbose=True): Review comment: All verbosity levels should be int because we can have multiple levels. ########## File path: python/tvm/auto_schedule/measure_record.py ########## @@ -0,0 +1,157 @@ +# 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. + +""" Serialization and other I/O support for tuning logs (measurement records). """ Review comment: ```suggestion """ Serialization and other I/O support for measurement records (tuning logs). """ ``` ########## File path: python/tvm/auto_schedule/auto_schedule.py ########## @@ -0,0 +1,195 @@ +# 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. + +""" +User interface for TVM Auto-scheduler. + +The basic schedule search process for TVM Auto-scheduler is designed to be: +`Program sampling` -> `Performance Tuning`. + +In `Program sampling`, we use some predefined precise or heuristic rules to generate several +initial schedules. Based on these initial starting points, we perform `Performance Tuning` which +uses cost model based evolutionary search to select schedules with the best performance. + +Candidate schedules are measured against the specific hardware target. +""" + +import tvm._ffi +from tvm.runtime import Object +from .measure import LocalBuilder, LocalRunner +from . import _ffi_api + + +@tvm._ffi.register_object("auto_schedule.HardwareParams") +class HardwareParams(Object): + """ The parameters of target hardware used to guide the search process of SearchPolicy. + + TODO(jcf94): This is considered to be merged with the new Target: + https://discuss.tvm.ai/t/rfc-tvm-target-specification/6844 + + Parameters + ---------- + num_cores : int + The number of device cores. + vector_unit_bytes : int + The width of vector units in bytes. + cache_line_bytes : int + The size of cache line in bytes. + """ + def __init__(self, num_cores, vector_unit_bytes, cache_line_bytes): + self.__init_handle_by_constructor__(_ffi_api.HardwareParams, num_cores, + vector_unit_bytes, cache_line_bytes) + + +@tvm._ffi.register_object("auto_schedule.SearchTask") +class SearchTask(Object): + """ The computation information and hardware parameters for a specific schedule search task. + + Parameters + ---------- + dag : ComputeDAG + The ComputeDAG for the corresponding compute declaration. + workload_key : str + The workload key for the corresponding compute declaration. + target : tvm.target.Target + The target device of this search task. + target_host : Optional[tvm.target.Target] + The target host device of this search task. + hardware_params : Optional[HardwareParams] + Hardware parameters used in this search task. + """ + def __init__(self, dag, workload_key, target, target_host=None, + hardware_params=None): + self.__init_handle_by_constructor__(_ffi_api.SearchTask, dag, + workload_key, target, target_host, + hardware_params) + + +@tvm._ffi.register_object("auto_schedule.SearchPolicy") +class SearchPolicy(Object): + """ The base class of search policies. """ + + +@tvm._ffi.register_object("auto_schedule.EmptyPolicy") +class EmptyPolicy(SearchPolicy): + """ This is an example empty search policy which will always generate + the init state of ComputeDAG. + """ + def __init__(self): + self.__init_handle_by_constructor__(_ffi_api.EmptyPolicy) + + +@tvm._ffi.register_object("auto_schedule.TuningOptions") +class TuningOptions(Object): + """ This controls the options of performance tuning. + + Parameters + ---------- + num_measure_trials: int = 0 + The number of measurement trials. + The search policy measures `num_measure_trials` schedules in total and returns the best one + among them. + With `num_measure_trials` == 0, the policy will do the schedule search but won't involve + measurement. + This can be used to get a runnable schedule quickly without auto-tuning. Review comment: ```suggestion With `num_measure_trials` == 0, the policy will do the schedule search but won't involve measurement. This can be used to get a runnable schedule quickly without auto-tuning. ``` ########## File path: python/tvm/auto_schedule/auto_schedule.py ########## @@ -0,0 +1,195 @@ +# 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. + +""" +User interface for TVM Auto-scheduler. + +The basic schedule search process for TVM Auto-scheduler is designed to be: +`Program sampling` -> `Performance Tuning`. + +In `Program sampling`, we use some predefined precise or heuristic rules to generate several +initial schedules. Based on these initial starting points, we perform `Performance Tuning` which +uses cost model based evolutionary search to select schedules with the best performance. + +Candidate schedules are measured against the specific hardware target. +""" + +import tvm._ffi +from tvm.runtime import Object +from .measure import LocalBuilder, LocalRunner +from . import _ffi_api + + +@tvm._ffi.register_object("auto_schedule.HardwareParams") +class HardwareParams(Object): + """ The parameters of target hardware used to guide the search process of SearchPolicy. + + TODO(jcf94): This is considered to be merged with the new Target: + https://discuss.tvm.ai/t/rfc-tvm-target-specification/6844 + + Parameters + ---------- + num_cores : int + The number of device cores. + vector_unit_bytes : int + The width of vector units in bytes. + cache_line_bytes : int + The size of cache line in bytes. + """ + def __init__(self, num_cores, vector_unit_bytes, cache_line_bytes): + self.__init_handle_by_constructor__(_ffi_api.HardwareParams, num_cores, + vector_unit_bytes, cache_line_bytes) + + +@tvm._ffi.register_object("auto_schedule.SearchTask") +class SearchTask(Object): + """ The computation information and hardware parameters for a specific schedule search task. + + Parameters + ---------- + dag : ComputeDAG + The ComputeDAG for the corresponding compute declaration. + workload_key : str + The workload key for the corresponding compute declaration. + target : tvm.target.Target + The target device of this search task. + target_host : Optional[tvm.target.Target] + The target host device of this search task. + hardware_params : Optional[HardwareParams] + Hardware parameters used in this search task. + """ + def __init__(self, dag, workload_key, target, target_host=None, + hardware_params=None): + self.__init_handle_by_constructor__(_ffi_api.SearchTask, dag, + workload_key, target, target_host, + hardware_params) + + +@tvm._ffi.register_object("auto_schedule.SearchPolicy") +class SearchPolicy(Object): + """ The base class of search policies. """ + + +@tvm._ffi.register_object("auto_schedule.EmptyPolicy") +class EmptyPolicy(SearchPolicy): + """ This is an example empty search policy which will always generate + the init state of ComputeDAG. + """ + def __init__(self): + self.__init_handle_by_constructor__(_ffi_api.EmptyPolicy) + + +@tvm._ffi.register_object("auto_schedule.TuningOptions") +class TuningOptions(Object): + """ This controls the options of performance tuning. + + Parameters + ---------- + num_measure_trials: int = 0 + The number of measurement trials. + The search policy measures `num_measure_trials` schedules in total and returns the best one + among them. + With `num_measure_trials` == 0, the policy will do the schedule search but won't involve + measurement. + This can be used to get a runnable schedule quickly without auto-tuning. + early_stopping: Optional[int] + Stop the tuning early if getting no improvement after n measurements. + num_measures_per_round: int = 64 + The number of schedules to be measured at each search round. + The whole schedule search process will try a total number of `num_measure_trials` in several + rounds. + verbose: boolean = True + Verbosity level. False for silent, True to output information during schedule search. Review comment: This should be int because we can have many levels. The previous comment from @MarisaKirisame is wrong. ########## File path: src/auto_schedule/measure.h ########## @@ -0,0 +1,438 @@ +/* + * 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. + */ + +/*! + * \file auto_schedule/measure.h + * \brief Distributed measurement infrastructure to measure the runtime costs of tensor programs. + * The flow of data structures is MeasureInput -> BuildeResult -> MeasureResult. Review comment: Copy the comments from the python file to replace these lines ########## File path: src/auto_schedule/auto_schedule.h ########## @@ -0,0 +1,116 @@ +/* + * 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. + */ + +/*! + * \file auto_schedule/auto_schedule.h + * \brief The user interface of the TVM Auto-scheduler. This is the entry structure to get + * schedule search requirements from upper level (Python API), and returns a high performance + * schedule after search process. + */ + +#ifndef TVM_AUTO_SCHEDULE_AUTO_SCHEDULE_H_ +#define TVM_AUTO_SCHEDULE_AUTO_SCHEDULE_H_ + +#include <utility> + +#include "measure.h" +#include "search_policy/search_policy.h" + +namespace tvm { +namespace auto_schedule { + +/*! \brief Tuning and measurement options. */ +class TuningOptionsNode : public Object { + public: + /*! \brief Number of total measurement trials. */ + int num_measure_trials; + /*! \brief Stops early the tuning if no improvement after n measurements. */ + int early_stopping; + /*! \brief The number of programs to be measured at each search round. */ + int num_measures_per_round; + /*! + * \brief Verbosity level. + * False for silent, true to output information during schedule searching. + */ + bool verbose; Review comment: should be int ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected]
