[Ansor][AutoTVM v2.0] Part 0: Ansor minimum system for auto schedule …

…generating (apache#5962)

* Code migration Start (#1)

* Init commit: Code migration Start

* Add loop_state.cc/h

* Add ComputeDAG basic test

* Split transform_step out & Update more UTs (#3)

* Split transform_step out

* Update GetProducers & GetConsumers

* Update UTs

* Add UT for CacheReadWrite & Some bug fix

* Add search_task, measure and serialization (#4)

* Add FollowSplit & FollowFusedSplit tests

* Update dag.InferBound & its UT

* Add search_task, measure and serialization

* Update Serialization UT

* Add MetaTileRewritePolicy (#5)

* Add feature

* Add cost_model, meta_tile_rewrite_policy

* Add MetaTileRewritePolicy basic UT

* Basic Python API for State (#6)

* Add Basic Python API for State

* Add UTs for State

* Add Python API: Measure & Task (#7)

* Update the return value of state operation

* Add task

* Copy measure.py & utils.py

* Fix LocalBuilder

* Fix LocalRunner

* Add ansor.auto_schedule() API; First AutoSchedule working version(#8)

* Add basic Python support for ansor.auto_schedule

* Update AutoSchedule API

* Bug fix for get the attach point of a fused iter

* Update UT after infer bug fix

* Bug fix & Add python serialization API (#10)

* Delete C++ UT hack since Python is ready

* Add ndarray.non_empty

* Update Serialization python API

* Improve code style, python wrapper and test cases (#11)

* Update c++ code style and unit test

* Update python State wrapper and test cases

* fix unit tests

* Add RPCRunner & OpenCL/CUDA test (#12)

* Add RPCRunner & OpenCL search test

* Add CUDA search test

* Add RPCRunner test

* rebase to upstream/master

* Add Ansor basic tutorial (#13)

* Add basic tutorial

* migrate feature extraction (#14)

* Add XGBModel & RPCRunnerWarpper (#15)

* Add XGBModel & RPCRunnerWarpper

* Revert "Add Parallel Granularity Mutation"

* Migrate workload_registry.py (#16)

* add workload registry

* update

* update

* add task scheduler (#17)

* Add conv2d cuda tutorial with workload registry (#18)

* add tune_test.py (the old tune_wkl.py) (#19)

* add tune_test.py (the old tune_wkl.py)

* update

* fix measure

* fix for gpu

* Code refine for tune_test.py & Add a pre load callback (#20)

* Bug fix for tutorials

* Add PreLoadMeasuredStates

* Add search_callback support for task tuner

* Code refine for tune_test.py

* Update

* Update

* Update

* Update

* Bug fix

* Add python custom sketch rule (#21)

* Add custom sketch rule

* Bug fix

* Ansor Relay Integration (without layout rewrite) (#22)

* relay integration

* Add tune_op_subgraph.py & Some code clean for tune_network.py (#23)

* Add single op tune scripts

* Add tune subgraph support

* Merge all op & all subgraph to one file

* Rename file

* add explicit_unroll_max_extent (#25)

* Add Index simplification & API update (#26)

* Add vectorized cooperative_fetching test

* Update math simplify for vectorized CF

* File rename

* Update tune_network

* API update

* Update PreLoadMeasuredStates & Some bug fix (#27)

* Add a threading wrapper to fix the test bug

* Set default TVM_USE_AUTO_SCHEDULER to false

* Update PreLoadMeasuredStates callback

* Add tensorize step for loop_state (#31)

* Add tensorize step

* State python api update (#33)

* Start to update api

* Add compute_dag to state

* API update

* kernel layout rewrite (#28)

* kernel layout rewrite

* remove some hacks

* add defuse_ops pass and move kernel_layout_rewrite pass after fuse_ops pass

* set TVM_RELAY_DISABLE_BUILD_CACHE for task extraction and prepare_layout_rewrite

* [cache flush] port cache flush to ansor (#32)

* Improve relay integration (#34)

* tmp checkpoint

* Improve relay integration

* Improve relay integration

* Fix xgb error & Simplify dispatcher (#35)

* Rename "MetaTileRewritePolicy" to "SketchPolicy". (#36)

* Rename "MetaTileRewritePolicy" to "SketchPolicy".

* Add a new class for auto_unroll_max_step, storage_offset in StageNode

* fix tune_op_subgraph.py

* rebase

* Migrate all node::make to noderef's construct function (#37)

* Start to move xxxnode::make to noderef()

* Update

* Update

* Finish transform_step

* Finish comute dag & auto schedule

* Update

* Update

* Update

* Update

* Update

* Code refine

* Code refine

* Code refine

* Update

* Update

* Some lint fix & Recover the double constructor of tvm::PrimExpr (#39)

* lint fix

* clang-format-fix

* pylint fix

* Update

* Recover the double constructor of tvm::PrimExpr

* Fix pylint

* pylint fix

* pylint fix

* Add MutateComputeLocation and MutateParallel in evolutionary search (#40)

* Add MutateComputeLocation and MutateParallel in evolutionary search

* fix lint

* Improve loop state python API (stage_tensors -> stage_ops) (#41)

* improve loop state python API (stage_tensors -> stage_ops)

* fix

* ComputeDAG bug fix & Add Custom TensorCore Matmul Example (#42)

* Bug Fix

* Sample example of Custom TensorCore Matmul

* Rever Commits, Start to build minimum Ansor system

* Code clean for minimum Ansor system

* Bug fix & Delete AccessAnalyzer

* Delete attachmap & Code clean

* Doc update

Update statenode::stages from vector to Array

* Headfile update & Python doc update

* clang-format fix

* pylint fix

* Update

* Doc update

* Update

* Bug fix after code merge to the new master

* clang-format fix

* Update

* Update

* Update std::vector to Array; Update verbosity setting; Some commemts
addressed

* std::vector->Array & std::string->String

* Add init_state to ComputeDAG

* Update

* Update some unordered_map to Map

* clang-format fix

* Comments addressed
Delete ReplayAndInferBound
Delete ReplaySteps & InferBoundCommon

* Lint fix

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Rename ansor namespace to auto_schedule

* Update

* Rename ThreadPool to ParallelFor

* Add parallel_for

* Remove ThreadPool

* Update python/tvm/auto_schedule/auto_schedule.py

* trigger CI

Co-authored-by: Lianmin Zheng <[email protected]>
Co-authored-by: Minmin Sun (孙敏敏) <[email protected]>
Co-authored-by: Zhao Wu <[email protected]>

CMakeLists.txt

@@ -200,6 +200,7 @@ assign_source_group("Include" ${GROUP_INCLUDE})
 
 # Source file lists
 file(GLOB_RECURSE COMPILER_SRCS
+    src/auto_schedule/*.cc
     src/node/*.cc
     src/ir/*.cc
     src/arith/*.cc

python/tvm/auto_schedule/__init__.py

@@ -0,0 +1,34 @@
+# 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.
+# pylint: disable=unused-import, redefined-builtin
+""" Namespace for TVM Auto-scheduler. """
+
+from . import compute_dag
+from . import measure
+from . import measure_record
+from . import loop_state
+from . import utils
+from . import workload_registry
+
+# Shortcut
+from .compute_dag import ComputeDAG
+from .auto_schedule import SearchTask, TuningOptions, HardwareParams, \
+    auto_schedule, EmptyPolicy
+from .measure import MeasureInput, LocalBuilder, LocalRunner
+from .measure_record import RecordToFile, RecordReader, load_best, \
+    load_records, save_records
+from .workload_registry import register_workload, make_workload_key

python/tvm/auto_schedule/_ffi_api.py

@@ -0,0 +1,22 @@
+# 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.
+
+""" Register FFI APIs from C++ for the namespace tvm.auto_schedule. """
+import tvm._ffi
+
+
+tvm._ffi._init_api("auto_schedule", __name__)

python/tvm/auto_schedule/auto_schedule.py

@@ -0,0 +1,194 @@
+# 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 policy
+
+    TODO(jcf94): This is considered to be merged with the new Target specification:
+    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: int = 1
+      Verbosity level. 0 for silent, 1 to output information during schedule search.
+    builder: Union[ProgramBuilder, str] = 'local'
+      ProgramBuilder which builds the program.
+    runner: Union[ProgramRunner, str] = 'local'
+      ProgramRunner which runs the program and measures time costs.
+    measure_callbacks: Optional[List[MeasureCallback]]
+      Callback functions called after each measurement.
+      Candidates:
+        - auto_schedule.RecordToFile
+    pre_search_callbacks: Optional[List[SearchCallback]]
+      Callback functions called before the search process.
+      Candidates:
+        - auto_schedule.PreloadMeasuredStates
+        - auto_schedule.PreloadCustomSketchRule
+        TODO(jcf94): Add these implementation in later PRs.
+    """
+    def __init__(self, num_measure_trials=0, early_stopping=None, num_measures_per_round=64,
+                 verbose=1, builder='local', runner='local', measure_callbacks=None,
+                 pre_search_callbacks=None):
+        if isinstance(builder, str):
+            if builder == 'local':
+                builder = LocalBuilder()
+            else:
+                raise ValueError("Invalid builder: " + builder)
+        elif not isinstance(builder, tvm.auto_schedule.measure.ProgramBuilder):
+            raise ValueError("Invalid builder: " + builder +
+                             " . TuningOptions expects a ProgramBuilder or string.")
+
+        if isinstance(runner, str):
+            if runner == 'local':
+                runner = LocalRunner()
+            else:
+                raise ValueError("Invalid runner: " + runner)
+        elif not isinstance(runner, tvm.auto_schedule.measure.ProgramRunner):
+            raise ValueError("Invalid runner: " + runner +
+                             " . TuningOptions expects a ProgramRunner or string.")
+
+        self.__init_handle_by_constructor__(
+            _ffi_api.TuningOptions, num_measure_trials, early_stopping if early_stopping else -1,
+            num_measures_per_round, verbose, builder, runner, measure_callbacks,
+            pre_search_callbacks)
+
+
+def auto_schedule(task, search_policy='default', tuning_options=None):
+    """ Do auto scheduling for a computation declaration.
+
+    The task parameter can be a `string` as workload_key, or directly
+    passing a `SearchTask` as input.
+
+    Parameters
+    ----------
+    task : SearchTask
+        The SearchTask for the computation declaration.
+    search_policy : Union[SearchPolicy, str] = 'default'
+        The search policy to be used for schedule search.
+    tuning_options : Optional[TuningOptions]
+        Tuning and measurement options.
+
+    Returns
+    -------
+        A `te.schedule` and the a list of `te.Tensor` to be used in `tvm.lower` or `tvm.build`.
+    """
+    if not isinstance(task, SearchTask):
+        raise ValueError("Invalid task: " + task +
+                         " . `auto_schedule.auto_schedule` expects a SearchTask.")
+
+    if isinstance(search_policy, str):
+        if search_policy == 'default':
+            # TODO(jcf94): This is an example policy for minimum system, will be upgrated to
+            # formal search policy later.
+            search_policy = EmptyPolicy()
+        else:
+            raise ValueError("Invalid search policy: " + search_policy)
+    elif not isinstance(search_policy, SearchPolicy):
+        raise ValueError("Invalid search policy: " + search_policy +
+                         " . `auto_schedule.auto_schedule` expects a SearchPolicy or a string.")
+
+    sch, tensors = _ffi_api.AutoSchedule(task, search_policy,
+                                         tuning_options if tuning_options else TuningOptions())
+    return sch, tensors