[Ansor][AutoTVM v2.0] Phase 0: Ansor minimum system for auto schedule generating

CMakeLists.txt

@@ -185,6 +185,7 @@ assign_source_group("Include" ${GROUP_INCLUDE})
 
 # Source file lists
 file(GLOB_RECURSE COMPILER_SRCS
+    src/ansor/*.cc
     src/node/*.cc
     src/ir/*.cc
     src/arith/*.cc

python/tvm/ansor/__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 Ansor auto-scheduler"""
+
+from . import compute_dag
+from . import measure
+from . import serialization
+from . import loop_state
+from . import utils
+from . import workload_registry
+
+# Shortcut
+from .compute_dag import ComputeDAG
+from .auto_schedule import SearchTask, TuneOption, HardwareParams, \
+    auto_schedule, EmptyPolicy
+from .measure import MeasureInput, LocalBuilder, LocalRunner
+from .serialization import LogToFile, LogReader, best_measure_pair_in_file, \
+    load_from_file, append_measure_records_to_file
+from .workload_registry import register_workload_by_func, make_workload_key_by_func

python/tvm/ansor/_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.ansor"""
+import tvm._ffi
+
+
+tvm._ffi._init_api("ansor", __name__)

python/tvm/ansor/auto_schedule.py

@@ -0,0 +1,206 @@
+# 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 Ansor auto-scheduler.
+
+The basic schedule search process for Ansor is designed to be:
+`Program sampling` -> `Performance Tuning`.
+
+In `Program sampling`, we use some predefined or heuristic rules to generate several initial
+schedules. Based on these initial start points, we have `Performance Tuning` to apply cost model
+and evolutionary search to seek for schedules with the best performance. Candidate schedules will
+be measured in the target hardware.
+"""
+
+import tvm._ffi
+from tvm.runtime import Object
+from .compute_dag import ComputeDAG
+from .measure import LocalBuilder, LocalRunner
+from . import _ffi_api
+
+
+@tvm._ffi.register_object("ansor.HardwareParams")
+class HardwareParams(Object):
+    """ The parameters of target hardware, this is used to guide the search process of
+    SearchPolicy.
+
+    TODO(...): This is considering to merge 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.
+    max_unroll_vec : int
+        The max length of an axis to be unrolled or vectorized.
+    max_innermost_split_factor : int
+        The max split factor for the innermost tile.
+    """
+    def __init__(self, num_cores, vector_unit_bytes, cache_line_bytes,
+                 max_unroll_vec, max_innermost_split_factor):
+        self.__init_handle_by_constructor__(_ffi_api.HardwareParams, num_cores,
+                                            vector_unit_bytes, cache_line_bytes,
+                                            max_unroll_vec, max_innermost_split_factor)
+
+
+@tvm._ffi.register_object("ansor.SearchTask")
+class SearchTask(Object):
+    """ The meta-information of a search task.
+
+    Parameters
+    ----------
+    dag : ComputeDAG
+        The ComputeDAG for target compute declaration.
+    workload_key : str
+        The workload key for target 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("ansor.SearchPolicy")
+class SearchPolicy(Object):
+    """ The base class for search policy  """
+
+
+@tvm._ffi.register_object("ansor.EmptyPolicy")
+class EmptyPolicy(SearchPolicy):
+    """ This is an example empty search policy which will always generate
+    the init state of target ComputeDAG.
+    """
+    def __init__(self):
+        self.__init_handle_by_constructor__(_ffi_api.EmptyPolicy)
+
+
+@tvm._ffi.register_object("ansor.TuneOption")
+class TuneOption(Object):
+    """ This controls the options of performance tuning.
+
+    Parameters
+    ----------
+    num_measure_trials: int = 0
+      The number of total schedule measure trials.
+      Ansor takes `num_measure_trials` state for measuring in total, and finally gets the best
+      schedule among them.
+      With `num_measure_trials` == 0, Ansor will do the schedule search but don't involve
+      measurement, this can be used if we want to quickly get a runnable schedule without
+      performance tuning.
+    early_stopping: int = -1
+      Stops early the tuning if no improvement get after n measurements.
+    num_measures_per_round: int = 64
+      The number of programs to be measured at each search round.
+      The whole schedule search process is designed to have several rounds to try a total
+      `num_measure_trials` schedules.
+      We have: `num_search_rounds` = `num_measure_trials` // `num_measures_per_round`
+    verbose: int = 1
+      Verbosity level. 0 for silent, 1 to output information during schedule search.
+    builder: Union[Builder, str] = 'local'
+      Builder which builds the program.
+    runner: Union[Runner, str] = 'local'
+      Runner which runs the program and measures time costs.
+    measure_callbacks: Optional[List[MeasureCallback]]
+      Callback functions called after each measure.
+      Candidates:
+        - ansor.LogToFile
+    pre_search_callbacks: Optional[List[SearchCallback]]
+      Callback functions called before the search process.
+      Candidates:
+        - ansor.PreloadMeasuredStates
+        - ansor.PreloadCustomSketchRule
+        TODO(jcf94): Add these implementation in later PRs.
+    """
+    def __init__(self, num_measure_trials=0, early_stopping=-1, 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)
+
+        if isinstance(runner, str):
+            if runner == 'local':
+                runner = LocalRunner()
+            else:
+                raise ValueError("Invalid runner: " + runner)
+
+        measure_callbacks = [] if measure_callbacks is None else measure_callbacks
+        pre_search_callbacks = [] if pre_search_callbacks is None else pre_search_callbacks
+
+        self.__init_handle_by_constructor__(
+            _ffi_api.TuneOption, num_measure_trials, early_stopping, num_measures_per_round,
+            verbose, builder, runner, measure_callbacks, pre_search_callbacks)
+
+
+def auto_schedule(task, target, target_host=None, search_policy='default',
+                  hardware_params=None, tune_option=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 : Union[SearchTask, str]
+        The target search task or workload key.
+    target : tvm.target.Target
+        The target device of this schedule search.
+    target_host : Optional[tvm.target.Target]
+        The target host device of this schedule search.
+    search_policy : Union[SearchPolicy, str] = 'default'
+        The search policy to be used for schedule search.
+    hardware_params : Optional[HardwareParams]
+        The hardware parameters of this schedule search.
+    tune_option : Optional[TuneOption]
+        Tuning and measurement options.
+
+    Returns
+    -------
+        A `te.schedule` and the target `te.Tensor`s to be used in `tvm.lower` or `tvm.build`
+    """
+    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)
+
+    tune_option = tune_option if tune_option else TuneOption()
+
+    if isinstance(task, str):
+        dag = ComputeDAG(task)
+        task = SearchTask(dag, task, target, target_host, hardware_params)
+    elif not isinstance(task, SearchTask):
+        raise ValueError("Invalid task: " + task + ". Expect a string or SearchTask")
+
+    sch, tensors = _ffi_api.AutoSchedule(task, search_policy, tune_option)
+    return sch, tensors