Revert commits

docs/conf.py

@@ -198,7 +198,6 @@
      '../tutorials/language',
      '../tutorials/optimize',
      '../tutorials/autotvm',
-     '../tutorials/ansor',
      '../tutorials/dev',
      '../tutorials/topi',
      '../tutorials/deployment',

include/tvm/relay/attrs/transform.h

@@ -296,19 +296,6 @@ struct LayoutTransformAttrs : public tvm::AttrsNode<LayoutTransformAttrs> {
   }
 };
 
-/*! \brief Attributes for KernelLayoutTransform operator */
-struct KernelLayoutTransformAttrs : public tvm::AttrsNode<KernelLayoutTransformAttrs> {
-  std::string src_layout;
-  std::string dst_layout;
-
-  TVM_DECLARE_ATTRS(KernelLayoutTransformAttrs, "relay.attrs.KernelLayoutTransformAttrs") {
-    TVM_ATTR_FIELD(src_layout)
-        .describe("The source layout of the tensor. (e.g. 1N32C112H112W)");
-    TVM_ATTR_FIELD(dst_layout)
-        .describe("The destination layout of the tensor. (e.g. 1N2C112H112W16c)");
-  }
-};
-
 /*! \brief Attributes for ShapeOf operator */
 struct ShapeOfAttrs : public tvm::AttrsNode<ShapeOfAttrs> {
   DataType dtype;

include/tvm/relay/transform.h

@@ -277,20 +277,6 @@ TVM_DLL Pass CanonicalizeOps();
  */
 TVM_DLL Pass AlterOpLayout();
 
-/*!
- * \brief Alternate the layouts of kernels.
- *
- * \return The pass.
- */
-TVM_DLL Pass KernelLayoutTransform();
-
-/*!
- * \brief The reverse of FuseOps.
- *
- * \return The pass.
- */
-TVM_DLL Pass DeFuseOps();
-
 /*!
  * \brief Given a dest layout, this pass transforms the expr such that most of the ops input data
  * layout is changed to the dest layout. In ideal situation, there are only 2 layout transforms, one

python/tvm/ansor/__init__.py

@@ -21,26 +21,14 @@
 from . import measure
 from . import serialization
 from . import loop_state
-from . import auto_schedule
 from . import utils
 from . import feature
-from . import workload_registry
-from . import task_scheduler
 
 # Shortcut
 from .compute_dag import ComputeDAG, LayoutRewriteLevel
-from .auto_schedule import SearchTask, SketchSearchPolicy, TuneOption, HardwareParams, \
-    PreloadMeasuredStates, PreloadCustomSketchRule, auto_schedule
+from .auto_schedule import SearchTask, HardwareParams
 from .measure import MeasureInput, LocalBuilder, LocalRunner, RPCRunner, LocalRPCMeasureContext
-from .cost_model import RandomModel
-from .cost_model.xgb_model import XGBModel
 from .serialization import LogToFile, LogReader, best_measure_pair_in_file, \
     load_from_file, write_measure_records_to_file
 from .workload_registry import register_workload_func, \
     workload_key_to_dag, make_workload_key_func
-from .task_scheduler import TaskScheduler, SimpleTaskScheduler
-from .dispatcher import DispatchContext, ApplyConfig, ApplyHistoryBest as apply_history_best, \
-    FallbackContext
-from .relay_integration import extract_from_program, extract_from_multiple_program, \
-    finish_layout_rewrite, prepare_layout_rewrite, auto_schedule_topi
-from .env import GLOBAL_SCOPE

python/tvm/ansor/auto_schedule.py

@@ -22,7 +22,6 @@
 import tvm._ffi
 from tvm.runtime import Object
 from .measure import LocalBuilder, LocalRunner
-from .cost_model import RandomModel, XGBModel
 from . import _ffi_api
 
 
@@ -64,206 +63,3 @@ def __init__(self, dag, workload_key, target, target_host=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  """
-    def continue_search(self, task, num_measure, verbose, measurer):
-        return _ffi_api.SearchPolicyContinueSearchOneRound(self, task,
-                                                           num_measure, verbose, measurer)
-
-    def set_task(self, task):
-        _ffi_api.SearchPolicySetTask(self, task)
-
-    def set_verbose(self, verbose):
-        _ffi_api.SearchPolicySetVerbose(self, verbose)
-
-    def run_callbacks(self, callbacks):
-        _ffi_api.SearchPolicyRunCallbacks(self, callbacks)
-
-
-@tvm._ffi.register_object("ansor.SketchSearchPolicy")
-class SketchSearchPolicy(SearchPolicy):
-    """  The search policy that searches in a hierarchical search space defined by sketches.
-    The policy randomly samples programs from the space defined by sketches
-    and use evolutionary search to fine-tune them.
-
-    Parameters
-    ----------
-    program_cost_model: CostModel
-        Cost model for programs
-    params: int
-        Parameters of the search policy. See `src/ansor/search_policy/sketch_search_policy.h`
-        to find the definitions. See code below to find the default values
-    seed: int
-        Random seed
-    """
-    def __init__(self,
-                 program_cost_model,
-                 params=None,
-                 seed=None):
-        # set default parameters
-        default_params = {
-            "eps_greedy": 0.05,
-
-            'evolutionary_search_population': 2048,
-            'evolutionary_search_num_iters': 15,
-            "evolutionary_search_mutation_prob": 0.85,
-            "evolutionary_search_use_measured_ratio": 0.2,
-
-            'cpu_multi_level_tiling_structure': 'SSRSRS',
-            'gpu_multi_level_tiling_structure': 'SSSRRSRS',
-
-            'disable_change_compute_location': 0,
-        }
-
-        if params is None:
-            params = default_params
-        else:
-            for key, value in default_params.items():
-                if key not in params:
-                    params[key] = value
-
-        self.__init_handle_by_constructor__(
-            _ffi_api.SketchSearchPolicy, program_cost_model, params,
-            seed or random.randint(1, 1 << 30))
-
-
-@tvm._ffi.register_object("ansor.SearchCallback")
-class SearchCallback(Object):
-    """Callback function before or after search process"""
-    pass
-
-
-@tvm._ffi.register_object("ansor.PreloadMeasuredStates")
-class PreloadMeasuredStates(SearchCallback):
-    """ A SearchCallback to load measured states from the log file for a search policy.
-    This can resume the state of the search policy.
-
-    Parameters
-    ----------
-    filename: str
-    """
-    def __init__(self, filename: str):
-        self.__init_handle_by_constructor__(
-            _ffi_api.PreloadMeasuredStates, filename)
-
-
-@tvm._ffi.register_object("ansor.PreloadCustomSketchRule")
-class PreloadCustomSketchRule(SearchCallback):
-    """
-    A SearchCallback for SketchSearchPolicy that allowing users to add
-    custom sketch rule.
-
-    Notes
-    -----
-    This is an advanced feature. Make sure you're clear how it
-    works and this should only be used in SketchSearchPolicy.
-
-    Parameters
-    ----------
-    meet_condition_func: Function
-        A function with `(policy, state, stage_id) -> int`
-    apply_func: Function
-        A function with `(policy, state, stage_id) -> [[State, int], ...]`
-    """
-    def __init__(self, meet_condition_func, apply_func):
-        self.__init_handle_by_constructor__(
-            _ffi_api.PreloadCustomSketchRule, meet_condition_func, apply_func)
-
-
-@tvm._ffi.register_object("ansor.TuneOption")
-class TuneOption(Object):
-    """ The options for tuning
-
-    Parameters
-    ----------
-    n_trials: int
-      Number of total measurement trials
-    early_stopping: int
-      Stops early the tuning if no improvement after n measurements
-    num_measure_per_iter: int
-      The number of programs to be measured at each iteration
-    verbose: int
-      Verbosity level. 0 means silent.
-    builder: Builder
-      Builder which builds the program
-    runner: Runner
-      Runner which runs the program and measure time costs
-    measure_callbacks: List[MeasureCallback]
-      Callback functions called after each measure
-      Candidates:
-        - ansor.LogToFile
-    pre_search_callbacks: List[SearchCallback]
-      Callback functions called before the search process
-      Candidates:
-        - ansor.PreloadMeasuredStates
-        - ansor.PreloadCustomSketchRule
-    """
-    def __init__(self, n_trials=0, early_stopping=-1, num_measure_per_iter=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 builder: " + runner)
-
-        if measure_callbacks is None:
-            measure_callbacks = []
-
-        if pre_search_callbacks is None:
-            pre_search_callbacks = []
-
-        self.__init_handle_by_constructor__(
-            _ffi_api.TuneOption, n_trials, early_stopping, num_measure_per_iter,
-            verbose, builder, runner, measure_callbacks, pre_search_callbacks)
-
-
-def auto_schedule(workload, target=None,
-                  target_host=None, search_policy='default',
-                  hardware_params=None, tune_option=None):
-    """ Do auto scheduling for a computation declaration.
-
-    The workload parameter can be a `string` as workload_key, or directly
-    passing a `SearchTask` as input.
-
-    Parameters
-    ----------
-    workload : Union[SearchTask, str]
-    target : Target
-    target_host : Target = None
-    search_policy : Union[SearchPolicy, str]
-    hardware_params : HardwareParams
-    tune_option : TuneOption
-
-    Returns
-    -------
-    sch : tvm.Schedule
-    tensors : List[Tensor]
-    """
-    if isinstance(search_policy, str):
-        if search_policy == 'default':
-            search_policy = SketchSearchPolicy(RandomModel())
-        else:
-            raise ValueError("Invalid search policy: " + search_policy)
-
-    if tune_option is None:
-        tune_option = TuneOption(n_trials=0)
-
-    if isinstance(workload, str):
-        sch, tensors = _ffi_api.AutoScheduleByWorkloadKey(
-            workload, target, target_host, search_policy, hardware_params, tune_option)
-        return sch, tensors
-    elif isinstance(workload, SearchTask):
-        sch, tensors = _ffi_api.AutoScheduleBySearchTask(workload, search_policy, tune_option)
-        return sch, tensors
-    else:
-        raise ValueError("Invalid workload: " + workload + ". Expect a string or SearchTask")