schedule_rule.h

Go to the documentation of this file.

/*
 * 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.
 */
 
#ifndef TVM_S_TIR_META_SCHEDULE_SCHEDULE_RULE_H_
#define TVM_S_TIR_META_SCHEDULE_SCHEDULE_RULE_H_
 
#include <tvm/ffi/container/array.h>
#include <tvm/ffi/container/map.h>
#include <tvm/ffi/function.h>
#include <tvm/ffi/optional.h>
#include <tvm/ffi/reflection/registry.h>
#include <tvm/ffi/string.h>
#include <tvm/ir/expr.h>
#include <tvm/runtime/object.h>
#include <tvm/s_tir/schedule/schedule.h>
 
namespace tvm {
namespace s_tir {
namespace meta_schedule {
 
class TuneContext;
class ScheduleRule;
 
class ScheduleRuleNode : public runtime::Object {
 public:
  virtual ~ScheduleRuleNode() = default;
 
  static void RegisterReflection() {
    namespace refl = tvm::ffi::reflection;
    refl::ObjectDef<ScheduleRuleNode>();
  }
 
  virtual void InitializeWithTuneContext(const TuneContext& context) = 0;
 
  virtual ffi::Array<s_tir::Schedule> Apply(const s_tir::Schedule& sch,
                                            const s_tir::SBlockRV& block) = 0;
 
  virtual ScheduleRule Clone() const = 0;
 
  static constexpr const bool _type_mutable = true;
  TVM_FFI_DECLARE_OBJECT_INFO("s_tir.meta_schedule.ScheduleRule", ScheduleRuleNode, Object);
};
 
class ScheduleRule : public runtime::ObjectRef {
 public:
  using FInitializeWithTuneContext = ffi::TypedFunction<void(const TuneContext&)>;
  using FApply = ffi::TypedFunction<ffi::Array<s_tir::Schedule>(const s_tir::Schedule&,
                                                                const s_tir::SBlockRV&)>;
  using FAsString = ffi::TypedFunction<ffi::String()>;
  using FClone = ffi::TypedFunction<ScheduleRule()>;
  TVM_DLL static ScheduleRule ApplyCustomRule();
  TVM_DLL static bool IsApplyCustomRule(const ScheduleRule& rule);
  TVM_DLL static ScheduleRule AutoInline(bool into_producer,          //
                                         bool into_consumer,          //
                                         bool inline_const_tensor,    //
                                         bool disallow_if_then_else,  //
                                         bool require_injective,      //
                                         bool require_ordered,        //
                                         ffi::Optional<ffi::Array<ffi::String>> disallow_op);
 
  TVM_DLL static ScheduleRule InlineConstantScalars();
 
  TVM_DLL static ScheduleRule MultiLevelTiling(
      ffi::String structure,                                      //
      ffi::Optional<ffi::Array<ffi::String>> tile_binds,          //
      ffi::Optional<Integer> max_innermost_factor,                //
      ffi::Optional<ffi::Array<Integer>> vector_load_lens,        //
      ffi::Optional<ffi::Map<ffi::String, ffi::Any>> reuse_read,  //
      ffi::Optional<ffi::Map<ffi::String, ffi::Any>> reuse_write,
      ffi::Optional<ffi::Function> filter_fn = std::nullopt);
 
  TVM_DLL static ScheduleRule MultiLevelTilingWithIntrin(
      ffi::String intrin_name, ffi::String structure,
      ffi::Optional<ffi::Array<ffi::String>> tile_binds,
      ffi::Optional<Integer> max_innermost_factor,
      ffi::Optional<ffi::Array<Integer>> vector_load_lens,
      ffi::Optional<ffi::Map<ffi::String, ffi::Any>> reuse_read,
      ffi::Optional<ffi::Map<ffi::String, ffi::Any>> reuse_write);
 
  TVM_DLL static ScheduleRule MultiLevelTilingTensorCore(
      ffi::Array<ffi::Map<ffi::String, ffi::String>> intrin_groups, ffi::String structure,
      ffi::Optional<ffi::Array<ffi::String>> tile_binds,
      ffi::Optional<Integer> max_innermost_factor,
      ffi::Optional<ffi::Array<Integer>> vector_load_lens,
      ffi::Optional<ffi::Map<ffi::String, ffi::Any>> reuse_read,
      ffi::Optional<ffi::Map<ffi::String, ffi::Any>> reuse_write, bool use_software_pipeline);
 
  TVM_DLL static ScheduleRule MultiLevelTilingWideVector(
      ffi::String structure, Integer vector_length_in_bits,
      ffi::Optional<Integer> max_innermost_factor,
      ffi::Optional<ffi::Map<ffi::String, ffi::Any>> reuse_read,
      ffi::Optional<ffi::Map<ffi::String, ffi::Any>> reuse_write);
 
  TVM_DLL static ScheduleRule AddRFactor(int max_jobs_per_core,  //
                                         ffi::Optional<Integer> max_innermost_factor);
  TVM_DLL static ScheduleRule CrossThreadReduction(ffi::Array<Integer> thread_extents);
  TVM_DLL static ScheduleRule RandomComputeLocation();
  TVM_DLL static ScheduleRule ParallelizeVectorizeUnroll(int max_jobs_per_core,                 //
                                                         int max_vectorize_extent,              //
                                                         ffi::Array<Integer> unroll_max_steps,  //
                                                         bool unroll_explicit);
  TVM_DLL static ScheduleRule AutoBind(int max_threadblocks, ffi::Array<Integer> thread_extents,
                                       int max_threads_per_block = -1);
  TVM_DLL static ScheduleRule PyScheduleRule(
      FInitializeWithTuneContext f_initialize_with_tune_context,  //
      FApply f_apply,                                             //
      FClone f_clone,                                             //
      FAsString f_as_string);
 
  TVM_DLL static ffi::Array<ScheduleRule, void> DefaultLLVM();
  TVM_DLL static ffi::Array<ScheduleRule, void> DefaultX86(const ffi::String& type);
  TVM_DLL static ffi::Array<ScheduleRule, void> DefaultCUDA();
  TVM_DLL static ffi::Array<ScheduleRule, void> DefaultCUDATensorCore();
  TVM_DLL static ffi::Array<ScheduleRule, void> DefaultHexagon();
  TVM_DLL static ffi::Array<ScheduleRule, void> DefaultARM(const ffi::String& type);
  TVM_DLL static ffi::Array<ScheduleRule, void> DefaultRISCV(int vlen);
 
  TVM_FFI_DEFINE_OBJECT_REF_METHODS_NULLABLE(ScheduleRule, ObjectRef, ScheduleRuleNode);
};
 
class PyScheduleRuleNode : public ScheduleRuleNode {
 public:
  using FInitializeWithTuneContext = ScheduleRule::FInitializeWithTuneContext;
  using FApply = ScheduleRule::FApply;
  using FClone = ScheduleRule::FClone;
  using FAsString = ScheduleRule::FAsString;
 
  FInitializeWithTuneContext f_initialize_with_tune_context;
  FApply f_apply;
  FAsString f_as_string;
  FClone f_clone;
 
  static void RegisterReflection() {
    // `f_initialize_with_tune_context` is not registered
    // `f_apply` is not registered
    // `f_as_string` is not registered
    // `f_clone` is not registered
    namespace refl = tvm::ffi::reflection;
    refl::ObjectDef<PyScheduleRuleNode>();
  }
 
  void InitializeWithTuneContext(const TuneContext& context) final;
  ffi::Array<s_tir::Schedule> Apply(const s_tir::Schedule& sch, const s_tir::SBlockRV& block) final;
  ScheduleRule Clone() const final;
  TVM_FFI_DECLARE_OBJECT_INFO_FINAL("s_tir.meta_schedule.PyScheduleRule", PyScheduleRuleNode,
                                    ScheduleRuleNode);
};
 
}  // namespace meta_schedule
}  // namespace s_tir
}  // namespace tvm
 
#endif  // TVM_S_TIR_META_SCHEDULE_SCHEDULE_RULE_H_