warning clean

mindspore/ccsrc/backend/optimizer/ascend/ir_fission/transdata_split.cc

@@ -14,10 +14,8 @@
  * limitations under the License.
  */
 #include "backend/optimizer/ascend/ir_fission/transdata_split.h"
-#include "backend/optimizer/ascend/ascend_helper.h"
-#include "backend/session/anf_runtime_algorithm.h"
+#include <set>
 #include "debug/anf_ir_dump.h"
-#include "utils/trace_base.h"
 
 namespace mindspore {
 namespace opt {

mindspore/ccsrc/backend/optimizer/ascend/mindir/bn_grad_unify_mindir.cc

@@ -14,15 +14,9 @@
  * limitations under the License.
  */
 #include "backend/optimizer/ascend/mindir/bn_grad_unify_mindir.h"
-
 #include <vector>
 #include <memory>
-
-#include "utils/utils.h"
-#include "utils/ms_context.h"
-#include "backend/optimizer/common/helper.h"
 #include "backend/session/anf_runtime_algorithm.h"
-#include "utils/trace_base.h"
 
 namespace mindspore {
 namespace opt {

mindspore/ccsrc/backend/optimizer/ascend/mindir/dynamic_reshape_unify_mindir.cc

@@ -14,15 +14,9 @@
  * limitations under the License.
  */
 #include "backend/optimizer/ascend/mindir/dynamic_reshape_unify_mindir.h"
-
 #include <vector>
 #include <memory>
-
-#include "utils/utils.h"
-#include "utils/ms_context.h"
-#include "backend/optimizer/common/helper.h"
 #include "backend/session/anf_runtime_algorithm.h"
-#include "utils/trace_base.h"
 
 namespace mindspore {
 namespace opt {

mindspore/ccsrc/backend/optimizer/gpu/adjust_depend_for_parallel_optimizer_recompute_all_gather_fusion.cc

@@ -1,149 +0,0 @@
-/**
- * Copyright 2021 Huawei Technologies Co., Ltd
- *
- * Licensed 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.
- */
-
-#include "backend/optimizer/gpu/adjust_depend_for_parallel_optimizer_recompute_all_gather_fusion.h"
-
-#include <algorithm>
-
-#include "utils/hash_map.h"
-#include "backend/session/anf_runtime_algorithm.h"
-#include "utils/utils.h"
-
-namespace mindspore {
-namespace opt {
-bool AdjustDependForParallelOptimizerRecomputeAllGatherFusion::Run(const FuncGraphPtr &graph) {
-  MS_EXCEPTION_IF_NULL(graph);
-  mindspore::HashMap<int64_t, bool> forward_allgather_recompute_value_in_fusion_group;
-  std::vector<AnfNodePtr> node_list = TopoSort(graph->get_return());
-  std::vector<int64_t> parallel_optimizer_recompute_allgather_fusion_ids;
-  std::vector<AnfNodePtr> parallel_optimizer_recompute_allgathers;
-  std::vector<AnfNodePtr> parallel_optimizer_recompute_first_fusion_allgathers;
-  int64_t unrecompute_max_fusion_id = -1;
-  int64_t recompute_min_fusion_id = 0;
-  for (auto &node : node_list) {
-    MS_EXCEPTION_IF_NULL(node);
-    if (!node->cast<CNodePtr>() || !AnfUtils::IsRealKernel(node)) {
-      continue;
-    }
-    auto cnode = node->cast<CNodePtr>();
-    if (!AnfAlgo::IsAllgather(cnode) || !AnfAlgo::IsFusion(cnode) || !AnfAlgo::IsFromParallelOptimizer(cnode)) {
-      continue;
-    }
-    if (AnfAlgo::IsRecompute(cnode)) {
-      int64_t fusion_id = AnfAlgo::GetNodeAttr<int64_t>(cnode, kAttrFusion);
-      if (std::find(parallel_optimizer_recompute_allgather_fusion_ids.begin(),
-                    parallel_optimizer_recompute_allgather_fusion_ids.end(),
-                    fusion_id) == parallel_optimizer_recompute_allgather_fusion_ids.end()) {
-        parallel_optimizer_recompute_allgather_fusion_ids.push_back(fusion_id);
-        if (recompute_min_fusion_id == 0 || fusion_id < recompute_min_fusion_id) {
-          recompute_min_fusion_id = fusion_id;
-        }
-        parallel_optimizer_recompute_first_fusion_allgathers.push_back(node);
-      } else {
-        parallel_optimizer_recompute_allgathers.push_back(node);
-      }
-    } else {
-      int64_t unrecompute_fusion_id = AnfAlgo::GetNodeAttr<int64_t>(cnode, kAttrFusion);
-      unrecompute_max_fusion_id = std::max(unrecompute_fusion_id, unrecompute_max_fusion_id);
-      bool would_be_recomputed =
-        AnfAlgo::HasNodeAttr(kAttrRecompute, cnode) && AnfAlgo::GetNodeAttr<bool>(cnode, kAttrRecompute);
-      auto [iter, inserted] =
-        forward_allgather_recompute_value_in_fusion_group.emplace(unrecompute_fusion_id, would_be_recomputed);
-      if (!inserted && iter->second != would_be_recomputed) {
-        MS_LOG(EXCEPTION) << "In same fusion group, the allgather recompute attribute should be equal. "
-                             "The normal node is:"
-                          << cnode->fullname_with_scope();
-      }
-    }
-  }
-  IncreaseAllgatherFusionId(parallel_optimizer_recompute_allgathers,
-                            parallel_optimizer_recompute_first_fusion_allgathers, unrecompute_max_fusion_id,
-                            recompute_min_fusion_id);
-  return AdjustAllgatherDepend(graph, parallel_optimizer_recompute_allgathers);
-}
-
-void AdjustDependForParallelOptimizerRecomputeAllGatherFusion::IncreaseAllgatherFusionId(
-  const std::vector<AnfNodePtr> &parallel_optimizer_recompute_allgathers,
-  const std::vector<AnfNodePtr> &parallel_optimizer_recompute_first_fusion_allgathers,
-  int64_t unrecompute_max_fusion_id, int64_t recompute_min_fusion_id) {
-  // means that there may some forward allgather and duplicated allgather would be fused.
-  if (recompute_min_fusion_id <= unrecompute_max_fusion_id) {
-    MS_LOG(WARNING) << "Increase the duplicated allgather fusion id";
-    for (auto &adjust_node : parallel_optimizer_recompute_first_fusion_allgathers) {
-      int64_t current_fusion_id = AnfAlgo::GetNodeAttr<int64_t>(adjust_node, kAttrFusion);
-      int64_t destination_fusion_id = current_fusion_id + unrecompute_max_fusion_id - recompute_min_fusion_id + 2;
-      AnfAlgo::SetNodeAttr(kAttrFusion, MakeValue(destination_fusion_id), adjust_node);
-    }
-    for (auto &adjust_node : parallel_optimizer_recompute_allgathers) {
-      int64_t current_fusion_id = AnfAlgo::GetNodeAttr<int64_t>(adjust_node, kAttrFusion);
-      int64_t destination_fusion_id = current_fusion_id + unrecompute_max_fusion_id - recompute_min_fusion_id + 2;
-      AnfAlgo::SetNodeAttr(kAttrFusion, MakeValue(destination_fusion_id), adjust_node);
-    }
-  }
-}
-
-bool AdjustDependForParallelOptimizerRecomputeAllGatherFusion::AdjustAllgatherDepend(
-  const FuncGraphPtr &graph, const std::vector<AnfNodePtr> &parallel_optimizer_recompute_allgathers) {
-  FuncGraphManagerPtr manager = graph->manager();
-  bool changed = false;
-  for (auto &node : parallel_optimizer_recompute_allgathers) {
-    auto cnode = node->cast<CNodePtr>();
-    auto depend_node = AnfAlgo::GetInputNode(cnode, 0);
-    if (IsPrimitiveCNode(depend_node, prim::kPrimDepend)) {
-      auto depend_cnode = depend_node->cast<CNodePtr>();
-      AnfNodeIndexSet allgather_node_set = manager->node_users()[cnode];
-      for (auto &node_pair : allgather_node_set) {
-        auto allgather_next_node = node_pair.first;
-        CNodePtr allgather_next_cnode = node_pair.first->cast<CNodePtr>();
-        if (allgather_next_cnode == nullptr || !IsValueNode<Primitive>(allgather_next_cnode->input(0))) {
-          continue;
-        }
-        std::vector<AnfNodePtr> inputs = {NewValueNode(std::make_shared<Primitive>(prim::kPrimDepend->name())),
-                                          allgather_next_node, AnfAlgo::GetInputNode(depend_cnode, 1)};
-        auto new_depend = graph->NewCNode(inputs);
-        new_depend->set_abstract(depend_node->abstract());
-        manager->SetEdge(node, 1, AnfAlgo::GetInputNode(depend_cnode, 0));
-        (void)manager->Replace(allgather_next_node, new_depend);
-        changed = true;
-      }
-    } else if (IsPrimitiveCNode(depend_node, prim::kPrimCast) &&
-               IsPrimitiveCNode(AnfAlgo::GetInputNode(depend_node->cast<CNodePtr>(), 0), prim::kPrimDepend)) {
-      auto cast_cnode = depend_node->cast<CNodePtr>();
-      auto cast_depend_node = AnfAlgo::GetInputNode(cast_cnode, 0);
-      auto cast_depend_cnode = cast_depend_node->cast<CNodePtr>();
-      AnfNodeIndexSet allgather_node_set = manager->node_users()[cnode];
-      for (auto &node_pair : allgather_node_set) {
-        auto allgather_next_node = node_pair.first;
-        CNodePtr allgather_next_cnode = node_pair.first->cast<CNodePtr>();
-        if (allgather_next_cnode == nullptr || !IsValueNode<Primitive>(allgather_next_cnode->input(0))) {
-          continue;
-        }
-        std::vector<AnfNodePtr> inputs = {NewValueNode(std::make_shared<Primitive>(prim::kPrimDepend->name())),
-                                          allgather_next_node, AnfAlgo::GetInputNode(cast_depend_cnode, 1)};
-        auto new_depend = graph->NewCNode(inputs);
-        new_depend->set_abstract(cast_depend_node->abstract());
-        manager->SetEdge(depend_node, 1, AnfAlgo::GetInputNode(cast_depend_cnode, 0));
-        (void)manager->Replace(allgather_next_node, new_depend);
-        changed = true;
-      }
-    } else {
-      MS_LOG(WARNING) << "The parallel optimizer recompute allgather has no depend edge";
-    }
-  }
-  return changed;
-}
-}  // namespace opt
-}  // namespace mindspore

mindspore/ccsrc/backend/optimizer/gpu/adjust_depend_for_parallel_optimizer_recompute_all_gather_fusion.h

@@ -1,44 +0,0 @@
-/**
- * Copyright 2021 Huawei Technologies Co., Ltd
- *
- * Licensed 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 MINDSPORE_CCSRC_BACKEND_OPTIMIZER_PASS_ADJUST_DEPEND_FOR_PARALLEL_OPTIMIZER_RECOMPUTE_ALL_GATHER_H_
-#define MINDSPORE_CCSRC_BACKEND_OPTIMIZER_PASS_ADJUST_DEPEND_FOR_PARALLEL_OPTIMIZER_RECOMPUTE_ALL_GATHER_H_
-#include <vector>
-#include <string>
-#include <utility>
-#include <memory>
-
-#include "backend/optimizer/common/optimizer.h"
-
-namespace mindspore {
-namespace opt {
-class AdjustDependForParallelOptimizerRecomputeAllGatherFusion : public Pass {
- public:
-  explicit AdjustDependForParallelOptimizerRecomputeAllGatherFusion(const std::string &name)
-      : Pass("adjust_depend_for_parallel_optimizer_recompute_all_gather") {}
-  ~AdjustDependForParallelOptimizerRecomputeAllGatherFusion() override = default;
-  bool Run(const FuncGraphPtr &graph) override;
-
- private:
-  bool AdjustAllgatherDepend(const FuncGraphPtr &graph,
-                             const std::vector<AnfNodePtr> &parallel_optimizer_recompute_allgathers);
-  void IncreaseAllgatherFusionId(const std::vector<AnfNodePtr> &parallel_optimizer_recompute_allgathers,
-                                 const std::vector<AnfNodePtr> &parallel_optimizer_recompute_first_fusion_allgathers,
-                                 int64_t unrecompute_max_fusion_id, int64_t recompute_min_fusion_id);
-};
-}  // namespace opt
-}  // namespace mindspore
-#endif  // MINDSPORE_CCSRC_BACKEND_OPTIMIZER_PASS_ADJUST_DEPEND_FOR_PARALLEL_OPTIMIZER_RECOMPUTE_ALL_GATHER_H_

mindspore/ccsrc/backend/optimizer/pass/adjust_depend_for_parallel_optimizer_recompute_all_gather.cc

@@ -15,7 +15,7 @@
  */
 
 #include "backend/optimizer/pass/adjust_depend_for_parallel_optimizer_recompute_all_gather.h"
-#include "utils/utils.h"
+#include <algorithm>
 #include "backend/session/anf_runtime_algorithm.h"
 
 namespace mindspore {

mindspore/ccsrc/backend/optimizer/pass/adjust_depend_for_parallel_optimizer_recompute_all_gather.h

@@ -26,20 +26,17 @@
 #include "ir/anf.h"
 #include "backend/optimizer/common/helper.h"
 #include "backend/optimizer/common/optimizer.h"
-#include "backend/optimizer/ascend/ascend_helper.h"
 
 namespace mindspore {
 namespace opt {
 class AdjustDependForParallelOptimizerRecomputeAllGather : public Pass {
  public:
   AdjustDependForParallelOptimizerRecomputeAllGather()
-      : Pass("adjust_depend_for_parallel_optimizer_recompute_all_gather"),
-        kernel_select_(std::make_shared<KernelSelect>()) {}
+      : Pass("adjust_depend_for_parallel_optimizer_recompute_all_gather") {}
   ~AdjustDependForParallelOptimizerRecomputeAllGather() override = default;
   bool Run(const FuncGraphPtr &graph) override;
 
  private:
-  KernelSelectPtr kernel_select_;
   bool AdjustAllgatherDepend(const FuncGraphPtr &graph,
                              const std::vector<AnfNodePtr> &parallel_optimizer_recompute_allgathers);
   void IncreaseAllgatherFusionId(const std::vector<AnfNodePtr> &parallel_optimizer_recompute_allgathers,

mindspore/ccsrc/backend/optimizer/pass/reduce_sum_optimizer.cc

@@ -16,15 +16,7 @@
 
 #include "backend/optimizer/pass/reduce_sum_optimizer.h"
 #include <vector>
-#include "backend/optimizer/common/helper.h"
 #include "backend/session/anf_runtime_algorithm.h"
-#include "utils/utils.h"
-#include "abstract/abstract_value.h"
-#include "base/core_ops.h"
-#include "ir/anf.h"
-#include "ir/dtype.h"
-#include "ir/scalar.h"
-#include "utils/anf_utils.h"
 
 namespace mindspore {
 namespace opt {
@@ -119,7 +111,6 @@ AnfNodePtr ReduceSumOptimizer::InsertAssistNode(const CNodePtr &cnode, const Ker
 // create a new assist value node to deal with the following two case:
 // 1: the axis_input is empty, the new tensor of the new value node should be 'range(shape.size())',
 // the shape is the first input'shape of ReduceSum;
-//
 // 2: the value of axis_input contain the value less 0,
 // the new tensor of the new value node should be "shape.size() + the_old_value_less_0",
 // the shape is the first input'shape of ReduceSum;

mindspore/ccsrc/backend/session/gpu_session.cc

@@ -57,7 +57,7 @@
 #include "backend/optimizer/gpu/concat_outputs_for_all_gather.h"
 #include "backend/optimizer/pass/getitem_tuple.h"
 #include "backend/optimizer/pass/optimize_updatestate.h"
-#include "backend/optimizer/gpu/adjust_depend_for_parallel_optimizer_recompute_all_gather_fusion.h"
+#include "backend/optimizer/pass/adjust_depend_for_parallel_optimizer_recompute_all_gather.h"
 #include "utils/ms_device_shape_transfer.h"
 #include "debug/anf_ir_dump.h"
 #include "debug/dump_proto.h"
@@ -200,8 +200,7 @@ void GPUSession::HardwareOptimize(const std::shared_ptr<KernelGraph> &kernel_gra
   pm->AddPass(std::make_shared<opt::AddReluV2Fusion>());
   pm->AddPass(std::make_shared<opt::AddReluGradV2Fusion>());
   pm->AddPass(std::make_shared<opt::AllReduceFusion>());
-  pm->AddPass(std::make_shared<opt::AdjustDependForParallelOptimizerRecomputeAllGatherFusion>(
-    "adjust_depend_for_parallel_optimizer_recompute_all_gather_fusion"));
+  pm->AddPass(std::make_shared<opt::AdjustDependForParallelOptimizerRecomputeAllGather>());
   pm->AddPass(std::make_shared<opt::AllGatherFusion>());
   pm->AddPass(std::make_shared<opt::ConcatOutputsForAllGather>());
   pm->AddPass(std::make_shared<opt::GetitemTuple>());

mindspore/ccsrc/runtime/hardware/gpu/gpu_device_context.cc

@@ -292,8 +292,7 @@ void GPUDeviceContext::OptimizeGraphWithDeviceInfo(const KernelGraphPtr &graph)
   pm->AddPass(std::make_shared<opt::AddReluV2Fusion>());
   pm->AddPass(std::make_shared<opt::AddReluGradV2Fusion>());
   pm->AddPass(std::make_shared<opt::AllReduceFusion>());
-  pm->AddPass(std::make_shared<opt::AdjustDependForParallelOptimizerRecomputeAllGatherFusion>(
-    "adjust_depend_for_parallel_optimizer_recompute_all_gather_fusion"));
+  pm->AddPass(std::make_shared<opt::AdjustDependForParallelOptimizerRecomputeAllGather>());
   pm->AddPass(std::make_shared<opt::AllGatherFusion>());
   pm->AddPass(std::make_shared<opt::ConcatOutputsForAllGather>());
   pm->AddPass(std::make_shared<opt::GetitemTuple>());

mindspore/ccsrc/runtime/hardware/gpu/optimizer.h

@@ -21,6 +21,7 @@
 #include "backend/optimizer/common/optimizer.h"
 #include "backend/optimizer/common/pass_manager.h"
 #include "backend/optimizer/common/common_backend_optimization.h"
+#include "backend/optimizer/pass/adjust_depend_for_parallel_optimizer_recompute_all_gather.h"
 #include "backend/optimizer/gpu/adam_weight_decay_fusion.h"
 #include "backend/optimizer/gpu/adam_fusion.h"
 #include "backend/optimizer/gpu/alltoall_fusion.h"
@@ -52,7 +53,6 @@
 #include "backend/optimizer/gpu/matmul_biasadd_fusion.h"
 #include "backend/optimizer/gpu/bce_with_logits_loss_fusion.h"
 #include "backend/optimizer/gpu/insert_cast_gpu.h"
-#include "backend/optimizer/gpu/adjust_depend_for_parallel_optimizer_recompute_all_gather_fusion.h"
 #include "backend/optimizer/gpu/neighbor_exchange_v2_fusion.h"
 
 #endif  // MINDSPORE_CCSRC_RUNTIME_HARDWARE_GPU_OPTIMIZER_H_