Cherry-pick PR 30103. Add Inplace strategy (Output reuse Input Varbase) in dygraph (#30103)

* add view strategy on squeeze,unsqueeze,reshape,flatten

* add squeeze unittest

* add unittests

* use View strategy as name rather than Reuse Allacation

* fix view api doc

* fix format

* use core.ops when input of reshape2 is Tensor

* fix test_cross_entropy_loss error because of reshape2

* fix test_cross_entropy_loss error because of reshape2

* add inplace strategy

* add elementwise_add sub

* let backward op not use inplace

* grad op do not use inplace

* fix memory increase error and add leaf error message

* delete selected_rows

* change op_function

* little change

* solve HandleViewBetweenInputAndOutput

* add unittest and leaf error message

* merge view error

* optimize op_function_generator format and support sum inplace op

* fix format of basic_engine

* fix format for framework

* little change of variable wrapper

* add reshape, squeeze, unsqueeze, scatter api

* add relu elu tanh softmax inplace api

* fix test_squeeze_op unittest

* fix test_relu_op unittest

* fix comment problems

* delete sample code of inplace api

* add reference of grad_pending_nodes in basic_engine

* fix unittest name

* add inplace apis into wlist

* fix error message

* add PADDLE_ENFORCE for set grad op twice

* fix head file error

Author: pangyoki

paddle/fluid/framework/details/op_registry.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <map>
 #include <memory>
 #include <string>
 #include <tuple>
@@ -247,8 +248,9 @@ struct OpInfoFiller<T, kGradOpBaseMaker> {
         const std::string& type,
         const imperative::NameVarBaseMap& var_base_map_in,
         const imperative::NameVarBaseMap& var_base_map_out,
-        const framework::AttributeMap& attrs) {
-      T maker(type, var_base_map_in, var_base_map_out, attrs);
+        const framework::AttributeMap& attrs,
+        const std::map<std::string, std::string>& inplace_map) {
+      T maker(type, var_base_map_in, var_base_map_out, attrs, inplace_map);
       return maker();
     };
   }

paddle/fluid/framework/grad_op_desc_maker.h

@@ -221,6 +221,10 @@ class SingleGradOpMaker<imperative::OpBase>
 
   std::shared_ptr<imperative::GradOpNode> operator()() const final {
     auto node = this->NewGradNode();
+    auto& inplace_map = this->GetInplaceMap();
+    if (!inplace_map.empty()) {
+      node->SetInplaceGradNameMap(inplace_map);
+    }
     {
       imperative::TracedGradOp traced_grad_op(node);
       try {

paddle/fluid/framework/type_defs.h

@@ -59,7 +59,8 @@ using DygraphGradOpMakerFN =
         const std::string& /*op_type*/,
         const imperative::NameVarBaseMap& /*var_base_map_in*/,
         const imperative::NameVarBaseMap& /*var_base_map_out*/,
-        const framework::AttributeMap& /*attributes*/)>;
+        const framework::AttributeMap& /*attributes*/,
+        const std::map<std::string, std::string>& /*inplace_map*/)>;
 
 using InferVarTypeFN =
     std::function<void(framework::InferVarTypeContext* /*context*/)>;

paddle/fluid/imperative/basic_engine.cc

@@ -108,7 +108,9 @@ void BasicEngine::CheckBackwardInputs(const OpBase& op) {
   }
 }
 
-void BasicEngine::PrepareGradAccumulators(const OpBase& op) {
+void BasicEngine::PrepareGradAccumulators(
+    const OpBase& op,
+    const std::vector<std::shared_ptr<GradOpNode>>& grad_pending_nodes) {
   for (const auto& pair : op.GetOutsMap()) {
     if (!pair.second.IsGrad()) {
       continue;
@@ -117,29 +119,94 @@ void BasicEngine::PrepareGradAccumulators(const OpBase& op) {
     for (const auto& var : pair.second) {
       if (!var) continue;
 
-      auto& accumulator = accumulators_[var.get()];
-      if (!accumulator) {
-        if (FLAGS_sort_sum_gradient) {
-          accumulator.reset(new SortedGradientAccumulator(var.get()));
-        } else {
-          accumulator.reset(new EagerGradientAccumulator(var.get()));
+      if (!var->HasGradNode()) {
+        auto& accumulator = accumulators_[var.get()];
+        if (!accumulator) {
+          if (FLAGS_sort_sum_gradient) {
+            accumulator.reset(new SortedGradientAccumulator(var.get()));
+          } else {
+            accumulator.reset(new EagerGradientAccumulator(var.get()));
+          }
         }
-      }
 
-      accumulator->IncreaseRefCnt();
+        accumulator->IncreaseRefCnt();
 
-      VLOG(3) << "Prepare to acccumulate variable grad " << var->Name() << "("
-              << var.get() << ")  with reference count "
-              << accumulator->RefCnt();
+        VLOG(3) << "Prepare to acccumulate variable grad " << var->Name() << "("
+                << var.get()
+                << ") that don't have grad node  with reference count "
+                << accumulator->RefCnt();
+
+        if (var->HasLeafHooks()) {
+          VLOG(3) << "Grad variable wrapper (" << var->Name()
+                  << ") has leaf grad hooks.";
+          PADDLE_ENFORCE_NE(
+              var->HasGradNode(), true,
+              platform::errors::PermissionDenied(
+                  "Only leaf Tensor's gradient can append hook to "
+                  "Gradientaccumulator."));
+          accumulator->SetPostHooks(var->GetLeafHooks());
+        }
+      } else {
+        // Because Inplace op overwrites the grad_node of the input grad_var. So
+        // only the information of grad_pending_node can be used to find the
+        // grad_node of grad_var.
+        bool find_grad_node_of_var = false;
+        for (auto& grad_pending_node : grad_pending_nodes) {
+          PADDLE_ENFORCE_NOT_NULL(
+              grad_pending_node,
+              platform::errors::NotFound("Grad pending node is nullptr."));
+          for (auto& grad_pending_op : *grad_pending_node) {
+            VLOG(6) << "Determine whether var (" << var->Name()
+                    << ") is the input var of grad_pending_op ("
+                    << grad_pending_op.Type() << ").";
+            grad_pending_op.EnforceHasInOut();
+            for (const auto& grad_pending_op_ins_pair :
+                 grad_pending_op.GetInsMap()) {
+              if (!grad_pending_op_ins_pair.second.IsGrad()) {
+                continue;
+              }
+              for (const auto& pending_in_var :
+                   grad_pending_op_ins_pair.second) {
+                if (var == pending_in_var) {
+                  VLOG(6) << "Var (" << var->Name()
+                          << ") is the input var of grad_pending_op ("
+                          << grad_pending_op.Type() << ").";
+                  find_grad_node_of_var = true;
+                  break;
+                }
+              }
+              if (find_grad_node_of_var) {
+                break;
+              }
+            }
+          }
 
-      if (var->HasLeafHooks()) {
-        VLOG(3) << "Grad variable wrapper (" << var->Name()
-                << ") has leaf grad hooks.";
-        PADDLE_ENFORCE_NE(var->HasGradNode(), true,
-                          platform::errors::PermissionDenied(
-                              "Only leaf Tensor's gradient can append hook to "
-                              "Gradientaccumulator."));
-        accumulator->SetPostHooks(var->GetLeafHooks());
+          if (find_grad_node_of_var) {
+            auto& accumulator =
+                accumulators_with_grad_node_[grad_pending_node][var.get()];
+
+            if (!accumulator) {
+              if (FLAGS_sort_sum_gradient) {
+                accumulator.reset(new SortedGradientAccumulator(var.get()));
+              } else {
+                accumulator.reset(new EagerGradientAccumulator(var.get()));
+              }
+            }
+
+            accumulator->IncreaseRefCnt();
+
+            VLOG(3) << "Prepare to acccumulate variable grad " << var->Name()
+                    << "(" << var.get()
+                    << ") that has grad node with reference count "
+                    << accumulator->RefCnt();
+            break;
+          }
+        }
+        PADDLE_ENFORCE_EQ(
+            find_grad_node_of_var, true,
+            platform::errors::NotFound(
+                "No grad node corresponding to grad Tensor (%s) was found.",
+                var->Name()));
       }
     }
   }
@@ -148,10 +215,13 @@ void BasicEngine::PrepareGradAccumulators(const OpBase& op) {
 void BasicEngine::PrepareDeps() {
   PADDLE_ENFORCE_EQ(
       node_deps_.empty(), true,
-      platform::errors::AlreadyExists("Op deps must be initialized here"));
+      platform::errors::AlreadyExists("Op deps must be initialized."));
   PADDLE_ENFORCE_EQ(
       accumulators_.empty(), true,
-      platform::errors::AlreadyExists("Accumulators must be initialized here"));
+      platform::errors::AlreadyExists("Accumulators must be initialized."));
+  PADDLE_ENFORCE_EQ(
+      accumulators_with_grad_node_.empty(), true,
+      platform::errors::AlreadyExists("Accumulators must be initialized."));
 
   std::queue<GradOpNode*> q;
   std::unordered_set<GradOpNode*> visited;
@@ -163,16 +233,17 @@ void BasicEngine::PrepareDeps() {
     auto* cur_node = q.front();
     q.pop();
 
+    const auto& grad_pending_nodes = cur_node->GradPendingNodes();
+
     for (auto& cur_op : *cur_node) {
       cur_op.EnforceHasInOut();
-      PrepareGradAccumulators(cur_op);
+      PrepareGradAccumulators(cur_op, grad_pending_nodes);
     }
 
-    const auto& grad_pending_nodes = cur_node->GradPendingNodes();
     for (auto& grad_pending_node : grad_pending_nodes) {
       PADDLE_ENFORCE_NOT_NULL(
           grad_pending_node,
-          platform::errors::NotFound("Grad pending node should not be null"));
+          platform::errors::NotFound("Grad pending node is nullptr."));
       ++node_deps_[grad_pending_node.get()];
       if (visited.count(grad_pending_node.get()) == 0) {
         visited.insert(grad_pending_node.get());
@@ -198,6 +269,8 @@ void BasicEngine::Execute() {
     auto shared_cur_node = std::move(q.front());
     q.pop();
 
+    auto& inplace_grad_name_map = shared_cur_node->InplaceGradNameMap();
+
     for (auto& cur_op : *shared_cur_node) {
       ++op_num;
 
@@ -222,11 +295,38 @@ void BasicEngine::Execute() {
             continue;
           }
 
-          auto iter = accumulators_.find(var.get());
-          PADDLE_ENFORCE_EQ(
-              iter != accumulators_.end(), true,
-              platform::errors::NotFound("Cannot find gradient of variable %s",
-                                         var->Name()));
+          std::unordered_map<VariableWrapper*,
+                             std::unique_ptr<GradientAccumulator>>::iterator
+              iter;
+          if (!var->HasGradNode()) {
+            VLOG(10) << "Find gradient of var (" << var->Name()
+                     << ") with no grad_node.";
+            iter = accumulators_.find(var.get());
+            PADDLE_ENFORCE_EQ(
+                iter != accumulators_.end(), true,
+                platform::errors::NotFound(
+                    "Cannot find gradient of variable %s", var->Name()));
+          } else {
+            bool flag_find_grad = false;
+            VLOG(10) << "Find gradient of var (" << var->Name()
+                     << ") with grad_node.";
+            for (auto& grad_pending_node :
+                 shared_cur_node->GradPendingNodes()) {
+              const auto& iter_grad_node =
+                  accumulators_with_grad_node_.find(grad_pending_node);
+              if (iter_grad_node != accumulators_with_grad_node_.end()) {
+                iter = iter_grad_node->second.find(var.get());
+                if (iter != iter_grad_node->second.end()) {
+                  flag_find_grad = true;
+                  break;
+                }
+              }
+            }
+            PADDLE_ENFORCE_EQ(
+                flag_find_grad, true,
+                platform::errors::NotFound(
+                    "Cannot find gradient of variable %s", var->Name()));
+          }
 
           // leaf_accumulators_ : hooks and accumulate-grad for leaf tensor
           if (var->IsLeafGrad()) {
@@ -245,6 +345,25 @@ void BasicEngine::Execute() {
             need_accu_var_list_.emplace_back(iter->second.get(), var);
             VLOG(10) << "create temporary var of " << var->Name()
                      << " for sum gradient within this graph!";
+          } else if (!inplace_grad_name_map.empty() &&
+                     inplace_grad_name_map.count(pair.first)) {
+            // When calculate Inplace grad op, create a new output var.
+            // If a tmp var has been created, there is no need to create it
+            // again.
+            for (auto& in_var :
+                 bwd_ins.at(inplace_grad_name_map.at(pair.first))) {
+              if (in_var == var) {
+                auto tmp_var = std::make_shared<VariableWrapper>(var->Name());
+                tmp_var->SetType(var->Type());
+                tmp_var->SetForwardDataType(var->ForwardDataType());
+                inplace_output_grad_var_list_.emplace_back(var, tmp_var);
+                var = tmp_var;
+                VLOG(10) << "Inplace grad op does not use the Inplace "
+                            "strategy, a temporary output var ("
+                         << var->Name() << ") will be created.";
+                break;
+              }
+            }
           }
         }
       }
@@ -280,6 +399,10 @@ void BasicEngine::Execute() {
                     cur_op.place());
       }
 
+      for (auto& pair : inplace_output_grad_var_list_) {
+        *pair.first = std::move(*pair.second);
+      }
+
       // Step 2: Sum Gradient of This graph
       for (auto& pair : need_accu_var_list_) {
         pair.first->SumGrad(std::move(pair.second), cur_op.id());
@@ -302,6 +425,7 @@ void BasicEngine::Execute() {
       }
 
       need_accu_var_list_.clear();
+      inplace_output_grad_var_list_.clear();
       leaf_accumulators_.clear();
 
       if (!retain_graph_) {
@@ -312,9 +436,9 @@ void BasicEngine::Execute() {
 
     // Step 3: Collect ready ops
     for (auto& grad_pending_node : shared_cur_node->GradPendingNodes()) {
-      PADDLE_ENFORCE_NOT_NULL(grad_pending_node,
-                              platform::errors::NotFound(
-                                  "Grad pending node should not be nullptr"));
+      PADDLE_ENFORCE_NOT_NULL(
+          grad_pending_node,
+          platform::errors::NotFound("Grad pending node is nullptr."));
       auto iter = node_deps_.find(grad_pending_node.get());
       if (iter == node_deps_.end()) {
         continue;
@@ -334,6 +458,7 @@ void BasicEngine::Clear() {
   init_node_.reset();
   node_deps_.clear();
   accumulators_.clear();
+  accumulators_with_grad_node_.clear();
   need_accu_var_list_.clear();
   leaf_accumulators_.clear();
 }

paddle/fluid/imperative/basic_engine.h

@@ -39,15 +39,33 @@ class BasicEngine : public Engine {
 
   void CheckBackwardInputs(const OpBase& op);
 
-  void PrepareGradAccumulators(const OpBase& op);
+  void PrepareGradAccumulators(
+      const OpBase& op,
+      const std::vector<std::shared_ptr<GradOpNode>>& grad_pending_nodes);
 
   void Clear();
 
  private:
   std::shared_ptr<GradOpNode> init_node_;
   std::unordered_map<GradOpNode*, size_t> node_deps_;
+  // The input and output of Inplace op are the same. If only `var` is used
+  // as the key, then the input and output of inplace op must be gradient
+  // accumulated. Therefore, add the `grad_node` as the key to prevent the
+  // problem of gradient accumulation in inplace op.
+  std::unordered_map<std::shared_ptr<GradOpNode>,
+                     std::unordered_map<VariableWrapper*,
+                                        std::unique_ptr<GradientAccumulator>>>
+      accumulators_with_grad_node_;
+  // Leaf var doesn't have grad_node, and leaf var with `stop_gradient=False`
+  // can't use Inplace strategy. If a var doesn't have grad_node, only use
+  // `var` as the key.
   std::unordered_map<VariableWrapper*, std::unique_ptr<GradientAccumulator>>
       accumulators_;
+  // The output grad var of Inplace grad op. Because Inplace grad op does not
+  // use the Inplace strategy, a new output grad var needs to be created.
+  std::vector<std::pair<std::shared_ptr<VariableWrapper>,
+                        std::shared_ptr<VariableWrapper>>>
+      inplace_output_grad_var_list_;
   std::vector<std::pair<GradientAccumulator*, std::shared_ptr<VariableWrapper>>>
       need_accu_var_list_;
   // leaf_accumulators_ is only for leaf tensor(hooks/accumulate grad)