Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

… frame_migration
PaddlePaddle · Jul 27, 2022 · 5eaaf2d · 5eaaf2d
2 parents 2799cd7 + 84d595f
commit 5eaaf2d
Show file tree

Hide file tree

Showing 92 changed files with 2,584 additions and 869 deletions.
diff --git a/cmake/xpu_kp.cmake b/cmake/xpu_kp.cmake
@@ -168,11 +168,13 @@ macro(compile_kernel COMPILE_ARGS)
   else()
     set(ABI_VERSION "-D_GLIBCXX_USE_CXX11_ABI=1")
   endif()
+  add_custom_target(
+    ${kernel_name}.xpu ALL
+    COMMAND ${CMAKE_COMMAND} -E copy ${kernel_path}/${kernel_name}.kps
+            kernel_build/${kernel_name}.xpu)
   add_custom_command(
     OUTPUT kernel_build/${kernel_name}.bin.o
     COMMAND ${CMAKE_COMMAND} -E make_directory kernel_build
-    COMMAND ${CMAKE_COMMAND} -E copy ${kernel_path}/${kernel_name}.kps
-            kernel_build/${kernel_name}.xpu
     COMMAND
       ${XPU_CLANG} --sysroot=${CXX_DIR} -std=c++11 ${ABI_VERSION} ${OPT_LEVEL}
       -fno-builtin -mcpu=xpu2 -fPIC ${XPU_CXX_DEFINES} ${XPU_CXX_FLAGS}
@@ -189,8 +191,6 @@ macro(compile_kernel COMPILE_ARGS)
   add_custom_command(
     OUTPUT kernel_build/${kernel_name}.host.o
     COMMAND ${CMAKE_COMMAND} -E make_directory kernel_build
-    COMMAND ${CMAKE_COMMAND} -E copy ${kernel_path}/${kernel_name}.kps
-            kernel_build/${kernel_name}.xpu
     COMMAND
       ${XPU_CLANG} --sysroot=${CXX_DIR} -std=c++11 ${ABI_VERSION} ${OPT_LEVEL}
       -fno-builtin -mcpu=xpu2 -fPIC ${XPU_CXX_DEFINES} ${XPU_CXX_FLAGS}

diff --git a/paddle/fluid/eager/auto_code_generator/final_state_generator/python_c_gen.py b/paddle/fluid/eager/auto_code_generator/final_state_generator/python_c_gen.py
@@ -57,6 +57,8 @@ def SkipAPIGeneration(forward_api_name):
     'adam',
     'adamw_',
     'adamw',
+    'average_accumulates',
+    'average_accumulates_',
     'decayed_adagrad_',
     'decayed_adagrad',
     'dgc_momentum_',

diff --git a/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc b/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc
@@ -297,7 +297,18 @@ void FuseElewiseAddActPass::RemoveIntermediateOut(Graph *graph) const {
       }
     }
   }
-  GraphSafeRemoveNodes(graph, need_removed_nodes);
+  details::RemovedVars *saved_removed_nodes = new details::RemovedVars;
+  GraphSafeRemoveNodes(graph, need_removed_nodes, saved_removed_nodes);
+  if (!saved_removed_nodes->empty()) {
+    // TODO(pangyoki): If kRemovedVars exists, merge saved_removed_nodes into
+    // RemovedVars.
+    PADDLE_ENFORCE_EQ(graph->Has(details::kRemovedVars),
+                      false,
+                      platform::errors::PreconditionNotMet(
+                          "Removed nodes are only saved for "
+                          "fuse_elewise_add_act_pass in temporary."));
+    graph->Set(details::kRemovedVars, saved_removed_nodes);
+  }
 }
 
 void FuseElewiseAddActPass::ReLinkNodes(Graph *graph,

diff --git a/paddle/fluid/framework/ir/graph.h b/paddle/fluid/framework/ir/graph.h
@@ -45,6 +45,8 @@ namespace details {
 // This attr is not recommended, because the graph should not dependence
 // the program once it is built.
 constexpr char kStaleProgramOpDescs[] = "stale_program_op_descs";
+constexpr char kRemovedVars[] = "removed_vars";
+typedef std::unordered_set<std::shared_ptr<ir::Node>> RemovedVars;
 }  //  namespace details
 
 namespace ir {

diff --git a/paddle/fluid/framework/ir/graph_helper.cc b/paddle/fluid/framework/ir/graph_helper.cc
@@ -549,6 +549,18 @@ static void GetGraphOpDesc(const std::vector<Node *> &nodes,
   }
 }
 
+template <class T = Node *>
+static void GetGraphVarDesc(const Graph &graph,
+                            const std::unordered_set<T> &nodes,
+                            std::vector<proto::VarDesc> *vars) {
+  for (T node : nodes) {
+    if (node->IsVar() && node->Var() &&
+        node->GetVarNodeBlockId() == graph.GetBlockId()) {
+      vars->emplace_back(*node->Var()->Proto());
+    }
+  }
+}
+
 static void GraphToBlock(const Graph &graph,
                          proto::BlockDesc *block,
                          const SortKind *sort_kind) {
@@ -562,11 +574,11 @@ static void GraphToBlock(const Graph &graph,
   }
 
   std::vector<proto::VarDesc> vars_in_graph;
-  for (Node *node : graph.Nodes()) {
-    if (node->IsVar() && node->Var() &&
-        node->GetVarNodeBlockId() == graph.GetBlockId()) {
-      vars_in_graph.emplace_back(*node->Var()->Proto());
-    }
+  GetGraphVarDesc<Node *>(graph, graph.Nodes(), &vars_in_graph);
+  if (graph.Has(details::kRemovedVars)) {
+    auto &removed_vars = graph.Get<details::RemovedVars>(details::kRemovedVars);
+    GetGraphVarDesc<std::shared_ptr<ir::Node>>(
+        graph, removed_vars, &vars_in_graph);
   }
 
   // add vars_in_graph to blcok

diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.cc b/paddle/fluid/framework/ir/graph_pattern_detector.cc
@@ -771,10 +771,18 @@ bool IsNthOutput(Node *var, Node *op, const std::string &argument, size_t nth) {
   return var->Name() == op->Op()->Output(argument)[nth];
 }
 
-void GraphSafeRemoveNodes(Graph *graph,
-                          const std::unordered_set<const Node *> &nodes) {
+void GraphSafeRemoveNodes(
+    Graph *graph,
+    const std::unordered_set<const Node *> &nodes,
+    std::unordered_set<std::shared_ptr<Node>> *saved_nodes) {
   for (auto *node : nodes) {
-    graph->RemoveNode(const_cast<Node *>(node));
+    if (saved_nodes != nullptr) {
+      // prevent unique_ptr node from being released
+      saved_nodes->insert(
+          std::move(graph->RemoveNode(const_cast<Node *>(node))));
+    } else {
+      graph->RemoveNode(const_cast<Node *>(node));
+    }
   }
 
   for (auto *node : graph->Nodes()) {

diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.h b/paddle/fluid/framework/ir/graph_pattern_detector.h
@@ -392,8 +392,10 @@ bool HasOutput(Node* op, const std::string& argument);
 bool IsNthOutput(Node* var, Node* op, const std::string& argument, size_t nth);
 
 // Graph safely remove some nodes, will automatically clean up the edges.
-void GraphSafeRemoveNodes(Graph* graph,
-                          const std::unordered_set<const Node*>& nodes);
+void GraphSafeRemoveNodes(
+    Graph* graph,
+    const std::unordered_set<const Node*>& nodes,
+    std::unordered_set<std::shared_ptr<Node>>* saved_nodes = nullptr);
 
 // Some pre-defined patterns those can be reused in multiple passes.
 // The related Fluid Layer or Op should be one pattern here for better re-usage

diff --git a/paddle/fluid/jit/all.h b/paddle/fluid/jit/all.h
@@ -0,0 +1,20 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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.
+
+#pragma once
+
+#include "base_function.h"
+#include "layer.h"
+#include "serializer.h"
+#include "serializer_utils.h"
diff --git a/paddle/fluid/jit/base_function.h b/paddle/fluid/jit/base_function.h
@@ -15,7 +15,6 @@
 #pragma once
 
 #include "paddle/phi/api/include/tensor.h"
-#include "paddle/phi/core/dense_tensor.h"
 
 namespace paddle {
 namespace jit {

diff --git a/paddle/fluid/jit/compilation_unit.cc b/paddle/fluid/jit/compilation_unit.cc
@@ -16,6 +16,8 @@
 
 #include "paddle/phi/core/enforce.h"
 
+#include "paddle/fluid/jit/base_function.h"
+
 namespace paddle {
 namespace jit {
 

diff --git a/paddle/fluid/jit/compilation_unit.h b/paddle/fluid/jit/compilation_unit.h
@@ -14,13 +14,14 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
 #include <unordered_map>
-
-#include "paddle/fluid/jit/base_function.h"
+#include <vector>
 
 namespace paddle {
 namespace jit {
+class BaseFunction;
 using Name2FunctionMap =
     std::unordered_map<std::string, std::shared_ptr<BaseFunction>>;
 

diff --git a/paddle/fluid/jit/function_schema.cc b/paddle/fluid/jit/function_schema.cc
@@ -14,6 +14,7 @@
 
 #include "paddle/fluid/jit/function_schema.h"
 
+#include "paddle/fluid/framework/program_desc.h"
 #include "paddle/phi/core/enforce.h"
 
 #include "paddle/fluid/jit/function_utils.h"
@@ -52,22 +53,21 @@ void FunctionSchema::AddOutputArg(const std::string& name) {
 FunctionInfo::FunctionInfo(const std::string& func_name,
                            const std::vector<std::string>& param_names,
                            const framework::ProgramDesc& program_desc)
-    : func_name_(func_name),
-      param_names_(param_names),
-      program_desc_(program_desc) {
+    : func_name_(func_name), param_names_(param_names) {
+  program_desc_.reset(new framework::ProgramDesc(program_desc));
   // Parse FunctionSchema
-  for (auto& in_name : program_desc_.GetFeedTargetNames()) {
+  for (auto& in_name : program_desc_->GetFeedTargetNames()) {
     schema_.AddInputArg(in_name);
   }
-  for (auto& out_name : program_desc_.GetFetchTargetNames()) {
+  for (auto& out_name : program_desc_->GetFetchTargetNames()) {
     schema_.AddOutputArg(out_name);
   }
 }
 
 const std::string& FunctionInfo::FunctionName() const { return func_name_; }
 
 const framework::ProgramDesc& FunctionInfo::ProgramDesc() const {
-  return program_desc_;
+  return *program_desc_.get();
 }
 
 const std::vector<std::string>& FunctionInfo::ParamNames() const {
@@ -83,7 +83,7 @@ const std::vector<std::string> FunctionInfo::OutputArgNames() const {
 }
 
 void FunctionInfo::RemoveDescFeedFetch() {
-  utils::RemoveFeedFetch(&program_desc_);
+  utils::RemoveFeedFetch(program_desc_.get());
 }
 
 }  // namespace jit

diff --git a/paddle/fluid/jit/function_schema.h b/paddle/fluid/jit/function_schema.h
@@ -14,15 +14,17 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
 #include <vector>
 
-#include "paddle/fluid/framework/program_desc.h"
-#include "paddle/fluid/framework/variable.h"
-
 namespace paddle {
+
+namespace framework {
+class ProgramDesc;
+}  // namespace framework
+
 namespace jit {
-using Variable = paddle::framework::Variable;
 
 class Argument {
  public:
@@ -75,7 +77,7 @@ class FunctionInfo {
  private:
   std::string func_name_;
   std::vector<std::string> param_names_;
-  framework::ProgramDesc program_desc_;
+  std::shared_ptr<framework::ProgramDesc> program_desc_;
   FunctionSchema schema_;
 };
 

diff --git a/paddle/fluid/jit/function_utils.cc b/paddle/fluid/jit/function_utils.cc
@@ -15,7 +15,9 @@
 #include "paddle/fluid/jit/function_utils.h"
 
 #include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/var_desc.h"
+#include "paddle/fluid/framework/variable.h"
 #include "paddle/phi/core/enforce.h"
 
 namespace paddle {
@@ -75,7 +77,7 @@ void ShareParamsIntoScope(const std::vector<std::string> &param_names,
   for (size_t i = 0; i < param_names.size(); ++i) {
     std::string name = param_names[i];
     auto &param = params_dict.find(name)->second;
-    auto &dense_tensor = param.Get<DenseTensor>();
+    auto &dense_tensor = param->Get<DenseTensor>();
     VLOG(3) << "share into scope: " << name;
     auto *var = scope->Var(name);
     auto *dst_tensor = var->GetMutable<DenseTensor>();

diff --git a/paddle/fluid/jit/function_utils.h b/paddle/fluid/jit/function_utils.h
@@ -18,18 +18,23 @@
 #include <unordered_map>
 #include <vector>
 
-#include "paddle/fluid/framework/scope.h"
-#include "paddle/fluid/framework/variable.h"
 #include "paddle/phi/api/include/tensor.h"
 #include "paddle/phi/common/place.h"
-#include "paddle/phi/core/dense_tensor.h"
 
 #include "paddle/fluid/jit/function_schema.h"
 
 namespace paddle {
+
+namespace framework {
+class Variable;
+class ProgramDesc;
+class Scope;
+}  // namespace framework
+
 namespace jit {
 using Variable = paddle::framework::Variable;
-using Name2VariableMap = std::unordered_map<std::string, Variable>;
+using Name2VariableMap =
+    std::unordered_map<std::string, std::shared_ptr<Variable>>;
 using DenseTensor = phi::DenseTensor;
 using Tensor = paddle::experimental::Tensor;
 

diff --git a/paddle/fluid/jit/layer.cc b/paddle/fluid/jit/layer.cc
@@ -14,17 +14,21 @@
 
 #include "paddle/fluid/jit/layer.h"
 
+#include "paddle/fluid/framework/variable.h"
+
+#include "paddle/fluid/jit/base_function.h"
+#include "paddle/fluid/jit/compilation_unit.h"
+#include "paddle/fluid/jit/function_schema.h"
+
 namespace paddle {
 namespace jit {
-Layer::Layer(const std::vector<std::shared_ptr<FunctionInfo>>& infos,
-             const Name2VariableMap& params_dict,
-             const phi::Place& place)
+Layer::Layer(const Name2VariableMap& params_dict, const phi::Place& place)
     : params_dict_(params_dict) {
-  VLOG(3) << "infos size: " << infos.size();
+  unit_.reset(new CompilationUnit());
 }
 
 std::shared_ptr<BaseFunction> Layer::Function(const std::string& name) const {
-  return unit_.Function(name);
+  return unit_->Function(name);
 }
 
 std::vector<Tensor> Layer::forward(const std::vector<Tensor>& inputs) {
@@ -42,15 +46,15 @@ void Layer::to(const phi::Place& place) {}
 
 void Layer::SetFunction(const std::string& name,
                         const std::shared_ptr<BaseFunction>& function) {
-  unit_.SetFunction(name, function);
+  unit_->SetFunction(name, function);
 }
 
 std::vector<std::string> Layer::FunctionNames() const {
-  return unit_.FunctionNames();
+  return unit_->FunctionNames();
 }
 
 const Name2FunctionMap& Layer::FunctionMap() const {
-  return unit_.FunctionMap();
+  return unit_->FunctionMap();
 }
 
 }  // namespace jit