Merge branch 'develop' of https://github.com/PaddlePaddle/paddle into backward_TensorCore

Author: AshburnLee

paddle/fluid/framework/details/CMakeLists.txt

@@ -134,11 +134,13 @@ set(IR_PASS_DEPS graph_viz_pass multi_devices_graph_pass
     modify_op_lock_and_record_event_pass
     coalesce_grad_tensor_pass fuse_all_reduce_op_pass backward_optimizer_op_deps_pass
     fuse_adam_op_pass fuse_sgd_op_pass fuse_momentum_op_pass
-    sync_batch_norm_pass runtime_context_cache_pass)
+    sync_batch_norm_pass runtime_context_cache_pass graph_to_program_pass)
 if(NOT APPLE AND NOT WIN32 AND (WITH_GPU OR WITH_ROCM))
   set(IR_PASS_DEPS ${IR_PASS_DEPS} fusion_group_pass)
 endif()
 cc_library(build_strategy SRCS build_strategy.cc DEPS pass_builder ${IR_PASS_DEPS})
+cc_test(build_strategy_test SRCS build_strategy_test.cc
+        DEPS build_strategy op_registry op_proto_maker graph)
 
 if (WITH_MKLDNN)
   target_link_libraries(build_strategy mkldnn_placement_pass)

paddle/fluid/framework/details/build_strategy.cc

@@ -20,6 +20,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/ir/multi_devices_graph_pass/multi_devices_graph_pass.h"
 
 DECLARE_bool(use_mkldnn);
+DECLARE_bool(convert_all_blocks);
 
 namespace paddle {
 namespace framework {
@@ -312,6 +313,11 @@ ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
                                 DeviceType use_device) const {
 #endif
   VLOG(1) << "apply all passes";
+  if (FLAGS_convert_all_blocks) {
+    PADDLE_ENFORCE_EQ(
+        graph->IsMainGraph(), true,
+        platform::errors::InvalidArgument("This graph is not main_graph"));
+  }
   // Create a default one if not finalized by user.
   CreatePassesFromStrategy(false);
 
@@ -432,7 +438,14 @@ ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
       }
     }
     VLOG(1) << "Start Apply Pass " << pass->Type();
-    graph = pass->Apply(graph);
+    if (FLAGS_convert_all_blocks) {
+      for (size_t i = 0; i < graph->SubGraphsSize(); ++i) {
+        VLOG(3) << "Apply Pass " << pass->Type() << "to SubGraph " << i;
+        pass->Apply(graph->GetSubGraph(i));
+      }
+    } else {
+      graph = pass->Apply(graph);
+    }
     VLOG(1) << "Finish Apply Pass " << pass->Type();
   }
   VLOG(1) << "All Passes Applied";

paddle/fluid/framework/details/build_strategy_test.cc

@@ -0,0 +1,300 @@
+//   Copyright (c) 2021 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.
+
+#include <algorithm>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#include "gtest/gtest-message.h"
+#include "gtest/gtest-test-part.h"
+#include "gtest/gtest.h"
+#include "gtest/gtest_pred_impl.h"
+
+#include "paddle/fluid/framework/details/build_strategy.h"
+#include "paddle/fluid/framework/op_proto_maker.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/var_type_inference.h"
+#include "paddle/fluid/platform/place.h"
+
+DECLARE_bool(convert_all_blocks);
+
+namespace paddle {
+namespace framework {
+
+class SumOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  void Make() {
+    AddInput("X", "").AsDuplicable();
+    AddOutput("Out", "").AsDuplicable();
+    AddComment("");
+  }
+};
+
+class SumOpWithKernel : public OperatorWithKernel {
+ public:
+  using OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {}
+  OpKernelType GetExpectedKernelType(
+      const ExecutionContext &ctx) const override {
+    return OpKernelType(proto::VarType::FP32, ctx.Input<Tensor>("X")->place());
+  }
+};
+
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_OP_WITHOUT_GRADIENT(sum, paddle::framework::SumOpWithKernel,
+                             paddle::framework::SumOpMaker);
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+static std::vector<platform::Place> CreatePlaces(size_t num, bool use_cuda) {
+  std::vector<platform::Place> result;
+  result.reserve(num);
+  for (size_t i = 0; i < num; ++i) {
+    if (use_cuda) {
+      result.emplace_back(platform::CUDAPlace(i));
+    } else {
+      result.emplace_back(platform::CPUPlace());
+    }
+  }
+  return result;
+}
+
+void BuildStrategyApply(BuildStrategy *build_strategy, ir::Graph *graph) {
+  std::string loss_name = "";
+  Scope scope;
+  std::vector<Scope *> scopes = {&scope};
+
+  auto places = CreatePlaces(1, false);
+  auto device = platform::Place2DeviceType(places[0]);
+
+#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
+  platform::NCCLCommunicator ctxs;
+#elif defined(PADDLE_WITH_XPU) && defined(PADDLE_WITH_XPU_BKCL)
+  platform::BKCLCommunicator ctxs;
+#endif
+
+  build_strategy->Apply(graph, places, loss_name, scopes, 1,
+#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
+                        device, &ctxs);
+#elif defined(PADDLE_WITH_XPU) && defined(PADDLE_WITH_XPU_BKCL)
+                        device, &ctxs);
+#else
+                        device);
+#endif
+}
+
+std::unique_ptr<ir::Graph> CreateGraph() {
+  ProgramDesc prog;
+  auto *op = prog.MutableBlock(0)->AppendOp();
+  op->SetType("sum");
+  op->SetInput("X", {"a1"});
+  op->SetOutput("Out", {"b1"});
+  op->SetAttr("op_role", 1);
+
+  prog.MutableBlock(0)->Var("a1")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("b1")->SetType(proto::VarType::LOD_TENSOR);
+
+  std::unique_ptr<ir::Graph> g(new ir::Graph(prog));
+  return g;
+}
+
+std::unique_ptr<ir::Graph> CreateMultiGraph() {
+  ProgramDesc prog;
+  prog.AppendBlock(prog.Block(0));
+  prog.AppendBlock(prog.Block(0));
+
+  // Set contents in block_0.
+  auto *op = prog.MutableBlock(0)->AppendOp();
+  op->SetType("sum");
+  op->SetInput("X", {"test_a", "test_b", "test_c"});
+  op->SetOutput("Out", {"test_out"});
+  op->SetAttr("op_role", 1);
+
+  prog.MutableBlock(0)->Var("test_a")->SetType(proto::VarType::SELECTED_ROWS);
+  prog.MutableBlock(0)->Var("test_b")->SetType(proto::VarType::SELECTED_ROWS);
+  prog.MutableBlock(0)->Var("test_c")->SetType(proto::VarType::SELECTED_ROWS);
+  prog.MutableBlock(0)->Var("test_out");
+  op->InferVarType(prog.MutableBlock(0));
+
+  prog.MutableBlock(0)->Var("test_b")->SetType(proto::VarType::LOD_TENSOR);
+  op->InferVarType(prog.MutableBlock(0));
+
+  // Set contents in block_1.
+  op = prog.MutableBlock(1)->AppendOp();
+  op->SetType("sum");
+  op->SetInput("X", {"a1"});
+  op->SetOutput("Out", {"b1"});
+  op->SetAttr("op_role", 1);
+
+  prog.MutableBlock(1)->Var("a1")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(1)->Var("b1")->SetType(proto::VarType::LOD_TENSOR);
+
+  // Set contents in block_2.
+  op = prog.MutableBlock(2)->AppendOp();
+  op->SetType("sum");
+  op->SetInput("X", {"a2"});
+  op->SetOutput("Out", {"b2"});
+  op->SetAttr("op_role", 1);
+
+  prog.MutableBlock(2)->Var("a2")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(2)->Var("b2")->SetType(proto::VarType::LOD_TENSOR);
+
+  std::unique_ptr<ir::Graph> g(new ir::Graph(prog));
+  return g;
+}
+
+inline bool CheckSubGraphSame(ir::Graph *g1, ir::Graph *g2) {
+  const auto &g1_nodes_set = g1->Nodes();
+  const auto &g2_nodes_set = g2->Nodes();
+
+  if (g1_nodes_set.size() != g2_nodes_set.size()) return false;
+
+  std::vector<ir::Node *> g1_nodes(g1_nodes_set.begin(), g1_nodes_set.end());
+  std::vector<ir::Node *> g2_nodes(g2_nodes_set.begin(), g2_nodes_set.end());
+
+  auto comp = [](ir::Node *n1, ir::Node *n2) {
+    return n1->Name() > n2->Name();
+  };
+  std::stable_sort(g1_nodes.begin(), g1_nodes.end(), comp);
+  std::stable_sort(g2_nodes.begin(), g2_nodes.end(), comp);
+
+  for (size_t i = 0; i < g1_nodes.size(); ++i) {
+    const auto &n1 = g1_nodes[i];
+    const auto &n2 = g2_nodes[i];
+
+    if (n1->NodeType() != n2->NodeType()) return false;
+    if (n1->Name() != n2->Name()) return false;
+
+    auto n1_inputs = n1->inputs;
+    auto n2_inputs = n2->inputs;
+    if (n1_inputs.size() != n2_inputs.size()) return false;
+
+    std::stable_sort(n1_inputs.begin(), n1_inputs.end(), comp);
+    std::stable_sort(n2_inputs.begin(), n2_inputs.end(), comp);
+    for (size_t i = 0; i < n1_inputs.size(); ++i) {
+      if (n1_inputs[i]->Name() != n2_inputs[i]->Name()) return false;
+    }
+
+    auto n1_outputs = n1->outputs;
+    auto n2_outputs = n2->outputs;
+    if (n1_outputs.size() != n2_outputs.size()) return false;
+
+    std::stable_sort(n1_outputs.begin(), n1_outputs.end(), comp);
+    std::stable_sort(n2_outputs.begin(), n2_outputs.end(), comp);
+    for (size_t i = 0; i < n1_outputs.size(); ++i) {
+      if (n1_outputs[i]->Name() != n2_outputs[i]->Name()) return false;
+    }
+
+    if (n1->IsVar()) {
+      const auto &var1 = n1->Var();
+      const auto &var2 = n2->Var();
+      if ((var1 == nullptr) != (var2 == nullptr)) return false;
+    }
+
+    if (n1->IsOp()) {
+      const auto &op1 = n1->Op();
+      const auto &op2 = n2->Op();
+      if ((op1 == nullptr) != (op2 == nullptr)) return false;
+
+      const auto &op1_input = op1->InputNames();
+      const auto &op2_input = op2->InputNames();
+      if (op1_input.size() != op2_input.size()) return false;
+      if (op1_input != op2_input) return false;
+
+      for (size_t i = 0; i < op1_input.size(); ++i) {
+        if (op1->Input(op1_input[i]) != op2->Input(op2_input[i])) return false;
+      }
+
+      const auto &op1_output = op1->OutputNames();
+      const auto &op2_output = op2->OutputNames();
+      if (op1_output.size() != op2_output.size()) return false;
+      if (op1_output != op2_output) return false;
+
+      for (size_t i = 0; i < op1_output.size(); ++i) {
+        if (op1->Output(op1_output[i]) != op2->Output(op2_output[i]))
+          return false;
+      }
+    }
+  }
+  return true;
+}
+
+inline bool CheckGraphSame(ir::Graph *g1, ir::Graph *g2) {
+  if (g1 == nullptr || g2 == nullptr) return true;
+
+  if (FLAGS_convert_all_blocks) {
+    if (g1->SubGraphsSize() != g2->SubGraphsSize()) return false;
+
+    for (size_t i = 0; i < g1->SubGraphsSize(); ++i) {
+      if (!CheckSubGraphSame(g1->GetSubGraph(i), g2->GetSubGraph(i)))
+        return false;
+    }
+  } else {
+    if (!CheckSubGraphSame(g1, g2)) return false;
+  }
+  return true;
+}
+
+TEST(BuildStrategy, Basic) {
+  BuildStrategy build_strategy;
+
+  ProgramDesc prog;
+  ir::Graph old_graph(prog), graph(prog);
+
+  BuildStrategyApply(&build_strategy, &graph);
+
+  ASSERT_TRUE(CheckGraphSame(&old_graph, &graph));
+}
+
+TEST(BuildStrategy, TestSingleGraph) {
+  BuildStrategy build_strategy;
+  auto graph = CreateGraph();
+  ir::Graph old_graph(graph->OriginProgram());
+
+  BuildStrategyApply(&build_strategy, graph.get());
+
+  // graph should not change for no pass here
+  ASSERT_TRUE(CheckGraphSame(&old_graph, graph.get()));
+}
+
+TEST(BuildStrategy, TestMultiGraph) {
+  // Set FLAGS_convert_all_blocks to true to make sure this test works.
+  bool flag_temp = FLAGS_convert_all_blocks;
+  FLAGS_convert_all_blocks = true;
+
+  BuildStrategy build_strategy;
+  auto graph = CreateMultiGraph();
+  ir::Graph old_graph(graph->OriginProgram());
+
+  BuildStrategyApply(&build_strategy, graph.get());
+
+  // graph should not change for no pass here
+  ASSERT_TRUE(CheckGraphSame(&old_graph, graph.get()));
+
+  // Recover FLAGS_convert_all_blocks.
+  FLAGS_convert_all_blocks = flag_temp;
+}
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/framework.proto

@@ -173,6 +173,8 @@ message VarDesc {
   // True if the variable is an input data and
   // have to check the feed data shape and dtype
   optional bool need_check_feed = 4 [ default = false ];
+  optional bool is_parameter = 5 [ default = false ];
+  optional bool stop_gradient = 6 [ default = false ];
 }
 
 message BlockDesc {