add common subexpression elimination (#44386)

paddle/fluid/framework/ir/CMakeLists.txt

@@ -115,6 +115,7 @@ pass_library(gpu_cpu_map_matmul_to_mul_pass inference)
 pass_library(dense_fc_to_sparse_pass inference)
 pass_library(dense_multihead_matmul_to_sparse_pass inference)
 pass_library(generate_pass DEPS pass_desc_proto)
+pass_library(common_subexpression_elimination_pass inference)
 target_link_libraries(generate_pass pass_desc_proto)
 
 if(WITH_TENSORRT)
@@ -326,6 +327,10 @@ cc_test(
   test_generate_pass_cc
   SRCS generate_pass_tester.cc
   DEPS generate_pass pass_desc_proto)
+cc_test(
+  test_common_subexpression_elimination_pass_cc
+  SRCS common_subexpression_elimination_pass_tester.cc
+  DEPS common_subexpression_elimination_pass)
 cc_test(
   test_delete_dropout_pass_cc
   SRCS delete_dropout_op_pass_test.cc

paddle/fluid/framework/ir/common_subexpression_elimination_pass.cc

+// 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.
+
+#include "paddle/fluid/framework/ir/common_subexpression_elimination_pass.h"
+#include <string>
+#include <type_traits>
+
+#include "paddle/fluid/framework/framework.pb.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+#include "paddle/fluid/framework/ir/node.h"
+#include "paddle/fluid/framework/op_version_registry.h"
+#include "paddle/fluid/framework/type_defs.h"
+#include "paddle/phi/core/enforce.h"
+#include "paddle/utils/variant.h"
+
+namespace {
+
+std::string NodeTypeToString(paddle::framework::ir::Node::Type type) {
+  if (type == paddle::framework::ir::Node::Type::kOperation) {
+    return "kOperation";
+  } else {
+    return "kVariable";
+  }
+}
+
+const std::unordered_set<std::string> commutative_operators{"mul",
+                                                            "bitwise_and",
+                                                            "bitwise_or",
+                                                            "equal_all",
+                                                            "equal",
+                                                            "not_equal",
+                                                            "logical_and",
+                                                            "logical_or",
+                                                            "elementwise_max",
+                                                            "elementwise_fmax",
+                                                            "elementwise_min",
+                                                            "elementwise_fmin",
+                                                            "elementwise_mul",
+                                                            "elementwise_add",
+                                                            "add_p",
+                                                            "max_p",
+                                                            "mul_p",
+                                                            "eq_p",
+                                                            "ne_p"};
+
+const std::unordered_set<std::string> nondeterministic_operators{
+    "dropout",
+    "dropout_nd",
+    "gaussian_random_batch_size_like",
+    "gaussian_random",
+    "randint",
+    "random_crop",
+    "random_routing",
+    "randperm",
+    "uniform_random_batch_size_like",
+    "uniform_random_inplace",
+    "uniform_random",
+    "fused_bias_dropout_residual_layer_norm"};
+
+const std::unordered_set<std::string> side_effect_operators{
+    "feed", "cast", "fetch", "fill_constant", "fill_constant_batch_size_like"};
+
+template <class T>
+inline void HashCombine(std::size_t *seed, const T &v) {
+  std::hash<T> hasher;
+  (*seed) ^= hasher(v) + 0x9e3779b9 + ((*seed) << 6) + ((*seed) >> 2);
+}
+
+}  // namespace
+
+namespace std {
+
+#define HASH_ATTRIBUTE(attr, id, type)         \
+  do {                                         \
+    if (attr.index() == id) {                  \
+      return std::hash<type>{}(get<id>(attr)); \
+    }                                          \
+  } while (0)
+
+#define HASH_VECTOR_ATTRIBUTE(attr, id, type) \
+  do {                                        \
+    if (attr.index() == id) {                 \
+      std::vector<type> vec = get<id>(attr);  \
+      size_t seed = 0;                        \
+      for (const auto &v : vec) {             \
+        HashCombine(&seed, v);                \
+      }                                       \
+      return seed;                            \
+    }                                         \
+  } while (0)
+
+template <>
+struct hash<paddle::framework::proto::VarType_Type> {
+  size_t operator()(const paddle::framework::proto::VarType_Type &attr) const {
+    using type = typename std::underlying_type<
+        paddle::framework::proto::VarType_Type>::type;
+    return std::hash<type>()(static_cast<type>(attr));
+  }
+};
+
+template <>
+struct hash<paddle::framework::Attribute> {
+  size_t operator()(const paddle::framework::Attribute &attr) const {
+    if (attr.index() == 0) {
+      return 0;
+    }
+    if (attr.index() == 7) {
+      return static_cast<size_t>(get<7>(attr));
+    }
+
+    HASH_ATTRIBUTE(attr, 1, int);
+    HASH_ATTRIBUTE(attr, 2, float);
+    HASH_ATTRIBUTE(attr, 3, std::string);
+    HASH_VECTOR_ATTRIBUTE(attr, 4, int);
+    HASH_VECTOR_ATTRIBUTE(attr, 5, float);
+    HASH_VECTOR_ATTRIBUTE(attr, 6, std::string);
+    HASH_ATTRIBUTE(attr, 8, std::vector<bool>);
+    HASH_ATTRIBUTE(attr, 9, paddle::framework::BlockDesc *);
+    HASH_ATTRIBUTE(attr, 10, int64_t);
+    HASH_VECTOR_ATTRIBUTE(attr, 11, paddle::framework::BlockDesc *);
+    HASH_VECTOR_ATTRIBUTE(attr, 12, int64_t);
+    HASH_VECTOR_ATTRIBUTE(attr, 13, double);
+    return 0;
+  }
+};
+}  // namespace std
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+void CommonSubexpressionEliminationPass::ApplyImpl(ir::Graph *graph) const {
+  PADDLE_ENFORCE_EQ(
+      graph->IsMainGraph(),
+      true,
+      platform::errors::InvalidArgument(
+          "CommonSubexpressionEliminationPass only accepts main graph"));
+
+  CommonSubexpressionEliminate(
+      graph, graph, [](Node *) -> Node * { return nullptr; });
+}
+
+void CommonSubexpressionEliminationPass::CommonSubexpressionEliminate(
+    ir::Graph *main_graph,
+    ir::Graph *graph,
+    std::function<Node *(Node *)> parent_exist_nodes) const {
+  const char *kSubBlock = "sub_block";
+  std::unordered_set<ir::Node *, HashOpNode, EqualOpNode> exist_nodes;
+  std::vector<Node *> nodes = TopologySortOperations(*graph);
+  for (Node *node : nodes) {
+    if (node->inputs.empty()) {
+      continue;
+    }
+    if (side_effect_operators.count(node->Name()) != 0) {
+      continue;
+    }
+    if (nondeterministic_operators.count(node->Name()) != 0) {
+      continue;
+    }
+
+    if (node->Op()->HasAttr(kSubBlock)) {
+      auto sub_block_id =
+          node->Op()->GetAttrIfExists<BlockDesc *>(kSubBlock)->ID();
+      CommonSubexpressionEliminate(
+          main_graph,
+          main_graph->GetSubGraph(sub_block_id),
+          [&exist_nodes, &parent_exist_nodes](Node *node) -> Node * {
+            auto exist_node = exist_nodes.find(node);
+            if (exist_node != exist_nodes.end()) {
+              return *exist_node;
+            }
+            return parent_exist_nodes(node);
+          });
+      continue;
+    }
+
+    Node *exist_node = parent_exist_nodes(node);
+    if (exist_node == nullptr) {
+      auto res = exist_nodes.insert(node);
+      if (!res.second) {
+        exist_node = *res.first;
+      }
+    }
+
+    if (exist_node != nullptr) {
+      for (size_t i = 0; i < exist_node->outputs.size(); ++i) {
+        Node *exist_node_output = exist_node->outputs[i];
+        Node *current_node_output = node->outputs[i];
+        std::vector<Node *> current_node_output_outputs =
+            current_node_output->outputs;
+        for (size_t i = 0; i < current_node_output_outputs.size(); ++i) {
+          IR_NODE_LINK_TO(exist_node_output, current_node_output_outputs[i]);
+        }
+      }
+      GraphSafeRemoveNodes(graph,
+                           std::unordered_set<const Node *>(
+                               node->outputs.begin(), node->outputs.end()));
+      GraphSafeRemoveNodes(graph, {node});
+    }
+  }
+}
+
+size_t HashOpNode::operator()(const Node *node) const {
+  PADDLE_ENFORCE_EQ(node->IsOp(),
+                    true,
+                    platform::errors::InvalidArgument(
+                        "HashOpNode only supports operation node type"));
+
+  size_t seed = 0;
+  std::vector<Node *> inputs(node->inputs);
+  if (commutative_operators.count(node->Name()) != 0) {
+    auto comparator = [](Node *a, Node *b) { return a->Name() > b->Name(); };
+    std::stable_sort(inputs.begin(), inputs.end(), comparator);
+  }
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    HashCombine(&seed, inputs[i]->id());
+    HashCombine(&seed, node->GraphId());
+  }
+  const std::string kDepVarName = std::string(Node::kControlDepVarName);
+  for (size_t i = 0; i < node->outputs.size(); ++i) {
+    if (node->outputs[i] == nullptr) {
+      continue;
+    }
+    if (node->outputs[i]->IsCtrlVar()) {
+      HashCombine(&seed, kDepVarName);
+    } else if (node->outputs[i]->IsVar()) {
+      HashCombine(&seed, node->outputs[i]->Var()->GetType());
+    }
+  }
+  OpDesc *desc = node->Op();
+  std::vector<std::string> attributes = desc->AttrNames();
+  sort(attributes.begin(), attributes.end());
+  for (const std::string &attribute : attributes) {
+    HashCombine(&seed, desc->GetAttr(attribute));
+  }
+  return seed;
+}
+
+bool EqualOpNode::operator()(const Node *lhs, const Node *rhs) const {
+  PADDLE_ENFORCE_EQ(lhs->IsOp() && rhs->IsOp(),
+                    true,
+                    platform::errors::InvalidArgument(
+                        "EqualOpNode only supports operation node type"));
+
+  if (lhs == nullptr && rhs == nullptr) {
+    return true;
+  }
+  if (lhs == nullptr || rhs == nullptr) {
+    return false;
+  }
+  if (lhs->NodeType() != rhs->NodeType()) {
+    return false;
+  }
+  if (lhs->Name() != rhs->Name()) {
+    return false;
+  }
+
+  std::vector<Node *> lhs_inputs(lhs->inputs);
+  std::vector<Node *> rhs_inputs(rhs->inputs);
+  if (commutative_operators.count(lhs->Name()) != 0) {
+    auto comparator = [](Node *a, Node *b) { return a->Name() > b->Name(); };
+    std::stable_sort(lhs_inputs.begin(), lhs_inputs.end(), comparator);
+    std::stable_sort(rhs_inputs.begin(), rhs_inputs.end(), comparator);
+  }
+
+  // compare inputs value
+  if (lhs_inputs.size() != rhs_inputs.size()) {
+    return false;
+  }
+  if (!std::equal(lhs_inputs.begin(), lhs_inputs.end(), rhs_inputs.begin())) {
+    return false;
+  }
+
+  // compare attribute
+  const OpDesc *lhs_desc = lhs->Op();
+  const OpDesc *rhs_desc = rhs->Op();
+  std::vector<std::string> lhs_attr_names = lhs_desc->AttrNames();
+  std::vector<std::string> rhs_attr_names = rhs_desc->AttrNames();
+  if (lhs_attr_names.size() != rhs_attr_names.size()) {
+    return false;
+  }
+  std::sort(lhs_attr_names.begin(), lhs_attr_names.end());
+  std::sort(rhs_attr_names.begin(), rhs_attr_names.end());
+  for (size_t i = 0; i < lhs_attr_names.size(); ++i) {
+    if (lhs_attr_names[i] != rhs_attr_names[i]) {
+      return false;
+    }
+    if (lhs_desc->GetAttr(lhs_attr_names[i]) !=
+        rhs_desc->GetAttr(rhs_attr_names[i])) {
+      return false;
+    }
+  }
+
+  // compare outputs value type
+  std::vector<Node *> lhs_outputs(lhs->outputs);
+  std::vector<Node *> rhs_outputs(rhs->outputs);
+  if (lhs_outputs.size() != rhs_outputs.size()) {
+    return false;
+  }
+  for (size_t i = 0; i < lhs_outputs.size(); ++i) {
+    if (!lhs_outputs[i]->IsVar() || !rhs_outputs[i]->IsVar()) {
+      return false;
+    }
+    if (lhs_outputs[i]->IsCtrlVar() != rhs_outputs[i]->IsCtrlVar()) {
+      return false;
+    }
+    if (lhs_outputs[i]->IsCtrlVar() && rhs_outputs[i]->IsCtrlVar()) {
+      continue;
+    }
+    if (lhs_outputs[i]->Var()->GetType() != rhs_outputs[i]->Var()->GetType()) {
+      return false;
+    }
+  }
+  return true;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(common_subexpression_elimination_pass,
+              paddle::framework::ir::CommonSubexpressionEliminationPass);
+REGISTER_PASS_CAPABILITY(common_subexpression_elimination_pass);

paddle/fluid/framework/ir/common_subexpression_elimination_pass.h

+// 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 <iostream>
+
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class Graph;
+
+class CommonSubexpressionEliminationPass : public FusePassBase {
+ public:
+  CommonSubexpressionEliminationPass() {}
+
+ protected:
+  void ApplyImpl(ir::Graph* graph) const override;
+
+ private:
+  void CommonSubexpressionEliminate(
+      ir::Graph* main_graph,
+      ir::Graph* graph,
+      std::function<Node*(Node*)> parent_exist_nodes) const;
+};
+
+struct HashOpNode {
+  size_t operator()(const Node* node) const;
+};
+
+struct EqualOpNode {
+  bool operator()(const Node* lhs, const Node* rhs) const;
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/common_subexpression_elimination_pass_tester.cc

+/* 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. */
+
+#include <gtest/gtest.h>
+
+#include <functional>
+#include <sstream>
+#include <string>
+#include <unordered_set>
+
+#include "paddle/fluid/framework/ir/common_subexpression_elimination_pass.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+#include "paddle/fluid/framework/ir/pass_tester_helper.h"
+#include "paddle/fluid/framework/op_version_registry.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+TEST(CommonSubexpressionEliminationPass, basic_test) {
+  // inputs                           operator            output
+  // --------------------------------------------------------------------
+  // (relu(a), b)                  elementwise_add ->      d
+  // (relu(a), c)                  elementwise_add ->      e
+  // (d, e)                        elementwise_add ->      f
+
+  Layers layers;
+  auto* a = layers.data("a", {1024, 768});
+  auto* b = layers.data("b", {1024, 768});
+  auto* c = layers.data("c", {1024, 768});
+  auto* d = layers.elementwise_add(layers.relu(a), b);
+  auto* e = layers.elementwise_add(layers.relu(a), c);
+  auto* f = layers.data("f", {1024, 768});
+  layers.elementwise_add(d, e, f, 0);
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(layers.main_program()));
+  auto pass =
+      PassRegistry::Instance().Get("common_subexpression_elimination_pass");
+  graph.reset(pass->Apply(graph.release()));
+  int num_nodes_after = GetNumOpNodes(graph, "relu");
+  PADDLE_ENFORCE_EQ(num_nodes_after,
+                    1,
+                    platform::errors::InvalidArgument(
+                        "Before the common subexpression elimination pass, "
+                        "there should be 1 "
+                        "relu op, but the result is %d",
+                        num_nodes_after));
+}
+
+TEST(CommonSubexpressionEliminationPass, commutative_operator_test) {
+  // inputs                           operator            output
+  // --------------------------------------------------------------------
+  // (a, b)                        elementwise_add ->      e
+  // (b, a)                        elementwise_add ->      f
+  // (e, c)                        elementwise_add ->      g
+  // (f, d)                        elementwise_add ->      h
+
+  Layers layers;
+  auto* a = layers.data("a", {1024, 768});
+  auto* b = layers.data("b", {1024, 768});
+  auto* c = layers.data("c", {1024, 768});
+  auto* d = layers.data("d", {1024, 768});
+  auto* e = layers.data("e", {1024, 768});
+  auto* f = layers.data("f", {1024, 768});
+  auto* g = layers.data("g", {1024, 768});
+  auto* h = layers.data("h", {1024, 768});
+
+  layers.elementwise_add(a, b, e, 0);
+  layers.elementwise_add(b, a, f, 0);
+  layers.elementwise_add(e, c, g, 0);
+  layers.elementwise_add(f, d, h, 0);
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(layers.main_program()));
+  auto pass =
+      PassRegistry::Instance().Get("common_subexpression_elimination_pass");
+  graph.reset(pass->Apply(graph.release()));
+  int num_nodes_after = GetNumOpNodes(graph, "elementwise_add");
+  PADDLE_ENFORCE_EQ(num_nodes_after,
+                    3,
+                    platform::errors::InvalidArgument(
+                        "Before the common subexpression elimination pass, "
+                        "there should be 3 "
+                        "elementwise_add op, but the result is %d",
+                        num_nodes_after));
+}
+
+TEST(CommonSubexpressionEliminationPass, nondeterministic_operator_test) {
+  // inputs                           operator            output
+  // --------------------------------------------------------------------
+  // (dropout(a), b)               elementwise_add ->      d
+  // (dropout(a), c)               elementwise_add ->      e
+  // (d, e)                        elementwise_add ->      f
+
+  Layers layers;
+  auto* a = layers.data("a", {1024, 768});
+  auto* b = layers.data("b", {1024, 768});
+  auto* c = layers.data("c", {1024, 768});
+  auto* d =
+      layers.elementwise_add(layers.dropout(a, 0.5, "downgrade_in_infer"), b);
+  auto* e =
+      layers.elementwise_add(layers.dropout(a, 0.5, "downgrade_in_infer"), c);
+  auto* f = layers.data("f", {1024, 768});
+  layers.elementwise_add(d, e, f, 0);
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(layers.main_program()));
+  auto pass =
+      PassRegistry::Instance().Get("common_subexpression_elimination_pass");
+  graph.reset(pass->Apply(graph.release()));
+  int num_nodes_after = GetNumOpNodes(graph, "dropout");
+  PADDLE_ENFORCE_EQ(num_nodes_after,
+                    2,
+                    platform::errors::InvalidArgument(
+                        "After the common subexpression elimination pass, "
+                        "there should still be 2 "
+                        "dropout op, but the result is %d",
+                        num_nodes_after));
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+USE_PASS(common_subexpression_elimination_pass);

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -282,6 +282,7 @@ CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
                   "conv_eltwiseadd_bn_fuse_pass",            //
                   "conv_transpose_bn_fuse_pass",             //
                   "conv_transpose_eltwiseadd_bn_fuse_pass",  //
+                  "common_subexpression_elimination_pass",   //
                   "is_test_pass",                            //
                   "constant_folding_pass",
                   // following pass should be located in the last, since