[Inference] rewrite identity_op_clean_pass (#55240)

* rewrite identity_op_clean_pass

* fix

* adjust identity_op_clean_pass order in gpu passes

* fix ut

paddle/fluid/framework/ir/CMakeLists.txt

@@ -434,6 +434,10 @@ cc_test(
   test_delete_assign_op_pass_cc
   SRCS delete_assign_op_pass_test.cc
   DEPS delete_assign_op_pass)
+cc_test(
+  test_identity_op_clean_pass_cc
+  SRCS identity_op_clean_pass_test.cc
+  DEPS identity_op_clean_pass)
 cc_test(
   test_delete_dropout_pass_cc
   SRCS delete_dropout_op_pass_test.cc

paddle/fluid/framework/ir/identity_op_clean_pass.cc

@@ -21,24 +21,33 @@ namespace paddle {
 namespace framework {
 namespace ir {
 
-class Graph;
+namespace patterns {
 
-void IdentityOpCleanPass::ApplyImpl(ir::Graph* graph) const {
-  FusePassBase::Init("identity_scale_op_clean", graph);
-
-  // pre_op -> useless_op_in -> useless_op -> useless_op_out
-  // ->
-  // pre_op -> useless_op_out
-  GraphPatternDetector detector;
-  auto useless_op_in =
-      detector.mutable_pattern()
-          ->NewNode("useless_op_in")
-          ->assert_has_n_outputs(1)
+// pre_op -> useless_op_in -> useless_op -> useless_op_out
+// ->
+// pre_op -> useless_op_out
+struct FindUselessOpPattern : public PatternBase {
+  FindUselessOpPattern(PDPattern* pattern, const std::string& name_scope);
+
+  // declare operator node's name
+  PATTERN_DECL_NODE(useless_op_in);
+  PATTERN_DECL_NODE(useless_op);
+  PATTERN_DECL_NODE(useless_op_out);
+};
+
+FindUselessOpPattern::FindUselessOpPattern(PDPattern* pattern,
+                                           const std::string& name_scope)
+    : PatternBase(pattern, name_scope, name_scope) {
+  auto* useless_op_in = pattern->NewNode(useless_op_in_repr())
+                            ->assert_is_var()
                             ->assert_var_not_persistable()
+                            ->assert_has_n_outputs(1)
                             ->assert_more([](Node* x) {
                               for (auto* op : x->inputs) {
-              auto op_type = op->Op()->Type();
-              if (op_type == "conditional_block" || op_type == "while") {
+                                CHECK_EQ(op->IsOp(), true);
+                                const auto& op_type = op->Op()->Type();
+                                if (op_type == "conditional_block" ||
+                                    op_type == "while" || op_type == "feed") {
                                   return false;
                                 }
                               }
@@ -46,79 +55,71 @@ void IdentityOpCleanPass::ApplyImpl(ir::Graph* graph) const {
                             });
 
   // This useless_op must have only one input and one output!
-  auto useless_op =
-      detector.mutable_pattern()
-          ->NewNode("useless_op")
+  auto* useless_op =
+      pattern->NewNode(useless_op_repr())
+          ->assert_is_op()
           ->assert_has_n_inputs(1)
           ->assert_has_n_outputs(1)
           ->assert_more([](Node* x) {
-            if (!x->IsOp()) {
-              return false;
-            }
-            if (x->Op()->Type() == "scale") {
+            const auto& op_type = x->Op()->Type();
+            if (op_type == "scale") {
               auto scale = x->Op()->GetAttrIfExists<float>("scale");
               auto bias = x->Op()->GetAttrIfExists<float>("bias");
-              if (bias == 0 && scale == 1) {
-                return true;
-              }
-            }
-            if (x->Op()->Type() == "cast") {
+              return bias == 0.f && scale == 1.f;
+            } else if (op_type == "cast") {
               auto in_dtype = x->Op()->GetAttrIfExists<int>("in_dtype");
               auto out_dtype = x->Op()->GetAttrIfExists<int>("out_dtype");
-              if (in_dtype == out_dtype) {
-                return true;
-              }
-            }
-            if (x->Op()->Type() == "c_identity") {
+              return in_dtype == out_dtype;
+            } else if (op_type == "c_identity") {
               return true;
+            } else if (op_type == "assign") {
+              const auto& in_name = x->Op()->Input("X")[0];
+              const auto& out_name = x->Op()->Output("Out")[0];
+              return in_name == out_name;
+            } else if (op_type == "concat") {
+              return x->Op()->Input("X").size() == 1;
             }
             // you can add more cases here.
             return false;
           });
-  auto useless_op_out = detector.mutable_pattern()->NewNode("useless_op_out");
+
+  auto* useless_op_out =
+      pattern->NewNode(useless_op_out_repr())->assert_is_var();
 
   useless_op->LinksFrom({useless_op_in}).LinksTo({useless_op_out});
+}
 
-  int found_subgraph_count = 0;
+}  // namespace patterns
+
+void IdentityOpCleanPass::ApplyImpl(ir::Graph* graph) const {
+  Init(name_scope_, graph);
+
+  GraphPatternDetector gpd;
+  patterns::FindUselessOpPattern pattern(gpd.mutable_pattern(), name_scope_);
+
+  int found_count = 0;
   GraphPatternDetector::handle_t handler =
       [&](const GraphPatternDetector::subgraph_t& subgraph, Graph* graph) {
-        Node* useless_op_var = subgraph.at(useless_op);
-        Node* useless_op_in_var = subgraph.at(useless_op_in);
-        Node* useless_op_out_var = subgraph.at(useless_op_out);
-        const std::string useless_op_in_name = useless_op_in_var->Name();
-        const std::string useless_op_out_name = useless_op_out_var->Name();
-        // Remove links in graph
-        GraphSafeRemoveNodes(graph, {useless_op_in_var, useless_op_var});
-        // Modify pre_op_desc
-        // Link pre_op directly to scale_out
-        for (auto& node : graph->Nodes()) {
-          if (node->IsOp()) {
-            auto* op_desc = node->Op();
-            auto out_vars_map = op_desc->Outputs();
-            for (auto out_var_map : out_vars_map) {
-              auto names = out_var_map.second;
-              bool reset = false;
-              for (size_t i = 0; i < names.size(); i++) {
-                if (names[i] == useless_op_in_name) {
-                  reset = true;
-                  names[i] = useless_op_out_name;
-                  break;
-                }
-              }
-              if (reset) {
-                op_desc->SetOutput(out_var_map.first, names);
-                op_desc->Flush();
-                IR_NODE_LINK_TO(node, useless_op_out_var);
-                break;
-              }
-            }
-          }
+        GET_IR_NODE_FROM_SUBGRAPH(useless_op_in, useless_op_in, pattern);
+        GET_IR_NODE_FROM_SUBGRAPH(useless_op, useless_op, pattern);
+        GET_IR_NODE_FROM_SUBGRAPH(useless_op_out, useless_op_out, pattern);
+        CHECK_EQ(useless_op_in->IsVar(), true);
+        CHECK_EQ(useless_op_out->IsVar(), true);
+        CHECK_EQ(useless_op->IsOp(), true);
+
+        for (auto* prev_op : useless_op_in->inputs) {
+          CHECK_EQ(prev_op->IsOp(), true);
+          prev_op->Op()->RenameOutput(useless_op_in->Var()->Name(),
+                                      useless_op_out->Var()->Name());
+          IR_NODE_LINK_TO(prev_op, useless_op_out);
         }
-        found_subgraph_count++;
+
+        GraphSafeRemoveNodes(graph, {useless_op_in, useless_op});
+        found_count++;
       };
 
-  detector(graph, handler);
-  AddStatis(found_subgraph_count);
+  gpd(graph, handler);
+  AddStatis(found_count);
 }
 
 }  // namespace ir

paddle/fluid/framework/ir/identity_op_clean_pass.h

@@ -27,7 +27,7 @@ class IdentityOpCleanPass : public FusePassBase {
   void ApplyImpl(ir::Graph* graph) const override;
 
  private:
-  virtual ~IdentityOpCleanPass() = default;
+  const std::string name_scope_{"identity_op_clean_pass"};
 };
 
 }  // namespace ir

paddle/fluid/framework/ir/identity_op_clean_pass_test.cc

+// Copyright (c) 2023 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 "paddle/fluid/framework/ir/pass.h"
+#include "paddle/fluid/framework/ir/pass_tester_helper.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+TEST(identity_op_clean_pass, assign) {
+  ProgramDesc program;
+  auto* x_var = program.MutableBlock(0)->Var("assign_x");
+  auto* out_var = program.MutableBlock(0)->Var("assign_out");
+  out_var->SetName(x_var->Name());
+  OpDesc* assign_op = program.MutableBlock(0)->AppendOp();
+  assign_op->SetType("assign");
+  assign_op->SetInput("X", {x_var->Name()});
+  assign_op->SetOutput("Out", {out_var->Name()});
+
+  std::unique_ptr<Graph> graph(new Graph(program));
+  auto pass = PassRegistry::Instance().Get("identity_op_clean_pass");
+  graph.reset(pass->Apply(graph.release()));
+  int assign_num = GetNumOpNodes(graph, "assign");
+  PADDLE_ENFORCE_EQ(
+      assign_num,
+      0,
+      platform::errors::PreconditionNotMet(
+          "graph should have 0 assign after identity_op_clean_pass, "
+          "but actually has %d.",
+          assign_num));
+}
+
+TEST(identity_op_clean_pass, scale) {
+  ProgramDesc program;
+  auto* x_var = program.MutableBlock(0)->Var("scale_x");
+  auto* out_var = program.MutableBlock(0)->Var("scale_out");
+  OpDesc* scale_op = program.MutableBlock(0)->AppendOp();
+  scale_op->SetType("scale");
+  scale_op->SetInput("X", {x_var->Name()});
+  scale_op->SetOutput("Out", {out_var->Name()});
+  scale_op->SetAttr("scale", 1.f);
+  scale_op->SetAttr("bias", 0.f);
+
+  std::unique_ptr<Graph> graph(new Graph(program));
+  auto pass = PassRegistry::Instance().Get("identity_op_clean_pass");
+  graph.reset(pass->Apply(graph.release()));
+  int scale_num = GetNumOpNodes(graph, "scale");
+  PADDLE_ENFORCE_EQ(
+      scale_num,
+      0,
+      platform::errors::PreconditionNotMet(
+          "graph should have 0 scale op after identity_op_clean_pass, "
+          "but actually has %d.",
+          scale_num));
+}
+
+TEST(identity_op_clean_pass, cast) {
+  ProgramDesc program;
+  auto* x_var = program.MutableBlock(0)->Var("cast_x");
+  auto* out_var = program.MutableBlock(0)->Var("cast_out");
+  OpDesc* cast_op = program.MutableBlock(0)->AppendOp();
+  cast_op->SetType("cast");
+  cast_op->SetInput("X", {x_var->Name()});
+  cast_op->SetOutput("Out", {out_var->Name()});
+  cast_op->SetAttr("in_dtype", 5);
+  cast_op->SetAttr("out_dtype", 5);
+
+  std::unique_ptr<Graph> graph(new Graph(program));
+  auto pass = PassRegistry::Instance().Get("identity_op_clean_pass");
+  graph.reset(pass->Apply(graph.release()));
+  int cast_num = GetNumOpNodes(graph, "cast");
+  PADDLE_ENFORCE_EQ(
+      cast_num,
+      0,
+      platform::errors::PreconditionNotMet(
+          "graph should have 0 cast after identity_op_clean_pass, "
+          "but actually has %d.",
+          cast_num));
+}
+
+TEST(identity_op_clean_pass, concat) {
+  ProgramDesc program;
+  auto* x_var = program.MutableBlock(0)->Var("concat_x");
+  auto* out_var = program.MutableBlock(0)->Var("concat_out");
+  OpDesc* concat_op = program.MutableBlock(0)->AppendOp();
+  concat_op->SetType("concat");
+  concat_op->SetInput("X", {x_var->Name()});
+  concat_op->SetOutput("Out", {out_var->Name()});
+
+  std::unique_ptr<Graph> graph(new Graph(program));
+  auto pass = PassRegistry::Instance().Get("identity_op_clean_pass");
+  graph.reset(pass->Apply(graph.release()));
+  int concat_num = GetNumOpNodes(graph, "concat");
+  PADDLE_ENFORCE_EQ(
+      concat_num,
+      0,
+      platform::errors::PreconditionNotMet(
+          "graph should have 0 concat after identity_op_clean_pass, "
+          "but actually has %d.",
+          concat_num));
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+USE_PASS(identity_op_clean_pass);

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -221,7 +221,6 @@ const std::vector<std::string> kCINNCompilerPasses{
 GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
   passes_.assign({
     "map_op_to_another_pass",                                           //
-        "identity_op_clean_pass",                                       //
         "is_test_pass",                                                 //
         "simplify_with_basic_ops_pass",                                 //
         "delete_quant_dequant_linear_op_pass",                          //
@@ -262,6 +261,7 @@ GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
         "conv_elementwise_add_fuse_pass",      //
 #endif                                         //
         "transpose_flatten_concat_fuse_pass",  //
+        "identity_op_clean_pass",              //
         "conv2d_fusion_layout_transfer_pass",  //
         "transfer_layout_elim_pass",
         "auto_mixed_precision_pass",  //

test/ir/inference/test_identity_clean_pass.py

@@ -20,6 +20,54 @@ from program_config import OpConfig, ProgramConfig, TensorConfig
 
 
 class TestIdentityScaleCleanPass(PassAutoScanTest):
+    def sample_predictor_configs(self, program_config):
+        config = self.create_inference_config(use_gpu=True)
+        yield config, ['relu', 'relu', 'scale'], (1e-5, 1e-5)
+
+    def sample_program_config(self, draw):
+        bias_after_scale = draw(st.booleans())
+        n = draw(st.integers(min_value=1, max_value=4))
+        c = draw(st.integers(min_value=1, max_value=20))
+        h = draw(st.integers(min_value=1, max_value=20))
+        w = draw(st.integers(min_value=1, max_value=20))
+
+        relu_op1 = OpConfig(
+            "relu", inputs={"X": ["relu_x"]}, outputs={"Out": ["relu_op1_out"]}
+        )
+        scale_op1 = OpConfig(
+            "scale",
+            inputs={"X": ["relu_op1_out"]},
+            outputs={"Out": ["scale_op1_out"]},
+            bias=0.0,
+            scale=1.0,
+            bias_after_scale=True,
+        )
+        scale_op2 = OpConfig(
+            "scale",
+            inputs={"X": ["scale_op1_out"]},
+            outputs={"Out": ["scale_op2_out"]},
+            bias=0.0,
+            scale=1.0,
+            bias_after_scale=True,
+        )
+        relu_op2 = OpConfig(
+            "relu",
+            inputs={"X": ["relu_op1_out"]},
+            outputs={"Out": ["relu_op2_out"]},
+        )
+        program_config = ProgramConfig(
+            ops=[relu_op1, relu_op2, scale_op1, scale_op2],
+            weights={},
+            inputs={"relu_x": TensorConfig(shape=[n, c, h, w])},
+            outputs=["scale_op2_out", "relu_op2_out"],
+        )
+        return program_config
+
+    def test(self):
+        self.run_and_statis(max_examples=25, passes=["identity_op_clean_pass"])
+
+
+class TestIdentityScaleCleanPass_V1(PassAutoScanTest):
     def sample_predictor_configs(self, program_config):
         config = self.create_inference_config(use_gpu=True)
         yield config, ['relu'], (1e-5, 1e-5)
@@ -31,22 +79,75 @@ class TestIdentityScaleCleanPass(PassAutoScanTest):
         h = draw(st.integers(min_value=1, max_value=20))
         w = draw(st.integers(min_value=1, max_value=20))
 
-        relu_op = OpConfig(
-            "relu", inputs={"X": ["relu_x"]}, outputs={"Out": ["relu_out"]}
+        relu_op1 = OpConfig(
+            "relu", inputs={"X": ["relu_x"]}, outputs={"Out": ["relu_op1_out"]}
         )
-        scale_op = OpConfig(
+        scale_op1 = OpConfig(
             "scale",
-            inputs={"X": ["relu_out"]},
-            outputs={"Out": ["scale_out"]},
+            inputs={"X": ["relu_op1_out"]},
+            outputs={"Out": ["scale_op1_out"]},
+            bias=0.0,
+            scale=1.0,
+            bias_after_scale=True,
+        )
+        scale_op2 = OpConfig(
+            "scale",
+            inputs={"X": ["scale_op1_out"]},
+            outputs={"Out": ["scale_op2_out"]},
             bias=0.0,
             scale=1.0,
             bias_after_scale=True,
         )
         program_config = ProgramConfig(
-            ops=[relu_op, scale_op],
+            ops=[relu_op1, scale_op1, scale_op2],
             weights={},
             inputs={"relu_x": TensorConfig(shape=[n, c, h, w])},
-            outputs=["scale_out"],
+            outputs=["scale_op2_out"],
+        )
+        return program_config
+
+    def test(self):
+        self.run_and_statis(max_examples=25, passes=["identity_op_clean_pass"])
+
+
+class TestIdentityScaleCleanPass_V2(PassAutoScanTest):
+    def sample_predictor_configs(self, program_config):
+        config = self.create_inference_config(use_gpu=True)
+        yield config, ['scale', 'relu'], (1e-5, 1e-5)
+
+    def sample_program_config(self, draw):
+        bias_after_scale = draw(st.booleans())
+        n = draw(st.integers(min_value=1, max_value=4))
+        c = draw(st.integers(min_value=1, max_value=20))
+        h = draw(st.integers(min_value=1, max_value=20))
+        w = draw(st.integers(min_value=1, max_value=20))
+
+        scale_op1 = OpConfig(
+            "scale",
+            inputs={"X": ["scale_op1_in"]},
+            outputs={"Out": ["scale_op1_out"]},
+            bias=0.0,
+            scale=1.0,
+            bias_after_scale=True,
+        )
+        scale_op2 = OpConfig(
+            "scale",
+            inputs={"X": ["scale_op1_out"]},
+            outputs={"Out": ["scale_op2_out"]},
+            bias=0.0,
+            scale=1.0,
+            bias_after_scale=True,
+        )
+        relu_op1 = OpConfig(
+            "relu",
+            inputs={"X": ["scale_op2_out"]},
+            outputs={"Out": ["relu_op1_out"]},
+        )
+        program_config = ProgramConfig(
+            ops=[scale_op1, scale_op2, relu_op1],
+            weights={},
+            inputs={"scale_op1_in": TensorConfig(shape=[n, c, h, w])},
+            outputs=["relu_op1_out"],
         )
         return program_config