Merge branch 'develop' of https://github.com/PaddlePaddle/paddle into add-reshape-reuse-input

test=develop

cmake/generic.cmake

@@ -261,6 +261,13 @@ function(cc_library TARGET_NAME)
         add_dependencies(${TARGET_NAME} mklml)
         target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
       endif()
+      # remove link to python, see notes at:
+      # https://github.com/pybind/pybind11/blob/master/docs/compiling.rst#building-manually
+      if("${cc_library_DEPS};" MATCHES "python;")
+        list(REMOVE_ITEM cc_library_DEPS python)
+        add_dependencies(${TARGET_NAME} python)
+        target_link_libraries(${TARGET_NAME} "-Wl,-undefined,dynamic_lookup")
+      endif()
       target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
       add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
     endif()

paddle/fluid/API.spec

@@ -116,6 +116,7 @@ paddle.fluid.layers.pad ArgSpec(args=['x', 'paddings', 'pad_value', 'name'], var
 paddle.fluid.layers.pad_constant_like ArgSpec(args=['x', 'y', 'pad_value', 'name'], varargs=None, keywords=None, defaults=(0.0, None))
 paddle.fluid.layers.label_smooth ArgSpec(args=['label', 'prior_dist', 'epsilon', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, 0.1, 'float32', None))
 paddle.fluid.layers.roi_pool ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0))
+paddle.fluid.layers.roi_align ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale', 'sampling_ratio', 'name'], varargs=None, keywords=None, defaults=(1, 1, 1.0, -1, None))
 paddle.fluid.layers.dice_loss ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,))
 paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR'))
 paddle.fluid.layers.image_resize_short ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',))

paddle/fluid/framework/details/var_handle.h

@@ -49,6 +49,8 @@ struct VarHandleBase {
 
   void AddOutput(OpHandleBase* out, ir::Node* node) {
     if (pending_ops_.find(out) == pending_ops_.end()) {
+      PADDLE_ENFORCE(out != nullptr, "The output of %s should not be nullptr",
+                     this->Node()->Name());
       pending_ops_.insert(out);
       node_->outputs.push_back(node);
     }

paddle/fluid/framework/ir/CMakeLists.txt

@@ -37,12 +37,17 @@ pass_library(embedding_fc_lstm_fuse_pass inference)
 pass_library(fc_gru_fuse_pass inference)
 pass_library(seq_concat_fc_fuse_pass inference)
 pass_library(conv_bn_fuse_pass inference)
+pass_library(seqconv_eltadd_relu_fuse_pass inference)
 if(WITH_MKLDNN)
     pass_library(mkldnn_placement_pass base)
+    pass_library(conv_bias_mkldnn_fuse_pass inference)
     pass_library(conv_relu_mkldnn_fuse_pass inference)
 endif()
 
 cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector )
+if(WITH_MKLDNN)
+  pass_library(conv_elementwise_add_mkldnn_fuse_pass inference)
+endif()
 
 set(GLOB_PASS_LIB ${PASS_LIBRARY} CACHE INTERNAL "Global PASS library")
 
@@ -56,4 +61,5 @@ cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS g
 cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
 if (WITH_MKLDNN)
     cc_test(test_conv_relu_mkldnn_fuse_pass SRCS conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass)
+    cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass)
 endif ()

paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc

+// Copyright (c) 2018 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/conv_bias_mkldnn_fuse_pass.h"
+#include <functional>
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+template <typename BinaryOperation>
+LoDTensor tensor_apply_eltwise(const LoDTensor& vec_a, const LoDTensor& vec_b,
+                               BinaryOperation f) {
+  PADDLE_ENFORCE_EQ(vec_a.dims(), vec_b.dims());
+  LoDTensor vec_y;
+  vec_y.Resize(vec_a.dims());
+  const float* a = vec_a.data<float>();
+  const float* b = vec_b.data<float>();
+  float* y = vec_y.mutable_data<float>(platform::CPUPlace());
+  for (int i = 0; i < vec_a.numel(); i++) {
+    y[i] = f(a[i], b[i]);
+  }
+  return vec_y;
+}
+
+std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  PADDLE_ENFORCE(graph.get());
+  FusePassBase::Init(name_scope_, graph.get());
+
+  auto* scope = param_scope();
+  PADDLE_ENFORCE(scope);
+
+  GraphPatternDetector gpd;
+  auto* conv_input =
+      gpd.mutable_pattern()
+          ->NewNode(patterns::PDNodeName(name_scope_, "conv_input"))
+          ->AsInput()
+          ->assert_is_op_input("conv2d", "Input");
+  patterns::ConvBias conv_bias_pattern(gpd.mutable_pattern(), name_scope_);
+  conv_bias_pattern(conv_input);
+  int found_conv_bias_count = 0;
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    VLOG(4) << "handle ConvBias fuse";
+    GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight,
+                              conv_bias_pattern);                      // Filter
+    GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_bias_pattern);  // tmp
+    GET_IR_NODE_FROM_SUBGRAPH(conv, conv, conv_bias_pattern);  // CONV op
+    // bias
+    GET_IR_NODE_FROM_SUBGRAPH(eltwise_bias, eltwise_bias, conv_bias_pattern);
+    // output
+    GET_IR_NODE_FROM_SUBGRAPH(eltwise_out, eltwise_out, conv_bias_pattern);
+    // elementwise_add op
+    GET_IR_NODE_FROM_SUBGRAPH(eltwise, eltwise, conv_bias_pattern);
+
+    PADDLE_ENFORCE(subgraph.count(conv_input));
+
+    // check if fuse can be done and if MKL-DNN should be used
+    FuseOptions fuse_option = FindFuseOption(*conv, *eltwise);
+    if (fuse_option == DO_NOT_FUSE || fuse_option == FUSE_NATIVE) {
+      VLOG(3) << "do not perform conv+bias fuse";
+      return;
+    }
+
+    auto* eltwise_bias_tensor =
+        scope->FindVar(eltwise_bias->Name())->GetMutable<LoDTensor>();
+
+    auto input_names = conv->Op()->InputNames();
+    bool has_bias = std::find(input_names.begin(), input_names.end(), "Bias") !=
+                    input_names.end();
+    if (has_bias && conv->Op()->Input("Bias").size() > 0) {
+      auto conv_bias_names = conv->Op()->Input("Bias");
+      // add eltwise bias to existing conv bias
+      PADDLE_ENFORCE_EQ(conv_bias_names.size(), 1);
+      auto* conv_bias_var = scope->FindVar(conv_bias_names[0]);
+      auto* conv_bias_tensor = conv_bias_var->GetMutable<LoDTensor>();
+      PADDLE_ENFORCE_EQ(conv_bias_tensor->dims(), eltwise_bias_tensor->dims());
+      *conv_bias_tensor = tensor_apply_eltwise(
+          *conv_bias_tensor, *eltwise_bias_tensor, std::plus<float>());
+
+      conv->Op()->SetOutput("Output",
+                            std::vector<std::string>({eltwise_out->Name()}));
+
+      GraphSafeRemoveNodes(graph.get(), {eltwise, conv_out});
+
+      IR_NODE_LINK_TO(conv, eltwise_out);
+    } else {
+      // take eltwise bias as conv bias
+      OpDesc desc;
+
+      desc.SetInput(
+          "Input", std::vector<std::string>({subgraph.at(conv_input)->Name()}));
+      desc.SetInput("Filter", std::vector<std::string>({conv_weight->Name()}));
+      desc.SetInput("Bias", std::vector<std::string>({eltwise_bias->Name()}));
+      desc.SetOutput("Output", std::vector<std::string>({eltwise_out->Name()}));
+      desc.SetType("conv2d");
+
+      for (auto& attr : conv->Op()->GetAttrMap()) {
+        desc.SetAttr(attr.first, attr.second);
+      }
+      auto conv_bias_node = g->CreateOpNode(&desc);
+
+      IR_NODE_LINK_TO(subgraph.at(conv_input), conv_bias_node);
+      IR_NODE_LINK_TO(conv_weight, conv_bias_node);
+      IR_NODE_LINK_TO(eltwise_bias, conv_bias_node);
+      IR_NODE_LINK_TO(conv_bias_node, eltwise_out);
+
+      GraphSafeRemoveNodes(graph.get(), {conv, eltwise, conv_out});
+    }
+
+    found_conv_bias_count++;
+  };
+  gpd(graph.get(), handler);
+  AddStatis(found_conv_bias_count);
+  return graph;
+}
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+REGISTER_PASS(conv_bias_mkldnn_fuse_pass,
+              paddle::framework::ir::ConvBiasFusePass);

paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h

+// Copyright (c) 2018 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 <string>
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+#include "paddle/fluid/framework/ir/pass.h"
+namespace paddle {
+namespace framework {
+namespace ir {
+/*
+* Fuse the Conv and Elementwise_add to a ConvBiasOp.
+*/
+class ConvBiasFusePass : public FusePassBase {
+ public:
+  virtual ~ConvBiasFusePass() {}
+
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
+  const std::string name_scope_{"conv_bias_mkldnn_fuse"};
+};
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc

+// Copyright (c) 2018 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/conv_elementwise_add_mkldnn_fuse_pass.h"
+#include <functional>
+#include <utility>
+
+#include "paddle/fluid/framework/ir/graph_traits.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+namespace {
+
+// The function keeps the graph consistent by replacing
+// a node 'from' in the set of inputs nodes
+// of the visited node by a node 'to'.
+void CorrectGraphEdges(Graph* graph, Node* from, Node* to) {
+  for (auto& node : GraphTraits::DFS(*graph)) {
+    auto from_in_inputs =
+        std::find(std::begin(node.inputs), std::end(node.inputs), from);
+
+    if (from_in_inputs != std::end(node.inputs)) {
+      IR_NODE_LINK_TO(to, (&node));
+
+      auto inputs = node.Op()->Inputs();
+
+      using input_type = VariableNameMap::value_type;
+
+      std::for_each(std::begin(inputs), std::end(inputs),
+                    [from, to, &node](const input_type& i) -> void {
+                      auto param_names = i.second;
+                      auto pi = std::find(std::begin(param_names),
+                                          std::end(param_names), from->Name());
+
+                      if (pi != std::end(param_names)) {
+                        node.Op()->SetInput(i.first, {to->Name()});
+                      }
+                    });
+    }
+  }
+}
+}  // namespace
+using graph_ptr = std::unique_ptr<ir::Graph>;
+
+graph_ptr ConvElementwiseAddMKLDNNFusePass::ApplyImpl(graph_ptr graph) const {
+  FusePassBase::Init(name_scope_, graph.get());
+
+  GraphPatternDetector gpd;
+  auto pattern = gpd.mutable_pattern();
+
+  patterns::Conv conv_pattern{pattern, name_scope_};
+  auto conv_output = conv_pattern();
+
+  patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope_};
+  elementwise_add_pattern(conv_output);
+
+  conv_output->AsIntermediate();
+
+  auto conv_op_has_bias = [](const Node& conv_op) -> std::pair<bool, Node*> {
+    auto bias_input_names = conv_op.Op()->Inputs();
+    auto bias_it = bias_input_names.find("Bias");
+
+    if (bias_it != std::end(bias_input_names)) {
+      bool has_bias = !bias_it->second.empty();
+
+      if (has_bias) {
+        auto conv_bias_names = bias_it->second;
+        auto conv_bias_names_it =
+            std::find_if(std::begin(conv_op.inputs), std::end(conv_op.inputs),
+                         [&conv_bias_names](Node* n) -> bool {
+                           return n->Name() == conv_bias_names[0];
+                         });
+        return std::make_pair(has_bias, *conv_bias_names_it);
+      }
+    }
+
+    return std::make_pair(false, nullptr);
+  };
+
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    GET_IR_NODE_FROM_SUBGRAPH(conv_op, conv_op, conv_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(conv_input, conv_input, conv_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(conv_filter, conv_filter, conv_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(conv_output, conv_output, conv_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op,
+                              elementwise_add_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_x, elementwise_add_x,
+                              elementwise_add_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out,
+                              elementwise_add_pattern);
+
+    if (FindFuseOption(*conv_op, *elementwise_add_op) != FUSE_MKLDNN) return;
+
+    OpDesc op_desc;
+    op_desc.SetType("conv2d");
+
+    op_desc.SetInput("Input", {conv_input->Name()});
+    op_desc.SetInput("Filter", {conv_filter->Name()});
+    op_desc.SetInput("ResidualData", {elementwise_add_x->Name()});
+    op_desc.SetOutput("Output", {conv_output->Name()});
+
+    bool has_bias;
+    Node* conv_bias;
+
+    std::tie(has_bias, conv_bias) = conv_op_has_bias(*conv_op);
+
+    if (has_bias) {
+      op_desc.SetInput("Bias", {conv_bias->Name()});
+    }
+
+    for (const auto& attr : conv_op->Op()->GetAttrMap()) {
+      op_desc.SetAttr(attr.first, attr.second);
+    }
+
+    op_desc.SetAttr("fuse_residual_connection", true);
+
+    auto fused_conv_op = g->CreateOpNode(&op_desc);
+
+    IR_NODE_LINK_TO(conv_input, fused_conv_op);
+    IR_NODE_LINK_TO(conv_filter, fused_conv_op);
+    IR_NODE_LINK_TO(elementwise_add_x, fused_conv_op);
+    IR_NODE_LINK_TO(fused_conv_op, conv_output);
+
+    if (has_bias) {
+      IR_NODE_LINK_TO(conv_bias, fused_conv_op);
+    }
+
+    CorrectGraphEdges(g, elementwise_add_out, conv_output);
+    GraphSafeRemoveNodes(g, {elementwise_add_out, conv_op, elementwise_add_op});
+  };
+
+  gpd(graph.get(), handler);
+
+  return graph;
+}
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(conv_elementwise_add_mkldnn_fuse_pass,
+              paddle::framework::ir::ConvElementwiseAddMKLDNNFusePass);

paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h

+// Copyright (c) 2018 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 <string>
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class ConvElementwiseAddMKLDNNFusePass : public FusePassBase {
+ public:
+  virtual ~ConvElementwiseAddMKLDNNFusePass() {}
+
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
+
+  const std::string name_scope_{"residual_connections_fuse_pass"};
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc

+// Copyright (c) 2018 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 <string>
+
+#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h"
+#include "paddle/fluid/framework/ir/graph_traits.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+namespace {
+constexpr int nodes_removed = 3;
+constexpr int nodes_added = 1;
+
+void SetOp(ProgramDesc* prog, const std::string& type,
+           const std::vector<std::pair<std::string, std::string>>& inputs,
+           const std::pair<std::string, std::string>& output) {
+  auto op = prog->MutableBlock(0)->AppendOp();
+  op->SetType(type);
+  op->SetAttr("use_mkldnn", true);
+
+  for (const auto& input : inputs) {
+    op->SetInput(input.first, {input.second});
+  }
+
+  op->SetOutput(output.first, {output.second});
+}
+
+struct IsReachable {
+  using func = std::function<bool(const std::string&, const std::string&)>;
+
+  auto operator()(const std::unique_ptr<ir::Graph>& graph) -> func {
+    auto find_node = [](const std::unique_ptr<ir::Graph>& graph,
+                        const std::string& name) -> Node* {
+      for (auto& node : GraphTraits::DFS(*graph)) {
+        if (name == node.Name()) {
+          return &node;
+        }
+      }
+
+      return nullptr;
+    };
+
+    return [&](std::string from, const std::string to) -> bool {
+      if (from == to) return true;
+
+      std::map<std::string, bool> visited;
+
+      for (auto& node : GraphTraits::DFS(*graph)) {
+        visited[node.Name()] = false;
+      }
+
+      visited[from] = true;
+
+      std::list<std::string> queue;
+      queue.push_back(from);
+
+      while (!queue.empty()) {
+        auto cur = find_node(graph, queue.front());
+        queue.pop_front();
+
+        if (cur == nullptr) return false;
+
+        for (auto n : cur->outputs) {
+          if (n->Name() == to) return true;
+
+          if (!visited[n->Name()]) {
+            visited[n->Name()] = true;
+            queue.push_back(n->Name());
+          }
+        }
+      }
+      return false;
+    };
+  }
+};
+
+void AssertOpsCount(const std::unique_ptr<ir::Graph>& graph) {
+  int conv_count = 0;
+  int elementwise_add_count = 0;
+
+  for (auto* node : graph->Nodes()) {
+    if (node->IsOp() && node->Op()->Type() == "conv2d") {
+      ++conv_count;
+    }
+    if (node->IsOp() && node->Op()->Type() == "elementwise_add") {
+      ++elementwise_add_count;
+    }
+  }
+  EXPECT_EQ(conv_count, 1);
+  EXPECT_EQ(elementwise_add_count, 0);
+}
+
+ProgramDesc BuildProgramDesc(const std::vector<std::string>& transient_vars,
+                             const std::vector<std::string>& persistent_vars) {
+  ProgramDesc prog;
+
+  auto add_var_to_prog = [&prog](const std::string& var_name) -> VarDesc* {
+    auto var = prog.MutableBlock(0)->Var(var_name);
+    var->SetType(proto::VarType::LOD_TENSOR);
+
+    return var;
+  };
+
+  for (const auto& v : transient_vars) {
+    add_var_to_prog(v);
+  }
+
+  for (const auto& v : persistent_vars) {
+    auto var = add_var_to_prog(v);
+    var->SetPersistable(true);
+  }
+
+  return prog;
+}
+}  // namespace
+
+TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionWithElementwiseAddRelu) {
+  auto prog =
+      BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"});
+
+  SetOp(&prog, "conv2d",
+        {{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}},
+        {"Output", "b"});
+  SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
+  SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+
+  IsReachable is_reachable;
+  EXPECT_TRUE(is_reachable(graph)("a", "relu"));
+
+  auto pass =
+      PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
+  int original_nodes_num = graph->Nodes().size();
+  graph = pass->Apply(std::move(graph));
+  int current_nodes_num = graph->Nodes().size();
+
+  EXPECT_TRUE(is_reachable(graph)("a", "relu"));
+
+  EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
+            current_nodes_num);
+
+  AssertOpsCount(graph);
+}
+
+TEST(ConvElementwiseAddMKLDNNFusePass,
+     ConvolutionWithElementwiseAddReluNoBias) {
+  auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"weights"});
+  SetOp(&prog, "conv2d", {{"Input", "a"}, {"Filter", "weights"}},
+        {"Output", "b"});
+  SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
+  SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+
+  IsReachable is_reachable;
+
+  EXPECT_TRUE(is_reachable(graph)("a", "relu"));
+
+  auto pass =
+      PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
+  int original_nodes_num = graph->Nodes().size();
+  graph = pass->Apply(std::move(graph));
+  int current_nodes_num = graph->Nodes().size();
+
+  EXPECT_TRUE(is_reachable(graph)("a", "relu"));
+
+  EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
+            current_nodes_num);
+
+  AssertOpsCount(graph);
+}
+
+TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionElementwiseAdd) {
+  auto prog = BuildProgramDesc({"a", "b", "c", "d"}, {"bias", "weights"});
+  SetOp(&prog, "conv2d",
+        {{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}},
+        {"Output", "b"});
+  SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+
+  IsReachable is_reachable;
+  EXPECT_TRUE(is_reachable(graph)("a", "d"));
+
+  auto pass =
+      PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
+  int original_nodes_num = graph->Nodes().size();
+  graph = pass->Apply(std::move(graph));
+  int current_nodes_num = graph->Nodes().size();
+
+  EXPECT_FALSE(is_reachable(graph)("a", "d"));
+
+  EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
+            current_nodes_num);
+  AssertOpsCount(graph);
+}
+
+TEST(ConvElementwiseAddMKLDNNFusePass, SigmoidConvolutionAddElementwiseRelu) {
+  auto prog =
+      BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"});
+  SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
+  SetOp(&prog, "conv2d",
+        {{"Input", "b"}, {"Bias", "bias"}, {"Filter", "weights"}},
+        {"Output", "c"});
+  SetOp(&prog, "elementwise_add", {{"X", "c"}, {"Y", "d"}}, {"Out", "e"});
+  SetOp(&prog, "relu", {{"X", "e"}}, {"Out", "f"});
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+
+  IsReachable is_reachable;
+
+  EXPECT_TRUE(is_reachable(graph)("a", "f"));
+
+  auto pass =
+      PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
+  int original_nodes_num = graph->Nodes().size();
+  graph = pass->Apply(std::move(graph));
+  int current_nodes_num = graph->Nodes().size();
+
+  EXPECT_TRUE(is_reachable(graph)("a", "f"));
+
+  EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
+            current_nodes_num);
+  AssertOpsCount(graph);
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+USE_PASS(conv_elementwise_add_mkldnn_fuse_pass);

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -761,6 +761,51 @@ PDNode *patterns::ConvReLU::operator()(
   return relu_out_var;
 }
 
+PDNode *patterns::SeqConvEltAddRelu::operator()(
+    paddle::framework::ir::PDNode *seqconv_input) {
+  // Create Operators
+  seqconv_input->assert_is_op_input("sequence_conv", "X");
+  auto *seqconv_op = pattern->NewNode(seqconv_repr())
+                         ->assert_is_op("sequence_conv")
+                         ->assert_op_attr<bool>("paddingTrainable", false)
+                         ->assert_op_attr<int>("contextStride", 1);
+
+  auto *eltadd_op =
+      pattern->NewNode(eltadd_repr())->assert_is_op("elementwise_add");
+  auto *relu_op = pattern->NewNode(relu_repr())->assert_is_op("relu");
+  // Create variables
+  // Filter
+  auto *seqconv_weight_var =
+      pattern->NewNode(seqconv_weight_repr())
+          ->AsInput()
+          ->assert_is_persistable_var()
+          ->assert_is_op_input("sequence_conv", "Filter");
+  // Bias
+  auto *eltadd_bias_var = pattern->NewNode(eltadd_bias_repr())
+                              ->AsInput()
+                              ->assert_is_op_input("elementwise_add");
+  // intermediate variable, will be removed in the IR after fuse.
+  auto *seqconv_out_var = pattern->NewNode(seqconv_out_repr())
+                              ->AsIntermediate()
+                              ->assert_is_only_output_of_op("sequence_conv")
+                              ->assert_is_op_input("elementwise_add");
+  auto *eltadd_out_var = pattern->NewNode(eltadd_out_repr())
+                             ->AsIntermediate()
+                             ->assert_is_only_output_of_op("elementwise_add")
+                             ->assert_is_only_input_of_op("relu");
+  // output
+  auto *relu_out_var = pattern->NewNode(relu_out_repr())
+                           ->AsOutput()
+                           ->assert_is_op_output("relu");
+
+  seqconv_op->LinksFrom({seqconv_input, seqconv_weight_var})
+      .LinksTo({seqconv_out_var});
+  eltadd_op->LinksFrom({seqconv_out_var, eltadd_bias_var})
+      .LinksTo({eltadd_out_var});
+  relu_op->LinksFrom({eltadd_out_var}).LinksTo({relu_out_var});
+  return relu_out_var;
+}
+
 PDNode *patterns::FC::operator()(paddle::framework::ir::PDNode *x,
                                  bool with_bias) {
   // Create shared nodes.
@@ -966,6 +1011,79 @@ PDNode *patterns::ElewiseAddActInplaceGrad::operator()(
   return ele_add_grad;
 }
 
+PDNode *patterns::ConvBias::operator()(
+    paddle::framework::ir::PDNode *conv_input) {
+  // Create Operators
+  conv_input->assert_is_op_input("conv2d", "Input");
+  auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d");
+  auto *eltiwse_op =
+      pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add");
+  // Create variables
+  // Filter
+  auto *conv_weight_var = pattern->NewNode(conv_weight_repr())
+                              ->AsInput()
+                              ->assert_is_persistable_var()
+                              ->assert_is_op_input("conv2d", "Filter");
+  // intermediate variable, will be removed in the IR after fuse.
+  auto *conv_out_var = pattern->NewNode(conv_out_repr())
+                           ->AsIntermediate()
+                           ->assert_is_only_output_of_op("conv2d")
+                           ->assert_is_op_input("elementwise_add");
+  // Bias stored in elementwise_add
+  auto *eltwise_bias_var = pattern->NewNode(eltwise_bias_repr())
+                               ->AsInput()
+                               ->assert_is_persistable_var()
+                               ->assert_is_op_input("elementwise_add", "Y");
+  // output
+  auto *eltwise_out_var = pattern->NewNode(eltwise_out_repr())
+                              ->AsOutput()
+                              ->assert_is_op_output("elementwise_add");
+  conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var});
+  eltiwse_op->LinksFrom({conv_out_var, eltwise_bias_var})
+      .LinksTo({eltwise_out_var});
+  return eltwise_out_var;
+}
+
+PDNode *patterns::Conv::operator()() {
+  auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");
+
+  auto input_var = pattern->NewNode(conv_input_repr())
+                       ->AsInput()
+                       ->assert_is_op_input("conv2d", "Input");
+
+  auto filter_var = pattern->NewNode(conv_filter_repr())
+                        ->AsInput()
+                        ->assert_is_op_input("conv2d", "Filter");
+
+  auto output_var = pattern->NewNode(conv_output_repr())
+                        ->AsOutput()
+                        ->assert_is_op_output("conv2d", "Output");
+
+  conv_op->LinksFrom({input_var, filter_var});
+  conv_op->LinksTo({output_var});
+
+  return output_var;
+}
+
+PDNode *patterns::ElementwiseAdd::operator()(PDNode *x_var) {
+  auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr())
+                                ->assert_is_op("elementwise_add");
+
+  x_var->assert_is_op_input("elementwise_add", "X");
+
+  auto y_var = pattern->NewNode(elementwise_add_x_repr())
+                   ->AsInput()
+                   ->assert_is_op_input("elementwise_add", "Y");
+
+  auto out_var = pattern->NewNode(elementwise_add_out_repr())
+                     ->AsOutput()
+                     ->assert_is_op_output("elementwise_add", "Out");
+
+  elementwise_add_op->LinksFrom({x_var, y_var});
+  elementwise_add_op->LinksTo({out_var});
+
+  return out_var;
+}
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -128,6 +128,15 @@ struct PDNode {
       const std::unordered_set<std::string>& op_types,
       const std::string& argument, int nth);
 
+  template <typename T>
+  PDNode* assert_op_attr(const std::string& attr_name, const T& attr) {
+    asserts_.emplace_back([=](Node* x) {
+      return x && x->IsOp() && x->Op()->HasAttr(attr_name) &&
+             boost::get<T>(x->Op()->GetAttr(attr_name)) == attr;
+    });
+    return this;
+  }
+
  private:
   PDNode(PDPattern* pattern, const std::string& name = "",
          Type type = Type::kVar)
@@ -434,6 +443,31 @@ struct ConvReLU : public PatternBase {
   PATTERN_DECL_NODE(relu_out);
 };
 
+// SEQCONV with Elementwise_Add ReLU
+// op: seqconv + elementwise_add + relu
+// named nodes:
+// seqconv_input, seqconv_weight,
+// seqconv_out, seqconv,
+// elementwise_add_bias, elementwise_add_out, elementwise_add
+// relu_out, relu
+struct SeqConvEltAddRelu : public PatternBase {
+  SeqConvEltAddRelu(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "seqconv_eltadd_relu") {}
+
+  PDNode* operator()(PDNode* seqconv_input);
+
+  // declare operator node's name
+  PATTERN_DECL_NODE(seqconv);
+  PATTERN_DECL_NODE(eltadd);
+  PATTERN_DECL_NODE(relu);
+  // declare variable node's name
+  PATTERN_DECL_NODE(seqconv_weight);
+  PATTERN_DECL_NODE(seqconv_out);
+  PATTERN_DECL_NODE(eltadd_bias);
+  PATTERN_DECL_NODE(eltadd_out);
+  PATTERN_DECL_NODE(relu_out);
+};
+
 // FC with bias
 // op: mul + elementwise_add
 // named nodes:
@@ -578,6 +612,65 @@ struct ElewiseAddActInplaceGrad : public PatternBase {
   PATTERN_DECL_NODE(d_ele_y);
   PATTERN_DECL_NODE(ele_y);
 };
+
+// Conv with Elementwise_add as bias
+// op: conv + elementwise_add
+// named nodes:
+// conv_input, conv_weight,
+// conv_out, conv,
+// eltwise_bias, eltwise_out,
+// elementwise_add
+struct ConvBias : public PatternBase {
+  ConvBias(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "conv_bias") {}
+  PDNode* operator()(PDNode* conv_input);
+  // declare operator node's name
+  PATTERN_DECL_NODE(conv);
+  PATTERN_DECL_NODE(eltwise);
+  // declare variable node's name
+  PATTERN_DECL_NODE(conv_weight);
+  PATTERN_DECL_NODE(conv_out);
+  PATTERN_DECL_NODE(eltwise_bias);
+  PATTERN_DECL_NODE(eltwise_out);
+};
+
+// Convolution op
+// Forward pass for convolution.
+// conv_input, conv_bias and conv_filter are inputs.
+// conv_output is a result of the operator.
+// residual_data is data used by skip connection.
+// If residual connection fusion is on, the formula is:
+// conv_output = conv_op(conv_filter, conv_input, conv_bias)
+//             + conv_residual_data
+// If the fusion is off, conv_residual_data is not added.
+struct Conv : public PatternBase {
+  Conv(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "convolution") {}
+
+  PDNode* operator()();
+
+  PATTERN_DECL_NODE(conv_op);
+  PATTERN_DECL_NODE(conv_input);
+  PATTERN_DECL_NODE(conv_filter);
+  PATTERN_DECL_NODE(conv_residual_data);
+  PATTERN_DECL_NODE(conv_output);
+};
+
+// ElementwiseAdd used in residual connections.
+// y_var is used and convolution output.
+// The operator is removed, when residual
+// connection fusion is on.
+struct ElementwiseAdd : public PatternBase {
+  ElementwiseAdd(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "elementwise_add") {}
+
+  PDNode* operator()(PDNode* x_var);
+
+  PATTERN_DECL_NODE(elementwise_add_op);
+  PATTERN_DECL_NODE(elementwise_add_x);
+  PATTERN_DECL_NODE(elementwise_add_y);
+  PATTERN_DECL_NODE(elementwise_add_out);
+};
 }  // namespace patterns
 
 // Link two ir::Nodes from each other.

paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc

+// Copyright (c) 2018 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/seqconv_eltadd_relu_fuse_pass.h"
+#include <string>
+#include "paddle/fluid/framework/lod_tensor.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope) {
+  GraphPatternDetector gpd;
+  auto* pattern = gpd.mutable_pattern();
+
+  PDNode* x = pattern->NewNode(patterns::PDNodeName(name_scope, "X"))
+                  ->assert_is_op_input("sequence_conv")
+                  ->assert_var_not_persistable();
+  patterns::SeqConvEltAddRelu fuse_pattern(pattern, name_scope);
+  fuse_pattern(x);
+
+  // Create New OpDesc
+  auto fuse_creator = [&](Node* seqconv, Node* input, Node* seqconv_weight,
+                          Node* eltadd_bias, Node* relu_out) {
+    OpDesc op_desc;
+    op_desc.SetType("fusion_seqconv_eltadd_relu");
+    op_desc.SetInput("X", {input->Name()});
+    op_desc.SetInput("Filter", {seqconv_weight->Name()});
+    op_desc.SetInput("Bias", {eltadd_bias->Name()});
+    op_desc.SetAttr("contextLength", seqconv->Op()->GetAttr("contextLength"));
+    op_desc.SetAttr("contextStart", seqconv->Op()->GetAttr("contextStart"));
+    op_desc.SetAttr("contextStride", seqconv->Op()->GetAttr("contextStride"));
+    PADDLE_ENFORCE(graph->Has(kParamScopeAttr));
+    auto* scope = graph->Get<Scope*>(kParamScopeAttr);
+    const std::string ColMat = patterns::UniqueKey("SeqConvColMat");
+    op_desc.SetOutput("ColMat", {ColMat});
+    op_desc.SetOutput("Out", {relu_out->Name()});
+    scope->Var(ColMat)->GetMutable<LoDTensor>();
+
+    auto* op = graph->CreateOpNode(&op_desc);
+    IR_NODE_LINK_TO(input, op);
+    IR_NODE_LINK_TO(seqconv_weight, op);
+    IR_NODE_LINK_TO(eltadd_bias, op);
+    IR_NODE_LINK_TO(op, relu_out);
+    return op;
+  };
+
+  int fusion_count{0};
+
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    VLOG(4) << "handle SeqConv EltAdd Relu fuse";
+    GET_IR_NODE_FROM_SUBGRAPH(seqconv, seqconv, fuse_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(seqconv_weight, seqconv_weight, fuse_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(seqconv_out, seqconv_out, fuse_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(eltadd, eltadd, fuse_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(eltadd_bias, eltadd_bias, fuse_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(eltadd_out, eltadd_out, fuse_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(relu, relu, fuse_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(relu_out, relu_out, fuse_pattern);
+
+    fuse_creator(seqconv, subgraph.at(x), seqconv_weight, eltadd_bias,
+                 relu_out);
+    std::unordered_set<const Node*> marked_nodes(
+        {seqconv, seqconv_out, eltadd, eltadd_out, relu});
+    GraphSafeRemoveNodes(graph, marked_nodes);
+    ++fusion_count;
+  };
+
+  gpd(graph, handler);
+
+  return fusion_count;
+}
+
+std::unique_ptr<ir::Graph> SeqConvEltAddReluFusePass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  FusePassBase::Init(name_scope_, graph.get());
+
+  int fusion_count = BuildFusion(graph.get(), name_scope_, param_scope());
+  AddStatis(fusion_count);
+
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(seqconv_eltadd_relu_fuse_pass,
+              paddle::framework::ir::SeqConvEltAddReluFusePass);

paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h

+// Copyright (c) 2018 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 <string>
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class SeqConvEltAddReluFusePass : public FusePassBase {
+ public:
+  virtual ~SeqConvEltAddReluFusePass() {}
+
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
+
+  const std::string name_scope_{"seqconv_eltadd_relu_fuse"};
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/parallel_executor.cc

@@ -299,6 +299,12 @@ void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes(
 }
 
 ParallelExecutor::~ParallelExecutor() {
+  const auto dev_ctxs =
+      platform::DeviceContextPool::Instance().GetAllDeviceContexts();
+  for (auto &dev_ctx : dev_ctxs) {
+    dev_ctx->Wait();
+  }
+
   if (member_->own_local_scope_) {
     for (size_t i = 1; i < member_->local_scopes_.size(); ++i) {
       Scope *local_scope = member_->local_scopes_[i];

paddle/fluid/inference/analysis/analyzer.cc

@@ -101,10 +101,12 @@ Analyzer::Analyzer() { Register("manager1", new DfgPassManagerImpl); }
 
 void Analyzer::Run(Argument* argument) {
   std::vector<std::string> passes;
+#ifdef PADDLE_WITH_MKLDNN
   if (use_mkldnn_) {
     VLOG(3) << "Adding MKL-DNN placement pass";
     passes.push_back("mkldnn_placement_pass");
   }
+#endif
   for (auto& pass : ir_passes_) {
     if (!disabled_ir_passes_.count(pass)) {
       passes.push_back(pass);

paddle/fluid/inference/analysis/analyzer.h

@@ -69,6 +69,7 @@ class Analyzer : public OrderedRegistry<PassManager> {
       // Manual update the passes here.
       "infer_clean_graph_pass",         //
       "attention_lstm_fuse_pass",       //
+      "seqconv_eltadd_relu_fuse_pass",  //
       "embedding_fc_lstm_fuse_pass",    //
       "fc_lstm_fuse_pass",              //
       "mul_lstm_fuse_pass",             //
@@ -79,7 +80,9 @@ class Analyzer : public OrderedRegistry<PassManager> {
       "conv_bn_fuse_pass",              //
       "conv_eltwiseadd_bn_fuse_pass",   //
 #ifdef PADDLE_WITH_MKLDNN
+      "conv_bias_mkldnn_fuse_pass",             //
       "conv_relu_mkldnn_fuse_pass",             //
+      "conv_elementwise_add_mkldnn_fuse_pass",  //
 #endif
   }};
 

paddle/fluid/inference/api/analysis_predictor.cc

@@ -77,10 +77,6 @@ bool AnalysisPredictor::Init(
     inference_program_ = program;
   }
 
-  if (config_._use_mkldnn) {
-    executor_->EnableMKLDNN(*inference_program_);
-  }
-
   executor_->Prepare(scope_.get(), *inference_program_, 0,
                      config_.use_feed_fetch_ops);
 

paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc

@@ -18,12 +18,12 @@ namespace paddle {
 namespace inference {
 
 using namespace framework;  // NOLINT
+static std::vector<float> result_data;
 
 struct DataRecord {
   std::vector<std::vector<std::vector<float>>> link_step_data_all;
   std::vector<size_t> lod;
   std::vector<std::vector<float>> rnn_link_data;
-  std::vector<float> result_data;
   size_t num_samples;  // total number of samples
   size_t batch_iter{0};
   size_t batch_size{1};
@@ -57,6 +57,7 @@ struct DataRecord {
     std::ifstream file(path);
     std::string line;
     int num_lines = 0;
+    result_data.clear();
     while (std::getline(file, line)) {
       num_lines++;
       std::vector<std::string> data;
@@ -135,13 +136,12 @@ TEST(Analyzer_rnn2, profile) {
 
   if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
     // the first inference result
-    DataRecord data(FLAGS_infer_data, FLAGS_batch_size);
     PADDLE_ENFORCE_GT(outputs.size(), 0);
     size_t size = GetSize(outputs[0]);
     PADDLE_ENFORCE_GT(size, 0);
     float *result = static_cast<float *>(outputs[0].data.data());
     for (size_t i = 0; i < size; i++) {
-      EXPECT_NEAR(result[i], data.result_data[i], 1e-3);
+      EXPECT_NEAR(result[i], result_data[i], 1e-3);
     }
   }
 }

paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc

@@ -183,7 +183,13 @@ TEST(Analyzer_seq_conv1, fuse_statis) {
   SetConfig(&cfg);
   int num_ops;
   auto predictor = CreatePaddlePredictor<AnalysisConfig>(cfg);
-  GetFuseStatis(predictor.get(), &num_ops);
+
+  auto fuse_statis = GetFuseStatis(predictor.get(), &num_ops);
+  ASSERT_TRUE(fuse_statis.count("fc_fuse"));
+  ASSERT_TRUE(fuse_statis.count("seqconv_eltadd_relu_fuse"));
+  EXPECT_EQ(fuse_statis.at("fc_fuse"), 2);
+  EXPECT_EQ(fuse_statis.at("seqconv_eltadd_relu_fuse"), 6);
+  EXPECT_EQ(num_ops, 32);
 }
 
 // Compare result of NativeConfig and AnalysisConfig

paddle/fluid/operators/CMakeLists.txt

@@ -86,7 +86,7 @@ function(op_library TARGET)
     # remove windows unsupported op, because windows has no nccl, no warpctc such ops.
     foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op" "hierarchical_sigmoid_op"
      "crf_decoding_op" "select_op" "lstmp_op" "gru_op" "fusion_gru_op" "lstm_op" "fusion_lstm_op" "cumsum_op"
-     "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op")
+      "fusion_seqconv_eltadd_relu_op" "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op")
         if ("${TARGET}" STREQUAL "${windows_unsupport_op}")
           return()
         endif()

paddle/fluid/operators/conv_mkldnn_op.cc

@@ -300,10 +300,10 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
     std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
     bool fuse_relu = ctx.Attr<bool>("fuse_relu");
-    bool fuse_eltwise = ctx.Attr<bool>("fuse_eltwise");
+    bool fuse_residual_conn = ctx.Attr<bool>("fuse_residual_connection");
     int groups = ctx.Attr<int>("groups");
 
-    // TODO: add support for dilation
+    // TODO(tpatejko): add support for dilation
     PADDLE_ENFORCE(
         dilations.size() == 2 && dilations[0] == 1 && dilations[1] == 1,
         "dilation in convolution is not implemented yet");
@@ -369,11 +369,11 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
           bias_tz, platform::MKLDNNGetDataType<T>(), memory::format::x);
       conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, bias_md, dst_md,
                                      strides, paddings, mkldnn_engine,
-                                     fuse_relu, fuse_eltwise);
+                                     fuse_relu, fuse_residual_conn);
     } else {
       conv_pd =
           ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides, paddings,
-                               mkldnn_engine, fuse_relu, fuse_eltwise);
+                               mkldnn_engine, fuse_relu, fuse_residual_conn);
     }
     // Save conv_pd/src_memory/weights_memory for backward pass
     dev_ctx.SetBlob(key_conv_pd, conv_pd);
@@ -386,8 +386,26 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     auto user_weights_memory_p = handler.AcquireWeightsMemory(
         user_weights_md, to_void_cast<T>(filter_data));
 
-    T* output_data =
+    T* output_data = nullptr;
+
+    if (fuse_residual_conn) {
+      auto residual_param = ctx.Input<Tensor>("ResidualData");
+      auto residual_param_data = residual_param->data<T>();
+
+      PADDLE_ENFORCE(
+          residual_param_data != nullptr,
+          "Provide data if you want MKLDNN conv+elementwise_add fusion");
+      PADDLE_ENFORCE_EQ(output->dims(), residual_param->dims(),
+                        "Output and elementwise parameter need to have the "
+                        "same dimension sizes");
+
+      output->ShareDataWith(*residual_param);
+      output_data = output->mutable_data<T>(ctx.GetPlace());
+    } else {
+      output_data =
           output->mutable_data<T>(ctx.GetPlace(), handler.GetDstMemorySize());
+    }
+
     // create reorder primitive if the input format is not the preferred one
     auto src_memory_p =
         handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline);
@@ -424,14 +442,15 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
 
  private:
   mkldnn::primitive_attr CreatePostOps(bool fuse_relu,
-                                       bool fuse_eltwise) const {
+                                       bool fuse_residual_conn) const {
     mkldnn::primitive_attr conv_attr;
     mkldnn::post_ops post_operations;
     // Fusion with Elementwise layer relies on adding a sum post-operation with
-    // the scale parameter. It is assumed that when fuse_eltwise is true, the
-    // Output tensor contains the data coming from residual connection. The
-    // result of this post_op is: Output = scale * Output + Conv_Out.
-    if (fuse_eltwise) {
+    // the scale parameter. It is assumed that when fuse_residual_connection is
+    // true, the output tensor contains the data coming from residual
+    // connection. The result of this post_op is:
+    // Output = scale * Output + Conv_Out.
+    if (fuse_residual_conn) {
       post_operations.append_sum(1.0f);
     }
     // Fusion with ReLU layer is executed through the PostOps feature. Create a
@@ -452,7 +471,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
                        const memory::desc& dst, const std::vector<int>& strides,
                        const std::vector<int>& paddings,
                        const mkldnn::engine& engine, const bool fuse_relu,
-                       const bool fuse_eltwise) const {
+                       const bool fuse_residual_conn) const {
     memory::dims stride_dims = {strides[0], strides[1]};
     memory::dims padding_dims = {paddings[0], paddings[1]};
 
@@ -461,7 +480,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         dst, stride_dims, padding_dims, padding_dims,
         mkldnn::padding_kind::zero);
 
-    mkldnn::primitive_attr conv_attr = CreatePostOps(fuse_relu, fuse_eltwise);
+    mkldnn::primitive_attr conv_attr =
+        CreatePostOps(fuse_relu, fuse_residual_conn);
 
     auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
         conv_desc, conv_attr, engine);
@@ -476,7 +496,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
                        const std::vector<int>& strides,
                        const std::vector<int>& paddings,
                        const mkldnn::engine& engine, const bool fuse_relu,
-                       const bool fuse_eltwise) const {
+                       const bool fuse_residual_conn) const {
     memory::dims stride_dims = {strides[0], strides[1]};
     memory::dims padding_dims = {paddings[0], paddings[1]};
 
@@ -485,7 +505,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         bias, dst, stride_dims, padding_dims, padding_dims,
         mkldnn::padding_kind::zero);
 
-    mkldnn::primitive_attr conv_attr = CreatePostOps(fuse_relu, fuse_eltwise);
+    mkldnn::primitive_attr conv_attr =
+        CreatePostOps(fuse_relu, fuse_residual_conn);
 
     auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
         conv_desc, conv_attr, engine);

paddle/fluid/operators/conv_op.cc

@@ -132,6 +132,11 @@ void Conv2DOpMaker::Make() {
             "(Tensor) The output tensor of convolution operator. "
             "The format of output tensor is also NCHW.")
       .Reuse("Input");
+  AddInput("ResidualData",
+           "(Tensor) Tensor with residual data "
+           "to which convolution output will be added."
+           "Used with fuse_residual_connection fusion.")
+      .AsDispensable();
   AddAttr<std::vector<int>>("strides",
                             "(vector<int> default:{1, 1}), the "
                             "strides(h_stride, w_stride) of "
@@ -164,10 +169,10 @@ void Conv2DOpMaker::Make() {
       .SetDefault(false);
   AddAttr<bool>("fuse_relu", "(bool, default false) Only used in mkldnn kernel")
       .SetDefault(false);
-  AddAttr<bool>("fuse_eltwise",
+  AddAttr<bool>("fuse_residual_connection",
                 "(bool, default false) Only used in mkldnn kernel. Used "
-                "whenever convolution output is connected via skip connection "
-                "to a previous layer.")
+                "whenever convolution output is as an input to residual "
+                "connection.")
       .SetDefault(false);
   AddAttr<std::string>(
       "data_format",

paddle/fluid/operators/detection/CMakeLists.txt

@@ -20,7 +20,7 @@ detection_library(box_coder_op SRCS box_coder_op.cc box_coder_op.cu)
 detection_library(iou_similarity_op SRCS iou_similarity_op.cc
 iou_similarity_op.cu)
 detection_library(mine_hard_examples_op SRCS mine_hard_examples_op.cc)
-detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc)
+detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc poly_util.cc gpc.cc)
 detection_library(prior_box_op SRCS prior_box_op.cc prior_box_op.cu)
 detection_library(anchor_generator_op SRCS anchor_generator_op.cc
 anchor_generator_op.cu)

paddle/fluid/operators/detection/generate_proposals_op.cc

@@ -12,10 +12,12 @@ 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 <cmath>
+#include <cstring>
 #include <string>
 #include <vector>
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/framework/var_type.h"
+#include "paddle/fluid/operators/detail/safe_ref.h"
 #include "paddle/fluid/operators/gather.h"
 #include "paddle/fluid/operators/math/math_function.h"
 
@@ -25,21 +27,17 @@ namespace operators {
 using Tensor = framework::Tensor;
 using LoDTensor = framework::LoDTensor;
 
-struct AppendProposalsFunctor {
-  LoDTensor *out_;
-  int64_t offset_;
-  Tensor *to_add_;
+static const double kBBoxClipDefault = std::log(1000.0 / 16.0);
 
-  AppendProposalsFunctor(LoDTensor *out, int64_t offset, Tensor *to_add)
-      : out_(out), offset_(offset), to_add_(to_add) {}
-
-  template <typename T>
-  void apply() const {
-    auto *out_data = out_->data<T>();
-    auto *to_add_data = to_add_->data<T>();
-    memcpy(out_data + offset_, to_add_data, to_add_->numel() * sizeof(T));
-  }
-};
+static void AppendProposals(Tensor *dst, int64_t offset, const Tensor &src) {
+  auto *out_data = dst->data<void>();
+  auto *to_add_data = src.data<void>();
+  size_t size_of_t = framework::SizeOfType(src.type());
+  offset *= size_of_t;
+  std::memcpy(
+      reinterpret_cast<void *>(reinterpret_cast<uintptr_t>(out_data) + offset),
+      to_add_data, src.numel() * size_of_t);
+}
 
 class GenerateProposalsOp : public framework::OperatorWithKernel {
  public:
@@ -75,8 +73,9 @@ class GenerateProposalsOp : public framework::OperatorWithKernel {
 };
 
 template <class T>
-void BoxCoder(const platform::DeviceContext &ctx, Tensor *all_anchors,
-              Tensor *bbox_deltas, Tensor *variances, Tensor *proposals) {
+static inline void BoxCoder(const platform::DeviceContext &ctx,
+                            Tensor *all_anchors, Tensor *bbox_deltas,
+                            Tensor *variances, Tensor *proposals) {
   T *proposals_data = proposals->mutable_data<T>(ctx.GetPlace());
 
   int64_t row = all_anchors->dims()[0];
@@ -108,11 +107,11 @@ void BoxCoder(const platform::DeviceContext &ctx, Tensor *all_anchors,
                       anchor_center_y;
       bbox_width = std::exp(std::min<T>(variances_data[i * len + 2] *
                                             bbox_deltas_data[i * len + 2],
-                                        std::log(1000.0 / 16.0))) *
+                                        kBBoxClipDefault)) *
                    anchor_width;
       bbox_height = std::exp(std::min<T>(variances_data[i * len + 3] *
                                              bbox_deltas_data[i * len + 3],
-                                         std::log(1000.0 / 16.0))) *
+                                         kBBoxClipDefault)) *
                     anchor_height;
     } else {
       bbox_center_x =
@@ -120,10 +119,10 @@ void BoxCoder(const platform::DeviceContext &ctx, Tensor *all_anchors,
       bbox_center_y =
           bbox_deltas_data[i * len + 1] * anchor_height + anchor_center_y;
       bbox_width = std::exp(std::min<T>(bbox_deltas_data[i * len + 2],
-                                        std::log(1000.0 / 16.0))) *
+                                        kBBoxClipDefault)) *
                    anchor_width;
       bbox_height = std::exp(std::min<T>(bbox_deltas_data[i * len + 3],
-                                         std::log(1000.0 / 16.0))) *
+                                         kBBoxClipDefault)) *
                     anchor_height;
     }
 
@@ -136,30 +135,32 @@ void BoxCoder(const platform::DeviceContext &ctx, Tensor *all_anchors,
 }
 
 template <class T>
-void ClipTiledBoxes(const platform::DeviceContext &ctx, const Tensor &im_info,
-                    Tensor *boxes) {
+static inline void ClipTiledBoxes(const platform::DeviceContext &ctx,
+                                  const Tensor &im_info, Tensor *boxes) {
   T *boxes_data = boxes->mutable_data<T>(ctx.GetPlace());
   const T *im_info_data = im_info.data<T>();
+  T zero(0);
   for (int64_t i = 0; i < boxes->numel(); ++i) {
     if (i % 4 == 0) {
       boxes_data[i] =
-          std::max(std::min(boxes_data[i], im_info_data[1] - 1), 0.0f);
+          std::max(std::min(boxes_data[i], im_info_data[1] - 1), zero);
     } else if (i % 4 == 1) {
       boxes_data[i] =
-          std::max(std::min(boxes_data[i], im_info_data[0] - 1), 0.0f);
+          std::max(std::min(boxes_data[i], im_info_data[0] - 1), zero);
     } else if (i % 4 == 2) {
       boxes_data[i] =
-          std::max(std::min(boxes_data[i], im_info_data[1] - 1), 0.0f);
+          std::max(std::min(boxes_data[i], im_info_data[1] - 1), zero);
     } else {
       boxes_data[i] =
-          std::max(std::min(boxes_data[i], im_info_data[0] - 1), 0.0f);
+          std::max(std::min(boxes_data[i], im_info_data[0] - 1), zero);
     }
   }
 }
 
 template <class T>
-void FilterBoxes(const platform::DeviceContext &ctx, Tensor *boxes,
-                 float min_size, const Tensor &im_info, Tensor *keep) {
+static inline void FilterBoxes(const platform::DeviceContext &ctx,
+                               Tensor *boxes, float min_size,
+                               const Tensor &im_info, Tensor *keep) {
   const T *im_info_data = im_info.data<T>();
   T *boxes_data = boxes->mutable_data<T>(ctx.GetPlace());
   T im_scale = im_info_data[2];
@@ -185,24 +186,24 @@ void FilterBoxes(const platform::DeviceContext &ctx, Tensor *boxes,
   keep->Resize({keep_len});
 }
 
-bool SortScorePairDescend(const std::pair<float, int> &pair1,
-                          const std::pair<float, int> &pair2) {
-  return pair1.first > pair2.first;
-}
-
 template <class T>
-void GetMaxScoreIndex(const std::vector<T> &scores,
-                      std::vector<std::pair<T, int>> *sorted_indices) {
+static inline std::vector<std::pair<T, int>> GetSortedScoreIndex(
+    const std::vector<T> &scores) {
+  std::vector<std::pair<T, int>> sorted_indices;
+  sorted_indices.reserve(scores.size());
   for (size_t i = 0; i < scores.size(); ++i) {
-    sorted_indices->push_back(std::make_pair(scores[i], i));
+    sorted_indices.emplace_back(scores[i], i);
   }
   // Sort the score pair according to the scores in descending order
-  std::stable_sort(sorted_indices->begin(), sorted_indices->end(),
-                   SortScorePairDescend);
+  std::stable_sort(sorted_indices.begin(), sorted_indices.end(),
+                   [](const std::pair<T, int> &a, const std::pair<T, int> &b) {
+                     return a.first < b.first;
+                   });
+  return sorted_indices;
 }
 
 template <class T>
-T BBoxArea(const T *box, const bool normalized) {
+static inline T BBoxArea(const T *box, bool normalized) {
   if (box[2] < box[0] || box[3] < box[1]) {
     // If coordinate values are is invalid
     // (e.g. xmax < xmin or ymax < ymin), return 0.
@@ -220,7 +221,7 @@ T BBoxArea(const T *box, const bool normalized) {
 }
 
 template <class T>
-T JaccardOverlap(const T *box1, const T *box2, const bool normalized) {
+static inline T JaccardOverlap(const T *box1, const T *box2, bool normalized) {
   if (box2[0] > box1[2] || box2[2] < box1[0] || box2[1] > box1[3] ||
       box2[3] < box1[1]) {
     return static_cast<T>(0.);
@@ -229,8 +230,8 @@ T JaccardOverlap(const T *box1, const T *box2, const bool normalized) {
     const T inter_ymin = std::max(box1[1], box2[1]);
     const T inter_xmax = std::min(box1[2], box2[2]);
     const T inter_ymax = std::min(box1[3], box2[3]);
-    const T inter_w = std::max(0.0f, inter_xmax - inter_xmin + 1);
-    const T inter_h = std::max(0.0f, inter_ymax - inter_ymin + 1);
+    const T inter_w = std::max(T(0), inter_xmax - inter_xmin + 1);
+    const T inter_h = std::max(T(0), inter_ymax - inter_ymin + 1);
     const T inter_area = inter_w * inter_h;
     const T bbox1_area = BBoxArea<T>(box1, normalized);
     const T bbox2_area = BBoxArea<T>(box2, normalized);
@@ -238,9 +239,21 @@ T JaccardOverlap(const T *box1, const T *box2, const bool normalized) {
   }
 }
 
+template <typename T>
+static inline Tensor VectorToTensor(const std::vector<T> &selected_indices,
+                                    int selected_num) {
+  Tensor keep_nms;
+  keep_nms.Resize({selected_num});
+  auto *keep_data = keep_nms.mutable_data<T>(platform::CPUPlace());
+  for (int i = 0; i < selected_num; ++i) {
+    keep_data[i] = selected_indices[i];
+  }
+  return keep_nms;
+}
+
 template <class T>
-Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox, Tensor *scores,
-           const T nms_threshold, const float eta) {
+static inline Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox,
+                         Tensor *scores, T nms_threshold, float eta) {
   PADDLE_ENFORCE_NOT_NULL(bbox);
   int64_t num_boxes = bbox->dims()[0];
   // 4: [xmin ymin xmax ymax]
@@ -248,20 +261,18 @@ Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox, Tensor *scores,
 
   std::vector<T> scores_data(num_boxes);
   std::copy_n(scores->data<T>(), num_boxes, scores_data.begin());
-  std::vector<std::pair<T, int>> sorted_indices;
-  GetMaxScoreIndex<T>(scores_data, &sorted_indices);
+  std::vector<std::pair<T, int>> sorted_indices =
+      GetSortedScoreIndex<T>(scores_data);
 
   std::vector<int> selected_indices;
   int selected_num = 0;
   T adaptive_threshold = nms_threshold;
   const T *bbox_data = bbox->data<T>();
-  bool flag;
   while (sorted_indices.size() != 0) {
-    int idx = sorted_indices.front().second;
-    flag = true;
-    for (size_t k = 0; k < selected_indices.size(); ++k) {
+    int idx = sorted_indices.back().second;
+    bool flag = true;
+    for (int kept_idx : selected_indices) {
       if (flag) {
-        const int kept_idx = selected_indices[k];
         T overlap = JaccardOverlap<T>(bbox_data + idx * box_size,
                                       bbox_data + kept_idx * box_size, false);
         flag = (overlap <= adaptive_threshold);
@@ -271,32 +282,29 @@ Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox, Tensor *scores,
     }
     if (flag) {
       selected_indices.push_back(idx);
-      selected_num++;
+      ++selected_num;
     }
-    sorted_indices.erase(sorted_indices.begin());
+    sorted_indices.erase(sorted_indices.end());
     if (flag && eta < 1 && adaptive_threshold > 0.5) {
       adaptive_threshold *= eta;
     }
   }
-  Tensor keep_nms;
-  keep_nms.Resize({selected_num});
-  int *keep_data = keep_nms.mutable_data<int>(ctx.GetPlace());
-  for (int i = 0; i < selected_num; ++i) {
-    keep_data[i] = selected_indices[i];
-  }
-
-  return keep_nms;
+  return VectorToTensor(selected_indices, selected_num);
 }
 
-template <typename DeviceContext, typename T>
+template <typename T>
 class GenerateProposalsKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
     auto *scores = context.Input<Tensor>("Scores");
     auto *bbox_deltas = context.Input<Tensor>("BboxDeltas");
     auto *im_info = context.Input<Tensor>("ImInfo");
-    auto *anchors = context.Input<Tensor>("Anchors");
-    auto *variances = context.Input<Tensor>("Variances");
+    auto anchors = detail::Ref(context.Input<Tensor>("Anchors"),
+                               "Cannot find input Anchors(%s) in scope",
+                               context.Inputs("Anchors")[0]);
+    auto variances = detail::Ref(context.Input<Tensor>("Variances"),
+                                 "Cannot find input Variances(%s) in scope",
+                                 context.Inputs("Variances")[0]);
 
     auto *rpn_rois = context.Output<LoDTensor>("RpnRois");
     auto *rpn_roi_probs = context.Output<LoDTensor>("RpnRoiProbs");
@@ -307,15 +315,16 @@ class GenerateProposalsKernel : public framework::OpKernel<T> {
     float min_size = context.Attr<float>("min_size");
     float eta = context.Attr<float>("eta");
 
-    auto &dev_ctx = context.template device_context<DeviceContext>();
+    auto &dev_ctx =
+        context.template device_context<platform::CPUDeviceContext>();
 
-    auto scores_dim = scores->dims();
+    auto &scores_dim = scores->dims();
     int64_t num = scores_dim[0];
     int64_t c_score = scores_dim[1];
     int64_t h_score = scores_dim[2];
     int64_t w_score = scores_dim[3];
 
-    auto bbox_dim = bbox_deltas->dims();
+    auto &bbox_dim = bbox_deltas->dims();
     int64_t c_bbox = bbox_dim[1];
     int64_t h_bbox = bbox_dim[2];
     int64_t w_bbox = bbox_dim[3];
@@ -330,17 +339,17 @@ class GenerateProposalsKernel : public framework::OpKernel<T> {
     scores_swap.mutable_data<T>({num, h_score, w_score, c_score},
                                 dev_ctx.GetPlace());
 
-    math::Transpose<DeviceContext, T, 4> trans;
+    math::Transpose<platform::CPUDeviceContext, T, 4> trans;
     std::vector<int> axis = {0, 2, 3, 1};
     trans(dev_ctx, *bbox_deltas, &bbox_deltas_swap, axis);
     trans(dev_ctx, *scores, &scores_swap, axis);
 
     framework::LoD lod;
-    std::vector<size_t> lod0(1, 0);
-    Tensor *anchor = const_cast<framework::Tensor *>(anchors);
-    anchor->Resize({anchors->numel() / 4, 4});
-    Tensor *var = const_cast<framework::Tensor *>(variances);
-    var->Resize({var->numel() / 4, 4});
+    lod.resize(1);
+    auto &lod0 = lod[0];
+    lod0.push_back(0);
+    anchors.Resize({anchors.numel() / 4, 4});
+    variances.Resize({variances.numel() / 4, 4});
 
     int64_t num_proposals = 0;
     for (int64_t i = 0; i < num; ++i) {
@@ -352,24 +361,17 @@ class GenerateProposalsKernel : public framework::OpKernel<T> {
       scores_slice.Resize({h_score * w_score * c_score, 1});
 
       std::pair<Tensor, Tensor> tensor_pair =
-          ProposalForOneImage(dev_ctx, im_info_slice, *anchor, *var,
+          ProposalForOneImage(dev_ctx, im_info_slice, anchors, variances,
                               bbox_deltas_slice, scores_slice, pre_nms_top_n,
                               post_nms_top_n, nms_thresh, min_size, eta);
-      Tensor proposals = tensor_pair.first;
-      Tensor scores = tensor_pair.second;
-
-      framework::VisitDataType(
-          framework::ToDataType(rpn_rois->type()),
-          AppendProposalsFunctor(rpn_rois, 4 * num_proposals, &proposals));
-      framework::VisitDataType(
-          framework::ToDataType(rpn_roi_probs->type()),
-          AppendProposalsFunctor(rpn_roi_probs, num_proposals, &scores));
+      Tensor &proposals = tensor_pair.first;
+      Tensor &scores = tensor_pair.second;
 
+      AppendProposals(rpn_rois, 4 * num_proposals, proposals);
+      AppendProposals(rpn_roi_probs, num_proposals, scores);
       num_proposals += proposals.dims()[0];
-      lod0.emplace_back(num_proposals);
+      lod0.push_back(num_proposals);
     }
-
-    lod.emplace_back(lod0);
     rpn_rois->set_lod(lod);
     rpn_roi_probs->set_lod(lod);
     rpn_rois->Resize({num_proposals, 4});
@@ -377,7 +379,7 @@ class GenerateProposalsKernel : public framework::OpKernel<T> {
   }
 
   std::pair<Tensor, Tensor> ProposalForOneImage(
-      const DeviceContext &ctx, const Tensor &im_info_slice,
+      const platform::CPUDeviceContext &ctx, const Tensor &im_info_slice,
       const Tensor &anchors, const Tensor &variances,
       const Tensor &bbox_deltas_slice,  // [M, 4]
       const Tensor &scores_slice,       // [N, 1]
@@ -392,8 +394,7 @@ class GenerateProposalsKernel : public framework::OpKernel<T> {
     for (int i = 0; i < scores_slice.numel(); ++i) {
       index[i] = i;
     }
-    std::function<bool(const int64_t &, const int64_t &)> compare =
-        [scores_data](const int64_t &i, const int64_t &j) {
+    auto compare = [scores_data](const int64_t &i, const int64_t &j) {
       return scores_data[i] > scores_data[j];
     };
 
@@ -452,33 +453,45 @@ class GenerateProposalsKernel : public framework::OpKernel<T> {
 class GenerateProposalsOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
-    AddInput("Scores", "The scores of anchors should be foreground.");
-    AddInput("BboxDeltas", "bbox_deltas.");
-    AddInput("ImInfo", "Information for image reshape.");
-    AddInput("Anchors", "All anchors.");
-    AddInput("Variances", " variances");
-
-    AddOutput("RpnRois", "Anchors.");
-    AddOutput("RpnRoiProbs", "Anchors.");
-    AddAttr<int>("pre_nms_topN", "pre_nms_topN");
-    AddAttr<int>("post_nms_topN", "post_nms_topN");
-    AddAttr<float>("nms_thresh", "nms_thres");
-    AddAttr<float>("min_size", "min size");
+    AddInput("Scores",
+             "(Tensor) The scores from conv is in shape (N, A, H, W), "
+             "N is batch size, A is number of anchors, "
+             "H and W are height and width of the feature map");
+    AddInput("BboxDeltas",
+             "(Tensor) Bounding box deltas from conv is in "
+             "shape (N, 4*A, H, W).");
+    AddInput("ImInfo",
+             "(Tensor) Information for image reshape is in shape (N, 3), "
+             "in format (height, width, scale)");
+    AddInput("Anchors",
+             "(Tensor) Bounding box anchors from anchor_generator_op "
+             "is in shape (A, H, W, 4).");
+    AddInput("Variances",
+             "(Tensor) Bounding box variances with same shape as `Anchors`.");
+
+    AddOutput("RpnRois",
+              "(LoDTensor), Output proposals with shape (rois_num, 4).");
+    AddOutput("RpnRoiProbs",
+              "(LoDTensor) Scores of proposals with shape (rois_num, 1).");
+    AddAttr<int>("pre_nms_topN",
+                 "Number of top scoring RPN proposals to keep before "
+                 "applying NMS.");
+    AddAttr<int>("post_nms_topN",
+                 "Number of top scoring RPN proposals to keep after "
+                 "applying NMS");
+    AddAttr<float>("nms_thresh", "NMS threshold used on RPN proposals.");
+    AddAttr<float>("min_size",
+                   "Proposal height and width both need to be greater "
+                   "than this min_size.");
     AddAttr<float>("eta", "The parameter for adaptive NMS.");
     AddComment(R"DOC(
-Generate Proposals OP
-
-This operator proposes rois according to each box with their probability to be a foreground object and 
-the box can be calculated by anchors. Bbox_deltais and scores are the output of RPN. Final proposals
-could be used to train detection net.
-
-Scores is the probability for each box to be an object. In format of (N, A, H, W) where N is batch size, A is number
-of anchors, H and W are height and width of the feature map.
-BboxDeltas is the differece between predicted box locatoin and anchor location. In format of (N, 4*A, H, W)
+This operator Generate bounding box proposals for Faster RCNN.
+The propoasls are generated for a list of images based on image
+score 'Scores', bounding box regression result 'BboxDeltas' as
+well as predefined bounding box shapes 'anchors'. Greedy
+non-maximum suppression is applied to generate the final bounding
+boxes.
 
-For generating proposals, this operator transposes and resizes scores and bbox_deltas in size of (H*W*A, 1) and (H*W*A, 4) and 
- calculate box locations as proposals candidates. Then clip boxes to image and remove predicted boxes with small area. 
-Finally, apply nms to get final proposals as output.
 )DOC");
   }
 };
@@ -490,6 +503,5 @@ namespace ops = paddle::operators;
 REGISTER_OPERATOR(generate_proposals, ops::GenerateProposalsOp,
                   ops::GenerateProposalsOpMaker,
                   paddle::framework::EmptyGradOpMaker);
-REGISTER_OP_CPU_KERNEL(
-    generate_proposals,
-    ops::GenerateProposalsKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(generate_proposals, ops::GenerateProposalsKernel<float>,
+                       ops::GenerateProposalsKernel<double>);

paddle/fluid/operators/detection/generate_proposals_op.cu

@@ -16,10 +16,13 @@ limitations under the License. */
 #include <string>
 #include <vector>
 #include "cub/cub.cuh"
+#include "paddle/fluid/framework/mixed_vector.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/memory/memory.h"
+#include "paddle/fluid/operators/detail/safe_ref.h"
 #include "paddle/fluid/operators/gather.cu.h"
 #include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/platform/for_range.h"
 
 namespace paddle {
 namespace operators {
@@ -36,36 +39,38 @@ namespace {
 
 int const kThreadsPerBlock = sizeof(uint64_t) * 8;
 
-template <typename T>
-__global__ void RangeInitKernel(const T start, const T delta, const int size,
-                                T *out) {
-  CUDA_1D_KERNEL_LOOP(i, size) { out[i] = start + i * delta; }
-}
+static const double kBBoxClipDefault = std::log(1000.0 / 16.0);
+
+struct RangeInitFunctor {
+  int start_;
+  int delta_;
+  int *out_;
+  __device__ void operator()(size_t i) { out_[i] = start_ + i * delta_; }
+};
 
 template <typename T>
-void SortDescending(const platform::CUDADeviceContext &ctx, const Tensor &value,
-                    Tensor *value_out, Tensor *index_out) {
-  int num = value.numel();
+static void SortDescending(const platform::CUDADeviceContext &ctx,
+                           const Tensor &value, Tensor *value_out,
+                           Tensor *index_out) {
+  int num = static_cast<int>(value.numel());
   Tensor index_in_t;
   int *idx_in = index_in_t.mutable_data<int>({num}, ctx.GetPlace());
-  int block = 512;
-  auto stream = ctx.stream();
-  RangeInitKernel<<<DIVUP(num, block), block, 0, stream>>>(0, 1, num, idx_in);
+  platform::ForRange<platform::CUDADeviceContext> for_range(ctx, num);
+  for_range(RangeInitFunctor{0, 1, idx_in});
+
   int *idx_out = index_out->mutable_data<int>({num}, ctx.GetPlace());
 
   const T *keys_in = value.data<T>();
   T *keys_out = value_out->mutable_data<T>({num}, ctx.GetPlace());
 
   // Determine temporary device storage requirements
-  void *d_temp_storage = NULL;
   size_t temp_storage_bytes = 0;
   cub::DeviceRadixSort::SortPairsDescending<T, int>(
-      d_temp_storage, temp_storage_bytes, keys_in, keys_out, idx_in, idx_out,
-      num);
+      nullptr, temp_storage_bytes, keys_in, keys_out, idx_in, idx_out, num);
 
   // Allocate temporary storage
   auto place = boost::get<platform::CUDAPlace>(ctx.GetPlace());
-  d_temp_storage = memory::Alloc(place, temp_storage_bytes);
+  void *d_temp_storage = memory::Alloc(place, temp_storage_bytes);
 
   // Run sorting operation
   cub::DeviceRadixSort::SortPairsDescending<T, int>(
@@ -76,22 +81,27 @@ void SortDescending(const platform::CUDADeviceContext &ctx, const Tensor &value,
 }
 
 template <typename T>
-__device__ __forceinline__ T Min(T x, T y) {
-  return x < y ? x : y;
-}
-
-template <typename T>
-__device__ __forceinline__ T Max(T x, T y) {
-  return x > y ? x : y;
-}
-
-template <typename T>
-__global__ void BoxDecodeAndClipKernel(const T *anchor, const T *deltas,
-                                       const T *var, const int *index,
-                                       const T *im_info, const int num,
-                                       T *proposals) {
-  T kBBoxClipDefault = log(1000.0 / 16.0);
-  CUDA_1D_KERNEL_LOOP(i, num) {
+struct BoxDecodeAndClipFunctor {
+  const T *anchor;
+  const T *deltas;
+  const T *var;
+  const int *index;
+  const T *im_info;
+
+  T *proposals;
+
+  BoxDecodeAndClipFunctor(const T *anchor, const T *deltas, const T *var,
+                          const int *index, const T *im_info, T *proposals)
+      : anchor(anchor),
+        deltas(deltas),
+        var(var),
+        index(index),
+        im_info(im_info),
+        proposals(proposals) {}
+
+  T bbox_clip_default{static_cast<T>(kBBoxClipDefault)};
+
+  __device__ void operator()(size_t i) {
     int k = index[i] * 4;
     T axmin = anchor[k];
     T aymin = anchor[k + 1];
@@ -108,17 +118,17 @@ __global__ void BoxDecodeAndClipKernel(const T *anchor, const T *deltas,
     T dxmax = deltas[k + 2];
     T dymax = deltas[k + 3];
 
-    T d_cx = 0., d_cy = 0., d_w = 0., d_h = 0.;
+    T d_cx, d_cy, d_w, d_h;
     if (var) {
       d_cx = cx + dxmin * w * var[k];
       d_cy = cy + dymin * h * var[k + 1];
-      d_w = exp(Min<T>(dxmax * var[k + 2], kBBoxClipDefault)) * w;
-      d_h = exp(Min<T>(dymax * var[k + 3], kBBoxClipDefault)) * h;
+      d_w = exp(Min(dxmax * var[k + 2], bbox_clip_default)) * w;
+      d_h = exp(Min(dymax * var[k + 3], bbox_clip_default)) * h;
     } else {
       d_cx = cx + dxmin * w;
       d_cy = cy + dymin * h;
-      d_w = exp(Min<T>(dxmax, kBBoxClipDefault)) * w;
-      d_h = exp(Min<T>(dymax, kBBoxClipDefault)) * h;
+      d_w = exp(Min(dxmax, bbox_clip_default)) * w;
+      d_h = exp(Min(dymax, bbox_clip_default)) * h;
     }
 
     T oxmin = d_cx - d_w * 0.5;
@@ -126,17 +136,21 @@ __global__ void BoxDecodeAndClipKernel(const T *anchor, const T *deltas,
     T oxmax = d_cx + d_w * 0.5 - 1.;
     T oymax = d_cy + d_h * 0.5 - 1.;
 
-    proposals[i * 4] = Max<T>(Min<T>(oxmin, im_info[1] - 1.), 0.);
-    proposals[i * 4 + 1] = Max<T>(Min<T>(oymin, im_info[0] - 1.), 0.);
-    proposals[i * 4 + 2] = Max<T>(Min<T>(oxmax, im_info[1] - 1.), 0.);
-    proposals[i * 4 + 3] = Max<T>(Min<T>(oymax, im_info[0] - 1.), 0.);
+    proposals[i * 4] = Max(Min(oxmin, im_info[1] - 1.), 0.);
+    proposals[i * 4 + 1] = Max(Min(oymin, im_info[0] - 1.), 0.);
+    proposals[i * 4 + 2] = Max(Min(oxmax, im_info[1] - 1.), 0.);
+    proposals[i * 4 + 3] = Max(Min(oymax, im_info[0] - 1.), 0.);
   }
-}
+
+  __device__ __forceinline__ T Min(T a, T b) const { return a > b ? b : a; }
+
+  __device__ __forceinline__ T Max(T a, T b) const { return a > b ? a : b; }
+};
 
 template <typename T, int BlockSize>
-__global__ void FilterBBoxes(const T *bboxes, const T *im_info,
-                             const T min_size, const int num, int *keep_num,
-                             int *keep) {
+static __global__ void FilterBBoxes(const T *bboxes, const T *im_info,
+                                    const T min_size, const int num,
+                                    int *keep_num, int *keep) {
   T im_h = im_info[0];
   T im_w = im_info[1];
   T im_scale = im_info[2];
@@ -181,7 +195,7 @@ __global__ void FilterBBoxes(const T *bboxes, const T *im_info,
   }
 }
 
-__device__ inline float IoU(const float *a, const float *b) {
+static __device__ inline float IoU(const float *a, const float *b) {
   float left = max(a[0], b[0]), right = min(a[2], b[2]);
   float top = max(a[1], b[1]), bottom = min(a[3], b[3]);
   float width = max(right - left + 1, 0.f), height = max(bottom - top + 1, 0.f);
@@ -191,7 +205,8 @@ __device__ inline float IoU(const float *a, const float *b) {
   return inter_s / (s_a + s_b - inter_s);
 }
 
-__global__ void NMSKernel(const int n_boxes, const float nms_overlap_thresh,
+static __global__ void NMSKernel(const int n_boxes,
+                                 const float nms_overlap_thresh,
                                  const float *dev_boxes, uint64_t *dev_mask) {
   const int row_start = blockIdx.y;
   const int col_start = blockIdx.x;
@@ -234,7 +249,7 @@ __global__ void NMSKernel(const int n_boxes, const float nms_overlap_thresh,
 }
 
 template <typename T>
-void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals,
+static void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals,
                 const Tensor &sorted_indices, const T nms_threshold,
                 Tensor *keep_out) {
   int boxes_num = proposals.dims()[0];
@@ -247,13 +262,10 @@ void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals,
 
   const T *boxes = proposals.data<T>();
   auto place = boost::get<platform::CUDAPlace>(ctx.GetPlace());
-  int size_bytes = boxes_num * col_blocks * sizeof(uint64_t);
-  uint64_t *d_mask =
-      reinterpret_cast<uint64_t *>(memory::Alloc(place, size_bytes));
-  NMSKernel<<<blocks, threads>>>(boxes_num, nms_threshold, boxes, d_mask);
-  uint64_t *h_mask = reinterpret_cast<uint64_t *>(
-      memory::Alloc(platform::CPUPlace(), size_bytes));
-  memory::Copy(platform::CPUPlace(), h_mask, place, d_mask, size_bytes, 0);
+  framework::Vector<uint64_t> mask(boxes_num * col_blocks);
+  NMSKernel<<<blocks, threads>>>(
+      boxes_num, nms_threshold, boxes,
+      mask.CUDAMutableData(boost::get<platform::CUDAPlace>(ctx.GetPlace())));
 
   std::vector<uint64_t> remv(col_blocks);
   memset(&remv[0], 0, sizeof(uint64_t) * col_blocks);
@@ -267,7 +279,7 @@ void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals,
     if (!(remv[nblock] & (1ULL << inblock))) {
       ++num_to_keep;
       keep_vec.push_back(i);
-      uint64_t *p = &h_mask[0] + i * col_blocks;
+      uint64_t *p = &mask[0] + i * col_blocks;
       for (int j = nblock; j < col_blocks; j++) {
         remv[j] |= p[j];
       }
@@ -276,12 +288,10 @@ void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals,
   int *keep = keep_out->mutable_data<int>({num_to_keep}, ctx.GetPlace());
   memory::Copy(place, keep, platform::CPUPlace(), keep_vec.data(),
                sizeof(int) * num_to_keep, 0);
-  memory::Free(place, d_mask);
-  memory::Free(platform::CPUPlace(), h_mask);
 }
 
 template <typename T>
-std::pair<Tensor, Tensor> ProposalForOneImage(
+static std::pair<Tensor, Tensor> ProposalForOneImage(
     const platform::CUDADeviceContext &ctx, const Tensor &im_info,
     const Tensor &anchors, const Tensor &variances,
     const Tensor &bbox_deltas,  // [M, 4]
@@ -300,18 +310,20 @@ std::pair<Tensor, Tensor> ProposalForOneImage(
   // 2. box decode and clipping
   Tensor proposals;
   proposals.mutable_data<T>({pre_nms_num, 4}, ctx.GetPlace());
-  int block = 512;
-  auto stream = ctx.stream();
-  BoxDecodeAndClipKernel<T><<<DIVUP(pre_nms_num, block), block, 0, stream>>>(
+
+  {
+    platform::ForRange<platform::CUDADeviceContext> for_range(ctx, pre_nms_num);
+    for_range(BoxDecodeAndClipFunctor<T>{
         anchors.data<T>(), bbox_deltas.data<T>(), variances.data<T>(),
-      index_sort.data<int>(), im_info.data<T>(), pre_nms_num,
-      proposals.data<T>());
+        index_sort.data<int>(), im_info.data<T>(), proposals.data<T>()});
+  }
 
   // 3. filter
   Tensor keep_index, keep_num_t;
   keep_index.mutable_data<int>({pre_nms_num}, ctx.GetPlace());
   keep_num_t.mutable_data<int>({1}, ctx.GetPlace());
   min_size = std::max(min_size, 1.0f);
+  auto stream = ctx.stream();
   FilterBBoxes<T, 512><<<1, 512, 0, stream>>>(
       proposals.data<T>(), im_info.data<T>(), min_size, pre_nms_num,
       keep_num_t.data<int>(), keep_index.data<int>());
@@ -355,8 +367,12 @@ class CUDAGenerateProposalsKernel : public framework::OpKernel<T> {
     auto *scores = context.Input<Tensor>("Scores");
     auto *bbox_deltas = context.Input<Tensor>("BboxDeltas");
     auto *im_info = context.Input<Tensor>("ImInfo");
-    auto *anchors = context.Input<Tensor>("Anchors");
-    auto *variances = context.Input<Tensor>("Variances");
+    auto anchors = detail::Ref(context.Input<Tensor>("Anchors"),
+                               "Cannot find input Anchors(%s) in scope",
+                               context.Inputs("Anchors")[0]);
+    auto variances = detail::Ref(context.Input<Tensor>("Variances"),
+                                 "Cannot find input Variances(%s) in scope",
+                                 context.Inputs("Variances")[0]);
 
     auto *rpn_rois = context.Output<LoDTensor>("RpnRois");
     auto *rpn_roi_probs = context.Output<LoDTensor>("RpnRoiProbs");
@@ -392,10 +408,8 @@ class CUDAGenerateProposalsKernel : public framework::OpKernel<T> {
     trans(dev_ctx, *bbox_deltas, &bbox_deltas_swap, axis);
     trans(dev_ctx, *scores, &scores_swap, axis);
 
-    Tensor *anchor = const_cast<framework::Tensor *>(anchors);
-    anchor->Resize({anchors->numel() / 4, 4});
-    Tensor *var = const_cast<framework::Tensor *>(variances);
-    var->Resize({var->numel() / 4, 4});
+    anchors.Resize({anchors.numel() / 4, 4});
+    variances.Resize({variances.numel() / 4, 4});
 
     rpn_rois->mutable_data<T>({bbox_deltas->numel() / 4, 4},
                               context.GetPlace());
@@ -417,12 +431,12 @@ class CUDAGenerateProposalsKernel : public framework::OpKernel<T> {
       scores_slice.Resize({h_score * w_score * c_score, 1});
 
       std::pair<Tensor, Tensor> box_score_pair =
-          ProposalForOneImage<T>(dev_ctx, im_info_slice, *anchor, *var,
+          ProposalForOneImage<T>(dev_ctx, im_info_slice, anchors, variances,
                                  bbox_deltas_slice, scores_slice, pre_nms_top_n,
                                  post_nms_top_n, nms_thresh, min_size, eta);
 
-      Tensor proposals = box_score_pair.first;
-      Tensor scores = box_score_pair.second;
+      Tensor &proposals = box_score_pair.first;
+      Tensor &scores = box_score_pair.second;
 
       memory::Copy(place, rpn_rois_data + num_proposals * 4, place,
                    proposals.data<T>(), sizeof(T) * proposals.numel(), 0);

paddle/fluid/operators/detection/gpc.cc

+// Copyright (c) 2018 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.
+
+/**
+ * @file src/gpc.cpp
+ * @author huhan02(com@baidu.com)
+ * @date 2015/12/18 14:17:30
+ * @brief
+ *
+ * @modified by sunyipeng
+ * @email sunyipeng@baidu.com
+ * @date 2018/6/12
+ **/
+
+#include "paddle/fluid/operators/detection/gpc.h"
+
+namespace gpc {
+
+typedef struct lmt_shape { /* Local minima table                */
+  double y;                /* Y coordinate at local minimum     */
+  edge_node *first_bound;  /* Pointer to bound list             */
+  struct lmt_shape *next;  /* Pointer to next local minimum     */
+} lmt_node;
+
+typedef struct sbt_t_shape { /* Scanbeam tree                     */
+  double y;                  /* Scanbeam node y value             */
+  struct sbt_t_shape *less;  /* Pointer to nodes with lower y     */
+  struct sbt_t_shape *more;  /* Pointer to nodes with higher y    */
+} sb_tree;
+
+typedef struct it_shape { /* Intersection table                */
+  edge_node *ie[2];       /* Intersecting edge (bundle) pair   */
+  gpc_vertex point;       /* Point of intersection             */
+  struct it_shape *next;  /* The next intersection table node  */
+} it_node;
+
+typedef struct st_shape { /* Sorted edge table                 */
+  edge_node *edge;        /* Pointer to AET edge               */
+  double xb;              /* Scanbeam bottom x coordinate      */
+  double xt;              /* Scanbeam top x coordinate         */
+  double dx;              /* Change in x for a unit y increase */
+  struct st_shape *prev;  /* Previous edge in sorted list      */
+} st_node;
+
+typedef struct bbox_shape { /* Contour axis-aligned bounding box */
+  double xmin;              /* Minimum x coordinate              */
+  double ymin;              /* Minimum y coordinate              */
+  double xmax;              /* Maximum x coordinate              */
+  double ymax;              /* Maximum y coordinate              */
+} bbox;
+
+/*
+===========================================================================
+                               Global Data
+===========================================================================
+*/
+
+/* Horizontal edge state transitions within scanbeam boundary */
+const h_state next_h_state[3][6] = {
+    /*        ABOVE     BELOW     CROSS */
+    /*        L   R     L   R     L   R */
+    /* NH */
+    {BH, TH, TH, BH, NH, NH},
+    /* BH */
+    {NH, NH, NH, NH, TH, TH},
+    /* TH */
+    {NH, NH, NH, NH, BH, BH}};
+
+/*
+===========================================================================
+                             Private Functions
+===========================================================================
+*/
+
+static void reset_it(it_node **it) {
+  it_node *itn;
+
+  while (*it) {
+    itn = (*it)->next;
+    gpc_free<it_node>(*it);
+    *it = itn;
+  }
+}
+
+static void reset_lmt(lmt_node **lmt) {
+  lmt_node *lmtn;
+
+  while (*lmt) {
+    lmtn = (*lmt)->next;
+    gpc_free<lmt_node>(*lmt);
+    *lmt = lmtn;
+  }
+}
+
+static void insert_bound(edge_node **b, edge_node *e) {
+  edge_node *existing_bound = NULL;
+
+  if (!*b) {
+    /* Link node e to the tail of the list */
+    *b = e;
+  } else {
+    /* Do primary sort on the x field */
+    if (e[0].bot.x < (*b)[0].bot.x) {
+      /* Insert a new node mid-list */
+      existing_bound = *b;
+      *b = e;
+      (*b)->next_bound = existing_bound;
+    } else {
+      if (e[0].bot.x == (*b)[0].bot.x) {
+        /* Do secondary sort on the dx field */
+        if (e[0].dx < (*b)[0].dx) {
+          /* Insert a new node mid-list */
+          existing_bound = *b;
+          *b = e;
+          (*b)->next_bound = existing_bound;
+        } else {
+          /* Head further down the list */
+          insert_bound(&((*b)->next_bound), e);
+        }
+      } else {
+        /* Head further down the list */
+        insert_bound(&((*b)->next_bound), e);
+      }
+    }
+  }
+}
+
+static edge_node **bound_list(lmt_node **lmt, double y) {
+  lmt_node *existing_node;
+
+  if (!*lmt) {
+    /* Add node onto the tail end of the LMT */
+    gpc_malloc<lmt_node>(*lmt, sizeof(lmt_node),
+                         const_cast<char *>("LMT insertion"));
+    (*lmt)->y = y;
+    (*lmt)->first_bound = NULL;
+    (*lmt)->next = NULL;
+    return &((*lmt)->first_bound);
+  } else if (y < (*lmt)->y) {
+    /* Insert a new LMT node before the current node */
+    existing_node = *lmt;
+    gpc_malloc<lmt_node>(*lmt, sizeof(lmt_node),
+                         const_cast<char *>("LMT insertion"));
+    (*lmt)->y = y;
+    (*lmt)->first_bound = NULL;
+    (*lmt)->next = existing_node;
+    return &((*lmt)->first_bound);
+  } else {
+    if (y > (*lmt)->y) {
+      /* Head further up the LMT */
+      return bound_list(&((*lmt)->next), y);
+    } else {
+      /* Use this existing LMT node */
+      return &((*lmt)->first_bound);
+    }
+  }
+}
+
+static void add_to_sbtree(int *entries, sb_tree **sbtree, double y) {
+  if (!*sbtree) {
+    /* Add a new tree node here */
+    gpc_malloc<sb_tree>(*sbtree, sizeof(sb_tree),
+                        const_cast<char *>("scanbeam tree insertion"));
+    (*sbtree)->y = y;
+    (*sbtree)->less = NULL;
+    (*sbtree)->more = NULL;
+    (*entries)++;
+  } else {
+    if ((*sbtree)->y > y) {
+      /* Head into the 'less' sub-tree */
+      add_to_sbtree(entries, &((*sbtree)->less), y);
+    } else {
+      if ((*sbtree)->y < y) {
+        /* Head into the 'more' sub-tree */
+        add_to_sbtree(entries, &((*sbtree)->more), y);
+      }
+    }
+  }
+}
+
+static void build_sbt(int *entries, double *sbt, sb_tree *sbtree) {
+  if (sbtree->less) {
+    build_sbt(entries, sbt, sbtree->less);
+  }
+  sbt[*entries] = sbtree->y;
+  (*entries)++;
+  if (sbtree->more) {
+    build_sbt(entries, sbt, sbtree->more);
+  }
+}
+
+static void free_sbtree(sb_tree **sbtree) {
+  if (*sbtree) {
+    free_sbtree(&((*sbtree)->less));
+    free_sbtree(&((*sbtree)->more));
+    gpc_free<sb_tree>(*sbtree);
+  }
+}
+
+static int count_optimal_vertices(gpc_vertex_list c) {
+  int result = 0;
+  int i = 0;
+
+  /* Ignore non-contributing contours */
+  if (c.num_vertices > 0) {
+    for (i = 0; i < c.num_vertices; i++) {
+      /* Ignore superfluous vertices embedded in horizontal edges */
+      if (gpc_optimal(c.vertex, i, c.num_vertices)) {
+        result++;
+      }
+    }
+  }
+  return result;
+}
+
+static edge_node *build_lmt(lmt_node **lmt, sb_tree **sbtree, int *sbt_entries,
+                            gpc_polygon *p, int type, gpc_op op) {
+  int c = 0;
+  int i = 0;
+  int min = 0;
+  int max = 0;
+  int num_edges = 0;
+  int v = 0;
+  int num_vertices = 0;
+  int total_vertices = 0;
+  int e_index = 0;
+  edge_node *e = NULL;
+  edge_node *edge_table = NULL;
+
+  for (c = 0; c < p->num_contours; c++) {
+    total_vertices += count_optimal_vertices(p->contour[c]);
+  }
+
+  /* Create the entire input polygon edge table in one go */
+  gpc_malloc<edge_node>(edge_table, total_vertices * sizeof(edge_node),
+                        const_cast<char *>("edge table creation"));
+
+  for (c = 0; c < p->num_contours; c++) {
+    if (p->contour[c].num_vertices < 0) {
+      /* Ignore the non-contributing contour and repair the vertex count */
+      p->contour[c].num_vertices = -p->contour[c].num_vertices;
+    } else {
+      /* Perform contour optimisation */
+      num_vertices = 0;
+      for (i = 0; i < p->contour[c].num_vertices; i++) {
+        if (gpc_optimal(p->contour[c].vertex, i, p->contour[c].num_vertices)) {
+          edge_table[num_vertices].vertex.x = p->contour[c].vertex[i].x;
+          edge_table[num_vertices].vertex.y = p->contour[c].vertex[i].y;
+
+          /* Record vertex in the scanbeam table */
+          add_to_sbtree(sbt_entries, sbtree, edge_table[num_vertices].vertex.y);
+
+          num_vertices++;
+        }
+      }
+
+      /* Do the contour forward pass */
+      for (min = 0; min < num_vertices; min++) {
+        /* If a forward local minimum... */
+        if (gpc_fwd_min(edge_table, min, num_vertices)) {
+          /* Search for the next local maximum... */
+          num_edges = 1;
+          max = gpc_next_index(min, num_vertices);
+          while (gpc_not_fmax(edge_table, max, num_vertices)) {
+            num_edges++;
+            max = gpc_next_index(max, num_vertices);
+          }
+
+          /* Build the next edge list */
+          e = &edge_table[e_index];
+          e_index += num_edges;
+          v = min;
+          e[0].bstate[BELOW] = UNBUNDLED;
+          e[0].bundle[BELOW][CLIP] = 0;
+          e[0].bundle[BELOW][SUBJ] = 0;
+          for (i = 0; i < num_edges; i++) {
+            e[i].xb = edge_table[v].vertex.x;
+            e[i].bot.x = edge_table[v].vertex.x;
+            e[i].bot.y = edge_table[v].vertex.y;
+
+            v = gpc_next_index(v, num_vertices);
+
+            e[i].top.x = edge_table[v].vertex.x;
+            e[i].top.y = edge_table[v].vertex.y;
+            e[i].dx = (edge_table[v].vertex.x - e[i].bot.x) /
+                      (e[i].top.y - e[i].bot.y);
+            e[i].type = type;
+            e[i].outp[ABOVE] = NULL;
+            e[i].outp[BELOW] = NULL;
+            e[i].next = NULL;
+            e[i].prev = NULL;
+            e[i].succ =
+                ((num_edges > 1) && (i < (num_edges - 1))) ? &(e[i + 1]) : NULL;
+            e[i].pred = ((num_edges > 1) && (i > 0)) ? &(e[i - 1]) : NULL;
+            e[i].next_bound = NULL;
+            e[i].bside[CLIP] = (op == GPC_DIFF) ? RIGHT : LEFT;
+            e[i].bside[SUBJ] = LEFT;
+          }
+          insert_bound(bound_list(lmt, edge_table[min].vertex.y), e);
+        }
+      }
+
+      /* Do the contour reverse pass */
+      for (min = 0; min < num_vertices; min++) {
+        /* If a reverse local minimum... */
+        if (gpc_rev_min(edge_table, min, num_vertices)) {
+          /* Search for the previous local maximum... */
+          num_edges = 1;
+          max = gpc_prev_index(min, num_vertices);
+          while (gpc_not_rmax(edge_table, max, num_vertices)) {
+            num_edges++;
+            max = gpc_prev_index(max, num_vertices);
+          }
+
+          /* Build the previous edge list */
+          e = &edge_table[e_index];
+          e_index += num_edges;
+          v = min;
+          e[0].bstate[BELOW] = UNBUNDLED;
+          e[0].bundle[BELOW][CLIP] = 0;
+          e[0].bundle[BELOW][SUBJ] = 0;
+          for (i = 0; i < num_edges; i++) {
+            e[i].xb = edge_table[v].vertex.x;
+            e[i].bot.x = edge_table[v].vertex.x;
+            e[i].bot.y = edge_table[v].vertex.y;
+
+            v = gpc_prev_index(v, num_vertices);
+
+            e[i].top.x = edge_table[v].vertex.x;
+            e[i].top.y = edge_table[v].vertex.y;
+            e[i].dx = (edge_table[v].vertex.x - e[i].bot.x) /
+                      (e[i].top.y - e[i].bot.y);
+            e[i].type = type;
+            e[i].outp[ABOVE] = NULL;
+            e[i].outp[BELOW] = NULL;
+            e[i].next = NULL;
+            e[i].prev = NULL;
+            e[i].succ =
+                ((num_edges > 1) && (i < (num_edges - 1))) ? &(e[i + 1]) : NULL;
+            e[i].pred = ((num_edges > 1) && (i > 0)) ? &(e[i - 1]) : NULL;
+            e[i].next_bound = NULL;
+            e[i].bside[CLIP] = (op == GPC_DIFF) ? RIGHT : LEFT;
+            e[i].bside[SUBJ] = LEFT;
+          }
+          insert_bound(bound_list(lmt, edge_table[min].vertex.y), e);
+        }
+      }
+    }
+  }
+  return edge_table;
+}  // NOLINT
+
+static void add_edge_to_aet(edge_node **aet, edge_node *edge, edge_node *prev) {
+  if (!*aet) {
+    /* Append edge onto the tail end of the AET */
+    *aet = edge;
+    edge->prev = prev;
+    edge->next = NULL;
+  } else {
+    /* Do primary sort on the xb field */
+    if (edge->xb < (*aet)->xb) {
+      /* Insert edge here (before the AET edge) */
+      edge->prev = prev;
+      edge->next = *aet;
+      (*aet)->prev = edge;
+      *aet = edge;
+    } else {
+      if (edge->xb == (*aet)->xb) {
+        /* Do secondary sort on the dx field */
+        if (edge->dx < (*aet)->dx) {
+          /* Insert edge here (before the AET edge) */
+          edge->prev = prev;
+          edge->next = *aet;
+          (*aet)->prev = edge;
+          *aet = edge;
+        } else {
+          /* Head further into the AET */
+          add_edge_to_aet(&((*aet)->next), edge, *aet);
+        }
+      } else {
+        /* Head further into the AET */
+        add_edge_to_aet(&((*aet)->next), edge, *aet);
+      }
+    }
+  }
+}
+
+static void add_intersection(it_node **it, edge_node *edge0, edge_node *edge1,
+                             double x, double y) {
+  it_node *existing_node;
+
+  if (!*it) {
+    /* Append a new node to the tail of the list */
+    gpc_malloc<it_node>(*it, sizeof(it_node),
+                        const_cast<char *>("IT insertion"));
+    (*it)->ie[0] = edge0;
+    (*it)->ie[1] = edge1;
+    (*it)->point.x = x;
+    (*it)->point.y = y;
+    (*it)->next = NULL;
+  } else {
+    if ((*it)->point.y > y) {
+      /* Insert a new node mid-list */
+      existing_node = *it;
+      gpc_malloc<it_node>(*it, sizeof(it_node),
+                          const_cast<char *>("IT insertion"));
+      (*it)->ie[0] = edge0;
+      (*it)->ie[1] = edge1;
+      (*it)->point.x = x;
+      (*it)->point.y = y;
+      (*it)->next = existing_node;
+    } else {
+      /* Head further down the list */
+      add_intersection(&((*it)->next), edge0, edge1, x, y);
+    }
+  }
+}
+
+static void add_st_edge(st_node **st, it_node **it, edge_node *edge,
+                        double dy) {
+  st_node *existing_node;
+  double den = 0.0;
+  double r = 0.0;
+  double x = 0.0;
+  double y = 0.0;
+
+  if (!*st) {
+    /* Append edge onto the tail end of the ST */
+    gpc_malloc<st_node>(*st, sizeof(st_node),
+                        const_cast<char *>("ST insertion"));
+    (*st)->edge = edge;
+    (*st)->xb = edge->xb;
+    (*st)->xt = edge->xt;
+    (*st)->dx = edge->dx;
+    (*st)->prev = NULL;
+  } else {
+    den = ((*st)->xt - (*st)->xb) - (edge->xt - edge->xb);
+
+    /* If new edge and ST edge don't cross */
+    if ((edge->xt >= (*st)->xt) || (edge->dx == (*st)->dx) ||
+        (fabs(den) <= DBL_EPSILON)) {
+      /* No intersection - insert edge here (before the ST edge) */
+      existing_node = *st;
+      gpc_malloc<st_node>(*st, sizeof(st_node),
+                          const_cast<char *>("ST insertion"));
+      (*st)->edge = edge;
+      (*st)->xb = edge->xb;
+      (*st)->xt = edge->xt;
+      (*st)->dx = edge->dx;
+      (*st)->prev = existing_node;
+    } else {
+      /* Compute intersection between new edge and ST edge */
+      r = (edge->xb - (*st)->xb) / den;
+      x = (*st)->xb + r * ((*st)->xt - (*st)->xb);
+      y = r * dy;
+
+      /* Insert the edge pointers and the intersection point in the IT */
+      add_intersection(it, (*st)->edge, edge, x, y);
+
+      /* Head further into the ST */
+      add_st_edge(&((*st)->prev), it, edge, dy);
+    }
+  }
+}
+
+static void build_intersection_table(it_node **it, edge_node *aet, double dy) {
+  st_node *st;
+  st_node *stp;
+  edge_node *edge = NULL;
+
+  /* Build intersection table for the current scanbeam */
+  reset_it(it);
+  st = NULL;
+
+  /* Process each AET edge */
+  for (edge = aet; edge; edge = edge->next) {
+    if ((edge->bstate[ABOVE] == BUNDLE_HEAD) || edge->bundle[ABOVE][CLIP] ||
+        edge->bundle[ABOVE][SUBJ]) {
+      add_st_edge(&st, it, edge, dy);
+    }
+  }
+
+  /* Free the sorted edge table */
+  while (st) {
+    stp = st->prev;
+    gpc_free<st_node>(st);
+    st = stp;
+  }
+}
+
+static int count_contours(polygon_node *polygon) {
+  int nc = 0;
+  int nv = 0;
+  vertex_node *v = NULL;
+  vertex_node *nextv = NULL;
+
+  for (nc = 0; polygon; polygon = polygon->next) {
+    if (polygon->active) {
+      /* Count the vertices in the current contour */
+      nv = 0;
+      for (v = polygon->proxy->v[LEFT]; v; v = v->next) {
+        nv++;
+      }
+
+      /* Record valid vertex counts in the active field */
+      if (nv > 2) {
+        polygon->active = nv;
+        nc++;
+      } else {
+        /* Invalid contour: just free the heap */
+        for (v = polygon->proxy->v[LEFT]; v; v = nextv) {
+          nextv = v->next;
+          gpc_free<vertex_node>(v);
+        }
+        polygon->active = 0;
+      }
+    }
+  }
+  return nc;
+}
+
+static void add_left(polygon_node *p, double x, double y) {
+  vertex_node *nv = NULL;
+
+  /* Create a new vertex node and set its fields */
+  gpc_malloc<vertex_node>(nv, sizeof(vertex_node),
+                          const_cast<char *>("vertex node creation"));
+  nv->x = x;
+  nv->y = y;
+
+  /* Add vertex nv to the left end of the polygon's vertex list */
+  nv->next = p->proxy->v[LEFT];
+
+  /* Update proxy->[LEFT] to point to nv */
+  p->proxy->v[LEFT] = nv;
+}
+
+static void merge_left(polygon_node *p, polygon_node *q, polygon_node *list) {
+  polygon_node *target = NULL;
+
+  /* Label contour as a hole */
+  q->proxy->hole = 1;
+
+  if (p->proxy != q->proxy) {
+    /* Assign p's vertex list to the left end of q's list */
+    p->proxy->v[RIGHT]->next = q->proxy->v[LEFT];
+    q->proxy->v[LEFT] = p->proxy->v[LEFT];
+
+    /* Redirect any p->proxy references to q->proxy */
+
+    for (target = p->proxy; list; list = list->next) {
+      if (list->proxy == target) {
+        list->active = 0;
+        list->proxy = q->proxy;
+      }
+    }
+  }
+}
+
+static void add_right(polygon_node *p, double x, double y) {
+  vertex_node *nv = NULL;
+
+  /* Create a new vertex node and set its fields */
+  gpc_malloc<vertex_node>(nv, sizeof(vertex_node),
+                          const_cast<char *>("vertex node creation"));
+  nv->x = x;
+  nv->y = y;
+  nv->next = NULL;
+
+  /* Add vertex nv to the right end of the polygon's vertex list */
+  p->proxy->v[RIGHT]->next = nv;
+
+  /* Update proxy->v[RIGHT] to point to nv */
+  p->proxy->v[RIGHT] = nv;
+}
+
+static void merge_right(polygon_node *p, polygon_node *q, polygon_node *list) {
+  polygon_node *target = NULL;
+
+  /* Label contour as external */
+  q->proxy->hole = 0;
+
+  if (p->proxy != q->proxy) {
+    /* Assign p's vertex list to the right end of q's list */
+    q->proxy->v[RIGHT]->next = p->proxy->v[LEFT];
+    q->proxy->v[RIGHT] = p->proxy->v[RIGHT];
+
+    /* Redirect any p->proxy references to q->proxy */
+    for (target = p->proxy; list; list = list->next) {
+      if (list->proxy == target) {
+        list->active = 0;
+        list->proxy = q->proxy;
+      }
+    }
+  }
+}
+
+static void add_local_min(polygon_node **p, edge_node *edge, double x,
+                          double y) {
+  polygon_node *existing_min = NULL;
+  vertex_node *nv = NULL;
+
+  existing_min = *p;
+
+  gpc_malloc<polygon_node>(*p, sizeof(polygon_node),
+                           const_cast<char *>("polygon node creation"));
+
+  /* Create a new vertex node and set its fields */
+  gpc_malloc<vertex_node>(nv, sizeof(vertex_node),
+                          const_cast<char *>("vertex node creation"));
+  nv->x = x;
+  nv->y = y;
+  nv->next = NULL;
+
+  /* Initialise proxy to point to p itself */
+  (*p)->proxy = (*p);
+  (*p)->active = 1;
+  (*p)->next = existing_min;
+
+  /* Make v[LEFT] and v[RIGHT] point to new vertex nv */
+  (*p)->v[LEFT] = nv;
+  (*p)->v[RIGHT] = nv;
+
+  /* Assign polygon p to the edge */
+  edge->outp[ABOVE] = *p;
+}
+
+static int count_tristrips(polygon_node *tn) {
+  int total = 0;
+
+  for (total = 0; tn; tn = tn->next) {
+    if (tn->active > 2) {
+      total++;
+    }
+  }
+  return total;
+}
+
+void add_vertex(vertex_node **t, double x, double y) {
+  if (!(*t)) {
+    gpc_malloc<vertex_node>(*t, sizeof(vertex_node),
+                            const_cast<char *>("tristrip vertex creation"));
+    (*t)->x = x;
+    (*t)->y = y;
+    (*t)->next = NULL;
+  } else {
+    /* Head further down the list */
+    add_vertex(&((*t)->next), x, y);
+  }
+}
+
+void gpc_vertex_create(edge_node *e, int p, int s, double x, double y) {
+  add_vertex(&(e->outp[p]->v[s]), x, y);
+  e->outp[p]->active++;
+}
+
+static void new_tristrip(polygon_node **tn, edge_node *edge, double x,
+                         double y) {
+  if (!(*tn)) {
+    gpc_malloc<polygon_node>(*tn, sizeof(polygon_node),
+                             const_cast<char *>("tristrip node creation"));
+    (*tn)->next = NULL;
+    (*tn)->v[LEFT] = NULL;
+    (*tn)->v[RIGHT] = NULL;
+    (*tn)->active = 1;
+    add_vertex(&((*tn)->v[LEFT]), x, y);
+    edge->outp[ABOVE] = *tn;
+  } else {
+    /* Head further down the list */
+    new_tristrip(&((*tn)->next), edge, x, y);
+  }
+}
+
+static bbox *create_contour_bboxes(gpc_polygon *p) {
+  bbox *box;
+  int c = 0;
+  int v = 0;
+
+  gpc_malloc<bbox>(box, p->num_contours * sizeof(bbox),
+                   const_cast<char *>("Bounding box creation"));
+
+  /* Construct contour bounding boxes */
+  for (c = 0; c < p->num_contours; c++) {
+    /* Initialise bounding box extent */
+    box[c].xmin = DBL_MAX;
+    box[c].ymin = DBL_MAX;
+    box[c].xmax = -DBL_MAX;
+    box[c].ymax = -DBL_MAX;
+
+    for (v = 0; v < p->contour[c].num_vertices; v++) {
+      /* Adjust bounding box */
+      if (p->contour[c].vertex[v].x < box[c].xmin) {
+        box[c].xmin = p->contour[c].vertex[v].x;
+      }
+      if (p->contour[c].vertex[v].y < box[c].ymin) {
+        box[c].ymin = p->contour[c].vertex[v].y;
+      }
+      if (p->contour[c].vertex[v].x > box[c].xmax) {
+        box[c].xmax = p->contour[c].vertex[v].x;
+      }
+      if (p->contour[c].vertex[v].y > box[c].ymax) {
+        box[c].ymax = p->contour[c].vertex[v].y;
+      }
+    }
+  }
+  return box;
+}
+
+static void minimax_test(gpc_polygon *subj, gpc_polygon *clip, gpc_op op) {
+  bbox *s_bbox;
+  bbox *c_bbox;
+  int s = 0;
+  int c = 0;
+  int *o_table = NULL;
+  int overlap = 0;
+
+  s_bbox = create_contour_bboxes(subj);
+  c_bbox = create_contour_bboxes(clip);
+
+  gpc_malloc<int>(o_table,
+                  subj->num_contours * clip->num_contours * sizeof(int),
+                  const_cast<char *>("overlap table creation"));
+
+  /* Check all subject contour bounding boxes against clip boxes */
+  for (s = 0; s < subj->num_contours; s++) {
+    for (c = 0; c < clip->num_contours; c++) {
+      o_table[c * subj->num_contours + s] =
+          (!((s_bbox[s].xmax < c_bbox[c].xmin) ||
+             (s_bbox[s].xmin > c_bbox[c].xmax))) &&
+          (!((s_bbox[s].ymax < c_bbox[c].ymin) ||
+             (s_bbox[s].ymin > c_bbox[c].ymax)));
+    }
+  }
+
+  /* For each clip contour, search for any subject contour overlaps */
+  for (c = 0; c < clip->num_contours; c++) {
+    overlap = 0;
+    for (s = 0; (!overlap) && (s < subj->num_contours); s++) {
+      overlap = o_table[c * subj->num_contours + s];
+    }
+
+    if (!overlap) {
+      /* Flag non contributing status by negating vertex count */
+      clip->contour[c].num_vertices = -clip->contour[c].num_vertices;
+    }
+  }
+
+  if (op == GPC_INT) {
+    /* For each subject contour, search for any clip contour overlaps */
+    for (s = 0; s < subj->num_contours; s++) {
+      overlap = 0;
+      for (c = 0; (!overlap) && (c < clip->num_contours); c++) {
+        overlap = o_table[c * subj->num_contours + s];
+      }
+
+      if (!overlap) {
+        /* Flag non contributing status by negating vertex count */
+        subj->contour[s].num_vertices = -subj->contour[s].num_vertices;
+      }
+    }
+  }
+
+  gpc_free<bbox>(s_bbox);
+  gpc_free<bbox>(c_bbox);
+  gpc_free<int>(o_table);
+}
+
+/*
+===========================================================================
+                             Public Functions
+===========================================================================
+*/
+
+void gpc_free_polygon(gpc_polygon *p) {
+  int c = 0;
+
+  for (c = 0; c < p->num_contours; c++) {
+    gpc_free<gpc_vertex>(p->contour[c].vertex);
+  }
+  gpc_free<int>(p->hole);
+  gpc_free<gpc_vertex_list>(p->contour);
+  p->num_contours = 0;
+}
+
+/*
+void gpc_read_polygon(FILE *fp, int read_hole_flags, gpc_polygon *p) {
+  int c = 0;
+  int v = 0;
+
+  fscanf(fp, "%d", &(p->num_contours));
+  gpc_malloc<int>(p->hole, p->num_contours * sizeof(int),
+                  (char *)"hole flag array creation");
+  gpc_malloc<gpc_vertex_list>(p->contour,
+                              p->num_contours * sizeof(gpc_vertex_list),
+                              (char *)"contour creation");
+  for (c = 0; c < p->num_contours; c++) {
+    fscanf(fp, "%d", &(p->contour[c].num_vertices));
+
+    if (read_hole_flags) {
+      fscanf(fp, "%d", &(p->hole[c]));
+    } else {
+      p->hole[c] = 0; // Assume all contours to be external
+    }
+
+    gpc_malloc<gpc_vertex>(p->contour[c].vertex,
+                           p->contour[c].num_vertices * sizeof(gpc_vertex),
+                           (char *)"vertex creation");
+    for (v = 0; v < p->contour[c].num_vertices; v++) {
+      fscanf(fp, "%lf %lf", &(p->contour[c].vertex[v].x),
+             &(p->contour[c].vertex[v].y));
+    }
+  }
+}
+
+void gpc_write_polygon(FILE *fp, int write_hole_flags, gpc_polygon *p) {
+  int c = 0;
+  int v = 0;
+
+  fprintf(fp, "%d\n", p->num_contours);
+  for (c = 0; c < p->num_contours; c++) {
+    fprintf(fp, "%d\n", p->contour[c].num_vertices);
+
+    if (write_hole_flags) {
+      fprintf(fp, "%d\n", p->hole[c]);
+    }
+
+    for (v = 0; v < p->contour[c].num_vertices; v++) {
+      fprintf(fp, "% .*lf % .*lf\n", DBL_DIG, p->contour[c].vertex[v].x,
+              DBL_DIG, p->contour[c].vertex[v].y);
+    }
+  }
+}
+*/
+
+void gpc_add_contour(gpc_polygon *p, gpc_vertex_list *new_contour, int hole) {
+  int *extended_hole = NULL;
+  int c = 0;
+  int v = 0;
+  gpc_vertex_list *extended_contour = NULL;
+
+  /* Create an extended hole array */
+  gpc_malloc<int>(extended_hole, (p->num_contours + 1) * sizeof(int),
+                  const_cast<char *>("contour hole addition"));
+
+  /* Create an extended contour array */
+  gpc_malloc<gpc_vertex_list>(extended_contour,
+                              (p->num_contours + 1) * sizeof(gpc_vertex_list),
+                              const_cast<char *>("contour addition"));
+
+  /* Copy the old contour and hole data into the extended arrays */
+  for (c = 0; c < p->num_contours; c++) {
+    extended_hole[c] = p->hole[c];
+    extended_contour[c] = p->contour[c];
+  }
+
+  /* Copy the new contour and hole onto the end of the extended arrays */
+  c = p->num_contours;
+  extended_hole[c] = hole;
+  extended_contour[c].num_vertices = new_contour->num_vertices;
+  gpc_malloc<gpc_vertex>(extended_contour[c].vertex,
+                         new_contour->num_vertices * sizeof(gpc_vertex),
+                         const_cast<char *>("contour addition"));
+  for (v = 0; v < new_contour->num_vertices; v++) {
+    extended_contour[c].vertex[v] = new_contour->vertex[v];
+  }
+
+  /* Dispose of the old contour */
+  gpc_free<gpc_vertex_list>(p->contour);
+  gpc_free<int>(p->hole);
+
+  /* Update the polygon information */
+  p->num_contours++;
+  p->hole = extended_hole;
+  p->contour = extended_contour;
+}
+
+// gpc_polygon_clip
+void gpc_polygon_clip(gpc_op op, gpc_polygon *subj, gpc_polygon *clip,
+                      gpc_polygon *result) {
+  sb_tree *sbtree = NULL;
+  it_node *it = NULL;
+  it_node *intersect = NULL;
+  edge_node *edge = NULL;
+  edge_node *prev_edge = NULL;
+  edge_node *next_edge = NULL;
+  edge_node *succ_edge = NULL;
+  edge_node *e0 = NULL;
+  edge_node *e1 = NULL;
+  edge_node *aet = NULL;
+  edge_node *c_heap = NULL;
+  edge_node *s_heap = NULL;
+  lmt_node *lmt = NULL;
+  lmt_node *local_min = NULL;
+  polygon_node *out_poly = NULL;
+  polygon_node *p = NULL;
+  polygon_node *q = NULL;
+  polygon_node *poly = NULL;
+  polygon_node *npoly = NULL;
+  polygon_node *cf = NULL;
+  vertex_node *vtx = NULL;
+  vertex_node *nv = NULL;
+  h_state horiz[2];
+  int in[2];
+  int exists[2];
+  int parity[2] = {LEFT, LEFT};
+  int c = 0;
+  int v = 0;
+  int contributing = 0;
+  int search = 0;
+  int scanbeam = 0;
+  int sbt_entries = 0;
+  int vclass = 0;
+  int bl = 0;
+  int br = 0;
+  int tl = 0;
+  int tr = 0;
+  double *sbt = NULL;
+  double xb = 0.0;
+  double px = 0.0;
+  double yb = 0.0;
+  double yt = 0.0;
+  double dy = 0.0;
+  double ix = 0.0;
+  double iy = 0.0;
+
+  /* Test for trivial NULL result cases */
+  if (((subj->num_contours == 0) && (clip->num_contours == 0)) ||
+      ((subj->num_contours == 0) && ((op == GPC_INT) || (op == GPC_DIFF))) ||
+      ((clip->num_contours == 0) && (op == GPC_INT))) {
+    result->num_contours = 0;
+    result->hole = NULL;
+    result->contour = NULL;
+    return;
+  }
+  /* Identify potentialy contributing contours */
+  if (((op == GPC_INT) || (op == GPC_DIFF)) && (subj->num_contours > 0) &&
+      (clip->num_contours > 0)) {
+    minimax_test(subj, clip, op);
+  }
+  /* Build LMT */
+  if (subj->num_contours > 0) {
+    s_heap = build_lmt(&lmt, &sbtree, &sbt_entries, subj, SUBJ, op);
+  }
+  if (clip->num_contours > 0) {
+    c_heap = build_lmt(&lmt, &sbtree, &sbt_entries, clip, CLIP, op);
+  }
+  /* Return a NULL result if no contours contribute */
+  if (lmt == NULL) {
+    result->num_contours = 0;
+    result->hole = NULL;
+    result->contour = NULL;
+    reset_lmt(&lmt);
+    gpc_free<edge_node>(s_heap);
+    gpc_free<edge_node>(c_heap);
+    return;
+  }
+
+  /* Build scanbeam table from scanbeam tree */
+  gpc_malloc<double>(sbt, sbt_entries * sizeof(double),
+                     const_cast<char *>("sbt creation"));
+  build_sbt(&scanbeam, sbt, sbtree);
+  scanbeam = 0;
+  free_sbtree(&sbtree);
+  /* Allow pointer re-use without causing memory leak */
+  if (subj == result) {
+    gpc_free_polygon(subj);
+  }
+  if (clip == result) {
+    gpc_free_polygon(clip);
+  }
+  /* Invert clip polygon for difference operation */
+  if (op == GPC_DIFF) {
+    parity[CLIP] = RIGHT;
+  }
+  local_min = lmt;
+
+  // Process each scanbeam
+  while (scanbeam < sbt_entries) {
+    /* Set yb and yt to the bottom and top of the scanbeam */
+    yb = sbt[scanbeam++];
+    if (scanbeam < sbt_entries) {
+      yt = sbt[scanbeam];
+      dy = yt - yb;
+    }
+    /* === SCANBEAM BOUNDARY PROCESSING ================================ */
+    /* If LMT node corresponding to yb exists */
+    if (local_min) {
+      if (local_min->y == yb) {
+        /* Add edges starting at this local minimum to the AET */
+        for (edge = local_min->first_bound; edge; edge = edge->next_bound) {
+          add_edge_to_aet(&aet, edge, NULL);
+        }
+        local_min = local_min->next;
+      }
+    }
+    /* Set dummy previous x value */
+    px = -DBL_MAX;
+    /* Create bundles within AET */
+    e0 = aet;
+    e1 = aet;
+    /* Set up bundle fields of first edge */
+    aet->bundle[ABOVE][aet->type] = (aet->top.y != yb);
+    aet->bundle[ABOVE][!aet->type] = 0;
+    aet->bstate[ABOVE] = UNBUNDLED;
+
+    for (next_edge = aet->next; next_edge; next_edge = next_edge->next) {
+      /* Set up bundle fields of next edge */
+      next_edge->bundle[ABOVE][next_edge->type] = (next_edge->top.y != yb);
+      next_edge->bundle[ABOVE][!next_edge->type] = 0;
+      next_edge->bstate[ABOVE] = UNBUNDLED;
+      /* Bundle edges above the scanbeam boundary if they coincide */
+      if (next_edge->bundle[ABOVE][next_edge->type]) {
+        if (gpc_eq(e0->xb, next_edge->xb) && gpc_eq(e0->dx, next_edge->dx) &&
+            (e0->top.y != yb)) {
+          next_edge->bundle[ABOVE][next_edge->type] ^=
+              e0->bundle[ABOVE][next_edge->type];
+          next_edge->bundle[ABOVE][!next_edge->type] =
+              e0->bundle[ABOVE][!next_edge->type];
+          next_edge->bstate[ABOVE] = BUNDLE_HEAD;
+          e0->bundle[ABOVE][CLIP] = 0;
+          e0->bundle[ABOVE][SUBJ] = 0;
+          e0->bstate[ABOVE] = BUNDLE_TAIL;
+        }
+        e0 = next_edge;
+      }
+    }
+    horiz[CLIP] = NH;
+    horiz[SUBJ] = NH;
+
+    // Process each edge at this scanbeam boundary
+    for (edge = aet; edge; edge = edge->next) {
+      exists[CLIP] =
+          edge->bundle[ABOVE][CLIP] + (edge->bundle[BELOW][CLIP] << 1);
+      exists[SUBJ] =
+          edge->bundle[ABOVE][SUBJ] + (edge->bundle[BELOW][SUBJ] << 1);
+      if (exists[CLIP] || exists[SUBJ]) {
+        /* Set bundle side */
+        edge->bside[CLIP] = parity[CLIP];
+        edge->bside[SUBJ] = parity[SUBJ];
+        /* Determine contributing status and quadrant occupancies */
+        switch (op) {
+          case GPC_DIFF:
+          case GPC_INT:
+            contributing = (exists[CLIP] && (parity[SUBJ] || horiz[SUBJ])) ||
+                           (exists[SUBJ] && (parity[CLIP] || horiz[CLIP])) ||
+                           (exists[CLIP] && exists[SUBJ] &&
+                            (parity[CLIP] == parity[SUBJ]));
+            br = (parity[CLIP]) && (parity[SUBJ]);
+            bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) &&
+                 (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]);
+            tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) &&
+                 (parity[SUBJ] ^ (horiz[SUBJ] != NH));
+            tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^
+                  edge->bundle[BELOW][CLIP]) &&
+                 (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^
+                  edge->bundle[BELOW][SUBJ]);
+            break;
+          case GPC_XOR:
+            contributing = exists[CLIP] || exists[SUBJ];
+            br = (parity[CLIP]) ^ (parity[SUBJ]);
+            bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) ^
+                 (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]);
+            tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) ^
+                 (parity[SUBJ] ^ (horiz[SUBJ] != NH));
+            tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^
+                  edge->bundle[BELOW][CLIP]) ^
+                 (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^
+                  edge->bundle[BELOW][SUBJ]);
+            break;
+          case GPC_UNION:
+            contributing = (exists[CLIP] && (!parity[SUBJ] || horiz[SUBJ])) ||
+                           (exists[SUBJ] && (!parity[CLIP] || horiz[CLIP])) ||
+                           (exists[CLIP] && exists[SUBJ] &&
+                            (parity[CLIP] == parity[SUBJ]));
+            br = (parity[CLIP]) || (parity[SUBJ]);
+            bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) ||
+                 (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]);
+            tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) ||
+                 (parity[SUBJ] ^ (horiz[SUBJ] != NH));
+            tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^
+                  edge->bundle[BELOW][CLIP]) ||
+                 (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^
+                  edge->bundle[BELOW][SUBJ]);
+            break;
+        }
+        // Update parity
+        parity[CLIP] ^= edge->bundle[ABOVE][CLIP];
+        parity[SUBJ] ^= edge->bundle[ABOVE][SUBJ];
+        /* Update horizontal state */
+        if (exists[CLIP]) {
+          horiz[CLIP] = next_h_state[horiz[CLIP]]
+                                    [((exists[CLIP] - 1) << 1) + parity[CLIP]];
+        }
+        if (exists[SUBJ]) {
+          horiz[SUBJ] = next_h_state[horiz[SUBJ]]
+                                    [((exists[SUBJ] - 1) << 1) + parity[SUBJ]];
+        }
+        vclass = tr + (tl << 1) + (br << 2) + (bl << 3);
+        if (contributing) {
+          xb = edge->xb;
+          switch (vclass) {
+            case EMN:
+            case IMN:
+              add_local_min(&out_poly, edge, xb, yb);
+              px = xb;
+              cf = edge->outp[ABOVE];
+              break;
+            case ERI:
+              if (xb != px) {
+                add_right(cf, xb, yb);
+                px = xb;
+              }
+              edge->outp[ABOVE] = cf;
+              cf = NULL;
+              break;
+            case ELI:
+              add_left(edge->outp[BELOW], xb, yb);
+              px = xb;
+              cf = edge->outp[BELOW];
+              break;
+            case EMX:
+              if (xb != px) {
+                add_left(cf, xb, yb);
+                px = xb;
+              }
+              merge_right(cf, edge->outp[BELOW], out_poly);
+              cf = NULL;
+              break;
+            case ILI:
+              if (xb != px) {
+                add_left(cf, xb, yb);
+                px = xb;
+              }
+              edge->outp[ABOVE] = cf;
+              cf = NULL;
+              break;
+            case IRI:
+              add_right(edge->outp[BELOW], xb, yb);
+              px = xb;
+              cf = edge->outp[BELOW];
+              edge->outp[BELOW] = NULL;
+              break;
+            case IMX:
+              if (xb != px) {
+                add_right(cf, xb, yb);
+                px = xb;
+              }
+              merge_left(cf, edge->outp[BELOW], out_poly);
+              cf = NULL;
+              edge->outp[BELOW] = NULL;
+              break;
+            case IMM:
+              if (xb != px) {
+                add_right(cf, xb, yb);
+                px = xb;
+              }
+              merge_left(cf, edge->outp[BELOW], out_poly);
+              edge->outp[BELOW] = NULL;
+              add_local_min(&out_poly, edge, xb, yb);
+              cf = edge->outp[ABOVE];
+              break;
+            case EMM:
+              if (xb != px) {
+                add_left(cf, xb, yb);
+                px = xb;
+              }
+              merge_right(cf, edge->outp[BELOW], out_poly);
+              edge->outp[BELOW] = NULL;
+              add_local_min(&out_poly, edge, xb, yb);
+              cf = edge->outp[ABOVE];
+              break;
+            case LED:
+              if (edge->bot.y == yb) {
+                add_left(edge->outp[BELOW], xb, yb);
+              }
+              edge->outp[ABOVE] = edge->outp[BELOW];
+              px = xb;
+              break;
+            case RED:
+              if (edge->bot.y == yb) {
+                add_right(edge->outp[BELOW], xb, yb);
+              }
+              edge->outp[ABOVE] = edge->outp[BELOW];
+              px = xb;
+              break;
+            default:
+              break;
+          } /* End of switch */
+        }   /* End of contributing conditional */
+      }     /* End of edge exists conditional */
+    }       // End of AET loop
+
+    /* Delete terminating edges from the AET, otherwise compute xt */
+    for (edge = aet; edge; edge = edge->next) {
+      if (edge->top.y == yb) {
+        prev_edge = edge->prev;
+        next_edge = edge->next;
+        if (prev_edge) {
+          prev_edge->next = next_edge;
+        } else {
+          aet = next_edge;
+        }
+        if (next_edge) {
+          next_edge->prev = prev_edge;
+        }
+        /* Copy bundle head state to the adjacent tail edge if required */
+        if ((edge->bstate[BELOW] == BUNDLE_HEAD) && prev_edge) {
+          if (prev_edge->bstate[BELOW] == BUNDLE_TAIL) {
+            prev_edge->outp[BELOW] = edge->outp[BELOW];
+            prev_edge->bstate[BELOW] = UNBUNDLED;
+            if (prev_edge->prev) {
+              if (prev_edge->prev->bstate[BELOW] == BUNDLE_TAIL) {
+                prev_edge->bstate[BELOW] = BUNDLE_HEAD;
+              }
+            }
+          }
+        }
+      } else {
+        if (edge->top.y == yt) {
+          edge->xt = edge->top.x;
+        } else {
+          edge->xt = edge->bot.x + edge->dx * (yt - edge->bot.y);
+        }
+      }
+    }
+
+    if (scanbeam < sbt_entries) {
+      /* === SCANBEAM INTERIOR PROCESSING ============================== */
+      build_intersection_table(&it, aet, dy);
+      /* Process each node in the intersection table */
+      for (intersect = it; intersect; intersect = intersect->next) {
+        e0 = intersect->ie[0];
+        e1 = intersect->ie[1];
+        /* Only generate output for contributing intersections */
+        if ((e0->bundle[ABOVE][CLIP] || e0->bundle[ABOVE][SUBJ]) &&
+            (e1->bundle[ABOVE][CLIP] || e1->bundle[ABOVE][SUBJ])) {
+          p = e0->outp[ABOVE];
+          q = e1->outp[ABOVE];
+          ix = intersect->point.x;
+          iy = intersect->point.y + yb;
+
+          in[CLIP] = (e0->bundle[ABOVE][CLIP] && !e0->bside[CLIP]) ||
+                     (e1->bundle[ABOVE][CLIP] && e1->bside[CLIP]) ||
+                     (!e0->bundle[ABOVE][CLIP] && !e1->bundle[ABOVE][CLIP] &&
+                      e0->bside[CLIP] && e1->bside[CLIP]);
+          in[SUBJ] = (e0->bundle[ABOVE][SUBJ] && !e0->bside[SUBJ]) ||
+                     (e1->bundle[ABOVE][SUBJ] && e1->bside[SUBJ]) ||
+                     (!e0->bundle[ABOVE][SUBJ] && !e1->bundle[ABOVE][SUBJ] &&
+                      e0->bside[SUBJ] && e1->bside[SUBJ]);
+
+          // Determine quadrant occupancies
+          switch (op) {
+            case GPC_DIFF:
+            case GPC_INT:
+              tr = (in[CLIP]) && (in[SUBJ]);
+              tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) &&
+                   (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]);
+              br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) &&
+                   (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]);
+              bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^
+                    e0->bundle[ABOVE][CLIP]) &&
+                   (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^
+                    e0->bundle[ABOVE][SUBJ]);
+              break;
+            case GPC_XOR:
+              tr = (in[CLIP]) ^ (in[SUBJ]);
+              tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) ^
+                   (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]);
+              br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) ^
+                   (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]);
+              bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^
+                    e0->bundle[ABOVE][CLIP]) ^
+                   (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^
+                    e0->bundle[ABOVE][SUBJ]);
+              break;
+            case GPC_UNION:
+              tr = (in[CLIP]) || (in[SUBJ]);
+              tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) ||
+                   (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]);
+              br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) ||
+                   (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]);
+              bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^
+                    e0->bundle[ABOVE][CLIP]) ||
+                   (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^
+                    e0->bundle[ABOVE][SUBJ]);
+              break;
+          }
+          vclass = tr + (tl << 1) + (br << 2) + (bl << 3);
+          switch (vclass) {
+            case EMN:
+              add_local_min(&out_poly, e0, ix, iy);
+              e1->outp[ABOVE] = e0->outp[ABOVE];
+              break;
+            case ERI:
+              if (p) {
+                add_right(p, ix, iy);
+                e1->outp[ABOVE] = p;
+                e0->outp[ABOVE] = NULL;
+              }
+              break;
+            case ELI:
+              if (q) {
+                add_left(q, ix, iy);
+                e0->outp[ABOVE] = q;
+                e1->outp[ABOVE] = NULL;
+              }
+              break;
+            case EMX:
+              if (p && q) {
+                add_left(p, ix, iy);
+                merge_right(p, q, out_poly);
+                e0->outp[ABOVE] = NULL;
+                e1->outp[ABOVE] = NULL;
+              }
+              break;
+            case IMN:
+              add_local_min(&out_poly, e0, ix, iy);
+              e1->outp[ABOVE] = e0->outp[ABOVE];
+              break;
+            case ILI:
+              if (p) {
+                add_left(p, ix, iy);
+                e1->outp[ABOVE] = p;
+                e0->outp[ABOVE] = NULL;
+              }
+              break;
+            case IRI:
+              if (q) {
+                add_right(q, ix, iy);
+                e0->outp[ABOVE] = q;
+                e1->outp[ABOVE] = NULL;
+              }
+              break;
+            case IMX:
+              if (p && q) {
+                add_right(p, ix, iy);
+                merge_left(p, q, out_poly);
+                e0->outp[ABOVE] = NULL;
+                e1->outp[ABOVE] = NULL;
+              }
+              break;
+            case IMM:
+              if (p && q) {
+                add_right(p, ix, iy);
+                merge_left(p, q, out_poly);
+                add_local_min(&out_poly, e0, ix, iy);
+                e1->outp[ABOVE] = e0->outp[ABOVE];
+              }
+              break;
+            case EMM:
+              if (p && q) {
+                add_left(p, ix, iy);
+                merge_right(p, q, out_poly);
+                add_local_min(&out_poly, e0, ix, iy);
+                e1->outp[ABOVE] = e0->outp[ABOVE];
+              }
+              break;
+            default:
+              break;
+          }  // End of switch
+        }    /* End of contributing intersection conditional */
+
+        /* Swap bundle sides in response to edge crossing */
+        if (e0->bundle[ABOVE][CLIP]) {
+          e1->bside[CLIP] = !e1->bside[CLIP];
+        }
+        if (e1->bundle[ABOVE][CLIP]) {
+          e0->bside[CLIP] = !e0->bside[CLIP];
+        }
+        if (e0->bundle[ABOVE][SUBJ]) {
+          e1->bside[SUBJ] = !e1->bside[SUBJ];
+        }
+        if (e1->bundle[ABOVE][SUBJ]) {
+          e0->bside[SUBJ] = !e0->bside[SUBJ];
+        }
+
+        /* Swap e0 and e1 bundles in the AET */
+        prev_edge = e0->prev;
+        next_edge = e1->next;
+        if (next_edge) {
+          next_edge->prev = e0;
+        }
+        if (e0->bstate[ABOVE] == BUNDLE_HEAD) {
+          search = 1;
+          while (search) {
+            prev_edge = prev_edge->prev;
+            if (prev_edge) {
+              if (prev_edge->bstate[ABOVE] != BUNDLE_TAIL) {
+                search = 0;
+              }
+            } else {
+              search = 0;
+            }
+          }
+        }
+        if (!prev_edge) {
+          aet->prev = e1;
+          e1->next = aet;
+          aet = e0->next;
+        } else {
+          prev_edge->next->prev = e1;
+          e1->next = prev_edge->next;
+          prev_edge->next = e0->next;
+        }
+        e0->next->prev = prev_edge;
+        e1->next->prev = e1;
+        e0->next = next_edge;
+      } /* End of IT loop*/
+
+      // Prepare for next scanbeam
+      for (edge = aet; edge; edge = next_edge) {
+        next_edge = edge->next;
+        succ_edge = edge->succ;
+        if ((edge->top.y == yt) && succ_edge) {
+          /* Replace AET edge by its successor */
+          succ_edge->outp[BELOW] = edge->outp[ABOVE];
+          succ_edge->bstate[BELOW] = edge->bstate[ABOVE];
+          succ_edge->bundle[BELOW][CLIP] = edge->bundle[ABOVE][CLIP];
+          succ_edge->bundle[BELOW][SUBJ] = edge->bundle[ABOVE][SUBJ];
+          prev_edge = edge->prev;
+          if (prev_edge) {
+            prev_edge->next = succ_edge;
+          } else {
+            aet = succ_edge;
+          }
+          if (next_edge) {
+            next_edge->prev = succ_edge;
+          }
+          succ_edge->prev = prev_edge;
+          succ_edge->next = next_edge;
+        } else {
+          /* Update this edge */
+          edge->outp[BELOW] = edge->outp[ABOVE];
+          edge->bstate[BELOW] = edge->bstate[ABOVE];
+          edge->bundle[BELOW][CLIP] = edge->bundle[ABOVE][CLIP];
+          edge->bundle[BELOW][SUBJ] = edge->bundle[ABOVE][SUBJ];
+          edge->xb = edge->xt;
+        }
+        edge->outp[ABOVE] = NULL;
+      }
+    }
+  } /* === END OF SCANBEAM PROCESSING ================================== */
+  // Generate result polygon from out_poly
+  result->contour = NULL;
+  result->hole = NULL;
+  result->num_contours = count_contours(out_poly);
+  if (result->num_contours > 0) {
+    gpc_malloc<int>(result->hole, result->num_contours * sizeof(int),
+                    const_cast<char *>("hole flag table creation"));
+    gpc_malloc<gpc_vertex_list>(result->contour,
+                                result->num_contours * sizeof(gpc_vertex_list),
+                                const_cast<char *>("contour creation"));
+
+    c = 0;
+    for (poly = out_poly; poly; poly = npoly) {
+      npoly = poly->next;
+      if (poly->active) {
+        result->hole[c] = poly->proxy->hole;
+        result->contour[c].num_vertices = poly->active;
+        gpc_malloc<gpc_vertex>(
+            result->contour[c].vertex,
+            result->contour[c].num_vertices * sizeof(gpc_vertex),
+            const_cast<char *>("vertex creation"));
+
+        v = result->contour[c].num_vertices - 1;
+        for (vtx = poly->proxy->v[LEFT]; vtx; vtx = nv) {
+          nv = vtx->next;
+          result->contour[c].vertex[v].x = vtx->x;
+          result->contour[c].vertex[v].y = vtx->y;
+          gpc_free<vertex_node>(vtx);
+          v--;
+        }
+        c++;
+      }
+      gpc_free<polygon_node>(poly);
+    }
+  } else {
+    for (poly = out_poly; poly; poly = npoly) {
+      npoly = poly->next;
+      gpc_free<polygon_node>(poly);
+    }
+  }
+
+  // Tidy up
+  reset_it(&it);
+  reset_lmt(&lmt);
+  gpc_free<edge_node>(c_heap);
+  gpc_free<edge_node>(s_heap);
+  gpc_free<double>(sbt);
+}  // NOLINT
+
+void gpc_free_tristrip(gpc_tristrip *t) {
+  int s = 0;
+  for (s = 0; s < t->num_strips; s++) {
+    gpc_free<gpc_vertex>(t->strip[s].vertex);
+  }
+  gpc_free<gpc_vertex_list>(t->strip);
+  t->num_strips = 0;
+}
+
+void gpc_polygon_to_tristrip(gpc_polygon *s, gpc_tristrip *t) {
+  gpc_polygon c;
+  c.num_contours = 0;
+  c.hole = NULL;
+  c.contour = NULL;
+  gpc_tristrip_clip(GPC_DIFF, s, &c, t);
+}
+
+// gpc_tristrip_clip
+void gpc_tristrip_clip(gpc_op op, gpc_polygon *subj, gpc_polygon *clip,
+                       gpc_tristrip *result) {
+  sb_tree *sbtree = NULL;
+  it_node *it = NULL;
+  it_node *intersect = NULL;
+  edge_node *edge = NULL;
+  edge_node *prev_edge = NULL;
+  edge_node *next_edge = NULL;
+  edge_node *succ_edge = NULL;
+  edge_node *e0 = NULL;
+  edge_node *e1 = NULL;
+  edge_node *aet = NULL;
+  edge_node *c_heap = NULL;
+  edge_node *s_heap = NULL;
+  edge_node *cf = NULL;
+  lmt_node *lmt = NULL;
+  lmt_node *local_min = NULL;
+  polygon_node *tlist = NULL;
+  polygon_node *tn = NULL;
+  polygon_node *tnn = NULL;
+  polygon_node *p = NULL;
+  polygon_node *q = NULL;
+  vertex_node *lt = NULL;
+  vertex_node *ltn = NULL;
+  vertex_node *rt = NULL;
+  vertex_node *rtn = NULL;
+  h_state horiz[2];
+  vertex_type cft = NUL;
+  int in[2];
+  int exists[2];
+  int parity[2] = {LEFT, LEFT};
+  int s = 0;
+  int v = 0;
+  int contributing = 0;
+  int search = 0;
+  int scanbeam = 0;
+  int sbt_entries = 0;
+  int vclass = 0;
+  int bl = 0;
+  int br = 0;
+  int tl = 0;
+  int tr = 0;
+  double *sbt = NULL;
+  double xb = 0.0;
+  double px = 0.0;
+  double nx = 0.0;
+  double yb = 0.0;
+  double yt = 0.0;
+  double dy = 0.0;
+  double ix = 0.0;
+  double iy = 0.0;
+
+  /* Test for trivial NULL result cases */
+  if (((subj->num_contours == 0) && (clip->num_contours == 0)) ||
+      ((subj->num_contours == 0) && ((op == GPC_INT) || (op == GPC_DIFF))) ||
+      ((clip->num_contours == 0) && (op == GPC_INT))) {
+    result->num_strips = 0;
+    result->strip = NULL;
+    return;
+  }
+
+  /* Identify potentialy contributing contours */
+  if (((op == GPC_INT) || (op == GPC_DIFF)) && (subj->num_contours > 0) &&
+      (clip->num_contours > 0)) {
+    minimax_test(subj, clip, op);
+  }
+  /* Build LMT */
+  if (subj->num_contours > 0) {
+    s_heap = build_lmt(&lmt, &sbtree, &sbt_entries, subj, SUBJ, op);
+  }
+  if (clip->num_contours > 0) {
+    c_heap = build_lmt(&lmt, &sbtree, &sbt_entries, clip, CLIP, op);
+  }
+  /* Return a NULL result if no contours contribute */
+  if (lmt == NULL) {
+    result->num_strips = 0;
+    result->strip = NULL;
+    reset_lmt(&lmt);
+    gpc_free<edge_node>(s_heap);
+    gpc_free<edge_node>(c_heap);
+    return;
+  }
+
+  /* Build scanbeam table from scanbeam tree */
+  gpc_malloc<double>(sbt, sbt_entries * sizeof(double),
+                     const_cast<char *>("sbt creation"));
+  build_sbt(&scanbeam, sbt, sbtree);
+  scanbeam = 0;
+  free_sbtree(&sbtree);
+
+  /* Invert clip polygon for difference operation */
+  if (op == GPC_DIFF) {
+    parity[CLIP] = RIGHT;
+  }
+  local_min = lmt;
+
+  // Process each scanbeam
+  while (scanbeam < sbt_entries) {
+    /* Set yb and yt to the bottom and top of the scanbeam */
+    yb = sbt[scanbeam++];
+    if (scanbeam < sbt_entries) {
+      yt = sbt[scanbeam];
+      dy = yt - yb;
+    }
+
+    /* === SCANBEAM BOUNDARY PROCESSING ================================ */
+    /* If LMT node corresponding to yb exists */
+    if (local_min) {
+      if (local_min->y == yb) {
+        /* Add edges starting at this local minimum to the AET */
+        for (edge = local_min->first_bound; edge; edge = edge->next_bound) {
+          add_edge_to_aet(&aet, edge, NULL);
+        }
+        local_min = local_min->next;
+      }
+    }
+    /* Set dummy previous x value */
+    /* Create bundles within AET */
+    px = -DBL_MAX;
+    e0 = aet;
+    e1 = aet;
+
+    /* Set up bundle fields of first edge */
+    aet->bundle[ABOVE][aet->type] = (aet->top.y != yb);
+    aet->bundle[ABOVE][!aet->type] = 0;
+    aet->bstate[ABOVE] = UNBUNDLED;
+
+    for (next_edge = aet->next; next_edge; next_edge = next_edge->next) {
+      /* Set up bundle fields of next edge */
+      next_edge->bundle[ABOVE][next_edge->type] = (next_edge->top.y != yb);
+      next_edge->bundle[ABOVE][!next_edge->type] = 0;
+      next_edge->bstate[ABOVE] = UNBUNDLED;
+
+      /* Bundle edges above the scanbeam boundary if they coincide */
+      if (next_edge->bundle[ABOVE][next_edge->type]) {
+        if (gpc_eq(e0->xb, next_edge->xb) && gpc_eq(e0->dx, next_edge->dx) &&
+            (e0->top.y != yb)) {
+          next_edge->bundle[ABOVE][next_edge->type] ^=
+              e0->bundle[ABOVE][next_edge->type];
+          next_edge->bundle[ABOVE][!next_edge->type] =
+              e0->bundle[ABOVE][!next_edge->type];
+          next_edge->bstate[ABOVE] = BUNDLE_HEAD;
+          e0->bundle[ABOVE][CLIP] = 0;
+          e0->bundle[ABOVE][SUBJ] = 0;
+          e0->bstate[ABOVE] = BUNDLE_TAIL;
+        }
+        e0 = next_edge;
+      }
+    }
+    horiz[CLIP] = NH;
+    horiz[SUBJ] = NH;
+
+    /* Process each edge at this scanbeam boundary */
+    for (edge = aet; edge; edge = edge->next) {
+      exists[CLIP] =
+          edge->bundle[ABOVE][CLIP] + (edge->bundle[BELOW][CLIP] << 1);
+      exists[SUBJ] =
+          edge->bundle[ABOVE][SUBJ] + (edge->bundle[BELOW][SUBJ] << 1);
+
+      if (exists[CLIP] || exists[SUBJ]) {
+        /* Set bundle side */
+        edge->bside[CLIP] = parity[CLIP];
+        edge->bside[SUBJ] = parity[SUBJ];
+
+        /* Determine contributing status and quadrant occupancies */
+        switch (op) {
+          case GPC_DIFF:
+          case GPC_INT:
+            contributing = (exists[CLIP] && (parity[SUBJ] || horiz[SUBJ])) ||
+                           (exists[SUBJ] && (parity[CLIP] || horiz[CLIP])) ||
+                           (exists[CLIP] && exists[SUBJ] &&
+                            (parity[CLIP] == parity[SUBJ]));
+            br = (parity[CLIP]) && (parity[SUBJ]);
+            bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) &&
+                 (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]);
+            tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) &&
+                 (parity[SUBJ] ^ (horiz[SUBJ] != NH));
+            tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^
+                  edge->bundle[BELOW][CLIP]) &&
+                 (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^
+                  edge->bundle[BELOW][SUBJ]);
+            break;
+          case GPC_XOR:
+            contributing = exists[CLIP] || exists[SUBJ];
+            br = (parity[CLIP]) ^ (parity[SUBJ]);
+            bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) ^
+                 (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]);
+            tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) ^
+                 (parity[SUBJ] ^ (horiz[SUBJ] != NH));
+            tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^
+                  edge->bundle[BELOW][CLIP]) ^
+                 (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^
+                  edge->bundle[BELOW][SUBJ]);
+            break;
+          case GPC_UNION:
+            contributing = (exists[CLIP] && (!parity[SUBJ] || horiz[SUBJ])) ||
+                           (exists[SUBJ] && (!parity[CLIP] || horiz[CLIP])) ||
+                           (exists[CLIP] && exists[SUBJ] &&
+                            (parity[CLIP] == parity[SUBJ]));
+            br = (parity[CLIP]) || (parity[SUBJ]);
+            bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) ||
+                 (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]);
+            tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) ||
+                 (parity[SUBJ] ^ (horiz[SUBJ] != NH));
+            tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^
+                  edge->bundle[BELOW][CLIP]) ||
+                 (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^
+                  edge->bundle[BELOW][SUBJ]);
+            break;
+        }
+
+        // Update parity
+        parity[CLIP] ^= edge->bundle[ABOVE][CLIP];
+        parity[SUBJ] ^= edge->bundle[ABOVE][SUBJ];
+
+        /* Update horizontal state */
+        if (exists[CLIP]) {
+          horiz[CLIP] = next_h_state[horiz[CLIP]]
+                                    [((exists[CLIP] - 1) << 1) + parity[CLIP]];
+        }
+        if (exists[SUBJ]) {
+          horiz[SUBJ] = next_h_state[horiz[SUBJ]]
+                                    [((exists[SUBJ] - 1) << 1) + parity[SUBJ]];
+        }
+        vclass = tr + (tl << 1) + (br << 2) + (bl << 3);
+
+        if (contributing) {
+          xb = edge->xb;
+          switch (vclass) {
+            case EMN:
+              new_tristrip(&tlist, edge, xb, yb);
+              cf = edge;
+              break;
+            case ERI:
+              edge->outp[ABOVE] = cf->outp[ABOVE];
+              if (xb != cf->xb) {
+                gpc_vertex_create(edge, ABOVE, RIGHT, xb, yb);
+              }
+              cf = NULL;
+              break;
+            case ELI:
+              gpc_vertex_create(edge, BELOW, LEFT, xb, yb);
+              edge->outp[ABOVE] = NULL;
+              cf = edge;
+              break;
+            case EMX:
+              if (xb != cf->xb) {
+                gpc_vertex_create(edge, BELOW, RIGHT, xb, yb);
+              }
+              edge->outp[ABOVE] = NULL;
+              cf = NULL;
+              break;
+            case IMN:
+              if (cft == LED) {
+                if (cf->bot.y != yb) {
+                  gpc_vertex_create(cf, BELOW, LEFT, cf->xb, yb);
+                }
+                new_tristrip(&tlist, cf, cf->xb, yb);
+              }
+              edge->outp[ABOVE] = cf->outp[ABOVE];
+              gpc_vertex_create(edge, ABOVE, RIGHT, xb, yb);
+              break;
+            case ILI:
+              new_tristrip(&tlist, edge, xb, yb);
+              cf = edge;
+              cft = ILI;
+              break;
+            case IRI:
+              if (cft == LED) {
+                if (cf->bot.y != yb) {
+                  gpc_vertex_create(cf, BELOW, LEFT, cf->xb, yb);
+                }
+                new_tristrip(&tlist, cf, cf->xb, yb);
+              }
+              gpc_vertex_create(edge, BELOW, RIGHT, xb, yb);
+              edge->outp[ABOVE] = NULL;
+              break;
+            case IMX:
+              gpc_vertex_create(edge, BELOW, LEFT, xb, yb);
+              edge->outp[ABOVE] = NULL;
+              cft = IMX;
+              break;
+            case IMM:
+              gpc_vertex_create(edge, BELOW, LEFT, xb, yb);
+              edge->outp[ABOVE] = cf->outp[ABOVE];
+              if (xb != cf->xb) {
+                gpc_vertex_create(cf, ABOVE, RIGHT, xb, yb);
+              }
+              cf = edge;
+              break;
+            case EMM:
+              gpc_vertex_create(edge, BELOW, RIGHT, xb, yb);
+              edge->outp[ABOVE] = NULL;
+              new_tristrip(&tlist, edge, xb, yb);
+              cf = edge;
+              break;
+            case LED:
+              if (edge->bot.y == yb) {
+                gpc_vertex_create(edge, BELOW, LEFT, xb, yb);
+              }
+              edge->outp[ABOVE] = edge->outp[BELOW];
+              cf = edge;
+              cft = LED;
+              break;
+            case RED:
+              edge->outp[ABOVE] = cf->outp[ABOVE];
+              if (cft == LED) {
+                if (cf->bot.y == yb) {
+                  gpc_vertex_create(edge, BELOW, RIGHT, xb, yb);
+                } else {
+                  if (edge->bot.y == yb) {
+                    gpc_vertex_create(cf, BELOW, LEFT, cf->xb, yb);
+                    gpc_vertex_create(edge, BELOW, RIGHT, xb, yb);
+                  }
+                }
+              } else {
+                gpc_vertex_create(edge, BELOW, RIGHT, xb, yb);
+                gpc_vertex_create(edge, ABOVE, RIGHT, xb, yb);
+              }
+              cf = NULL;
+              break;
+            default:
+              break;
+          } /* End of switch */
+        }   /* End of contributing conditional */
+      }     /* End of edge exists conditional */
+    }       // End of AET loop
+
+    /* Delete terminating edges from the AET, otherwise compute xt */
+    for (edge = aet; edge; edge = edge->next) {
+      if (edge->top.y == yb) {
+        prev_edge = edge->prev;
+        next_edge = edge->next;
+        if (prev_edge) {
+          prev_edge->next = next_edge;
+        } else {
+          aet = next_edge;
+        }
+        if (next_edge) {
+          next_edge->prev = prev_edge;
+        }
+
+        /* Copy bundle head state to the adjacent tail edge if required */
+        if ((edge->bstate[BELOW] == BUNDLE_HEAD) && prev_edge) {
+          if (prev_edge->bstate[BELOW] == BUNDLE_TAIL) {
+            prev_edge->outp[BELOW] = edge->outp[BELOW];
+            prev_edge->bstate[BELOW] = UNBUNDLED;
+            if (prev_edge->prev) {
+              if (prev_edge->prev->bstate[BELOW] == BUNDLE_TAIL) {
+                prev_edge->bstate[BELOW] = BUNDLE_HEAD;
+              }
+            }
+          }
+        }
+      } else {
+        if (edge->top.y == yt) {
+          edge->xt = edge->top.x;
+        } else {
+          edge->xt = edge->bot.x + edge->dx * (yt - edge->bot.y);
+        }
+      }
+    }
+
+    if (scanbeam < sbt_entries) {
+      /* === SCANBEAM INTERIOR PROCESSING ============================== */
+      build_intersection_table(&it, aet, dy);
+      /* Process each node in the intersection table */
+      for (intersect = it; intersect; intersect = intersect->next) {
+        e0 = intersect->ie[0];
+        e1 = intersect->ie[1];
+
+        /* Only generate output for contributing intersections */
+        if ((e0->bundle[ABOVE][CLIP] || e0->bundle[ABOVE][SUBJ]) &&
+            (e1->bundle[ABOVE][CLIP] || e1->bundle[ABOVE][SUBJ])) {
+          p = e0->outp[ABOVE];
+          q = e1->outp[ABOVE];
+          ix = intersect->point.x;
+          iy = intersect->point.y + yb;
+
+          in[CLIP] = (e0->bundle[ABOVE][CLIP] && !e0->bside[CLIP]) ||
+                     (e1->bundle[ABOVE][CLIP] && e1->bside[CLIP]) ||
+                     (!e0->bundle[ABOVE][CLIP] && !e1->bundle[ABOVE][CLIP] &&
+                      e0->bside[CLIP] && e1->bside[CLIP]);
+          in[SUBJ] = (e0->bundle[ABOVE][SUBJ] && !e0->bside[SUBJ]) ||
+                     (e1->bundle[ABOVE][SUBJ] && e1->bside[SUBJ]) ||
+                     (!e0->bundle[ABOVE][SUBJ] && !e1->bundle[ABOVE][SUBJ] &&
+                      e0->bside[SUBJ] && e1->bside[SUBJ]);
+
+          switch (op) {  // Determine quadrant occupancies
+            case GPC_DIFF:
+            case GPC_INT:
+              tr = (in[CLIP]) && (in[SUBJ]);
+              tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) &&
+                   (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]);
+              br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) &&
+                   (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]);
+              bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^
+                    e0->bundle[ABOVE][CLIP]) &&
+                   (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^
+                    e0->bundle[ABOVE][SUBJ]);
+              break;
+            case GPC_XOR:
+              tr = (in[CLIP]) ^ (in[SUBJ]);
+              tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) ^
+                   (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]);
+              br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) ^
+                   (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]);
+              bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^
+                    e0->bundle[ABOVE][CLIP]) ^
+                   (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^
+                    e0->bundle[ABOVE][SUBJ]);
+              break;
+            case GPC_UNION:
+              tr = (in[CLIP]) || (in[SUBJ]);
+              tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) ||
+                   (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]);
+              br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) ||
+                   (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]);
+              bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^
+                    e0->bundle[ABOVE][CLIP]) ||
+                   (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^
+                    e0->bundle[ABOVE][SUBJ]);
+              break;
+          }
+
+          vclass = tr + (tl << 1) + (br << 2) + (bl << 3);
+          switch (vclass) {
+            case EMN:
+              new_tristrip(&tlist, e1, ix, iy);
+              e0->outp[ABOVE] = e1->outp[ABOVE];
+              break;
+            case ERI:
+              if (p) {
+                gpc_p_edge(prev_edge, e0, ABOVE);
+                gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy);
+                gpc_vertex_create(e0, ABOVE, RIGHT, ix, iy);
+                e1->outp[ABOVE] = e0->outp[ABOVE];
+                e0->outp[ABOVE] = NULL;
+              }
+              break;
+            case ELI:
+              if (q) {
+                gpc_n_edge(next_edge, e1, ABOVE);
+                gpc_vertex_create(e1, ABOVE, LEFT, ix, iy);
+                gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy);
+                e0->outp[ABOVE] = e1->outp[ABOVE];
+                e1->outp[ABOVE] = NULL;
+              }
+              break;
+            case EMX:
+              if (p && q) {
+                gpc_vertex_create(e0, ABOVE, LEFT, ix, iy);
+                e0->outp[ABOVE] = NULL;
+                e1->outp[ABOVE] = NULL;
+              }
+              break;
+            case IMN:
+              gpc_p_edge(prev_edge, e0, ABOVE);
+              gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy);
+              gpc_n_edge(next_edge, e1, ABOVE);
+              gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy);
+              new_tristrip(&tlist, prev_edge, px, iy);
+              e1->outp[ABOVE] = prev_edge->outp[ABOVE];
+              gpc_vertex_create(e1, ABOVE, RIGHT, ix, iy);
+              new_tristrip(&tlist, e0, ix, iy);
+              next_edge->outp[ABOVE] = e0->outp[ABOVE];
+              gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy);
+              break;
+            case ILI:
+              if (p) {
+                gpc_vertex_create(e0, ABOVE, LEFT, ix, iy);
+                gpc_n_edge(next_edge, e1, ABOVE);
+                gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy);
+                e1->outp[ABOVE] = e0->outp[ABOVE];
+                e0->outp[ABOVE] = NULL;
+              }
+              break;
+            case IRI:
+              if (q) {
+                gpc_vertex_create(e1, ABOVE, RIGHT, ix, iy);
+                gpc_p_edge(prev_edge, e0, ABOVE);
+                gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy);
+                e0->outp[ABOVE] = e1->outp[ABOVE];
+                e1->outp[ABOVE] = NULL;
+              }
+              break;
+            case IMX:
+              if (p && q) {
+                gpc_vertex_create(e0, ABOVE, RIGHT, ix, iy);
+                gpc_vertex_create(e1, ABOVE, LEFT, ix, iy);
+                e0->outp[ABOVE] = NULL;
+                e1->outp[ABOVE] = NULL;
+                gpc_p_edge(prev_edge, e0, ABOVE);
+                gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy);
+                new_tristrip(&tlist, prev_edge, px, iy);
+                gpc_n_edge(next_edge, e1, ABOVE);
+                gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy);
+                next_edge->outp[ABOVE] = prev_edge->outp[ABOVE];
+                gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy);
+              }
+              break;
+            case IMM:
+              if (p && q) {
+                gpc_vertex_create(e0, ABOVE, RIGHT, ix, iy);
+                gpc_vertex_create(e1, ABOVE, LEFT, ix, iy);
+                gpc_p_edge(prev_edge, e0, ABOVE);
+                gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy);
+                new_tristrip(&tlist, prev_edge, px, iy);
+                gpc_n_edge(next_edge, e1, ABOVE);
+                gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy);
+                e1->outp[ABOVE] = prev_edge->outp[ABOVE];
+                gpc_vertex_create(e1, ABOVE, RIGHT, ix, iy);
+                new_tristrip(&tlist, e0, ix, iy);
+                next_edge->outp[ABOVE] = e0->outp[ABOVE];
+                gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy);
+              }
+              break;
+            case EMM:
+              if (p && q) {
+                gpc_vertex_create(e0, ABOVE, LEFT, ix, iy);
+                new_tristrip(&tlist, e1, ix, iy);
+                e0->outp[ABOVE] = e1->outp[ABOVE];
+              }
+              break;
+            default:
+              break;
+          } /* End of switch */
+        }   /* End of contributing intersection conditional */
+
+        // Swap bundle sides in response to edge crossing
+        if (e0->bundle[ABOVE][CLIP]) {
+          e1->bside[CLIP] = !e1->bside[CLIP];
+        }
+        if (e1->bundle[ABOVE][CLIP]) {
+          e0->bside[CLIP] = !e0->bside[CLIP];
+        }
+        if (e0->bundle[ABOVE][SUBJ]) {
+          e1->bside[SUBJ] = !e1->bside[SUBJ];
+        }
+        if (e1->bundle[ABOVE][SUBJ]) {
+          e0->bside[SUBJ] = !e0->bside[SUBJ];
+        }
+
+        /* Swap e0 and e1 bundles in the AET */
+        prev_edge = e0->prev;
+        next_edge = e1->next;
+        if (e1->next) {
+          e1->next->prev = e0;
+        }
+
+        if (e0->bstate[ABOVE] == BUNDLE_HEAD) {
+          search = 1;
+          while (search) {
+            prev_edge = prev_edge->prev;
+            if (prev_edge) {
+              if (prev_edge->bundle[ABOVE][CLIP] ||
+                  prev_edge->bundle[ABOVE][SUBJ] ||
+                  (prev_edge->bstate[ABOVE] == BUNDLE_HEAD)) {
+                search = 0;
+              }
+            } else {
+              search = 0;
+            }
+          }
+        }
+        if (!prev_edge) {
+          e1->next = aet;
+          aet = e0->next;
+        } else {
+          e1->next = prev_edge->next;
+          prev_edge->next = e0->next;
+        }
+        e0->next->prev = prev_edge;
+        e1->next->prev = e1;
+        e0->next = next_edge;
+      } /* End of IT loop*/
+
+      /* Prepare for next scanbeam */
+      for (edge = aet; edge; edge = next_edge) {
+        next_edge = edge->next;
+        succ_edge = edge->succ;
+
+        if ((edge->top.y == yt) && succ_edge) {
+          /* Replace AET edge by its successor */
+          succ_edge->outp[BELOW] = edge->outp[ABOVE];
+          succ_edge->bstate[BELOW] = edge->bstate[ABOVE];
+          succ_edge->bundle[BELOW][CLIP] = edge->bundle[ABOVE][CLIP];
+          succ_edge->bundle[BELOW][SUBJ] = edge->bundle[ABOVE][SUBJ];
+          prev_edge = edge->prev;
+          if (prev_edge) {
+            prev_edge->next = succ_edge;
+          } else {
+            aet = succ_edge;
+          }
+          if (next_edge) {
+            next_edge->prev = succ_edge;
+          }
+          succ_edge->prev = prev_edge;
+          succ_edge->next = next_edge;
+        } else {
+          /* Update this edge */
+          edge->outp[BELOW] = edge->outp[ABOVE];
+          edge->bstate[BELOW] = edge->bstate[ABOVE];
+          edge->bundle[BELOW][CLIP] = edge->bundle[ABOVE][CLIP];
+          edge->bundle[BELOW][SUBJ] = edge->bundle[ABOVE][SUBJ];
+          edge->xb = edge->xt;
+        }
+        edge->outp[ABOVE] = NULL;
+      }
+    }
+  } /* === END OF SCANBEAM PROCESSING ================================== */
+
+  // Generate result tristrip from tlist
+  result->strip = NULL;
+  result->num_strips = count_tristrips(tlist);
+  if (result->num_strips > 0) {
+    gpc_malloc<gpc_vertex_list>(result->strip,
+                                result->num_strips * sizeof(gpc_vertex_list),
+                                const_cast<char *>("tristrip list creation"));
+
+    s = 0;
+    for (tn = tlist; tn; tn = tnn) {
+      tnn = tn->next;
+      if (tn->active > 2) {
+        /* Valid tristrip: copy the vertices and free the heap */
+        result->strip[s].num_vertices = tn->active;
+        gpc_malloc<gpc_vertex>(result->strip[s].vertex,
+                               tn->active * sizeof(gpc_vertex),
+                               const_cast<char *>("tristrip creation"));
+        v = 0;
+        if (0) {
+          lt = tn->v[RIGHT];
+          rt = tn->v[LEFT];
+        } else {
+          lt = tn->v[LEFT];
+          rt = tn->v[RIGHT];
+        }
+        while (lt || rt) {
+          if (lt) {
+            ltn = lt->next;
+            result->strip[s].vertex[v].x = lt->x;
+            result->strip[s].vertex[v].y = lt->y;
+            v++;
+            gpc_free<vertex_node>(lt);
+            lt = ltn;
+          }
+          if (rt) {
+            rtn = rt->next;
+            result->strip[s].vertex[v].x = rt->x;
+            result->strip[s].vertex[v].y = rt->y;
+            v++;
+            gpc_free<vertex_node>(rt);
+            rt = rtn;
+          }
+        }
+        s++;
+      } else {
+        /* Invalid tristrip: just free the heap */
+        for (lt = tn->v[LEFT]; lt; lt = ltn) {
+          ltn = lt->next;
+          gpc_free<vertex_node>(lt);
+        }
+        for (rt = tn->v[RIGHT]; rt; rt = rtn) {
+          rtn = rt->next;
+          gpc_free<vertex_node>(rt);
+        }
+      }
+      gpc_free<polygon_node>(tn);
+    }
+  }
+  // Tidy up
+  reset_it(&it);
+  reset_lmt(&lmt);
+  gpc_free<edge_node>(c_heap);
+  gpc_free<edge_node>(s_heap);
+  gpc_free<double>(sbt);
+}  // NOLINT
+
+}  // namespace gpc
+
+/* vim: set expandtab ts=4 sw=4 sts=4 tw=100: */

paddle/fluid/operators/detection/gpc.h

+// Copyright (c) 2018 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.
+
+/***************************************************************************
+ *
+ * Copyright (c) 2015 Baidu.com, Inc. All Rights Reserved
+ *
+ **************************************************************************/
+
+/**
+ * @file include/gpc.h
+ * @author huhan02(com@baidu.com)
+ * @date 2015/12/18 13:52:10
+ * @brief
+ *
+ * @modified by sunyipeng
+ * @email sunyipeng@baidu.com
+ * @date 2018/6/12
+ **/
+
+#ifndef PADDLE_FLUID_OPERATORS_DETECTION_GPC_H_  // GPC_H_
+#define PADDLE_FLUID_OPERATORS_DETECTION_GPC_H_  // GPC_H_
+
+#include <float.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+namespace gpc {
+
+typedef enum {  // Set operation type
+  GPC_DIFF,     // Difference
+  GPC_INT,      // Intersection
+  GPC_XOR,      // Exclusive or
+  GPC_UNION     // Union
+} gpc_op;
+
+typedef struct {  // Polygon vertex structure
+  double x;       // Vertex x component
+  double y;       // vertex y component
+} gpc_vertex;
+
+typedef struct {       // Vertex list structure
+  int num_vertices;    // Number of vertices in list
+  gpc_vertex *vertex;  // Vertex array pointer
+} gpc_vertex_list;
+
+typedef struct {             // Polygon set structure
+  int num_contours;          // Number of contours in polygon
+  int *hole;                 // Hole  external contour flags
+  gpc_vertex_list *contour;  // Contour array pointer
+} gpc_polygon;
+
+typedef struct {           // Tristrip set structure
+  int num_strips;          // Number of tristrips
+  gpc_vertex_list *strip;  // Tristrip array pointer
+} gpc_tristrip;
+
+typedef enum { LEFT, RIGHT } gpc_left_right;
+
+typedef enum { ABOVE, BELOW } gpc_above_below;
+
+typedef enum { CLIP, SUBJ } gpc_clip_subj;
+
+typedef enum {      /* Edge intersection classes         */
+               NUL, /* Empty non-intersection            */
+               EMX, /* External maximum                  */
+               ELI, /* External left intermediate        */
+               TED, /* Top edge                          */
+               ERI, /* External right intermediate       */
+               RED, /* Right edge                        */
+               IMM, /* Internal maximum and minimum      */
+               IMN, /* Internal minimum                  */
+               EMN, /* External minimum                  */
+               EMM, /* External maximum and minimum      */
+               LED, /* Left edge                         */
+               ILI, /* Internal left intermediate        */
+               BED, /* Bottom edge                       */
+               IRI, /* Internal right intermediate       */
+               IMX, /* Internal maximum                  */
+               FUL  /* Full non-intersection             */
+} vertex_type;
+
+typedef enum {     /* Horizontal edge states            */
+               NH, /* No horizontal edge                */
+               BH, /* Bottom horizontal edge            */
+               TH  /* Top horizontal edge               */
+} h_state;
+
+typedef enum {              /* Edge bundle state                 */
+               UNBUNDLED,   /* Isolated edge not within a bundle */
+               BUNDLE_HEAD, /* Bundle head node                  */
+               BUNDLE_TAIL  /* Passive bundle tail node          */
+} bundle_state;
+
+typedef struct v_shape { /* Internal vertex list datatype     */
+  double x;              /* X coordinate component            */
+  double y;              /* Y coordinate component            */
+  struct v_shape *next;  /* Pointer to next vertex in list    */
+} vertex_node;
+
+typedef struct p_shape { /* Internal contour / tristrip type  */
+  int active;            /* Active flag / vertex count        */
+  int hole;              /* Hole / external contour flag      */
+  vertex_node *v[2];     /* Left and right vertex list ptrs   */
+  struct p_shape *next;  /* Pointer to next polygon contour   */
+  struct p_shape *proxy; /* Pointer to actual structure used  */
+} polygon_node;
+
+typedef struct edge_shape {
+  gpc_vertex vertex;             /* Piggy-backed contour vertex data  */
+  gpc_vertex bot;                /* Edge lower (x, y) coordinate      */
+  gpc_vertex top;                /* Edge upper (x, y) coordinate      */
+  double xb;                     /* Scanbeam bottom x coordinate      */
+  double xt;                     /* Scanbeam top x coordinate         */
+  double dx;                     /* Change in x for a unit y increase */
+  int type;                      /* Clip / subject edge flag          */
+  int bundle[2][2];              /* Bundle edge flags                 */
+  int bside[2];                  /* Bundle left / right indicators    */
+  bundle_state bstate[2];        /* Edge bundle state                 */
+  polygon_node *outp[2];         /* Output polygon / tristrip pointer */
+  struct edge_shape *prev;       /* Previous edge in the AET          */
+  struct edge_shape *next;       /* Next edge in the AET              */
+  struct edge_shape *pred;       /* Edge connected at the lower end   */
+  struct edge_shape *succ;       /* Edge connected at the upper end   */
+  struct edge_shape *next_bound; /* Pointer to next bound in LMT      */
+} edge_node;
+
+inline bool gpc_eq(float a, float b) { return (fabs(a - b) <= 1e-6); }
+
+inline bool gpc_prev_index(float a, float b) { return (fabs(a - b) <= 1e-6); }
+
+inline int gpc_prev_index(int i, int n) { return ((i - 1 + n) % n); }
+
+inline int gpc_next_index(int i, int n) { return ((i + 1) % n); }
+
+inline int gpc_optimal(gpc_vertex *v, int i, int n) {
+  return (v[(i + 1) % n].y != v[i].y || v[(i - 1 + n) % n].y != v[i].y);
+}
+
+inline int gpc_fwd_min(edge_node *v, int i, int n) {
+  return (v[(i + 1) % n].vertex.y > v[i].vertex.y &&
+          v[(i - 1 + n) % n].vertex.y >= v[i].vertex.y);
+}
+
+inline int gpc_not_fmax(edge_node *v, int i, int n) {
+  return (v[(i + 1) % n].vertex.y > v[i].vertex.y);
+}
+
+inline int gpc_rev_min(edge_node *v, int i, int n) {
+  return (v[(i + 1) % n].vertex.y >= v[i].vertex.y &&
+          v[(i - 1 + n) % n].vertex.y > v[i].vertex.y);
+}
+
+inline int gpc_not_rmax(edge_node *v, int i, int n) {
+  return (v[(i - 1 + n) % n].vertex.y > v[i].vertex.y);
+}
+
+// inline void gpc_p_edge(edge_node *d, edge_node *e, int p, double i, double j)
+// {
+inline void gpc_p_edge(edge_node *d, edge_node *e, int p) {
+  d = e;
+  do {
+    d = d->prev;
+  } while (!d->outp[p]);
+  // i = d->bot.x + d->dx * (j - d->bot.y);
+}
+
+// inline void gpc_n_edge(edge_node *d, edge_node *e, int p, double i, double j)
+// {
+inline void gpc_n_edge(edge_node *d, edge_node *e, int p) {
+  d = e;
+  do {
+    d = d->next;
+  } while (!d->outp[p]);
+  // i = d->bot.x + d->dx * (j - d->bot.y);
+}
+
+template <typename T>
+void gpc_malloc(T *&p, int b, char *s) {
+  if (b > 0) {
+    p = (T *)malloc(b);
+
+    if (!p) {
+      fprintf(stderr, "gpc malloc failure: %s\n", s);
+      exit(0);
+    }
+  } else {
+    p = NULL;
+  }
+}
+template <typename T>
+void gpc_free(T *&p) {
+  if (p) {
+    free(p);
+    p = NULL;
+  }
+}
+
+/*
+===========================================================================
+                       Public Function Prototypes
+===========================================================================
+*/
+
+void add_vertex(vertex_node **t, double x, double y);
+
+void gpc_vertex_create(edge_node *e, int p, int s, double x, double y);
+
+/*
+void gpc_read_polygon(FILE *infile_ptr, int read_hole_flags,
+                      gpc_polygon *polygon);
+
+void gpc_write_polygon(FILE *outfile_ptr, int write_hole_flags,
+                       gpc_polygon *polygon);
+*/
+void gpc_add_contour(gpc_polygon *polygon, gpc_vertex_list *contour, int hole);
+
+void gpc_polygon_clip(gpc_op set_operation, gpc_polygon *subject_polygon,
+                      gpc_polygon *clip_polygon, gpc_polygon *result_polygon);
+
+void gpc_tristrip_clip(gpc_op set_operation, gpc_polygon *subject_polygon,
+                       gpc_polygon *clip_polygon,
+                       gpc_tristrip *result_tristrip);
+
+void gpc_polygon_to_tristrip(gpc_polygon *polygon, gpc_tristrip *tristrip);
+
+void gpc_free_polygon(gpc_polygon *polygon);
+
+void gpc_free_tristrip(gpc_tristrip *tristrip);
+
+}  // namespace gpc
+
+#endif  // PADDLE_FLUID_OPERATORS_DETECTION_GPC_H_
+/* vim: set expandtab ts=4 sw=4 sts=4 tw=100: */

paddle/fluid/operators/detection/multiclass_nms_op.cc

@@ -9,10 +9,11 @@ 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/op_registry.h"
+#include "paddle/fluid/operators/detection/poly_util.h"
 
 namespace paddle {
 namespace operators {
@@ -20,9 +21,6 @@ namespace operators {
 using Tensor = framework::Tensor;
 using LoDTensor = framework::LoDTensor;
 
-constexpr int64_t kOutputDim = 6;
-constexpr int64_t kBBoxSize = 4;
-
 class MultiClassNMSOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -42,10 +40,15 @@ class MultiClassNMSOp : public framework::OperatorWithKernel {
                       "The rank of Input(BBoxes) must be 3.");
     PADDLE_ENFORCE_EQ(score_dims.size(), 3,
                       "The rank of Input(Scores) must be 3.");
-    PADDLE_ENFORCE_EQ(box_dims[2], 4,
-                      "The 2nd dimension of Input(BBoxes) must be 4, "
+    PADDLE_ENFORCE(box_dims[2] == 4 || box_dims[2] == 8 || box_dims[2] == 16 ||
+                       box_dims[2] == 24 || box_dims[2] == 32,
+                   "The 2nd dimension of Input(BBoxes) must be 4 or 8, "
                    "represents the layout of coordinate "
-                      "[xmin, ymin, xmax, ymax]");
+                   "[xmin, ymin, xmax, ymax] or "
+                   "4 points: [x1, y1, x2, y2, x3, y3, x4, y4] or "
+                   "8 points: [xi, yi] i= 1,2,...,8 or "
+                   "12 points: [xi, yi] i= 1,2,...,12 or "
+                   "16 points: [xi, yi] i= 1,2,...,16");
     PADDLE_ENFORCE_EQ(box_dims[1], score_dims[2],
                       "The 1st dimensiong of Input(BBoxes) must be equal to "
                       "3rd dimension of Input(Scores), which represents the "
@@ -53,7 +56,7 @@ class MultiClassNMSOp : public framework::OperatorWithKernel {
 
     // Here the box_dims[0] is not the real dimension of output.
     // It will be rewritten in the computing kernel.
-    ctx->SetOutputDim("Out", {box_dims[1], 6});
+    ctx->SetOutputDim("Out", {box_dims[1], box_dims[2] + 2});
   }
 
  protected:
@@ -128,6 +131,21 @@ static inline T JaccardOverlap(const T* box1, const T* box2,
   }
 }
 
+template <class T>
+T PolyIoU(const T* box1, const T* box2, const size_t box_size,
+          const bool normalized) {
+  T bbox1_area = PolyArea<T>(box1, box_size, normalized);
+  T bbox2_area = PolyArea<T>(box2, box_size, normalized);
+  T inter_area = PolyOverlapArea<T>(box1, box2, box_size, normalized);
+  if (bbox1_area == 0 || bbox2_area == 0 || inter_area == 0) {
+    // If coordinate values are is invalid
+    // if area size <= 0,  return 0.
+    return T(0.);
+  } else {
+    return inter_area / (bbox1_area + bbox2_area - inter_area);
+  }
+}
+
 template <typename T>
 class MultiClassNMSKernel : public framework::OpKernel<T> {
  public:
@@ -137,6 +155,8 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
     // The total boxes for each instance.
     int64_t num_boxes = bbox.dims()[0];
     // 4: [xmin ymin xmax ymax]
+    // 8: [x1 y1 x2 y2 x3 y3 x4 y4]
+    // 16, 24, or 32: [x1 y1 x2 y2 ...  xn yn], n = 8, 12 or 16
     int64_t box_size = bbox.dims()[1];
 
     std::vector<T> scores_data(num_boxes);
@@ -154,8 +174,19 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
       for (size_t k = 0; k < selected_indices->size(); ++k) {
         if (keep) {
           const int kept_idx = (*selected_indices)[k];
-          T overlap = JaccardOverlap<T>(bbox_data + idx * box_size,
+          T overlap = T(0.);
+          // 4: [xmin ymin xmax ymax]
+          if (box_size == 4) {
+            overlap = JaccardOverlap<T>(bbox_data + idx * box_size,
                                         bbox_data + kept_idx * box_size, true);
+          }
+          // 8: [x1 y1 x2 y2 x3 y3 x4 y4] or 16, 24, 32
+          if (box_size == 8 || box_size == 16 || box_size == 24 ||
+              box_size == 32) {
+            overlap =
+                PolyIoU<T>(bbox_data + idx * box_size,
+                           bbox_data + kept_idx * box_size, box_size, true);
+          }
           keep = overlap <= adaptive_threshold;
         } else {
           break;
@@ -228,7 +259,9 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
   void MultiClassOutput(const Tensor& scores, const Tensor& bboxes,
                         const std::map<int, std::vector<int>>& selected_indices,
                         Tensor* outs) const {
-    int predict_dim = scores.dims()[1];
+    int64_t predict_dim = scores.dims()[1];
+    int64_t box_size = bboxes.dims()[1];
+    int64_t out_dim = bboxes.dims()[1] + 2;
     auto* scores_data = scores.data<T>();
     auto* bboxes_data = bboxes.data<T>();
     auto* odata = outs->data<T>();
@@ -240,11 +273,11 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
       const std::vector<int>& indices = it.second;
       for (size_t j = 0; j < indices.size(); ++j) {
         int idx = indices[j];
-        const T* bdata = bboxes_data + idx * kBBoxSize;
-        odata[count * kOutputDim] = label;           // label
-        odata[count * kOutputDim + 1] = sdata[idx];  // score
-        // xmin, ymin, xmax, ymax
-        std::memcpy(odata + count * kOutputDim + 2, bdata, 4 * sizeof(T));
+        const T* bdata = bboxes_data + idx * box_size;
+        odata[count * out_dim] = label;           // label
+        odata[count * out_dim + 1] = sdata[idx];  // score
+        // xmin, ymin, xmax, ymax or multi-points coordinates
+        std::memcpy(odata + count * out_dim + 2, bdata, box_size * sizeof(T));
         count++;
       }
     }
@@ -261,6 +294,7 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
     int64_t class_num = score_dims[1];
     int64_t predict_dim = score_dims[2];
     int64_t box_dim = boxes->dims()[2];
+    int64_t out_dim = boxes->dims()[2] + 2;
 
     std::vector<std::map<int, std::vector<int>>> all_indices;
     std::vector<size_t> batch_starts = {0};
@@ -283,7 +317,7 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
       T* od = outs->mutable_data<T>({1}, ctx.GetPlace());
       od[0] = -1;
     } else {
-      outs->mutable_data<T>({num_kept, kOutputDim}, ctx.GetPlace());
+      outs->mutable_data<T>({num_kept, out_dim}, ctx.GetPlace());
       for (int64_t i = 0; i < batch_size; ++i) {
         Tensor ins_score = scores->Slice(i, i + 1);
         ins_score.Resize({class_num, predict_dim});
@@ -311,10 +345,11 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
     AddInput("BBoxes",
-             "(Tensor) A 3-D Tensor with shape [N, M, 4] represents the "
+             "(Tensor) A 3-D Tensor with shape "
+             "[N, M, 4 or 8 16 24 32] represents the "
              "predicted locations of M bounding bboxes, N is the batch size. "
              "Each bounding box has four coordinate values and the layout is "
-             "[xmin, ymin, xmax, ymax].");
+             "[xmin, ymin, xmax, ymax], when box size equals to 4.");
     AddInput("Scores",
              "(Tensor) A 3-D Tensor with shape [N, C, M] represents the "
              "predicted confidence predictions. N is the batch size, C is the "
@@ -351,8 +386,12 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker {
     AddOutput("Out",
               "(LoDTensor) A 2-D LoDTensor with shape [No, 6] represents the "
               "detections. Each row has 6 values: "
-              "[label, confidence, xmin, ymin, xmax, ymax], No is the total "
-              "number of detections in this mini-batch. For each instance, "
+              "[label, confidence, xmin, ymin, xmax, ymax] or "
+              "(LoDTensor) A 2-D LoDTensor with shape [No, 10] represents the "
+              "detections. Each row has 10 values: "
+              "[label, confidence, x1, y1, x2, y2, x3, y3, x4, y4]. No is the "
+              "total number of detections in this mini-batch."
+              "For each instance, "
               "the offsets in first dimension are called LoD, the number of "
               "offset is N + 1, if LoD[i + 1] - LoD[i] == 0, means there is "
               "no detected bbox.");

paddle/fluid/operators/detection/poly_util.cc

+/* Copyright (c) 2018 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. */
+
+#ifndef POLY_UTIL_CC_
+#define POLY_UTIL_CC_
+
+#include "paddle/fluid/operators/detection/poly_util.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using gpc::gpc_polygon_clip;
+using gpc::gpc_free_polygon;
+
+template <class T>
+void Array2PointVec(const T*& box, const size_t box_size,
+                    std::vector<Point_<T>>& vec) {
+  size_t pts_num = box_size / 2;
+  vec.resize(pts_num);
+  for (size_t i = 0; i < pts_num; i++) {
+    vec.at(i).x = box[2 * i];
+    vec.at(i).y = box[2 * i + 1];
+  }
+}
+
+template <class T>
+void Array2Poly(const T*& box, const size_t box_size, gpc::gpc_polygon& poly) {
+  size_t pts_num = box_size / 2;
+  poly.num_contours = 1;
+  poly.hole = (int*)malloc(sizeof(int));
+  poly.hole[0] = 0;
+  poly.contour = (gpc::gpc_vertex_list*)malloc(sizeof(gpc::gpc_vertex_list));
+  poly.contour->num_vertices = pts_num;
+  poly.contour->vertex =
+      (gpc::gpc_vertex*)malloc(sizeof(gpc::gpc_vertex) * pts_num);
+  for (size_t i = 0; i < pts_num; ++i) {
+    poly.contour->vertex[i].x = box[2 * i];
+    poly.contour->vertex[i].y = box[2 * i + 1];
+  }
+}
+
+template <class T>
+void PointVec2Poly(const std::vector<Point_<T>>& vec, gpc::gpc_polygon& poly) {
+  int pts_num = vec.size();
+  poly.num_contours = 1;
+  poly.hole = (int*)malloc(sizeof(int));
+  poly.hole[0] = 0;
+  poly.contour = (gpc::gpc_vertex_list*)malloc(sizeof(gpc::gpc_vertex_list));
+  poly.contour->num_vertices = pts_num;
+  poly.contour->vertex =
+      (gpc::gpc_vertex*)malloc(sizeof(gpc::gpc_vertex) * pts_num);
+  for (size_t i = 0; i < pts_num; ++i) {
+    poly.contour->vertex[i].x = vec[i].x;
+    poly.contour->vertex[i].y = vec[i].y;
+  }
+}
+
+template <class T>
+void Poly2PointVec(const gpc::gpc_vertex_list& contour,
+                   std::vector<Point_<T>>& vec) {
+  int pts_num = contour.num_vertices;
+  vec.resize(pts_num);
+  for (int i = 0; i < pts_num; i++) {
+    vec.at(i).x = contour.vertex[i].x;
+    vec.at(i).y = contour.vertex[i].y;
+  }
+}
+
+template <class T>
+T GetContourArea(std::vector<Point_<T>>& vec) {
+  size_t pts_num = vec.size();
+  if (pts_num < 3) return T(0.);
+  T area = T(0.);
+  for (size_t i = 0; i < pts_num; ++i) {
+    area += vec[i].x * vec[(i + 1) % pts_num].y -
+            vec[i].y * vec[(i + 1) % pts_num].x;
+  }
+  return std::fabs(area / 2.0);
+}
+
+template <class T>
+T PolyArea(const T* box, const size_t box_size, const bool normalized) {
+  // If coordinate values are is invalid
+  // if area size <= 0,  return 0.
+  std::vector<Point_<T>> vec;
+  Array2PointVec<T>(box, box_size, vec);
+  return GetContourArea<T>(vec);
+}
+
+template <class T>
+T PolyOverlapArea(const T* box1, const T* box2, const size_t box_size,
+                  const bool normalized) {
+  gpc::gpc_polygon poly1;
+  gpc::gpc_polygon poly2;
+  Array2Poly<T>(box1, box_size, poly1);
+  Array2Poly<T>(box2, box_size, poly2);
+  gpc::gpc_polygon respoly;
+  gpc::gpc_op op = gpc::GPC_INT;
+  gpc::gpc_polygon_clip(op, &poly2, &poly1, &respoly);
+
+  T inter_area = T(0.);
+  int contour_num = respoly.num_contours;
+  for (int i = 0; i < contour_num; ++i) {
+    std::vector<Point_<T>> resvec;
+    Poly2PointVec<T>(respoly.contour[i], resvec);
+    // inter_area += std::fabs(cv::contourArea(resvec)) + 0.5f *
+    // (cv::arcLength(resvec, true));
+    inter_area += GetContourArea<T>(resvec);
+  }
+
+  gpc::gpc_free_polygon(&poly1);
+  gpc::gpc_free_polygon(&poly2);
+  gpc::gpc_free_polygon(&respoly);
+  return inter_area;
+}
+
+}  // namespace operators
+}  // namespace paddle
+
+#endif

paddle/fluid/operators/detection/poly_util.h

+/* Copyright (c) 2018 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. */
+
+#ifndef POLY_UTIL_H_
+#define POLY_UTIL_H_
+
+#include <vector>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/detection/gpc.h"
+
+namespace paddle {
+namespace operators {
+
+template <class T>
+class Point_ {
+ public:
+  // default constructor
+  Point_() {}
+  Point_(T _x, T _y) {}
+  Point_(const Point_& pt) {}
+
+  Point_& operator=(const Point_& pt);
+  // conversion to another data type
+  // template<typename _T> operator Point_<_T>() const;
+  // conversion to the old-style C structures
+  // operator Vec<T, 2>() const;
+
+  // checks whether the point is inside the specified rectangle
+  // bool inside(const Rect_<T>& r) const;
+  T x;  //!< x coordinate of the point
+  T y;  //!< y coordinate of the point
+};
+
+template <class T>
+void Array2PointVec(const T*& box, const size_t box_size,
+                    std::vector<Point_<T>>& vec);
+
+template <class T>
+void Array2Poly(const T*& box, const size_t box_size, gpc::gpc_polygon& poly);
+
+template <class T>
+void PointVec2Poly(const std::vector<Point_<T>>& vec, gpc::gpc_polygon& poly);
+
+template <class T>
+void Poly2PointVec(const gpc::gpc_vertex_list& contour,
+                   std::vector<Point_<T>>& vec);
+
+template <class T>
+T GetContourArea(std::vector<Point_<T>>& vec);
+
+template <class T>
+T PolyArea(const T* box, const size_t box_size, const bool normalized);
+
+template <class T>
+T PolyOverlapArea(const T* box1, const T* box2, const size_t box_size,
+                  const bool normalized);
+}  // namespace operators
+}  // namespace paddle
+
+#include "paddle/fluid/operators/detection/poly_util.cc"
+
+#endif  // POLY_UTIL_H_

paddle/fluid/operators/detection/polygon_box_transform_op.cc

@@ -41,9 +41,9 @@ class PolygonBoxTransformCPUKernel : public framework::OpKernel<T> {
         for (int id_w = 0; id_w < width; ++id_w) {
           id = id_n * height * width + width * id_h + id_w;
           if (id_n % 2 == 0) {
-            out_data[id] = id_w - in_data[id];
+            out_data[id] = id_w * 4 - in_data[id];
           } else {
-            out_data[id] = id_h - in_data[id];
+            out_data[id] = id_h * 4 - in_data[id];
           }
         }
       }

paddle/fluid/operators/detection/polygon_box_transform_op.cu

@@ -32,9 +32,9 @@ __global__ void PolygonBoxTransformKernel(const int n, const int h, const int w,
   if (id_n < n && id_h < h && id_w < w) {
     int id = id_n * h * w + w * id_h + id_w;
     if (id_n % 2 == 0) {
-      output[id] = id_w - input[id];
+      output[id] = id_w * 4 - input[id];
     } else {
-      output[id] = id_h - input[id];
+      output[id] = id_h * 4 - input[id];
     }
   }
 }

paddle/fluid/operators/distributed/grpc_client.cc

@@ -86,7 +86,7 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep,
     // stub context
     s->response_call_back_ = nullptr;
 
-    platform::RecordEvent record_event(method, p_ctx);
+    platform::RecordRPCEvent record_event(method, p_ctx);
 
     auto call = s->stub_g_.PrepareUnaryCall(
         s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_);
@@ -143,7 +143,7 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep,
     // stub context
     s->response_call_back_ = ProcGetResponse;
 
-    platform::RecordEvent record_event(method, p_ctx);
+    platform::RecordRPCEvent record_event(method, p_ctx);
 
     auto call = s->stub_g_.PrepareUnaryCall(
         s->context_.get(), "/sendrecv.SendRecvService/GetVariable", buf, &cq_);
@@ -191,7 +191,7 @@ VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep,
     // stub context
     s->response_call_back_ = ProcGetResponse;
 
-    platform::RecordEvent record_event(method, p_ctx);
+    platform::RecordRPCEvent record_event(method, p_ctx);
 
     auto call = s->stub_g_.PrepareUnaryCall(
         s->context_.get(), "/sendrecv.SendRecvService/PrefetchVariable", req,
@@ -221,7 +221,7 @@ VarHandlePtr GRPCClient::AsyncSendBatchBarrier(const std::string& ep,
   sendrecv::VariableMessage req;
   req.set_varname(BATCH_BARRIER_MESSAGE);
 
-  platform::RecordEvent record_event(method, nullptr);
+  platform::RecordRPCEvent record_event(method, nullptr);
 
   auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
@@ -246,7 +246,7 @@ VarHandlePtr GRPCClient::AsyncSendFetchBarrier(const std::string& ep,
   sendrecv::VariableMessage req;
   req.set_varname(FETCH_BARRIER_MESSAGE);
 
-  platform::RecordEvent record_event(method, nullptr);
+  platform::RecordRPCEvent record_event(method, nullptr);
 
   auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
@@ -271,7 +271,7 @@ VarHandlePtr GRPCClient::AsyncSendComplete(const std::string& ep,
   sendrecv::VariableMessage req;
   req.set_varname(COMPLETE_MESSAGE);
 
-  platform::RecordEvent record_event(method, nullptr);
+  platform::RecordRPCEvent record_event(method, nullptr);
 
   auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
@@ -301,7 +301,7 @@ VarHandlePtr GRPCClient::AsyncCheckpointNotify(const std::string& ep,
   req.set_varname(CHECKPOINT_SAVE_MESSAGE);
   req.set_out_varname(dir);
 
-  platform::RecordEvent record_event(method, nullptr);
+  platform::RecordRPCEvent record_event(method, nullptr);
 
   auto rpc = s->stub_->AsyncCheckpointNotify(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));

paddle/fluid/operators/distributed/grpc_serde.cc

@@ -36,7 +36,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
                            const platform::DeviceContext& ctx,
                            ::grpc::ByteBuffer* msg,
                            const std::string& out_name) {
-  platform::RecordEvent record_event("serial", &ctx);
+  platform::RecordRPCEvent record_event("serial", &ctx);
   // Default DestroyCallback does nothing, When using GPU
   // the CPU buffer need to be freed.
   DestroyCallback destroy_callback = [](void* backing) {};
@@ -148,7 +148,7 @@ void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg,
                                const platform::DeviceContext& ctx,
                                const framework::Scope* scope,
                                framework::Variable** var) {
-  platform::RecordEvent record_event("deserial", &ctx);
+  platform::RecordRPCEvent record_event("deserial", &ctx);
   operators::distributed::GRPCVariableResponse resp(scope, &ctx);
   PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!");
   *var = resp.GetVar();

paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.cc

+/* Copyright (c) 2018 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/operators/fusion_seqconv_eltadd_relu_op.h"
+#include <algorithm>  // for min, max
+#include <string>
+#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/fc_compute.h"
+
+namespace paddle {
+namespace operators {
+
+void FusionSeqConvEltAddReluOp::InferShape(
+    framework::InferShapeContext* ctx) const {
+  PADDLE_ENFORCE(ctx->HasInput("X"),
+                 "Input(X) of FusionSeqConvEltAddReluOp should not be null.");
+  PADDLE_ENFORCE(
+      ctx->HasInput("Filter"),
+      "Input(Filter) of FusionSeqConvEltAddReluOp should not be null.");
+  PADDLE_ENFORCE(
+      ctx->HasInput("Bias"),
+      "Input(Bias) of FusionSeqConvEltAddReluOp should not be null.");
+  PADDLE_ENFORCE(
+      ctx->HasOutput("Out"),
+      "Output(Out) of FusionSeqConvEltAddReluOp should not be null.");
+  PADDLE_ENFORCE(
+      ctx->HasOutput("ColMat"),
+      "Output(ColMat) of FusionSeqConvEltAddReluOp should not be null.");
+
+  auto x_dims = ctx->GetInputDim("X");
+  auto w_dims = ctx->GetInputDim("Filter");
+  int context_length = ctx->Attrs().Get<int>("contextLength");
+  PADDLE_ENFORCE(
+      ctx->Attrs().Get<int>("contextStride") == 1,
+      "Currently, FusionSeqConvEltAddReluOp only supports contextStride=1.");
+  PADDLE_ENFORCE(x_dims.size() == 2 && w_dims.size() == 2,
+                 "Input(X, Filter) should be 2-D tensor.");
+  PADDLE_ENFORCE(x_dims.size() == 2 && w_dims.size() == 2,
+                 "Input(X, Filter) should be 2-D tensor.");
+  PADDLE_ENFORCE(w_dims[0] == context_length * x_dims[1],
+                 "Filter's height should be context_length * "
+                 "input_hidden_size .");
+  PADDLE_ENFORCE_GT(context_length + ctx->Attrs().Get<int>("contextStart"), 0,
+                    "contextStart size should be smaller than contextLength.");
+
+  ctx->SetOutputDim("Out", {x_dims[0], w_dims[1]});
+  ctx->SetOutputDim("ColMat", {x_dims[0], w_dims[0]});
+  ctx->ShareLoD("X", "Out");
+}
+
+framework::OpKernelType FusionSeqConvEltAddReluOp::GetExpectedKernelType(
+    const framework::ExecutionContext& ctx) const {
+  return framework::OpKernelType(
+      framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
+      ctx.device_context());
+}
+
+void FusionSeqConvEltAddReluOpMaker::Make() {
+  AddInput("X",
+           "(LoDTensor) the input is a LodTensor, which support "
+           "variable-time length input sequence. The underlying tensor in "
+           "this LoDTensor is a matrix with shape (T X M), where T is the "
+           "total time steps in this mini-batch, M is the dim size of x.");
+  // PaddingData only support false yet, should be ensured at pass.
+  AddInput("Filter",
+           "(Tensor) same as the input(Filter) of sequence conv op is an "
+           "learnable parameter."
+           "This is a tensor with shape (K, N), where K is the "
+           "context_length * dim size of x, N is the output feature size.");
+  AddInput("Bias",
+           "(Tensor) the learnable weights. shape (1, N), where N is the "
+           "output feature size");
+  AddOutput(
+      "Out",
+      "(LoDTensor) the output(Out) is a LodTensor, which support "
+      "variable-time length output sequence. The underlying tensor in "
+      "this LoDTensor is a matrix with shape (T, N), where, T is the "
+      "total time steps in this mini-batch, N is the output feature size.");
+  AddOutput("ColMat",
+            "(Tensor) (T, K), where T is where T is the "
+            "total time steps in this mini-batch, K is height of Filter")
+      .AsIntermediate();
+  AddAttr<int>("contextLength",
+               "(int) the contextLength of FusionSeqConvEltAddReluOp is the "
+               "height of the convolution kernel.")
+      .GreaterThan(0);
+  AddAttr<int>("contextStart",
+               "(int, default:0) the contextStart of FusionSeqConvEltAddReluOp "
+               "represents the beginning of the convolution of the number of "
+               "rows of sequence, which can be negative. The negative number "
+               "means to pad contextStart time-steps of zeros or learnable "
+               "parameters at the beginning of each instance. The positive "
+               "number means to skip contextStart time-steps of each "
+               "instance.")
+      .SetDefault(0);
+  AddAttr<int>(
+      "contextStride",
+      "(int, default:1) the contextStride of FusionSeqConvEltAddReluOp "
+      "represents the stride length of convolution kernel. "
+      "Currently, FusionSeqConvEltAddReluOp only supports"
+      "contextStride=1.")
+      .SetDefault(1)
+      .GreaterThan(0);
+  AddComment(R"DOC(
+Fusion Sequence Conv and ElementwiseAdd Operator.
+)DOC");
+}
+
+template <typename T>
+class FusionSeqConvEltAddReluKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    using DeviceContext = paddle::platform::CPUDeviceContext;
+    auto* x = ctx.Input<LoDTensor>("X");
+    auto* w = ctx.Input<Tensor>("Filter");
+    auto* b = ctx.Input<Tensor>("Bias");
+    auto* y = ctx.Output<LoDTensor>("Out");
+    auto* col = ctx.Output<Tensor>("ColMat");
+
+    auto x_lod = x->lod();
+    auto x_dims = x->dims();
+    auto w_dims = w->dims();
+    PADDLE_ENFORCE_EQ(b->numel(), w_dims[1],
+                      "bias size should be equal to output feature size.");
+    PADDLE_ENFORCE_EQ(x_lod.size(), 1UL,
+                      "Only support one level sequence now.");
+
+    const T* x_data = x->data<T>();
+    const T* w_data = w->data<T>();
+    const T* b_data = b->data<T>();
+    T* y_data = y->mutable_data<T>(ctx.GetPlace());
+    T* col_data = col->mutable_data<T>(ctx.GetPlace());
+
+    int context_start = ctx.Attr<int>("contextStart");
+    int context_length = ctx.Attr<int>("contextLength");
+    int up_pad = std::max(0, -context_start);
+    int down_pad = std::max(0, context_start + context_length - 1);
+    // im2col
+    int src_mat_w = static_cast<int>(x_dims[1]);
+    int src_mat_w_sz = src_mat_w * sizeof(T);
+    int col_mat_w = static_cast<int>(w_dims[0]);
+    int col_mat_w_sz = col_mat_w * sizeof(T);
+    for (int i = 0; i < static_cast<int>(x_lod[0].size()) - 1; ++i) {
+      int st = x_lod[0][i];
+      int ed = x_lod[0][i + 1];
+      const T* src_data = x_data + st * src_mat_w;
+      T* dst_data = col_data + st * col_mat_w;
+      int seq_len = ed - st;
+      if (seq_len > up_pad + down_pad) {
+        // zero all up_pad and fill data
+        std::memset(dst_data, 0, up_pad * col_mat_w_sz);
+        dst_data = dst_data + up_pad * src_mat_w;
+        int copy_size = col_mat_w_sz - up_pad * src_mat_w_sz;
+        for (int j = 0; j < up_pad; ++j) {
+          // blas.VCOPY?
+          std::memcpy(dst_data, src_data, copy_size);
+          dst_data += (col_mat_w - src_mat_w);
+          copy_size += src_mat_w_sz;
+        }
+        // fill data
+        for (int j = 0; j < seq_len - up_pad - down_pad; ++j) {
+          std::memcpy(dst_data, src_data, copy_size);
+          dst_data += col_mat_w;
+          src_data += src_mat_w;
+        }
+        // zero all down_pad and fill data
+        std::memset(dst_data, 0, down_pad * col_mat_w_sz);
+        copy_size -= src_mat_w_sz;
+        for (int j = 0; j < down_pad; ++j) {
+          std::memcpy(dst_data, src_data, copy_size);
+          dst_data += col_mat_w;
+          src_data += src_mat_w;
+          copy_size -= src_mat_w_sz;
+        }
+      } else {
+        PADDLE_ENFORCE_GE(context_length, up_pad + down_pad + 1);
+        std::memset(dst_data, 0, seq_len * col_mat_w_sz);
+        dst_data = dst_data + up_pad * src_mat_w;
+        int zero_sz = up_pad * src_mat_w_sz;
+        int cur_src_sz = seq_len * src_mat_w_sz;
+        for (int j = 0; j < std::min(up_pad, seq_len); ++j) {
+          int copy_size = std::min(cur_src_sz, col_mat_w_sz - zero_sz);
+          std::memcpy(dst_data, src_data, copy_size);
+          dst_data += (col_mat_w - src_mat_w);
+          zero_sz -= src_mat_w_sz;
+        }
+        // from bottom
+        dst_data = col_data + ed * col_mat_w;
+        src_data = x_data + st * src_mat_w;
+        zero_sz = down_pad * src_mat_w_sz;
+        for (int j = 1; j <= std::min(down_pad, seq_len); ++j) {
+          int copy_size = std::min(cur_src_sz, col_mat_w_sz - zero_sz);
+          std::memcpy(dst_data - (zero_sz + copy_size) / sizeof(T),
+                      src_data + std::max(seq_len - j - up_pad, 0) * src_mat_w,
+                      copy_size);
+          dst_data -= col_mat_w;
+          zero_sz -= src_mat_w_sz;
+        }
+      }
+    }
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
+    auto blas = math::GetBlas<DeviceContext, T>(dev_ctx);
+    math::FCCompute<DeviceContext, T>(blas, x_dims[0], w_dims[1], w_dims[0],
+                                      col_data, w_data, y_data, b_data, true);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(fusion_seqconv_eltadd_relu, ops::FusionSeqConvEltAddReluOp,
+                  ops::FusionSeqConvEltAddReluOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+
+REGISTER_OP_CPU_KERNEL(fusion_seqconv_eltadd_relu,
+                       ops::FusionSeqConvEltAddReluKernel<float>,
+                       ops::FusionSeqConvEltAddReluKernel<double>);

paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.h

+/* Copyright (c) 2018 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 "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using LoDTensor = framework::LoDTensor;
+using Tensor = framework::Tensor;
+
+class FusionSeqConvEltAddReluOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override;
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override;
+};
+
+class FusionSeqConvEltAddReluOpMaker
+    : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override;
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/gather.h

@@ -39,11 +39,9 @@ void CPUGather(const platform::DeviceContext& ctx, const Tensor& src,
   PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()));
   // check index of shape 1-D
   PADDLE_ENFORCE(index.dims().size() == 1);
-  int index_size = index.dims()[0];
+  int64_t index_size = index.dims()[0];
 
   auto src_dims = src.dims();
-  framework::DDim output_dims(src_dims);
-  output_dims[0] = index_size;
 
   const T* p_src = src.data<T>();
   const int* p_index = index.data<int>();
@@ -55,7 +53,7 @@ void CPUGather(const platform::DeviceContext& ctx, const Tensor& src,
 
   const size_t slice_bytes = slice_size * sizeof(T);
 
-  for (int i = 0; i < index_size; ++i) {
+  for (int64_t i = 0; i < index_size; ++i) {
     int index_ = p_index[i];
     memcpy(p_output + i * slice_size, p_src + index_ * slice_size, slice_bytes);
   }

paddle/fluid/operators/math/CMakeLists.txt

@@ -76,5 +76,5 @@ cc_test(concat_test SRCS concat_test.cc DEPS concat)
 cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info)
 cc_library(jit_kernel 
     SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_lstm.cc
-    DEPS cpu_info cblas activation_functions)
+    DEPS cpu_info cblas)
 cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel)

paddle/fluid/operators/math/fc_compute.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/jit_kernel.h"
 
 DECLARE_int32(paddle_num_threads);
 
@@ -30,20 +31,25 @@ inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
   if (B == NULL) {
     return;
   }
+  if (relu) {
+    const auto& vaddrelu = jitkernel::KernelPool::Instance()
+                               .template Get<jitkernel::VAddReluKernel<T>>(N);
+    for (int i = 0; i < M; i++) {
+      T* dst = Y + i * N;
+      vaddrelu->Compute(B, dst, dst);
+    }
+  } else {
+    const auto& vadd = jitkernel::KernelPool::Instance()
+                           .template Get<jitkernel::VAddKernel<T>>(N);
 
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for if (FLAGS_paddle_num_threads > 1)
 #endif
     for (int i = 0; i < M; i++) {
-    blas.AXPY(N, static_cast<T>(1), B, Y + i * N);
+      T* dst = Y + i * N;
+      vadd->Compute(B, dst, dst);
     }
-
-  if (!relu) {
-    return;
   }
-
-  // TODO(TJ): fuse relu
-  LOG(FATAL) << "Not implemented!";
 }
 
 }  // namespace math

paddle/fluid/operators/math/jit_kernel.h

@@ -86,6 +86,12 @@ class VAddBiasKernel : public Kernel {
   virtual void Compute(const T a, const T *x, T *y) const = 0;
 };
 
+template <typename T>
+class VAddReluKernel : public Kernel {
+ public:
+  virtual void Compute(const T *x, const T *y, T *z) const = 0;
+};
+
 template <typename T>
 class VActKernel : public Kernel {
  public:

paddle/fluid/operators/math/jit_kernel_blas.cc

@@ -378,11 +378,99 @@ class VIdentityKernelImpl : public VIdentityKernel<T> {
   void Compute(const T* x, T* y) const override {}
 };
 
+/* VAddRelu JitKernel */
+template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+class VAddReluKernelImpl : public VAddReluKernel<T> {
+ public:
+  explicit VAddReluKernelImpl(int d) : VAddReluKernel<T>() { this->num_ = d; }
+  void Compute(const T* x, const T* y, T* z) const override {
+    for (int i = 0; i < this->num_; ++i) {
+      z[i] = x[i] + y[i];
+      z[i] = z[i] > 0 ? z[i] : 0;
+    }
+  }
+};
+
+#define INTRI8_FLOAT(isa)                               \
+  template <>                                           \
+  void VAddReluKernelImpl<float, isa, kEQ8>::Compute(   \
+      const float* x, const float* y, float* z) const { \
+    __m256 tmpx = _mm256_loadu_ps(x);                   \
+    __m256 tmpy = _mm256_loadu_ps(y);                   \
+    tmpy = _mm256_add_ps(tmpx, tmpy);                   \
+    tmpy = _mm256_max_ps(tmpy, _mm256_setzero_ps());    \
+    _mm256_storeu_ps(z, tmpy);                          \
+  }
+
+#define INTRI16_FLOAT(isa)                              \
+  template <>                                           \
+  void VAddReluKernelImpl<float, isa, kEQ16>::Compute(  \
+      const float* x, const float* y, float* z) const { \
+    __m256 zeros = _mm256_setzero_ps();                 \
+    __m256 tmp0 = _mm256_loadu_ps(x);                   \
+    __m256 tmp1 = _mm256_loadu_ps(y);                   \
+    tmp0 = _mm256_add_ps(tmp0, tmp1);                   \
+    tmp0 = _mm256_max_ps(tmp0, zeros);                  \
+    tmp1 = _mm256_loadu_ps(x + 8);                      \
+    __m256 tmp2 = _mm256_loadu_ps(y + 8);               \
+    tmp1 = _mm256_add_ps(tmp1, tmp2);                   \
+    tmp1 = _mm256_max_ps(tmp1, zeros);                  \
+    _mm256_storeu_ps(z, tmp0);                          \
+    _mm256_storeu_ps(z + 8, tmp1);                      \
+  }
+
+#define INTRI_COMMON_FLOAT(isa, block)                             \
+  template <>                                                      \
+  VAddReluKernelImpl<float, isa, block>::VAddReluKernelImpl(int d) \
+      : VAddReluKernel<float>() {                                  \
+    this->num_ = d;                                                \
+    this->end_ = d - d % AVX_FLOAT_BLOCK;                          \
+    this->rest_ = d - this->end_;                                  \
+  }                                                                \
+  template <>                                                      \
+  void VAddReluKernelImpl<float, isa, block>::Compute(             \
+      const float* x, const float* y, float* z) const {            \
+    __m256 zeros = _mm256_setzero_ps();                            \
+    for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) {        \
+      __m256 tmpx = _mm256_loadu_ps(x + i);                        \
+      __m256 tmpy = _mm256_loadu_ps(y + i);                        \
+      tmpy = _mm256_add_ps(tmpx, tmpy);                            \
+      tmpy = _mm256_max_ps(tmpy, zeros);                           \
+      _mm256_storeu_ps(z + i, tmpy);                               \
+    }                                                              \
+    for (int i = this->end_; i < this->num_; ++i) {                \
+      z[i] = x[i] + y[i];                                          \
+      z[i] = z[i] > 0 ? z[i] : 0;                                  \
+    }                                                              \
+  }
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+INTRI16_FLOAT(jit::avx);
+INTRI_COMMON_FLOAT(jit::avx, kGT16);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+INTRI16_FLOAT(jit::avx2);
+INTRI_COMMON_FLOAT(jit::avx2, kGT16);
+#endif
+#ifdef __AVX512F__
+// TODO(TJ): refine avx512
+INTRI8_FLOAT(jit::avx512f);
+INTRI16_FLOAT(jit::avx512f);
+INTRI_COMMON_FLOAT(jit::avx512f, kGT16);
+#endif
+
+#undef INTRI8_FLOAT
+#undef INTRI16_FLOAT
+#undef INTRI_COMMON_FLOAT
+
 REGISTER_JITKERNEL(vmul, VMulKernel);
 REGISTER_JITKERNEL(vadd, VAddKernel);
 REGISTER_JITKERNEL(vscal, VScalKernel);
 REGISTER_JITKERNEL(vaddb, VAddBiasKernel);
 REGISTER_JITKERNEL(vrelu, VReluKernel);
+REGISTER_JITKERNEL(vaddrelu, VAddReluKernel);
 REGISTER_JITKERNEL(videntity, VIdentityKernel);
 
 }  // namespace jitkernel

paddle/fluid/operators/math/jit_kernel_exp.cc

@@ -27,13 +27,6 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 namespace math {
-
-#ifdef __AVX__
-namespace detail {
-__m256 Exp(__m256 a);
-}  // namespace detail
-#endif
-
 namespace jitkernel {
 namespace jit = platform::jit;
 
@@ -69,37 +62,186 @@ FOR_EACH_ISA(MKL_FLOAT, kGT16);
 FOR_EACH_ISA_BLOCK(MKL_DOUBLE);
 #endif
 
-#define INTRI8_FLOAT(isa)                                                  \
+namespace detail {
+
+#ifdef __AVX__
+
+#define ALIGN32 __attribute__((aligned(32)))
+
+#define _PS256_CONST(Name, Val)                                      \
+  static const float _ps256_##Name[8] ALIGN32 = {Val, Val, Val, Val, \
+                                                 Val, Val, Val, Val}
+
+#define _PI256_CONST(Name, Val)                                    \
+  static const int _pi256_##Name[8] ALIGN32 = {Val, Val, Val, Val, \
+                                               Val, Val, Val, Val}
+
+_PI256_CONST(0x7f, 0x7f);
+_PS256_CONST(one, 1.f);
+_PS256_CONST(0p5, 0.5f);
+_PS256_CONST(exp_hi, 88.3762626647949f);
+_PS256_CONST(exp_lo, -88.3762626647949f);
+_PS256_CONST(cephes_LOG2EF, 1.44269504088896341);
+_PS256_CONST(cephes_exp_C1, 0.693359375);
+_PS256_CONST(cephes_exp_C2, -2.12194440e-4);
+_PS256_CONST(cephes_exp_p0, 1.9875691500E-4);
+_PS256_CONST(cephes_exp_p1, 1.3981999507E-3);
+_PS256_CONST(cephes_exp_p2, 8.3334519073E-3);
+_PS256_CONST(cephes_exp_p3, 4.1665795894E-2);
+_PS256_CONST(cephes_exp_p4, 1.6666665459E-1);
+_PS256_CONST(cephes_exp_p5, 5.0000001201E-1);
+
+typedef union imm_xmm_union {
+  __m256i imm;
+  __m128i xmm[2];
+} imm_xmm_union;
+
+#define COPY_IMM_TO_XMM(imm_, xmm0_, xmm1_) \
+  {                                         \
+    imm_xmm_union u ALIGN32;                \
+    u.imm = imm_;                           \
+    xmm0_ = u.xmm[0];                       \
+    xmm1_ = u.xmm[1];                       \
+  }
+
+#define COPY_XMM_TO_IMM(xmm0_, xmm1_, imm_) \
+  {                                         \
+    imm_xmm_union u ALIGN32;                \
+    u.xmm[0] = xmm0_;                       \
+    u.xmm[1] = xmm1_;                       \
+    imm_ = u.imm;                           \
+  }
+
+#define AVX2_BITOP_USING_SSE2(fn)                           \
+  static inline __m256i avx2_mm256_##fn(__m256i x, int y) { \
+    /* use SSE2 to perform the bitop AVX2 */                \
+    __m128i x1, x2;                                         \
+    __m256i ret;                                            \
+    COPY_IMM_TO_XMM(x, x1, x2);                             \
+    x1 = _mm_##fn(x1, y);                                   \
+    x2 = _mm_##fn(x2, y);                                   \
+    COPY_XMM_TO_IMM(x1, x2, ret);                           \
+    return ret;                                             \
+  }
+
+#define AVX2_INTOP_USING_SSE2(fn)                                    \
+  static inline __m256i avx2_mm256_add_epi32(__m256i x, __m256i y) { \
+    /* use SSE2 to perform the AVX2 integer operation */             \
+    __m128i x1, x2;                                                  \
+    __m128i y1, y2;                                                  \
+    __m256i ret;                                                     \
+    COPY_IMM_TO_XMM(x, x1, x2);                                      \
+    COPY_IMM_TO_XMM(y, y1, y2);                                      \
+    x1 = _mm_##fn(x1, y1);                                           \
+    x2 = _mm_##fn(x2, y2);                                           \
+    COPY_XMM_TO_IMM(x1, x2, ret);                                    \
+    return ret;                                                      \
+  }
+
+AVX2_BITOP_USING_SSE2(slli_epi32);
+AVX2_INTOP_USING_SSE2(add_epi32);
+
+#define AVXEXP_BASE                                                            \
+  __m256 tmp = _mm256_setzero_ps(), fx;                                        \
+  __m256 one = *reinterpret_cast<const __m256*>(_ps256_one);                   \
+  __m256i imm0;                                                                \
+  x = _mm256_min_ps(x, *reinterpret_cast<const __m256*>(_ps256_exp_hi));       \
+  x = _mm256_max_ps(x, *reinterpret_cast<const __m256*>(_ps256_exp_lo));       \
+  /* express exp(x) as exp(g + n*log(2)) */                                    \
+  fx = _mm256_mul_ps(x,                                                        \
+                     *reinterpret_cast<const __m256*>(_ps256_cephes_LOG2EF));  \
+  fx = _mm256_add_ps(fx, *reinterpret_cast<const __m256*>(_ps256_0p5));        \
+  tmp = _mm256_floor_ps(fx);                                                   \
+  /* if greater, substract 1 */                                                \
+  __m256 mask = _mm256_cmp_ps(tmp, fx, _CMP_GT_OS);                            \
+  mask = _mm256_and_ps(mask, one);                                             \
+  fx = _mm256_sub_ps(tmp, mask);                                               \
+  tmp = _mm256_mul_ps(fx,                                                      \
+                      *reinterpret_cast<const __m256*>(_ps256_cephes_exp_C1)); \
+  __m256 z = _mm256_mul_ps(                                                    \
+      fx, *reinterpret_cast<const __m256*>(_ps256_cephes_exp_C2));             \
+  x = _mm256_sub_ps(x, tmp);                                                   \
+  x = _mm256_sub_ps(x, z);                                                     \
+  z = _mm256_mul_ps(x, x);                                                     \
+  __m256 y = *reinterpret_cast<const __m256*>(_ps256_cephes_exp_p0);           \
+  y = _mm256_mul_ps(y, x);                                                     \
+  y = _mm256_add_ps(y,                                                         \
+                    *reinterpret_cast<const __m256*>(_ps256_cephes_exp_p1));   \
+  y = _mm256_mul_ps(y, x);                                                     \
+  y = _mm256_add_ps(y,                                                         \
+                    *reinterpret_cast<const __m256*>(_ps256_cephes_exp_p2));   \
+  y = _mm256_mul_ps(y, x);                                                     \
+  y = _mm256_add_ps(y,                                                         \
+                    *reinterpret_cast<const __m256*>(_ps256_cephes_exp_p3));   \
+  y = _mm256_mul_ps(y, x);                                                     \
+  y = _mm256_add_ps(y,                                                         \
+                    *reinterpret_cast<const __m256*>(_ps256_cephes_exp_p4));   \
+  y = _mm256_mul_ps(y, x);                                                     \
+  y = _mm256_add_ps(y,                                                         \
+                    *reinterpret_cast<const __m256*>(_ps256_cephes_exp_p5));   \
+  y = _mm256_mul_ps(y, z);                                                     \
+  y = _mm256_add_ps(y, x);                                                     \
+  y = _mm256_add_ps(y, one);                                                   \
+  /* build 2^n */                                                              \
+  imm0 = _mm256_cvttps_epi32(fx)
+
+__m256 ExpAVX(__m256 x) {
+  AVXEXP_BASE;
+  // two AVX2 instructions using SSE2
+  imm0 = avx2_mm256_add_epi32(imm0,
+                              *reinterpret_cast<const __m256i*>(_pi256_0x7f));
+  imm0 = avx2_mm256_slli_epi32(imm0, 23);
+  __m256 pow2n = _mm256_castsi256_ps(imm0);
+  y = _mm256_mul_ps(y, pow2n);
+  return y;
+}
+#endif
+
+#ifdef __AVX2__
+__m256 ExpAVX2(__m256 x) {
+  AVXEXP_BASE;
+  // two AVX2 instructions
+  imm0 = _mm256_add_epi32(imm0, *reinterpret_cast<const __m256i*>(_pi256_0x7f));
+  imm0 = _mm256_slli_epi32(imm0, 23);
+  __m256 pow2n = _mm256_castsi256_ps(imm0);
+  y = _mm256_mul_ps(y, pow2n);
+  return y;
+}
+#endif
+
+}  // namespace detail
+
+#define INTRI8_FLOAT(isa, expisa)                                          \
   template <>                                                              \
   void VExpKernelImpl<float, isa, kEQ8>::Compute(const float* x, float* y) \
       const {                                                              \
     __m256 tmp = _mm256_loadu_ps(x);                                       \
-    _mm256_storeu_ps(y, detail::Exp(tmp));                                 \
+    _mm256_storeu_ps(y, expisa(tmp));                                      \
   }
 
-#define INTRI16_FLOAT(isa)                                                  \
+#define INTRI16_FLOAT(isa, expisa)                                          \
   template <>                                                               \
   void VExpKernelImpl<float, isa, kEQ16>::Compute(const float* x, float* y) \
       const {                                                               \
     __m256 tmp0 = _mm256_loadu_ps(x);                                       \
     __m256 tmp1 = _mm256_loadu_ps(x + 8);                                   \
-    tmp0 = detail::Exp(tmp0);                                               \
-    tmp1 = detail::Exp(tmp1);                                               \
+    tmp0 = expisa(tmp0);                                                    \
+    tmp1 = expisa(tmp1);                                                    \
     _mm256_storeu_ps(y, tmp0);                                              \
     _mm256_storeu_ps(y + 8, tmp1);                                          \
   }
 
 #ifdef __AVX__
-INTRI8_FLOAT(jit::avx);
-INTRI16_FLOAT(jit::avx);
+INTRI8_FLOAT(jit::avx, detail::ExpAVX);
+INTRI16_FLOAT(jit::avx, detail::ExpAVX);
 #endif
 #ifdef __AVX2__
-INTRI8_FLOAT(jit::avx2);
-INTRI16_FLOAT(jit::avx2);
+INTRI8_FLOAT(jit::avx2, detail::ExpAVX2);
+INTRI16_FLOAT(jit::avx2, detail::ExpAVX2);
 #endif
 #ifdef __AVX512F__
-INTRI8_FLOAT(jit::avx512f);
-INTRI16_FLOAT(jit::avx512f);
+INTRI8_FLOAT(jit::avx512f, detail::ExpAVX2);
+INTRI16_FLOAT(jit::avx512f, detail::ExpAVX2);
 #endif
 // TODO(TJ): eq16 test and complete avx512
 
@@ -135,26 +277,27 @@ class VSigmoidKernelImpl : public VSigmoidKernel<T> {
   std::shared_ptr<const VExpKernel<T>> vexp_;
 };
 
-#define INTRI_SIGMOID(tmp, min, max)              \
+#define INTRI_SIGMOID(tmp, min, max, expisa)      \
   tmp = _mm256_max_ps(tmp, min);                  \
   tmp = _mm256_min_ps(tmp, max);                  \
   tmp = _mm256_sub_ps(_mm256_set1_ps(0.0f), tmp); \
-  tmp = detail::Exp(tmp);                         \
+  tmp = expisa(tmp);                              \
   tmp = _mm256_add_ps(_mm256_set1_ps(1.0f), tmp); \
   tmp = _mm256_div_ps(_mm256_set1_ps(1.0f), tmp)
 
-#define INTRI8_FLOAT(isa)                                                      \
+#define INTRI8_FLOAT(isa, expisa)                                              \
   template <>                                                                  \
   void VSigmoidKernelImpl<float, isa, kEQ8>::Compute(const float* x, float* y) \
       const {                                                                  \
+    /* TODO(TJ): try to use static const*/                                     \
     __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);                        \
     __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);                        \
     __m256 tmp = _mm256_loadu_ps(x);                                           \
-    INTRI_SIGMOID(tmp, min, max);                                              \
+    INTRI_SIGMOID(tmp, min, max, expisa);                                      \
     _mm256_storeu_ps(y, tmp);                                                  \
   }
 
-#define INTRI16_FLOAT(isa)                                              \
+#define INTRI16_FLOAT(isa, expisa)                                      \
   template <>                                                           \
   void VSigmoidKernelImpl<float, isa, kEQ16>::Compute(const float* x,   \
                                                       float* y) const { \
@@ -162,13 +305,13 @@ class VSigmoidKernelImpl : public VSigmoidKernel<T> {
     __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);                 \
     __m256 tmp0 = _mm256_loadu_ps(x);                                   \
     __m256 tmp1 = _mm256_loadu_ps(x + 8);                               \
-    INTRI_SIGMOID(tmp0, min, max);                                      \
-    INTRI_SIGMOID(tmp1, min, max);                                      \
+    INTRI_SIGMOID(tmp0, min, max, expisa);                              \
+    INTRI_SIGMOID(tmp1, min, max, expisa);                              \
     _mm256_storeu_ps(y, tmp0);                                          \
     _mm256_storeu_ps(y + 8, tmp1);                                      \
   }
 
-#define INTRI_GT8LT16_FLOAT(isa)                                             \
+#define INTRI_GT8LT16_FLOAT(isa, expisa)                                     \
   template <>                                                                \
   VSigmoidKernelImpl<float, isa, kGT8LT16>::VSigmoidKernelImpl(int d)        \
       : VSigmoidKernel<float>() {                                            \
@@ -184,7 +327,7 @@ class VSigmoidKernelImpl : public VSigmoidKernel<T> {
     __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);                      \
     __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);                      \
     __m256 tmp = _mm256_loadu_ps(x);                                         \
-    INTRI_SIGMOID(tmp, min, max);                                            \
+    INTRI_SIGMOID(tmp, min, max, expisa);                                    \
     _mm256_storeu_ps(y, tmp);                                                \
     const float min_ = SIGMOID_THRESHOLD_MIN;                                \
     const float max_ = SIGMOID_THRESHOLD_MAX;                                \
@@ -198,7 +341,7 @@ class VSigmoidKernelImpl : public VSigmoidKernel<T> {
     }                                                                        \
   }
 
-#define INTRI_GT16_FLOAT(isa)                                                \
+#define INTRI_GT16_FLOAT(isa, expisa)                                        \
   template <>                                                                \
   VSigmoidKernelImpl<float, isa, kGT16>::VSigmoidKernelImpl(int d)           \
       : VSigmoidKernel<float>() {                                            \
@@ -215,7 +358,7 @@ class VSigmoidKernelImpl : public VSigmoidKernel<T> {
     __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);                      \
     for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) {                  \
       __m256 tmp = _mm256_loadu_ps(x + i);                                   \
-      INTRI_SIGMOID(tmp, min, max);                                          \
+      INTRI_SIGMOID(tmp, min, max, expisa);                                  \
       _mm256_storeu_ps(y + i, tmp);                                          \
     }                                                                        \
     const float min_ = SIGMOID_THRESHOLD_MIN;                                \
@@ -231,22 +374,20 @@ class VSigmoidKernelImpl : public VSigmoidKernel<T> {
   }
 
 #ifdef __AVX__
-INTRI8_FLOAT(jit::avx);
-INTRI16_FLOAT(jit::avx);
-INTRI_GT8LT16_FLOAT(jit::avx);
-INTRI_GT16_FLOAT(jit::avx);
+INTRI8_FLOAT(jit::avx, detail::ExpAVX);
+INTRI16_FLOAT(jit::avx, detail::ExpAVX);
+INTRI_GT8LT16_FLOAT(jit::avx, detail::ExpAVX);
+INTRI_GT16_FLOAT(jit::avx, detail::ExpAVX);
 #endif
 #ifdef __AVX2__
-INTRI8_FLOAT(jit::avx2);
-INTRI16_FLOAT(jit::avx2);
-// INTRI_GT8LT16_FLOAT(jit::avx2);
-// INTRI_GT16_FLOAT(jit::avx2);
+INTRI8_FLOAT(jit::avx2, detail::ExpAVX2);
+INTRI16_FLOAT(jit::avx2, detail::ExpAVX2);
+// maybe use avx at gt8lt16 and gt16
 #endif
 #ifdef __AVX512F__
-INTRI8_FLOAT(jit::avx512f);
-INTRI16_FLOAT(jit::avx512f);
-// INTRI_GT8LT16_FLOAT(jit::avx512f);
-// INTRI_GT16_FLOAT(jit::avx512f);
+INTRI8_FLOAT(jit::avx512f, detail::ExpAVX2);
+INTRI16_FLOAT(jit::avx512f, detail::ExpAVX2);
+// maybe use avx2 at gt8lt16 and gt16
 #endif
 
 #undef INTRI8_FLOAT
@@ -280,36 +421,36 @@ class VTanhKernelImpl : public VTanhKernel<T> {
   std::shared_ptr<const VAddBiasKernel<T>> vaddbias_;
 };
 
-#define INTRI_VTANH(tmp)                                   \
+#define INTRI_VTANH(tmp, expisa)                           \
   tmp = _mm256_mul_ps(_mm256_set1_ps(-2.0f), tmp);         \
   tmp = _mm256_min_ps(tmp, _mm256_set1_ps(EXP_MAX_INPUT)); \
-  tmp = detail::Exp(tmp);                                  \
+  tmp = expisa(tmp);                                       \
   tmp = _mm256_add_ps(_mm256_set1_ps(1.0f), tmp);          \
   tmp = _mm256_div_ps(_mm256_set1_ps(2.0f), tmp);          \
   tmp = _mm256_sub_ps(tmp, _mm256_set1_ps(1.0f))
 
-#define INTRI8_FLOAT(isa)                                                   \
+#define INTRI8_FLOAT(isa, expisa)                                           \
   template <>                                                               \
   void VTanhKernelImpl<float, isa, kEQ8>::Compute(const float* x, float* y) \
       const {                                                               \
     __m256 tmp = _mm256_loadu_ps(x);                                        \
-    INTRI_VTANH(tmp);                                                       \
+    INTRI_VTANH(tmp, expisa);                                               \
     _mm256_storeu_ps(y, tmp);                                               \
   }
 
-#define INTRI16_FLOAT(isa)                                                   \
+#define INTRI16_FLOAT(isa, expisa)                                           \
   template <>                                                                \
   void VTanhKernelImpl<float, isa, kEQ16>::Compute(const float* x, float* y) \
       const {                                                                \
     __m256 tmp0 = _mm256_loadu_ps(x);                                        \
     __m256 tmp1 = _mm256_loadu_ps(x + 8);                                    \
-    INTRI_VTANH(tmp0);                                                       \
-    INTRI_VTANH(tmp1);                                                       \
+    INTRI_VTANH(tmp0, expisa);                                               \
+    INTRI_VTANH(tmp1, expisa);                                               \
     _mm256_storeu_ps(y, tmp0);                                               \
     _mm256_storeu_ps(y + 8, tmp1);                                           \
   }
 
-#define INTRI_GT8LT16_FLOAT(isa)                                              \
+#define INTRI_GT8LT16_FLOAT(isa, expisa)                                      \
   template <>                                                                 \
   VTanhKernelImpl<float, isa, kGT8LT16>::VTanhKernelImpl(int d)               \
       : VTanhKernel<float>() {                                                \
@@ -327,7 +468,7 @@ class VTanhKernelImpl : public VTanhKernel<T> {
   void VTanhKernelImpl<float, isa, kGT8LT16>::Compute(const float* x,         \
                                                       float* y) const {       \
     __m256 tmp = _mm256_loadu_ps(x);                                          \
-    INTRI_VTANH(tmp);                                                         \
+    INTRI_VTANH(tmp, expisa);                                                 \
     _mm256_storeu_ps(y, tmp);                                                 \
     x += AVX_FLOAT_BLOCK;                                                     \
     y += AVX_FLOAT_BLOCK;                                                     \
@@ -337,7 +478,7 @@ class VTanhKernelImpl : public VTanhKernel<T> {
     vaddbias_->Compute(-1.f, y, y);                                           \
   }
 
-#define INTRI_GT16_FLOAT(isa)                                                 \
+#define INTRI_GT16_FLOAT(isa, expisa)                                         \
   template <>                                                                 \
   VTanhKernelImpl<float, isa, kGT16>::VTanhKernelImpl(int d)                  \
       : VTanhKernel<float>() {                                                \
@@ -356,7 +497,7 @@ class VTanhKernelImpl : public VTanhKernel<T> {
       const {                                                                 \
     for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) {                   \
       __m256 tmp = _mm256_loadu_ps(x + i);                                    \
-      INTRI_VTANH(tmp);                                                       \
+      INTRI_VTANH(tmp, expisa);                                               \
       _mm256_storeu_ps(y + i, tmp);                                           \
     }                                                                         \
     x += this->end_;                                                          \
@@ -368,19 +509,19 @@ class VTanhKernelImpl : public VTanhKernel<T> {
   }
 
 #ifdef __AVX__
-INTRI8_FLOAT(jit::avx);
-INTRI16_FLOAT(jit::avx);
-INTRI_GT8LT16_FLOAT(jit::avx);
-INTRI_GT16_FLOAT(jit::avx);
+INTRI8_FLOAT(jit::avx, detail::ExpAVX);
+INTRI16_FLOAT(jit::avx, detail::ExpAVX);
+INTRI_GT8LT16_FLOAT(jit::avx, detail::ExpAVX);
+INTRI_GT16_FLOAT(jit::avx, detail::ExpAVX);
 #endif
 #ifdef __AVX2__
-INTRI8_FLOAT(jit::avx2);
-INTRI16_FLOAT(jit::avx2);
+INTRI8_FLOAT(jit::avx2, detail::ExpAVX2);
+INTRI16_FLOAT(jit::avx2, detail::ExpAVX2);
 // maybe use avx at gt8lt16 and gt16
 #endif
 #ifdef __AVX512F__
-INTRI8_FLOAT(jit::avx512f);
-INTRI16_FLOAT(jit::avx512f);
+INTRI8_FLOAT(jit::avx512f, detail::ExpAVX2);
+INTRI16_FLOAT(jit::avx512f, detail::ExpAVX2);
 // maybe use avx at gt8lt16 and gt16
 #endif
 

paddle/fluid/operators/math/jit_kernel_lstm.cc

@@ -25,13 +25,18 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 namespace math {
-#ifdef __AVX__
+namespace jitkernel {
 namespace detail {
-__m256 Exp(__m256 a);
-}  // namespace detail
+#ifdef __AVX__
+__m256 ExpAVX(__m256 x);
 #endif
 
-namespace jitkernel {
+#ifdef __AVX2__
+__m256 ExpAVX2(__m256 x);
+#endif
+
+}  // namespace detail
+
 namespace jit = platform::jit;
 
 #ifdef __AVX__
@@ -43,43 +48,72 @@ class AVXAct {
   virtual __m256 Compute(__m256 x) const = 0;
 };
 
-template <act_type type>
+template <act_type type, jit::cpu_isa_t isa>
 class AVXActImpl : public AVXAct {
  public:
   __m256 Compute(__m256 x) const override { PADDLE_THROW("Unkown type!"); }
 };
 
-template <>
-__m256 AVXActImpl<kSigmoid>::Compute(__m256 x) const {
-  __m256 ones = _mm256_set1_ps(1.0f);
-  x = _mm256_max_ps(x, _mm256_set1_ps(SIGMOID_THRESHOLD_MIN));
-  x = _mm256_min_ps(x, _mm256_set1_ps(SIGMOID_THRESHOLD_MAX));
-  x = _mm256_sub_ps(_mm256_set1_ps(0.0f), x);
-  x = detail::Exp(x);
-  x = _mm256_add_ps(ones, x);
-  return _mm256_div_ps(ones, x);
-}
+#define AVX_SIGMOID(isa, expisa)                                 \
+  template <>                                                    \
+  __m256 AVXActImpl<kSigmoid, isa>::Compute(__m256 x) const {    \
+    __m256 ones = _mm256_set1_ps(1.0f);                          \
+    x = _mm256_max_ps(x, _mm256_set1_ps(SIGMOID_THRESHOLD_MIN)); \
+    x = _mm256_min_ps(x, _mm256_set1_ps(SIGMOID_THRESHOLD_MAX)); \
+    x = _mm256_sub_ps(_mm256_set1_ps(0.0f), x);                  \
+    x = expisa(x);                                               \
+    x = _mm256_add_ps(ones, x);                                  \
+    return _mm256_div_ps(ones, x);                               \
+  }
 
-template <>
-__m256 AVXActImpl<kTanh>::Compute(__m256 x) const {
-  __m256 ones = _mm256_set1_ps(1.0f);
-  x = _mm256_mul_ps(_mm256_set1_ps(-2.0f), x);
-  x = _mm256_min_ps(x, _mm256_set1_ps(EXP_MAX_INPUT));
-  x = detail::Exp(x);
-  x = _mm256_add_ps(ones, x);
-  x = _mm256_div_ps(_mm256_set1_ps(2.0f), x);
-  return _mm256_sub_ps(x, ones);
-}
+#define AVX_TANH(isa, expisa)                              \
+  template <>                                              \
+  __m256 AVXActImpl<kTanh, isa>::Compute(__m256 x) const { \
+    __m256 ones = _mm256_set1_ps(1.0f);                    \
+    x = _mm256_mul_ps(_mm256_set1_ps(-2.0f), x);           \
+    x = _mm256_min_ps(x, _mm256_set1_ps(EXP_MAX_INPUT));   \
+    x = expisa(x);                                         \
+    x = _mm256_add_ps(ones, x);                            \
+    x = _mm256_div_ps(_mm256_set1_ps(2.0f), x);            \
+    return _mm256_sub_ps(x, ones);                         \
+  }
 
-template <>
-__m256 AVXActImpl<kRelu>::Compute(__m256 x) const {
-  return _mm256_max_ps(x, _mm256_setzero_ps());
-}
+#define AVX_RELU(isa)                                      \
+  template <>                                              \
+  __m256 AVXActImpl<kRelu, isa>::Compute(__m256 x) const { \
+    return _mm256_max_ps(x, _mm256_setzero_ps());          \
+  }
+
+#define AVX_IDENTITY(isa)                                      \
+  template <>                                                  \
+  __m256 AVXActImpl<kIdentity, isa>::Compute(__m256 x) const { \
+    return x;                                                  \
+  }
+
+#define FOR_EACH_AVX_ISA(macro_) \
+  macro_(jit::avx);              \
+  macro_(jit::avx2);             \
+  macro_(jit::avx512f)
+
+FOR_EACH_AVX_ISA(AVX_RELU);
+FOR_EACH_AVX_ISA(AVX_IDENTITY);
+
+AVX_SIGMOID(jit::avx, detail::ExpAVX);
+AVX_TANH(jit::avx, detail::ExpAVX);
+
+#ifdef __AVX2__
+AVX_SIGMOID(jit::avx2, detail::ExpAVX2);
+AVX_SIGMOID(jit::avx512f, detail::ExpAVX2);
+AVX_TANH(jit::avx2, detail::ExpAVX2);
+AVX_TANH(jit::avx512f, detail::ExpAVX2);
+#endif
+
+#undef FOR_EACH_AVX_ISA
+#undef AVX_IDENTITY
+#undef AVX_RELU
+#undef AVX_TANH
+#undef AVX_SIGMOID
 
-template <>
-__m256 AVXActImpl<kIdentity>::Compute(__m256 x) const {
-  return x;
-}
 #endif
 
 template <typename T>
@@ -119,23 +153,6 @@ class LSTMKernelImpl : public LSTMKernel<T> {
     act_cell_d_ = GetActKernel<T>(act_cell, d);
     vmul_d_ = KernelPool::Instance().template Get<VMulKernel<T>>(d);
     vadd_d_ = KernelPool::Instance().template Get<VAddKernel<T>>(d);
-#ifdef __AVX__
-    auto GetAVXAct = [&](const std::string& type) -> std::unique_ptr<AVXAct> {
-      if (type == "sigmoid") {
-        return std::unique_ptr<AVXAct>(new AVXActImpl<kSigmoid>());
-      } else if (type == "relu") {
-        return std::unique_ptr<AVXAct>(new AVXActImpl<kRelu>());
-      } else if (type == "tanh") {
-        return std::unique_ptr<AVXAct>(new AVXActImpl<kTanh>());
-      } else if (type == "identity" || type == "") {
-        return std::unique_ptr<AVXAct>(new AVXActImpl<kIdentity>());
-      }
-      PADDLE_THROW("Not support type: %s", type);
-    };
-    avx_act_gate_ = GetAVXAct(act_gate);
-    avx_act_cand_ = GetAVXAct(act_cand);
-    avx_act_cell_ = GetAVXAct(act_cell);
-#endif
   }
 
   void ComputeCtHt(T* gates, const T* ct_1, T* ct, T* ht, const T* wp_data,
@@ -176,6 +193,27 @@ class LSTMKernelImpl : public LSTMKernel<T> {
 };
 
 #define INTRI8_FLOAT(isa)                                                      \
+  template <>                                                                  \
+  LSTMKernelImpl<float, isa, kEQ8>::LSTMKernelImpl(                            \
+      const std::string& act_gate, const std::string& act_cand,                \
+      const std::string& act_cell, int d)                                      \
+      : LSTMKernel<float>() {                                                  \
+    auto GetAVXAct = [&](const std::string& type) -> std::unique_ptr<AVXAct> { \
+      if (type == "sigmoid") {                                                 \
+        return std::unique_ptr<AVXAct>(new AVXActImpl<kSigmoid, isa>());       \
+      } else if (type == "relu") {                                             \
+        return std::unique_ptr<AVXAct>(new AVXActImpl<kRelu, isa>());          \
+      } else if (type == "tanh") {                                             \
+        return std::unique_ptr<AVXAct>(new AVXActImpl<kTanh, isa>());          \
+      } else if (type == "identity" || type == "") {                           \
+        return std::unique_ptr<AVXAct>(new AVXActImpl<kIdentity, isa>());      \
+      }                                                                        \
+      PADDLE_THROW("Not support type: %s", type);                              \
+    };                                                                         \
+    avx_act_gate_ = GetAVXAct(act_gate);                                       \
+    avx_act_cand_ = GetAVXAct(act_cand);                                       \
+    avx_act_cell_ = GetAVXAct(act_cell);                                       \
+  }                                                                            \
   template <>                                                                  \
   void LSTMKernelImpl<float, isa, kEQ8>::ComputeCtHt(                          \
       float* gates, const float* ct_1, float* ct, float* ht,                   \
@@ -195,6 +233,20 @@ class LSTMKernelImpl : public LSTMKernel<T> {
     /* H_t = act_cell(C_t) * ogated */                                         \
     o = _mm256_mul_ps(avx_act_cell_->Compute(f), avx_act_gate_->Compute(o));   \
     _mm256_storeu_ps(ht, o);                                                   \
+  }                                                                            \
+  template <>                                                                  \
+  void LSTMKernelImpl<float, isa, kEQ8>::ComputeC1H1(                          \
+      float* gates, float* ct, float* ht, const float* wp_data) const {        \
+    __m256 c, i, o;                                                            \
+    c = _mm256_loadu_ps(gates);                                                \
+    i = _mm256_loadu_ps(gates + 8);                                            \
+    o = _mm256_loadu_ps(gates + 24);                                           \
+    /* C_t = igated * cgated*/                                                 \
+    c = _mm256_mul_ps(avx_act_gate_->Compute(i), avx_act_cand_->Compute(c));   \
+    _mm256_storeu_ps(ct, c);                                                   \
+    /* H_t = act_cell(C_t) * ogated */                                         \
+    o = _mm256_mul_ps(avx_act_cell_->Compute(c), avx_act_gate_->Compute(o));   \
+    _mm256_storeu_ps(ht, o);                                                   \
   }
 
 // TODO(TJ): optimize keq16

paddle/fluid/operators/math/jit_kernel_test.cc

@@ -712,6 +712,63 @@ TEST(JitKernel, vadd) {
   }
 }
 
+void vaddrelu_ref(const int n, const float* x, const float* y, float* z) {
+  for (int i = 0; i < n; ++i) {
+    z[i] = x[i] + y[i];
+    z[i] = z[i] > 0 ? z[i] : 0;
+  }
+}
+void vaddrelu_better(
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VAddKernel<float>>& vadd,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VReluKernel<float>>& vrelu,
+    const float* x, const float* y, float* z) {
+  vadd->Compute(x, y, z);
+  vrelu->Compute(z, z);
+}
+
+TEST(JitKernel, vaddrelu) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
+    std::vector<float> x(d), y(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data());
+    RandomVec<float>(d, y.data());
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VAddReluKernel<float>>(d);
+    const auto& vadd =
+        jit::KernelPool::Instance().template Get<jit::VAddKernel<float>>(d);
+    const auto& vrelu =
+        jit::KernelPool::Instance().template Get<jit::VReluKernel<float>>(d);
+    const float* x_data = x.data();
+    const float* y_data = y.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vadd_ref(d, x_data, y_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vaddrelu_better(vadd, vrelu, x_data, y_data, zref_data);
+    }
+    auto tmkle = GetCurrentUS();
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, y_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, better takes: " << (tmkle - tmkls) / repeat << " us, "
+            << "tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
 TEST(JitKernel, pool) {
   namespace jit = paddle::operators::math::jitkernel;
   const int frame_size = 4;

paddle/fluid/operators/roi_align_op.cc

+/* Copyright (c) 2018 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/operators/roi_align_op.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+
+class ROIAlignOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of ROIAlignOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("ROIs"),
+                   "Input(ROIs) of ROIAlignOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of ROIAlignOp should not be null.");
+    auto input_dims = ctx->GetInputDim("X");
+    auto rois_dims = ctx->GetInputDim("ROIs");
+
+    PADDLE_ENFORCE(input_dims.size() == 4,
+                   "The format of input tensor is NCHW.");
+    PADDLE_ENFORCE(rois_dims.size() == 2,
+                   "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)"
+                   "given as [[x1, y1, x2, y2], …].");
+    PADDLE_ENFORCE(rois_dims[1] == 4,
+                   "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)"
+                   "given as [[x1, y1, x2, y2], …].");
+    int pooled_height = ctx->Attrs().Get<int>("pooled_height");
+    int pooled_width = ctx->Attrs().Get<int>("pooled_width");
+    float spatial_scale = ctx->Attrs().Get<float>("spatial_scale");
+
+    PADDLE_ENFORCE_GT(pooled_height, 0,
+                      "The pooled output height must greater than 0");
+    PADDLE_ENFORCE_GT(pooled_width, 0,
+                      "The pooled output width must greater than 0");
+    PADDLE_ENFORCE_GT(spatial_scale, 0.0f,
+                      "The spatial scale must greater than 0");
+
+    auto out_dims = input_dims;
+    out_dims[0] = rois_dims[0];
+    out_dims[1] = input_dims[1];
+    out_dims[2] = pooled_height;
+    out_dims[3] = pooled_width;
+
+    ctx->SetOutputDim("Out", out_dims);
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
+        ctx.device_context());
+  }
+};
+
+class ROIAlignGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "The GRAD@Out of ROIAlignGradOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName("X")),
+                   "The GRAD@X of ROIAlignGradOp should not be null.");
+    ctx->SetOutputsDim(framework::GradVarName("X"), ctx->GetInputsDim("X"));
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
+        ctx.device_context());
+  }
+};
+
+class ROIAlignOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(Tensor), "
+             "The input of ROIAlignOp. "
+             "The format of input tensor is NCHW. Where N is batch size, "
+             "C is the number of input channels, "
+             "H is the height of the feature, and "
+             "W is the width of the feature.");
+    AddInput("ROIs",
+             "(LoDTensor), "
+             "ROIs (Regions of Interest) to pool over. "
+             "should be a 2-D LoDTensor of shape (num_rois, 4)"
+             "given as [[x1, y1, x2, y2], …]. "
+             "(x1, y1) is the top left coordinates, and "
+             "(x2, y2) is the bottom right coordinates.");
+    AddOutput("Out",
+              "(Tensor), "
+              "The output of ROIAlignOp is a 4-D tensor with shape "
+              "(num_rois, channels, pooled_h, pooled_w).");
+    AddAttr<float>("spatial_scale",
+                   "(float, default 1.0), "
+                   "Multiplicative spatial scale factor "
+                   "to translate ROI coords from their input scale "
+                   "to the scale used when pooling.")
+        .SetDefault(1.0);
+    AddAttr<int>("pooled_height",
+                 "(int, default 1), "
+                 "The pooled output height.")
+        .SetDefault(1);
+    AddAttr<int>("pooled_width",
+                 "(int, default 1), "
+                 "The pooled output width.")
+        .SetDefault(1);
+    AddAttr<int>("sampling_ratio",
+                 "(int,default -1),"
+                 "number of sampling points in the interpolation grid"
+                 "If <=0, then grid points are adaptive to roi_width "
+                 "and pooled_w, likewise for height")
+        .SetDefault(-1);
+    AddComment(R"DOC(
+**RoIAlign Operator**
+
+Region of interest align (also known as RoI align) is to perform
+bilinear interpolation on inputs of nonuniform sizes to obtain 
+fixed-size feature maps (e.g. 7*7)
+
+Dividing each region proposal into equal-sized sections with
+the pooled_width and pooled_height. Location remains the origin
+result.
+
+In each ROI bin, the value of the four regularly sampled locations 
+are computed directly through bilinear interpolation. The output is
+the mean of four locations.
+Thus avoid the misaligned problem.   
+    )DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(roi_align, ops::ROIAlignOp, ops::ROIAlignOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+REGISTER_OPERATOR(roi_align_grad, ops::ROIAlignGradOp);
+REGISTER_OP_CPU_KERNEL(
+    roi_align,
+    ops::CPUROIAlignOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::CPUROIAlignOpKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    roi_align_grad,
+    ops::CPUROIAlignGradOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::CPUROIAlignGradOpKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/roi_align_op.cu

+/* Copyright (c) 2016 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/operators/roi_align_op.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+
+static constexpr int kNumCUDAThreads = 512;
+static constexpr int kNumMaxinumNumBlocks = 4096;
+
+static inline int NumBlocks(const int N) {
+  return std::min((N + kNumCUDAThreads - 1) / kNumCUDAThreads,
+                  kNumMaxinumNumBlocks);
+}
+
+#define CUDA_1D_KERNEL_LOOP(i, n)                              \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \
+       i += blockDim.x * gridDim.x)
+
+template <class T>
+__device__ T BilinearInterpolate(const T* input_data, const int height,
+                                 const int width, T y, T x) {
+  if (y < -1.0 || y > height || x < -1.0 || x > width) {
+    return 0;
+  }
+  y = y <= 0 ? 0 : y;
+  x = x <= 0 ? 0 : x;
+  int y_low = static_cast<int>(y);
+  int x_low = static_cast<int>(x);
+  int y_high;
+  int x_high;
+  if (y_low >= height - 1) {
+    y_high = y_low = height - 1;
+    y = static_cast<T>(y_low);
+  } else {
+    y_high = y_low + 1;
+  }
+  if (x_low >= width - 1) {
+    x_high = x_low = width - 1;
+    x = static_cast<T>(x_low);
+  } else {
+    x_high = x_low + 1;
+  }
+  T ly = y - y_low, lx = x - x_low;
+  T hy = 1. - ly, hx = 1. - lx;
+
+  T v1 = input_data[y_low * width + x_low];
+  T v2 = input_data[y_low * width + x_high];
+  T v3 = input_data[y_high * width + x_low];
+  T v4 = input_data[y_high * width + x_high];
+  T w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx;
+
+  T val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4);
+  return val;
+}
+
+template <class T>
+__device__ void BilinearInterpolateGradient(const int height, const int width,
+                                            T y, T x, T* w1, T* w2, T* w3,
+                                            T* w4, int* x_low, int* x_high,
+                                            int* y_low, int* y_high) {
+  if (y < -1.0 || y > height || x < -1.0 || x > width) {
+    return;
+  }
+
+  y = y <= 0 ? 0 : y;
+  x = x <= 0 ? 0 : x;
+  *y_low = static_cast<int>(y);
+  *x_low = static_cast<int>(x);
+  if (*y_low >= height - 1) {
+    *y_high = *y_low = height - 1;
+    y = static_cast<T>(*y_low);
+  } else {
+    *y_high = *y_low + 1;
+  }
+  if (*x_low >= width - 1) {
+    *x_high = *x_low = width - 1;
+    x = static_cast<T>(*x_low);
+  } else {
+    *x_high = *x_low + 1;
+  }
+  T ly = y - *y_low, lx = x - *x_low;
+  T hy = 1. - ly, hx = 1. - lx;
+  *w1 = hy * hx, *w2 = hy * lx, *w3 = ly * hx, *w4 = ly * lx;
+
+  return;
+}
+
+template <class T>
+__global__ void GPUROIAlignForward(
+    const int nthreads, const T* input_data, const T* input_rois,
+    const float spatial_scale, const int channels, const int height,
+    const int width, const int pooled_height, const int pooled_width,
+    const int sampling_ratio, int* roi_batch_id_data, T* output_data) {
+  CUDA_1D_KERNEL_LOOP(i, nthreads) {
+    int pw = i % pooled_width;
+    int ph = (i / pooled_width) % pooled_height;
+    int c = (i / pooled_width / pooled_height) % channels;
+    int n = i / pooled_width / pooled_height / channels;
+
+    const T* offset_input_rois = input_rois + n * kROISize;
+    int roi_batch_ind = roi_batch_id_data[n];
+
+    T roi_xmin = offset_input_rois[0] * spatial_scale;
+    T roi_ymin = offset_input_rois[1] * spatial_scale;
+    T roi_xmax = offset_input_rois[2] * spatial_scale;
+    T roi_ymax = offset_input_rois[3] * spatial_scale;
+
+    T roi_width = max(roi_xmax - roi_xmin, static_cast<T>(1.));
+    T roi_height = max(roi_ymax - roi_ymin, static_cast<T>(1.));
+    T bin_size_h = static_cast<T>(roi_height) / static_cast<T>(pooled_height);
+    T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
+
+    const T* offset_input_data =
+        input_data + (roi_batch_ind * channels + c) * height * width;
+
+    int roi_bin_grid_h = (sampling_ratio > 0)
+                             ? sampling_ratio
+                             : ceil(roi_height / pooled_height);
+    int roi_bin_grid_w =
+        (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width);
+    const T count = roi_bin_grid_h * roi_bin_grid_w;
+    T output_val = 0;
+    for (int iy = 0; iy < roi_bin_grid_h; iy++) {
+      const T y = roi_ymin + ph * bin_size_h +
+                  static_cast<T>(iy + .5f) * bin_size_h /
+                      static_cast<T>(roi_bin_grid_h);
+      for (int ix = 0; ix < roi_bin_grid_w; ix++) {
+        const T x = roi_xmin + pw * bin_size_w +
+                    static_cast<T>(ix + .5f) * bin_size_w /
+                        static_cast<T>(roi_bin_grid_w);
+        T val = BilinearInterpolate(offset_input_data, height, width, y, x);
+        output_val += val;
+      }
+    }
+    output_val /= count;
+    output_data[i] = output_val;
+  }
+}
+
+template <typename T>
+__global__ void GPUROIAlignBackward(const int nthreads, const T* input_rois,
+                                    const T* out_grad, const int num_rois,
+                                    const float spatial_scale,
+                                    const int channels, const int height,
+                                    const int width, const int pooled_height,
+                                    const int pooled_width,
+                                    const int sampling_ratio,
+                                    int* roi_batch_id_data, T* input_grad) {
+  CUDA_1D_KERNEL_LOOP(i, nthreads) {
+    int pw = i % pooled_width;
+    int ph = (i / pooled_width) % pooled_height;
+    int c = (i / pooled_width / pooled_height) % channels;
+    int n = i / pooled_width / pooled_height / channels;
+    const T* offset_input_rois = input_rois + n * kROISize;
+    int roi_batch_ind = roi_batch_id_data[n];
+
+    T roi_xmin = offset_input_rois[0] * spatial_scale;
+    T roi_ymin = offset_input_rois[1] * spatial_scale;
+    T roi_xmax = offset_input_rois[2] * spatial_scale;
+    T roi_ymax = offset_input_rois[3] * spatial_scale;
+
+    T roi_width = max(roi_xmax - roi_xmin, static_cast<T>(1.));
+    T roi_height = max(roi_ymax - roi_ymin, static_cast<T>(1.));
+    T bin_size_h = static_cast<T>(roi_height) / static_cast<T>(pooled_height);
+    T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
+
+    T* offset_input_grad =
+        input_grad + (roi_batch_ind * channels + c) * height * width;
+
+    const T* offset_out_grad =
+        out_grad + (n * channels + c) * pooled_height * pooled_width;
+    const T out_grad_this_bin = offset_out_grad[ph * pooled_width + pw];
+
+    int roi_bin_grid_h = (sampling_ratio > 0)
+                             ? sampling_ratio
+                             : ceil(roi_height / pooled_height);
+    int roi_bin_grid_w =
+        (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width);
+
+    const T count = roi_bin_grid_h * roi_bin_grid_w;
+    for (int iy = 0; iy < roi_bin_grid_h; iy++) {
+      const T y = roi_ymin + ph * bin_size_h +
+                  static_cast<T>(iy + .5f) * bin_size_h /
+                      static_cast<T>(roi_bin_grid_h);
+      for (int ix = 0; ix < roi_bin_grid_w; ix++) {
+        const T x = roi_xmin + pw * bin_size_w +
+                    static_cast<T>(ix + .5f) * bin_size_w /
+                        static_cast<T>(roi_bin_grid_w);
+        T w1 = 0, w2 = 0, w3 = 0, w4 = 0;
+        int x_low = -1, x_high = -1, y_low = -1, y_high = -1;
+        BilinearInterpolateGradient(height, width, y, x, &w1, &w2, &w3, &w4,
+                                    &x_low, &x_high, &y_low, &y_high);
+        T diff1 = out_grad_this_bin * w1 / count;
+        T diff2 = out_grad_this_bin * w2 / count;
+        T diff3 = out_grad_this_bin * w3 / count;
+        T diff4 = out_grad_this_bin * w4 / count;
+        if (x_low >= 0 && x_high >= 0 && y_low >= 0 && y_high >= 0) {
+          platform::CudaAtomicAdd(offset_input_grad + y_low * width + x_low,
+                                  diff1);
+          platform::CudaAtomicAdd(offset_input_grad + y_low * width + x_high,
+                                  diff2);
+          platform::CudaAtomicAdd(offset_input_grad + y_high * width + x_low,
+                                  diff3);
+          platform::CudaAtomicAdd(offset_input_grad + y_high * width + x_high,
+                                  diff4);
+        }
+      }
+    }
+  }
+}
+
+template <typename Place, typename T>
+class GPUROIAlignOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* in = ctx.Input<Tensor>("X");
+    auto* rois = ctx.Input<LoDTensor>("ROIs");
+    auto* out = ctx.Output<Tensor>("Out");
+
+    auto pooled_height = ctx.Attr<int>("pooled_height");
+    auto pooled_width = ctx.Attr<int>("pooled_width");
+    auto spatial_scale = ctx.Attr<float>("spatial_scale");
+    auto sampling_ratio = ctx.Attr<int>("sampling_ratio");
+
+    auto in_dims = in->dims();
+    int batch_size = in_dims[0];
+    int channels = in_dims[1];
+    int height = in_dims[2];
+    int width = in_dims[3];
+
+    int rois_num = rois->dims()[0];
+
+    if (rois_num == 0) return;
+
+    int output_size = out->numel();
+    int blocks = NumBlocks(output_size);
+    int threads = kNumCUDAThreads;
+
+    Tensor roi_batch_id_list;
+    roi_batch_id_list.Resize({rois_num});
+    int* roi_batch_id_data =
+        roi_batch_id_list.mutable_data<int>(platform::CPUPlace());
+    auto rois_lod = rois->lod().back();
+    int rois_batch_size = rois_lod.size() - 1;
+    PADDLE_ENFORCE_EQ(
+        rois_batch_size, batch_size,
+        "The rois_batch_size and imgs batch_size must be the same.");
+    int rois_num_with_lod = rois_lod[rois_batch_size];
+    PADDLE_ENFORCE_EQ(rois_num, rois_num_with_lod,
+                      "The rois_num from input and lod must be the same.");
+    for (int n = 0; n < rois_batch_size; ++n) {
+      for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) {
+        roi_batch_id_data[i] = n;
+      }
+    }
+    Tensor roi_batch_id_list_gpu;
+    framework::TensorCopySync(roi_batch_id_list, ctx.GetPlace(),
+                              &roi_batch_id_list_gpu);
+    GPUROIAlignForward<
+        T><<<blocks, threads, 0, ctx.cuda_device_context().stream()>>>(
+        output_size, in->data<T>(), rois->data<T>(), spatial_scale, channels,
+        height, width, pooled_height, pooled_width, sampling_ratio,
+        roi_batch_id_list_gpu.data<int>(),
+        out->mutable_data<T>(ctx.GetPlace()));
+  }
+};
+
+template <typename Place, typename T>
+class GPUROIAlignGradOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* in = ctx.Input<Tensor>("X");
+    auto* rois = ctx.Input<LoDTensor>("ROIs");
+
+    auto* out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* in_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
+
+    auto pooled_height = ctx.Attr<int>("pooled_height");
+    auto pooled_width = ctx.Attr<int>("pooled_width");
+    auto spatial_scale = ctx.Attr<float>("spatial_scale");
+    auto sampling_ratio = ctx.Attr<int>("sampling_ratio");
+
+    int rois_num = rois->dims()[0];
+    int channels = in->dims()[1];
+    int height = in->dims()[2];
+    int width = in->dims()[3];
+
+    if (!in_grad) {
+      return;
+    }
+    Tensor roi_batch_id_list;
+    roi_batch_id_list.Resize({rois_num});
+    int* roi_batch_id_data =
+        roi_batch_id_list.mutable_data<int>(platform::CPUPlace());
+    auto rois_lod = rois->lod().back();
+    int rois_batch_size = rois_lod.size() - 1;
+    for (int n = 0; n < rois_batch_size; ++n) {
+      for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) {
+        roi_batch_id_data[i] = n;
+      }
+    }
+    Tensor roi_batch_id_list_gpu;
+    framework::TensorCopySync(roi_batch_id_list, ctx.GetPlace(),
+                              &roi_batch_id_list_gpu);
+
+    in_grad->mutable_data<T>(ctx.GetPlace());
+    math::SetConstant<Place, T> set_zero;
+    set_zero(ctx.cuda_device_context(), in_grad, static_cast<T>(0));
+
+    int output_grad_size = out_grad->numel();
+    int blocks = NumBlocks(output_grad_size);
+    int threads = kNumCUDAThreads;
+
+    if (output_grad_size > 0) {
+      GPUROIAlignBackward<
+          T><<<blocks, threads, 0, ctx.cuda_device_context().stream()>>>(
+          output_grad_size, rois->data<T>(), out_grad->data<T>(), rois_num,
+          spatial_scale, channels, height, width, pooled_height, pooled_width,
+          sampling_ratio, roi_batch_id_list_gpu.data<int>(),
+          in_grad->mutable_data<T>(ctx.GetPlace()));
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    roi_align,
+    ops::GPUROIAlignOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GPUROIAlignOpKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    roi_align_grad,
+    ops::GPUROIAlignGradOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GPUROIAlignGradOpKernel<paddle::platform::CUDADeviceContext, double>);

paddle/fluid/operators/roi_align_op.h

+/* Copyright (c) 2018 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 <algorithm>
+#include <limits>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+
+static constexpr int kROISize = 4;
+
+template <class T>
+void PreCalcForBilinearInterpolate(
+    const platform::DeviceContext& ctx, const int height, const int width,
+    const int pooled_height, const int pooled_width, const int iy_upper,
+    const int ix_upper, T roi_ymin, T roi_xmin, T bin_size_h, T bin_size_w,
+    int roi_bin_grid_h, int roi_bin_grid_w, Tensor* pre_pos, Tensor* pre_w) {
+  int pre_calc_index = 0;
+  int* pre_pos_data = pre_pos->mutable_data<int>(ctx.GetPlace());
+  T* pre_w_data = pre_w->mutable_data<T>(ctx.GetPlace());
+  for (int ph = 0; ph < pooled_height; ph++) {
+    for (int pw = 0; pw < pooled_width; pw++) {
+      for (int iy = 0; iy < iy_upper; iy++) {
+        // calculate y of sample points
+        T y = roi_ymin + ph * bin_size_h +
+              static_cast<T>(iy + .5f) * bin_size_h /
+                  static_cast<T>(roi_bin_grid_h);
+        // calculate x of samle points
+        for (int ix = 0; ix < ix_upper; ix++) {
+          T x = roi_xmin + pw * bin_size_w +
+                static_cast<T>(ix + .5f) * bin_size_w /
+                    static_cast<T>(roi_bin_grid_w);
+          // deal with elements out of map
+          if (y < -1.0 || y > height || x < -1.0 || x > width) {
+            for (int i = 0; i < kROISize; ++i) {
+              pre_pos_data[i + pre_calc_index * kROISize] = 0;
+              pre_w_data[i + pre_calc_index * kROISize] = 0;
+            }
+            pre_calc_index += 1;
+            continue;
+          }
+          y = y <= 0 ? 0 : y;
+          x = x <= 0 ? 0 : x;
+
+          int y_low = static_cast<int>(y);
+          int x_low = static_cast<int>(x);
+          int y_high;
+          int x_high;
+          if (y_low >= height - 1) {
+            y_high = y_low = height - 1;
+            y = static_cast<T>(y_low);
+          } else {
+            y_high = y_low + 1;
+          }
+          if (x_low >= width - 1) {
+            x_high = x_low = width - 1;
+            x = static_cast<T>(x_low);
+          } else {
+            x_high = x_low + 1;
+          }
+          T ly = y - y_low, lx = x - x_low;
+          T hy = 1. - ly, hx = 1. - lx;
+          pre_pos_data[pre_calc_index * kROISize] = y_low * width + x_low;
+          pre_pos_data[pre_calc_index * kROISize + 1] = y_low * width + x_high;
+          pre_pos_data[pre_calc_index * kROISize + 2] = y_high * width + x_low;
+          pre_pos_data[pre_calc_index * kROISize + 3] = y_high * width + x_high;
+          pre_w_data[pre_calc_index * kROISize] = hy * hx;
+          pre_w_data[pre_calc_index * kROISize + 1] = hy * lx;
+          pre_w_data[pre_calc_index * kROISize + 2] = ly * hx;
+          pre_w_data[pre_calc_index * kROISize + 3] = ly * lx;
+          pre_calc_index += 1;
+        }
+      }
+    }
+  }
+}
+
+template <class T>
+void bilinear_interpolate_gradient(const int height, const int width, T y, T x,
+                                   const T out_grad_this_bin, const T count,
+                                   T* batch_grad_data) {
+  int x_low, y_low, x_high, y_high;
+  T w1, w2, w3, w4;
+  if (y < -1.0 || y > height || x < -1.0 || x > width) {
+    w1 = w2 = w3 = w4 = 0;
+    x_low = x_high = y_low = y_high = -1;
+    return;
+  }
+  y = y <= 0 ? 0 : y;
+  x = x <= 0 ? 0 : x;
+  y_low = static_cast<int>(y);
+  x_low = static_cast<int>(x);
+  if (y_low >= height - 1) {
+    y_high = y_low = height - 1;
+    y = static_cast<T>(y_low);
+  } else {
+    y_high = y_low + 1;
+  }
+
+  if (x_low >= width - 1) {
+    x_high = x_low = width - 1;
+    x = static_cast<T>(x_low);
+  } else {
+    x_high = x_low + 1;
+  }
+
+  T ly = y - y_low, lx = x - x_low;
+  T hy = 1. - ly, hx = 1. - lx;
+  w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx;
+  T diff1 = out_grad_this_bin * w1 / count;
+  T diff2 = out_grad_this_bin * w2 / count;
+  T diff3 = out_grad_this_bin * w3 / count;
+  T diff4 = out_grad_this_bin * w4 / count;
+  if (x_low >= 0 && x_high >= 0 && y_low >= 0 && y_high >= 0) {
+    *(batch_grad_data + y_low * width + x_low) += diff1;
+    *(batch_grad_data + y_low * width + x_high) += diff2;
+    *(batch_grad_data + y_high * width + x_low) += diff3;
+    *(batch_grad_data + y_high * width + x_high) += diff4;
+  }
+}
+
+template <typename DeviceContext, typename T>
+class CPUROIAlignOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* in = ctx.Input<framework::Tensor>("X");
+    auto* rois = ctx.Input<framework::LoDTensor>("ROIs");
+    auto* out = ctx.Output<framework::Tensor>("Out");
+    auto pooled_height = ctx.Attr<int>("pooled_height");
+    auto pooled_width = ctx.Attr<int>("pooled_width");
+    auto spatial_scale = ctx.Attr<float>("spatial_scale");
+    auto sampling_ratio = ctx.Attr<int>("sampling_ratio");
+
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
+
+    auto in_dims = in->dims();
+    int batch_size = in_dims[0];
+    int channels = in_dims[1];
+    int height = in_dims[2];
+    int width = in_dims[3];
+    int rois_num = rois->dims()[0];
+
+    auto in_stride = framework::stride(in_dims);
+    auto roi_stride = framework::stride(rois->dims());
+    auto out_stride = framework::stride(out->dims());
+
+    const T* input_data = in->data<T>();
+    framework::Tensor roi_batch_id_list;
+    roi_batch_id_list.Resize({rois_num});
+    int* roi_batch_id_data =
+        roi_batch_id_list.mutable_data<int>(ctx.GetPlace());
+
+    auto rois_lod = rois->lod().back();
+    int rois_batch_size = rois_lod.size() - 1;
+    PADDLE_ENFORCE_EQ(
+        rois_batch_size, batch_size,
+        "The rois_batch_size and imgs batch_size must be the same.");
+    int rois_num_with_lod = rois_lod[rois_batch_size];
+    PADDLE_ENFORCE_EQ(rois_num, rois_num_with_lod,
+                      "The rois_num from input and lod must be the same.");
+    for (int n = 0; n < rois_batch_size; ++n) {
+      for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) {
+        roi_batch_id_data[i] = n;
+      }
+    }
+    T* output_data = out->mutable_data<T>(ctx.GetPlace());
+    const T* rois_data = rois->data<T>();
+    for (int n = 0; n < rois_num; ++n) {
+      int roi_batch_id = roi_batch_id_data[n];
+      T roi_xmin = rois_data[0] * spatial_scale;
+      T roi_ymin = rois_data[1] * spatial_scale;
+      T roi_xmax = rois_data[2] * spatial_scale;
+      T roi_ymax = rois_data[3] * spatial_scale;
+
+      T roi_width = std::max(roi_xmax - roi_xmin, static_cast<T>(1.));
+      T roi_height = std::max(roi_ymax - roi_ymin, static_cast<T>(1.));
+      T bin_size_h = static_cast<T>(roi_height) / static_cast<T>(pooled_height);
+      T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
+      const T* batch_data = input_data + roi_batch_id * in_stride[0];
+
+      int roi_bin_grid_h = (sampling_ratio > 0)
+                               ? sampling_ratio
+                               : ceil(roi_height / pooled_height);
+      int roi_bin_grid_w = (sampling_ratio > 0)
+                               ? sampling_ratio
+                               : ceil(roi_width / pooled_width);
+      const T count = roi_bin_grid_h * roi_bin_grid_w;
+      Tensor pre_pos;
+      Tensor pre_w;
+      int pre_size = count * out_stride[1];
+      pre_pos.Resize({pre_size, kROISize});
+      pre_w.Resize({pre_size, kROISize});
+
+      PreCalcForBilinearInterpolate(
+          dev_ctx, height, width, pooled_height, pooled_width, roi_bin_grid_h,
+          roi_bin_grid_w, roi_ymin, roi_xmin, bin_size_h, bin_size_w,
+          roi_bin_grid_h, roi_bin_grid_w, &pre_pos, &pre_w);
+      const int* pre_pos_data = pre_pos.data<int>();
+      const T* pre_w_data = pre_w.data<T>();
+      for (int c = 0; c < channels; c++) {
+        int pre_calc_index = 0;
+        for (int ph = 0; ph < pooled_height; ph++) {
+          for (int pw = 0; pw < pooled_width; pw++) {
+            const int pool_index = ph * pooled_width + pw;
+            T output_val = 0;
+            for (int iy = 0; iy < roi_bin_grid_h; iy++) {
+              for (int ix = 0; ix < roi_bin_grid_w; ix++) {
+                for (int i = 0; i < kROISize; i++) {
+                  int pos = pre_pos_data[pre_calc_index * kROISize + i];
+                  T w = pre_w_data[pre_calc_index * kROISize + i];
+                  output_val += w * batch_data[pos];
+                }
+                pre_calc_index += 1;
+              }
+            }
+            output_val /= count;
+            output_data[pool_index] = output_val;
+          }
+        }
+        batch_data += in_stride[1];
+        output_data += out_stride[1];
+      }
+      rois_data += roi_stride[0];
+    }
+  }
+};
+
+template <typename DeviceContext, typename T>
+class CPUROIAlignGradOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* in = ctx.Input<framework::Tensor>("X");
+    auto* rois = ctx.Input<framework::LoDTensor>("ROIs");
+    auto* out_grad =
+        ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto* in_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+
+    auto pooled_height = ctx.Attr<int>("pooled_height");
+    auto pooled_width = ctx.Attr<int>("pooled_width");
+    auto spatial_scale = ctx.Attr<float>("spatial_scale");
+    auto sampling_ratio = ctx.Attr<int>("sampling_ratio");
+    auto in_dims = in->dims();
+    if (!in_grad) {
+      return;
+    }
+    int channels = in_dims[1];
+    int height = in_dims[2];
+    int width = in_dims[3];
+    int rois_num = rois->dims()[0];
+    Tensor roi_batch_id_list;
+    roi_batch_id_list.Resize({rois_num});
+    int* roi_batch_id_data =
+        roi_batch_id_list.mutable_data<int>(ctx.GetPlace());
+
+    auto rois_lod = rois->lod().back();
+    int rois_batch_size = rois_lod.size() - 1;
+    for (int n = 0; n < rois_batch_size; ++n) {
+      for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) {
+        roi_batch_id_data[i] = n;
+      }
+    }
+
+    const T* rois_data = rois->data<T>();
+    const T* out_grad_data = out_grad->data<T>();
+    T* in_grad_data = in_grad->mutable_data<T>(ctx.GetPlace());
+
+    auto in_stride = framework::stride(in->dims());
+    auto roi_stride = framework::stride(rois->dims());
+    auto out_stride = framework::stride(out_grad->dims());
+
+    for (int n = 0; n < rois_num; ++n) {
+      int roi_batch_idx = roi_batch_id_data[n];
+      T roi_xmin = rois_data[0] * spatial_scale;
+      T roi_ymin = rois_data[1] * spatial_scale;
+      T roi_xmax = rois_data[2] * spatial_scale;
+      T roi_ymax = rois_data[3] * spatial_scale;
+      T roi_width = std::max(roi_xmax - roi_xmin, static_cast<T>(1.));
+      T roi_height = std::max(roi_ymax - roi_ymin, static_cast<T>(1.));
+      T bin_size_h = static_cast<T>(roi_height) / static_cast<T>(pooled_height);
+      T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
+      for (int c = 0; c < channels; ++c) {
+        T* batch_grad_data =
+            in_grad_data + roi_batch_idx * in_stride[0] + c * in_stride[1];
+        const T* batch_out_grad_data =
+            out_grad_data + n * out_stride[0] + c * out_stride[1];
+        for (int ph = 0; ph < pooled_height; ++ph) {
+          for (int pw = 0; pw < pooled_width; ++pw) {
+            int pool_index = ph * pooled_width + pw;
+            T out_grad_this_bin = batch_out_grad_data[pool_index];
+            int roi_bin_grid_h = (sampling_ratio > 0)
+                                     ? sampling_ratio
+                                     : ceil(roi_height / pooled_height);
+            int roi_bin_grid_w = (sampling_ratio > 0)
+                                     ? sampling_ratio
+                                     : ceil(roi_width / pooled_width);
+            T count = roi_bin_grid_h * roi_bin_grid_w;
+            for (int iy = 0; iy < roi_bin_grid_h; iy++) {
+              const T y = roi_ymin + ph * bin_size_h +
+                          static_cast<T>(iy + .5f) * bin_size_h /
+                              static_cast<T>(roi_bin_grid_h);
+              for (int ix = 0; ix < roi_bin_grid_w; ix++) {
+                const T x = roi_xmin + pw * bin_size_w +
+                            static_cast<T>(ix + .5f) * bin_size_w /
+                                static_cast<T>(roi_bin_grid_w);
+                bilinear_interpolate_gradient(height, width, y, x,
+                                              out_grad_this_bin, count,
+                                              batch_grad_data);
+              }
+            }
+          }
+        }
+      }
+      rois_data += roi_stride[0];
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/roi_pool_op.cc

@@ -174,4 +174,4 @@ REGISTER_OP_CPU_KERNEL(
 REGISTER_OP_CPU_KERNEL(
     roi_pool_grad,
     ops::CPUROIPoolGradOpKernel<paddle::platform::CPUDeviceContext, float>,
-    ops::CPUROIPoolOpKernel<paddle::platform::CPUDeviceContext, double>);
+    ops::CPUROIPoolGradOpKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/roi_pool_op.cu

@@ -249,4 +249,4 @@ REGISTER_OP_CUDA_KERNEL(
 REGISTER_OP_CUDA_KERNEL(
     roi_pool_grad,
     ops::GPUROIPoolGradOpKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::GPUROIPoolOpKernel<paddle::platform::CUDADeviceContext, double>);
+    ops::GPUROIPoolGradOpKernel<paddle::platform::CUDADeviceContext, double>);

paddle/fluid/platform/device_context.cc

@@ -35,6 +35,16 @@ platform::DeviceContext* DeviceContextPool::Get(const platform::Place& place) {
   return it->second.get();
 }
 
+const std::vector<const DeviceContext*>
+DeviceContextPool::GetAllDeviceContexts() const {
+  std::vector<const DeviceContext*> all_device_ctx;
+  all_device_ctx.reserve(device_contexts_.size());
+  for (auto& dev_ctx : device_contexts_) {
+    all_device_ctx.emplace_back(dev_ctx.second.get());
+  }
+  return all_device_ctx;
+}
+
 DeviceContextPool::DeviceContextPool(
     const std::vector<platform::Place>& places) {
   PADDLE_ENFORCE_GT(places.size(), 0);

paddle/fluid/platform/device_context.h

@@ -217,6 +217,9 @@ class DeviceContextPool {
   /*! \brief  Return handle of single device context. */
   platform::DeviceContext* Get(const platform::Place& place);
 
+  /*! \brief  Return all the device contexts. */
+  const std::vector<const DeviceContext*> GetAllDeviceContexts() const;
+
   template <typename Place>
   const typename DefaultDeviceContextType<Place>::TYPE* GetByPlace(
       const Place& place) {

paddle/fluid/platform/profiler.cc

@@ -30,6 +30,8 @@ limitations under the License. */
 #include "paddle/fluid/platform/device_tracer.h"
 #include "paddle/fluid/string/printf.h"
 
+DEFINE_bool(enable_rpc_profiler, false, "Enable rpc profiler or not.");
+
 namespace paddle {
 namespace platform {
 
@@ -193,6 +195,13 @@ RecordEvent::~RecordEvent() {
   PopEvent(name_, dev_ctx_);
 }
 
+RecordRPCEvent::RecordRPCEvent(const std::string& name,
+                               const DeviceContext* dev_ctx) {
+  if (FLAGS_enable_rpc_profiler) {
+    event_.reset(new platform::RecordEvent(name, dev_ctx));
+  }
+}
+
 RecordBlock::RecordBlock(int block_id)
     : is_enabled_(false), start_ns_(PosixInNsec()) {
   std::lock_guard<std::mutex> l(profiler_mu);

paddle/fluid/platform/profiler.h

@@ -87,6 +87,16 @@ struct RecordEvent {
   std::string full_name_;
 };
 
+class RecordRPCEvent {
+ public:
+  // dev_ctx can be set to nullptr if device is cpu.
+  RecordRPCEvent(const std::string& name, const DeviceContext* dev_ctx);
+  ~RecordRPCEvent() {}
+
+ private:
+  std::unique_ptr<RecordEvent> event_;
+};
+
 struct RecordBlock {
   explicit RecordBlock(int block_id);
   ~RecordBlock();

python/paddle/fluid/__init__.py

@@ -120,6 +120,7 @@ def __bootstrap__():
         read_env_flags.append('rpc_deadline')
         read_env_flags.append('rpc_server_profile_period')
         read_env_flags.append('rpc_server_profile_path')
+        read_env_flags.append('enable_rpc_profiler')
 
     if core.is_compiled_with_cuda():
         read_env_flags += [

python/paddle/fluid/layers/nn.py

@@ -96,6 +96,7 @@ __all__ = [
     'pad_constant_like',
     'label_smooth',
     'roi_pool',
+    'roi_align',
     'dice_loss',
     'image_resize',
     'image_resize_short',
@@ -5435,6 +5436,54 @@ def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0):
     return pool_out
 
 
+@templatedoc()
+def roi_align(input,
+              rois,
+              pooled_height=1,
+              pooled_width=1,
+              spatial_scale=1.0,
+              sampling_ratio=-1,
+              name=None):
+    """
+    ${comment}
+
+    Args:
+        input (Variable): ${x_comment}
+        rois (Variable): ROIs (Regions of Interest) to pool over.
+        pooled_height (integer): ${pooled_height_comment} Default: 1
+        pooled_width (integer): ${pooled_width_comment} Default: 1
+        spatial_scale (float): ${spatial_scale_comment} Default: 1.0
+        sampling_ratio(intger): ${sampling_ratio_comment} Default: -1
+
+    Returns:
+        Variable: ${out_comment}.
+    Examples:
+        .. code-block:: python
+
+            align_out = fluid.layers.roi_align(input=x, 
+                                               rois=rois, 
+                                               pooled_height=7, 
+                                               pooled_width=7,
+                                               spatial_scale=0.5,
+                                               sampling_ratio=-1)
+    """
+    helper = LayerHelper('roi_align', **locals())
+    dtype = helper.input_dtype()
+    align_out = helper.create_tmp_variable(dtype)
+    helper.append_op(
+        type="roi_align",
+        inputs={"X": input,
+                "ROIs": rois},
+        outputs={"Out": align_out},
+        attrs={
+            "pooled_height": pooled_height,
+            "pooled_width": pooled_width,
+            "spatial_scale": spatial_scale,
+            "sampling_ratio": sampling_ratio
+        })
+    return align_out
+
+
 def dice_loss(input, label, epsilon=0.00001):
     """
     Dice loss for comparing the similarity of two batch of data,

python/paddle/fluid/tests/unittests/test_fusion_seqconv_eltadd_relu_op.py

+#   Copyright (c) 2018 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.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+import random
+from op_test import OpTest
+from test_seq_conv import seqconv
+
+
+class TestSeqConvEltAddRelu(OpTest):
+    def set_conf(self):
+        pass
+
+    def setUp(self):
+        self.op_type = 'fusion_seqconv_eltadd_relu'
+        self.lod = [[6, 4]]
+        self.in_fea_size = 16
+        self.out_fea_size = 8
+        self.context_length = 4
+        self.context_stride = 1
+        self.context_start = 0
+        self.set_conf()
+
+        assert self.context_stride == 1
+
+        T = sum(self.lod[0])
+        x = np.random.uniform(-1, 1, [T, self.in_fea_size]).astype('float32')
+        w = np.random.uniform(
+            -1, 1, [self.in_fea_size * self.context_length,
+                    self.out_fea_size]).astype('float32')
+        b = np.random.uniform(-2, 1, [1, self.out_fea_size]).astype('float32')
+        out = seqconv(x, self.lod, w, self.context_length, self.context_start)
+        out = np.maximum(out + b, 0)
+
+        self.inputs = {'X': (x, self.lod), 'Filter': w, 'Bias': b}
+        self.attrs = {
+            'contextStart': self.context_start,
+            'contextLength': self.context_length,
+            'contextStride': self.context_stride
+        }
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestSeqConvEltAddReluBS1(TestSeqConvEltAddRelu):
+    def set_conf(self):
+        self.lod = [[10]]
+
+
+class TestSeqConvEltAddReluBS1Case2(TestSeqConvEltAddRelu):
+    def set_conf(self):
+        self.lod = [[2]]
+
+
+class TestSeqConvEltAddReluCase1(TestSeqConvEltAddRelu):
+    def set_conf(self):
+        self.lod = [[3, 5, 1, 6]]
+        self.context_length = 3
+        self.context_start = -2
+
+
+class TestSeqConvEltAddReluCase2(TestSeqConvEltAddRelu):
+    def set_conf(self):
+        self.lod = [[10, 1, 2, 4, 1, 5, 6]]
+        self.in_fea_size = 2
+        self.context_length = 4
+        self.context_start = -1
+
+
+class TestSeqConvEltAddReluCase3(TestSeqConvEltAddRelu):
+    def set_conf(self):
+        self.lod = [[10, 1, 2, 4, 1, 5, 6]]
+        self.context_length = 5
+        self.context_start = -4
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_layers.py

@@ -465,6 +465,16 @@ class TestBook(unittest.TestCase):
             self.assertIsNotNone(output)
         print(str(program))
 
+    def test_roi_align(self):
+        program = Program()
+        with program_guard(program):
+            x = layers.data(name="x", shape=[256, 30, 30], dtype="float32")
+            rois = layers.data(
+                name="rois", shape=[4], dtype="float32", lod_level=1)
+            output = layers.roi_align(x, rois, 14, 14, 0.5, 2)
+            self.assertIsNotNone(output)
+        print(str(program))
+
     def test_resize_bilinear(self):
         program = Program()
         with program_guard(program):

python/paddle/fluid/tests/unittests/test_polygon_box_transform.py

@@ -37,7 +37,7 @@ def PolygonBoxRestore(input):
     indexes = indexes.repeat(
         [batch_size], axis=0)  # [batch_size, geo_channels/2, 2, h, w]
     return indexes.reshape(
-        input.shape) - input  # [batch_size, geo_channels, h, w]
+        input.shape) * 4 - input  # [batch_size, geo_channels, h, w]
 
 
 class TestPolygonBoxRestoreOp(OpTest):

python/paddle/fluid/tests/unittests/test_roi_align_op.py

+#    Copyright (c) 2018 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.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+import math
+import sys
+from op_test import OpTest
+
+
+class TestROIAlignOp(OpTest):
+    def set_data(self):
+        self.init_test_case()
+        self.make_rois()
+        self.calc_roi_align()
+        self.inputs = {'X': self.x, 'ROIs': (self.rois[:, 1:5], self.rois_lod)}
+        self.attrs = {
+            'spatial_scale': self.spatial_scale,
+            'pooled_height': self.pooled_height,
+            'pooled_width': self.pooled_width,
+            'sampling_ratio': self.sampling_ratio
+        }
+
+        self.outputs = {'Out': self.out_data}
+
+    def init_test_case(self):
+        self.batch_size = 3
+        self.channels = 3
+        self.height = 8
+        self.width = 6
+
+        # n, c, h, w
+        self.x_dim = (self.batch_size, self.channels, self.height, self.width)
+
+        self.spatial_scale = 1.0 / 2.0
+        self.pooled_height = 2
+        self.pooled_width = 2
+        self.sampling_ratio = -1
+
+        self.x = np.random.random(self.x_dim).astype('float32')
+
+    def pre_calc(self, x_i, roi_xmin, roi_ymin, roi_bin_grid_h, roi_bin_grid_w,
+                 bin_size_h, bin_size_w):
+        count = roi_bin_grid_h * roi_bin_grid_w
+        bilinear_pos = np.zeros(
+            [self.channels, self.pooled_height, self.pooled_width, count, 4],
+            np.float32)
+        bilinear_w = np.zeros(
+            [self.pooled_height, self.pooled_width, count, 4], np.float32)
+        for ph in range(self.pooled_width):
+            for pw in range(self.pooled_height):
+                c = 0
+                for iy in range(roi_bin_grid_h):
+                    y = roi_ymin + ph * bin_size_h + (iy + 0.5) * \
+                        bin_size_h / roi_bin_grid_h
+                    for ix in range(roi_bin_grid_w):
+                        x = roi_xmin + pw * bin_size_w + (ix + 0.5) * \
+                            bin_size_w / roi_bin_grid_w
+                        if y < -1.0 or y > self.height or \
+                               x < -1.0 or x > self.width:
+                            continue
+                        if y <= 0:
+                            y = 0
+                        if x <= 0:
+                            x = 0
+                        y_low = int(y)
+                        x_low = int(x)
+                        if y_low >= self.height - 1:
+                            y = y_high = y_low = self.height - 1
+                        else:
+                            y_high = y_low + 1
+                        if x_low >= self.width - 1:
+                            x = x_high = x_low = self.width - 1
+                        else:
+                            x_high = x_low + 1
+                        ly = y - y_low
+                        lx = x - x_low
+                        hy = 1 - ly
+                        hx = 1 - lx
+                        for ch in range(self.channels):
+                            bilinear_pos[ch, ph, pw, c, 0] = x_i[ch, y_low,
+                                                                 x_low]
+                            bilinear_pos[ch, ph, pw, c, 1] = x_i[ch, y_low,
+                                                                 x_high]
+                            bilinear_pos[ch, ph, pw, c, 2] = x_i[ch, y_high,
+                                                                 x_low]
+                            bilinear_pos[ch, ph, pw, c, 3] = x_i[ch, y_high,
+                                                                 x_high]
+                        bilinear_w[ph, pw, c, 0] = hy * hx
+                        bilinear_w[ph, pw, c, 1] = hy * lx
+                        bilinear_w[ph, pw, c, 2] = ly * hx
+                        bilinear_w[ph, pw, c, 3] = ly * lx
+                        c = c + 1
+        return bilinear_pos, bilinear_w
+
+    def calc_roi_align(self):
+        self.out_data = np.zeros(
+            (self.rois_num, self.channels, self.pooled_height,
+             self.pooled_width)).astype('float32')
+
+        for i in range(self.rois_num):
+            roi = self.rois[i]
+            roi_batch_id = int(roi[0])
+            x_i = self.x[roi_batch_id]
+            roi_xmin = roi[1] * self.spatial_scale
+            roi_ymin = roi[2] * self.spatial_scale
+            roi_xmax = roi[3] * self.spatial_scale
+            roi_ymax = roi[4] * self.spatial_scale
+            roi_width = max(roi_xmax - roi_xmin, 1)
+            roi_height = max(roi_ymax - roi_ymin, 1)
+            bin_size_h = float(roi_height) / float(self.pooled_height)
+            bin_size_w = float(roi_width) / float(self.pooled_width)
+            roi_bin_grid_h = self.sampling_ratio if self.sampling_ratio > 0 else \
+                                 math.ceil(roi_height / self.pooled_height)
+            roi_bin_grid_w = self.sampling_ratio if self.sampling_ratio > 0 else \
+                                 math.ceil(roi_width / self.pooled_width)
+            count = int(roi_bin_grid_h * roi_bin_grid_w)
+            pre_size = count * self.pooled_width * self.pooled_height
+            bilinear_pos, bilinear_w = self.pre_calc(x_i, roi_xmin, roi_ymin,
+                                                     int(roi_bin_grid_h),
+                                                     int(roi_bin_grid_w),
+                                                     bin_size_h, bin_size_w)
+            for ch in range(self.channels):
+                align_per_bin = (bilinear_pos[ch] * bilinear_w).sum(axis=-1)
+                output_val = align_per_bin.mean(axis=-1)
+                self.out_data[i, ch, :, :] = output_val
+
+    def make_rois(self):
+        rois = []
+        self.rois_lod = [[]]
+        for bno in range(self.batch_size):
+            self.rois_lod[0].append(bno + 1)
+            for i in range(bno + 1):
+                x1 = np.random.random_integers(
+                    0, self.width // self.spatial_scale - self.pooled_width)
+                y1 = np.random.random_integers(
+                    0, self.height // self.spatial_scale - self.pooled_height)
+
+                x2 = np.random.random_integers(x1 + self.pooled_width,
+                                               self.width // self.spatial_scale)
+                y2 = np.random.random_integers(
+                    y1 + self.pooled_height, self.height // self.spatial_scale)
+
+                roi = [bno, x1, y1, x2, y2]
+                rois.append(roi)
+        self.rois_num = len(rois)
+        self.rois = np.array(rois).astype("float32")
+
+    def setUp(self):
+        self.op_type = "roi_align"
+        self.set_data()
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')

python/paddle/fluid/tests/unittests/test_seq_conv.py

@@ -20,6 +20,53 @@ import random
 from op_test import OpTest
 
 
+def seqconv(x,
+            lod,
+            filter,
+            context_length,
+            context_start,
+            padding_trainable=False,
+            padding_data=None):
+    [T, M] = x.shape
+    col = np.zeros((T, context_length * M)).astype('float32')
+    offset = [0]
+    for seq_len in lod[0]:
+        offset.append(offset[-1] + seq_len)
+    begin_pad = np.max([0, -context_start])
+    for i in range(len(offset) - 1):
+        for j in range(context_length):
+            in_begin = offset[i] + context_start + j
+            in_end = offset[i + 1] + context_start + j
+            out_begin = offset[i]
+            out_end = offset[i + 1]
+            if in_begin < offset[i]:
+                pad_size = np.min(
+                    [offset[i] - in_begin, offset[i + 1] - offset[i]])
+                if padding_trainable:
+                    sub_w = padding_data[j:j + pad_size, :]
+                    col[offset[i]:offset[i] + pad_size, j * M:(j + 1) *
+                        M] = sub_w
+                out_begin = offset[i] + pad_size
+                in_begin = offset[i]
+
+            if in_end > offset[i + 1]:
+                pad_size = np.min(
+                    [in_end - offset[i + 1], offset[i + 1] - offset[i]])
+                if padding_trainable:
+                    sub_w = padding_data[begin_pad + context_start + j -
+                                         pad_size:begin_pad + context_start +
+                                         j, :]
+                    col[offset[i + 1] - pad_size:offset[i + 1], j * M:(j + 1) *
+                        M] = sub_w
+                in_end = offset[i + 1]
+                out_end = offset[i + 1] - pad_size
+            if in_end <= in_begin:
+                continue
+            in_sub = x[in_begin:in_end, :]
+            col[out_begin:out_end, j * M:(j + 1) * M] += in_sub
+    return np.dot(col, filter)
+
+
 class TestSeqProject(OpTest):
     def setUp(self):
         self.init_test_case()
@@ -66,57 +113,9 @@ class TestSeqProject(OpTest):
             'paddingTrainable': self.padding_trainable,
             'contextStride': self.context_stride
         }
-        out = np.zeros(
-            (self.input_size[0], self.output_represention)).astype('float32')
+        out = seqconv(x, self.lod, w, self.context_length, self.context_start,
+                      self.padding_trainable, self.pad_data)
         self.outputs = {'Out': out}
-        self.compute()
-
-    def compute(self):
-        x, lod = self.inputs['X']
-        filter = self.inputs['Filter']
-        pading_data = self.pad_data
-        out = np.zeros((self.input_size[0], self.context_length *
-                        self.input_size[1])).astype('float32')
-        offset = [0]
-        for seq_len in lod[0]:
-            offset.append(offset[-1] + seq_len)
-        begin_pad = np.max([0, -self.context_start])
-
-        for i in range(len(offset) - 1):
-            for j in range(self.context_length):
-                in_begin = offset[i] + self.context_start + j
-                in_end = offset[i + 1] + self.context_start + j
-                out_begin = offset[i]
-                out_end = offset[i + 1]
-                if in_begin < offset[i]:
-                    pad_size = np.min(
-                        [offset[i] - in_begin, offset[i + 1] - offset[i]])
-                    if self.padding_trainable:
-                        sub_w = pading_data[j:j + pad_size, :]
-                        out[offset[i]:offset[i] + pad_size, j * self.input_size[
-                            1]:(j + 1) * self.input_size[1]] = sub_w
-                    out_begin = offset[i] + pad_size
-                    in_begin = offset[i]
-
-                if in_end > offset[i + 1]:
-                    pad_size = np.min(
-                        [in_end - offset[i + 1], offset[i + 1] - offset[i]])
-                    if self.padding_trainable:
-                        sub_w = pading_data[begin_pad + self.context_start + j -
-                                            pad_size:begin_pad +
-                                            self.context_start + j, :]
-                        out[offset[i + 1] - pad_size:offset[i + 1], j * self.
-                            input_size[1]:(j + 1) * self.input_size[1]] = sub_w
-                    in_end = offset[i + 1]
-                    out_end = offset[i + 1] - pad_size
-                if in_end <= in_begin:
-                    continue
-
-                in_sub = x[in_begin:in_end, :]
-                out[out_begin:out_end, j * self.input_size[1]:(j + 1) *
-                    self.input_size[1]] += in_sub
-
-        np.dot(out, filter, out=self.outputs['Out'])
 
     def test_check_output(self):
         self.check_output()

python/paddle/fluid/transpiler/inference_transpiler.py

@@ -74,7 +74,7 @@ class InferenceTranspiler(object):
         '''
         Transpile the program fusing elementwise_add into conv for MKLDNN
         program. Elementwise add following convolution OP can be fused by adding
-        'fuse_eltwise' attribute to convolution OP and replacing its output
+        'fuse_residual_connection' attribute to convolution OP and replacing its output
         Tensor with second parameter of elementwise_add.
         The result of fuse is:
             - before:
@@ -92,7 +92,8 @@ class InferenceTranspiler(object):
             if current_op.type in ['conv2d']:
                 next_op = self.block.ops[i + 1]
                 if next_op.type == 'elementwise_add':
-                    self._fuse_conv_eltwise(current_op, next_op)
+                    self._fuse_conv_eltwise(i, current_op, next_op)
+                    self.block._remove_op(i + 1)  # Remove old conv
                     self.block._remove_op(i + 1)  # Remove elementwise_add
             i = i + 1
         self._adjust_input()
@@ -444,7 +445,7 @@ class InferenceTranspiler(object):
             outputs={"Output": out_var},
             attrs=attrs)
 
-    def _fuse_conv_eltwise(self, conv_op, eltwise_op):
+    def _fuse_conv_eltwise(self, index, conv_op, eltwise_op):
         '''
         fuse the conv op with elementwise_add
 
@@ -454,9 +455,30 @@ class InferenceTranspiler(object):
         :type eltwise_op: Operator
         '''
 
-        conv_op._set_attr("fuse_eltwise", True)
-        self.input_map[conv_op.output("Output")[0]] = eltwise_op.input("Y")[0]
-        self.input_map[eltwise_op.output("Out")[0]] = eltwise_op.input("Y")[0]
+        eltwise_input = "X"
+        if eltwise_op.input("X")[0] == conv_op.output("Output")[0]:
+            eltwise_input = "Y"
+
+        residual_var = self.block.vars[eltwise_op.input(eltwise_input)[0]]
+        out_var = self.block.vars[eltwise_op.output("Out")[0]]
+        filter_var = self.block.vars[conv_op.input("Filter")[0]]
+        in_var = self.block.vars[conv_op.input("Input")[0]]
+        bias_var = self.block.vars[conv_op.input("Bias")[0]]
+
+        conv_op._set_attr("fuse_residual_connection", True)
+        attrs = {name: conv_op.attr(name) for name in conv_op.attr_names}
+
+        self.block._insert_op(
+            index,
+            type="conv2d",
+            inputs={
+                "Input": in_var,
+                "Filter": filter_var,
+                "Bias": bias_var,
+                "ResidualData": residual_var
+            },
+            outputs={"Output": out_var},
+            attrs=attrs)
 
     def _adjust_input(self):
         for i in range(len(self.block.ops)):