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

README.md

@@ -2,8 +2,8 @@
 
 
 [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
-[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html)
-[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html)
+[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.0/getstarted/index_en.html)
+[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html)
 [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
 
@@ -19,7 +19,7 @@ Our vision is to enable deep learning for everyone via PaddlePaddle.
 Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
 
 
-### Latest PaddlePaddle Release: [Fluid 1.0.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.0.0)
+### Latest PaddlePaddle Release: [Fluid 1.0.1](https://github.com/PaddlePaddle/Paddle/tree/release/1.0.0)
 ### Install Latest Stable Release:
 ```
 # Linux CPU
@@ -27,9 +27,9 @@ pip install paddlepaddle
 # Linux GPU cuda9cudnn7
 pip install paddlepaddle-gpu
 # Linux GPU cuda8cudnn7
-pip install paddlepaddle-gpu==0.15.0.post87
+pip install paddlepaddle-gpu==1.0.1.post87
 # Linux GPU cuda8cudnn5
-pip install paddlepaddle-gpu==0.15.0.post85
+pip install paddlepaddle-gpu==1.0.1.post85
 
 # For installation on other platform, refer to http://paddlepaddle.org/
 ```

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()
@@ -311,6 +318,8 @@ function(cc_test TARGET_NAME)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
+    # No unit test should exceed 10 minutes.
+    set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600)
   endif()
 endfunction(cc_test)
 
@@ -629,6 +638,8 @@ function(py_test TARGET_NAME)
              PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
              ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
              WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+    # No unit test should exceed 10 minutes.
+    set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600)
   endif()
 endfunction()
 

paddle/fluid/API.spec

@@ -61,12 +61,12 @@ paddle.fluid.layers.cos_sim ArgSpec(args=['X', 'Y'], varargs=None, keywords=None
 paddle.fluid.layers.cross_entropy ArgSpec(args=['input', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100))
 paddle.fluid.layers.square_error_cost ArgSpec(args=['input', 'label'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', 'num_chunk_types', 'excluded_chunk_types'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None))
+paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None, None))
 paddle.fluid.layers.conv2d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None))
 paddle.fluid.layers.conv3d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None))
 paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'param_attr', 'bias_attr', 'use_cudnn'], varargs=None, keywords=None, defaults=(None, None, False))
-paddle.fluid.layers.softmax ArgSpec(args=['input', 'param_attr', 'bias_attr', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(None, None, True, None))
+paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(False, None))
+paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
 paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None))
 paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None))
 paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False))
@@ -97,8 +97,8 @@ paddle.fluid.layers.warpctc ArgSpec(args=['input', 'label', 'blank', 'norm_by_ti
 paddle.fluid.layers.sequence_reshape ArgSpec(args=['input', 'new_dim'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None))
-paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples'], varargs=None, keywords=None, defaults=(None, None, None, None))
-paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None))
+paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'name'], varargs=None, keywords=None, defaults=(0, None))
 paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
@@ -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

@@ -10,7 +10,7 @@ function(pass_library TARGET DEST)
     set(oneValueArgs "")
     set(multiValueArgs SRCS DEPS)
     cmake_parse_arguments(op_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-    cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass ${op_library_DEPS})
+    cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS})
     # add more DEST here, such as train, dist and collect USE_PASS into a file automatically.
     if (${DEST} STREQUAL "base" OR ${DEST} STREQUAL "inference")
         message(STATUS "add pass ${TARGET} ${DEST}")
@@ -25,13 +25,11 @@ cc_library(graph_helper SRCS graph_helper.cc DEPS graph)
 cc_library(pass SRCS pass.cc DEPS graph node graph_helper)
 cc_library(graph_traits SRCS graph_traits.cc DEPS graph)
 cc_library(graph_pattern_detector SRCS graph_pattern_detector.cc DEPS graph graph_helper graph_traits)
+cc_library(fuse_pass_base SRCS fuse_pass_base.cc DEPS pass)
 
 pass_library(graph_to_program_pass base)
 pass_library(graph_viz_pass base)
 pass_library(fc_fuse_pass inference)
-if (WITH_MKLDNN)
-    pass_library(conv_relu_mkldnn_fuse_pass inference)
-endif ()
 pass_library(attention_lstm_fuse_pass inference)
 pass_library(infer_clean_graph_pass inference)
 pass_library(fc_lstm_fuse_pass inference)
@@ -39,6 +37,13 @@ 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)
+    pass_library(conv_elementwise_add_mkldnn_fuse_pass inference)
+endif()
 
 cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector )
 
@@ -54,4 +59,5 @@ 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/attention_lstm_fuse_pass.cc

@@ -262,7 +262,7 @@ std::unique_ptr<ir::Graph> AttentionLSTMFusePass::ApplyImpl(
   std::unordered_set<std::string> specified_vars({"data_lod_attention",
                                                   "cell_init", "hidden_init",
                                                   "data", "week", "minute"});
-  int count = 0;
+  size_t count = 0;
   for (auto* node : graph->Nodes()) {
     if (node->IsVar() && specified_vars.count(node->Name())) {
       ++count;

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_bn_fuse_pass.cc

@@ -126,12 +126,21 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
     // conv, batch_norm,
     // conv_weight, conv_out,
     // bn_scale, bn_bias, bn_mean, bn_variance,
-    // bn_out, bn_mean_out, bn_variance_out, bn_saved_mean, bn_saved_variance
+    // bn_out, bn_mean_out, bn_variance_out, bn_saved_mean,
+    // bn_saved_variance
     GET_CONV_BN_NODES(conv_bn_pattern);
 
+    // check if fuse can be done and if MKL-DNN should be used
+    FuseOptions fuse_option = FindFuseOption(*conv, *batch_norm);
+    if (fuse_option == DO_NOT_FUSE) {
+      VLOG(3) << "do not perform conv+bn fuse";
+      return;
+    }
+
     // Create eltwise_y (conv bias) variable
     VarDesc eltwise_y_in_desc(
         patterns::PDNodeName(name_scope_, "eltwise_y_in"));
+    eltwise_y_in_desc.SetPersistable(true);
     auto* eltwise_y_in_node = g->CreateVarNode(&eltwise_y_in_desc);
     auto* eltwise_y_in_tensor =
         scope->Var(eltwise_y_in_node->Name())->GetMutable<LoDTensor>();
@@ -151,27 +160,59 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
                                *bn_mean, *bn_variance, eltwise_y_in_tensor,
                                epsilon);
 
-    // Create an elementwise add node
-    OpDesc desc;
-    desc.SetInput("X", std::vector<std::string>({conv_out->Name()}));
-    desc.SetInput("Y", std::vector<std::string>({eltwise_y_in_node->Name()}));
-    desc.SetOutput("Out", std::vector<std::string>({bn_out->Name()}));
-    desc.SetType("elementwise_add");
-    desc.SetAttr("axis", 1);
-    bool a = boost::get<bool>(conv->Op()->GetAttr("use_mkldnn"));
-    desc.SetAttr("use_mkldnn", a);
-    auto eltwise_op = g->CreateOpNode(&desc);  // OpDesc will be copied.
-
-    GraphSafeRemoveNodes(graph.get(), {bn_scale, bn_bias, bn_mean, bn_variance,
-                                       batch_norm, bn_mean_out, bn_variance_out,
-                                       bn_saved_mean, bn_saved_variance});
-
-    PADDLE_ENFORCE(subgraph.count(conv_input));
-    IR_NODE_LINK_TO(conv_out, eltwise_op);
-    IR_NODE_LINK_TO(eltwise_y_in_node, eltwise_op);
-    IR_NODE_LINK_TO(eltwise_op, bn_out);
-
-    found_conv_bn_count++;
+    // with MKL-DNN fuse conv+bn into conv with bias
+    // without MKL-DNN fuse conv+bn into conv+elementwise_add
+    if (fuse_option == FUSE_MKLDNN) {
+      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) {
+        // reuse existing conv bias node
+        auto conv_bias_names = conv->Op()->Input("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_y_in_tensor->dims());
+
+        auto eigen_conv_bias = EigenVector<float>::From(*conv_bias_tensor);
+        eigen_conv_bias += EigenVector<float>::From(*eltwise_y_in_tensor);
+      } else {
+        // add new conv_bias node
+        conv->Op()->SetInput(
+            "Bias", std::vector<std::string>({eltwise_y_in_node->Name()}));
+        IR_NODE_LINK_TO(eltwise_y_in_node, conv);
+      }
+      conv->Op()->SetOutput("Output",
+                            std::vector<std::string>({bn_out->Name()}));
+
+      GraphSafeRemoveNodes(
+          graph.get(),
+          {conv_out, bn_scale, bn_bias, bn_mean, bn_variance, batch_norm,
+           bn_mean_out, bn_variance_out, bn_saved_mean, bn_saved_variance});
+
+      IR_NODE_LINK_TO(conv, bn_out);
+      found_conv_bn_count++;
+    } else {  // fuse_option == FUSE_NATIVE
+      // create an elementwise add node.
+      OpDesc desc;
+      desc.SetInput("X", std::vector<std::string>({conv_out->Name()}));
+      desc.SetInput("Y", std::vector<std::string>({eltwise_y_in_node->Name()}));
+      desc.SetOutput("Out", std::vector<std::string>({bn_out->Name()}));
+      desc.SetType("elementwise_add");
+      desc.SetAttr("axis", 1);
+      auto eltwise_op = g->CreateOpNode(&desc);  // OpDesc will be copied.
+
+      GraphSafeRemoveNodes(
+          graph.get(),
+          {bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, bn_mean_out,
+           bn_variance_out, bn_saved_mean, bn_saved_variance});
+
+      IR_NODE_LINK_TO(conv_out, eltwise_op);
+      IR_NODE_LINK_TO(eltwise_y_in_node, eltwise_op);
+      IR_NODE_LINK_TO(eltwise_op, bn_out);
+      found_conv_bn_count++;
+    }
   };
 
   gpd(graph.get(), handler);
@@ -237,7 +278,6 @@ std::unique_ptr<ir::Graph> ConvEltwiseAddBNFusePass::ApplyImpl(
         {bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, bn_mean_out,
          bn_variance_out, bn_saved_mean, bn_saved_variance, eltwise_out});
 
-    PADDLE_ENFORCE(subgraph.count(conv_input));
     IR_NODE_LINK_TO(eltwise, bn_out);
 
     found_conv_bn_count++;

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/conv_relu_mkldnn_fuse_pass.cc

@@ -46,6 +46,12 @@ std::unique_ptr<ir::Graph> ConvReLUFusePass::ApplyImpl(
     GET_IR_NODE_FROM_SUBGRAPH(relu_out, relu_out, conv_relu_pattern);  // Out
     GET_IR_NODE_FROM_SUBGRAPH(relu, relu, conv_relu_pattern);  // ReLU op
 
+    FuseOptions fuse_option = FindFuseOption(*conv, *relu);
+    if (fuse_option == DO_NOT_FUSE) {
+      VLOG(3) << "do not perform conv+relu fuse";
+      return;
+    }
+
     // Transform Conv node into ConvReLU node.
     OpDesc* desc = conv->Op();
     desc->SetOutput("Output", std::vector<std::string>({relu_out->Name()}));

paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc

@@ -20,17 +20,19 @@ namespace paddle {
 namespace framework {
 namespace ir {
 
-void SetOp(ProgramDesc* prog, const std::string& type,
+void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
            const std::vector<std::string>& inputs,
-           const std::vector<std::string>& outputs) {
+           const std::vector<std::string>& outputs, bool use_mkldnn = false) {
   auto* op = prog->MutableBlock(0)->AppendOp();
   op->SetType(type);
   if (type == "conv2d") {
-    op->SetAttr("use_mkldnn", true);
+    op->SetAttr("use_mkldnn", use_mkldnn);
+    op->SetAttr("name", name);
     op->SetInput("Input", {inputs[0]});
     op->SetInput("Filter", {inputs[1]});
     op->SetInput("Bias", {inputs[2]});
   } else if (type == "relu") {
+    op->SetAttr("use_mkldnn", use_mkldnn);
     op->SetInput("X", inputs);
   }
   op->SetOutput("Out", outputs);
@@ -43,7 +45,8 @@ void SetOp(ProgramDesc* prog, const std::string& type,
 ProgramDesc BuildProgramDesc() {
   ProgramDesc prog;
   for (auto& v :
-       std::vector<std::string>({"a", "b", "c", "weights", "bias", "f", "g"})) {
+       std::vector<std::string>({"a", "b", "c", "weights", "bias", "f", "g",
+                                 "h", "weights2", "bias2", "k", "l"})) {
     auto* var = prog.MutableBlock(0)->Var(v);
     var->SetType(proto::VarType::SELECTED_ROWS);
     if (v == "weights" || v == "bias") {
@@ -51,14 +54,24 @@ ProgramDesc BuildProgramDesc() {
     }
   }
 
-  SetOp(&prog, "OP0", std::vector<std::string>({"a"}),
+  SetOp(&prog, "OP0", "op0", std::vector<std::string>({"a"}),
         std::vector<std::string>({"b"}));
-  SetOp(&prog, "OP1", std::vector<std::string>({"b"}),
+  SetOp(&prog, "OP1", "op1", std::vector<std::string>({"b"}),
         std::vector<std::string>({"c"}));
-  SetOp(&prog, "conv2d", std::vector<std::string>({"c", "weights", "bias"}),
-        std::vector<std::string>({"f"}));
-  SetOp(&prog, "relu", std::vector<std::string>({"f"}),
-        std::vector<std::string>({"g"}));
+  // conv+relu, both with MKL-DNN
+  SetOp(&prog, "conv2d", "conv1",
+        std::vector<std::string>({"c", "weights", "bias"}),
+        std::vector<std::string>({"f"}), true);
+  SetOp(&prog, "relu", "relu1", std::vector<std::string>({"f"}),
+        std::vector<std::string>({"g"}), true);
+  SetOp(&prog, "OP3", "op3", std::vector<std::string>({"g"}),
+        std::vector<std::string>({"h"}));
+  // conv+relu, only one with MKL-DNN
+  SetOp(&prog, "conv2d", "conv2",
+        std::vector<std::string>({"h", "weights2", "bias2"}),
+        std::vector<std::string>({"k"}), true);
+  SetOp(&prog, "relu", "relu2", std::vector<std::string>({"k"}),
+        std::vector<std::string>({"l"}));
 
   return prog;
 }
@@ -88,10 +101,16 @@ TEST(ConvReLUFusePass, basic) {
       auto* op = node->Op();
       ASSERT_TRUE(op->HasAttr("use_mkldnn"));
       EXPECT_TRUE(boost::get<bool>(op->GetAttr("use_mkldnn")));
-      ASSERT_TRUE(op->HasAttr("fuse_relu"));
-      bool fuse_relu = boost::get<bool>(op->GetAttr("fuse_relu"));
-      if (fuse_relu) {
-        ++conv_relu_count;
+      // check if only "conv1" convolution is fused
+      auto op_name = boost::get<std::string>(op->GetAttr("name"));
+      if (op_name == "conv1") {
+        ASSERT_TRUE(op->HasAttr("fuse_relu"));
+        bool fuse_relu = boost::get<bool>(op->GetAttr("fuse_relu"));
+        if (fuse_relu) {
+          ++conv_relu_count;
+        }
+      } else if (op_name == "conv2") {
+        ASSERT_FALSE(op->HasAttr("fuse_relu"));
       }
     }
   }

paddle/fluid/framework/ir/fuse_pass_base.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/fuse_pass_base.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+void FusePassBase::Init(const std::string& repr, Graph* graph) const {
+  repr_ = repr;
+  graph_ = graph;
+}
+
+Scope* FusePassBase::param_scope() const {
+  PADDLE_ENFORCE(graph_->Has(kParamScopeAttr));
+  return graph_->Get<framework::Scope*>(kParamScopeAttr);
+}
+
+void FusePassBase::AddStatis(int count_of_fused) const {
+  PADDLE_ENFORCE(graph_);
+  PADDLE_ENFORCE(!repr_.empty());
+  if (!graph_->Has(kFuseStatisAttr)) {
+    graph_->Set(kFuseStatisAttr, new std::unordered_map<std::string, int>);
+  }
+  auto& info =
+      graph_->Get<std::unordered_map<std::string, int>>(kFuseStatisAttr);
+  info[repr_] = count_of_fused;
+}
+
+FuseOptions FusePassBase::FindFuseOption(const Node& node1,
+                                         const Node& node2) const {
+#ifdef PADDLE_WITH_MKLDNN
+  bool node1_mkldnn = node1.Op()->HasAttr("use_mkldnn") &&
+                      boost::get<bool>(node1.Op()->GetAttr("use_mkldnn"));
+  bool node2_mkldnn = node2.Op()->HasAttr("use_mkldnn") &&
+                      boost::get<bool>(node2.Op()->GetAttr("use_mkldnn"));
+  if (node1_mkldnn && node2_mkldnn)
+    return FUSE_MKLDNN;
+  else if (!node1_mkldnn && !node2_mkldnn)
+    return FUSE_NATIVE;
+  else
+    return DO_NOT_FUSE;
+#else
+  return FUSE_NATIVE;
+#endif
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/fuse_pass_base.h

@@ -25,32 +25,24 @@ namespace ir {
 static const char kParamScopeAttr[] = "__param_scope__";
 static const char kFuseStatisAttr[] = "__fuse_statis__";
 
+enum FuseOptions {
+  DO_NOT_FUSE,  // fusing will not be done
+  FUSE_NATIVE,  // fusing will be done without MKL-DNN
+  FUSE_MKLDNN   // fusing will be done with MKL-DNN
+};
+
 class FusePassBase : public Pass {
  public:
-  void Init(const std::string& repr, Graph* graph) const {
-    repr_ = repr;
-    graph_ = graph;
-  }
-
-  Scope* param_scope() const {
-    PADDLE_ENFORCE(graph_->Has(kParamScopeAttr));
-    return graph_->Get<framework::Scope*>(kParamScopeAttr);
-  }
-
-  void AddStatis(int count_of_fused) const {
-    PADDLE_ENFORCE(graph_);
-    PADDLE_ENFORCE(!repr_.empty());
-    if (!graph_->Has(kFuseStatisAttr)) {
-      graph_->Set(kFuseStatisAttr, new std::unordered_map<std::string, int>);
-    }
-    auto& info =
-        graph_->Get<std::unordered_map<std::string, int>>(kFuseStatisAttr);
-    info[repr_] = count_of_fused;
-  }
+  void Init(const std::string& repr, Graph* graph) const;
+  Scope* param_scope() const;
+  void AddStatis(int count_of_fused) const;
 
   virtual ~FusePassBase() {}
 
  protected:
+  virtual FuseOptions FindFuseOption(const Node& node1,
+                                     const Node& node2) const;
+
   mutable Graph* graph_;
   mutable std::string repr_;
 };

paddle/fluid/framework/ir/graph_helper_test.cc

@@ -200,15 +200,15 @@ TEST(GraphHelperTest, GraphNum) {
 
   Graph g(prog);
   BuildZeroGraph(&g);
-  ASSERT_EQ(GraphNum(g), 0);
+  ASSERT_EQ(GraphNum(g), 0UL);
 
   Graph g2(prog);
   BuildOneGraph(&g2);
-  ASSERT_EQ(GraphNum(g2), 1);
+  ASSERT_EQ(GraphNum(g2), 1UL);
 
   Graph g3(prog);
   BuildTwoGraphs(&g3);
-  ASSERT_EQ(GraphNum(g3), 2);
+  ASSERT_EQ(GraphNum(g3), 2UL);
 }
 
 }  // namespace ir

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -259,6 +259,8 @@ GraphPatternDetector::DetectPatterns() {
   return result;
 }
 
+// TODO(Superjomn) enhance the function as it marks unique unique as duplicates
+// see https://github.com/PaddlePaddle/Paddle/issues/13550
 void GraphPatternDetector::UniquePatterns(
     std::vector<GraphPatternDetector::subgraph_t> *subgraphs) {
   if (subgraphs->empty()) return;
@@ -759,6 +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.
@@ -964,6 +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/graph_test.cc

@@ -124,7 +124,7 @@ TEST(GraphTest, Basic) {
       ASSERT_EQ(n->outputs.size(), 0UL);
     }
   }
-  ASSERT_EQ(nodes.size(), 5);
+  ASSERT_EQ(nodes.size(), 5UL);
 }
 
 TEST(GraphTest, WriteAfterRead) {

paddle/fluid/framework/ir/mkldnn_placement_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/mkldnn_placement_pass.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+std::unique_ptr<ir::Graph> MKLDNNPlacementPass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  VLOG(3) << "Aplies MKL-DNN placement strategy.";
+  for (const Node* n : graph->Nodes()) {
+    if (n->IsOp() && n->Op()->HasAttr("use_mkldnn")) {
+      n->Op()->SetAttr("use_mkldnn", true);
+    }
+  }
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(mkldnn_placement_pass,
+              paddle::framework::ir::MKLDNNPlacementPass);

paddle/fluid/framework/ir/mkldnn_placement_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 "paddle/fluid/framework/ir/pass.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class MKLDNNPlacementPass : public Pass {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override;
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

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/op_desc.cc

@@ -85,10 +85,6 @@ class CompileTimeInferShapeContext : public InferShapeContext {
       VLOG(3) << "input " << in << " is not LodTensor";
       return;
     }
-    PADDLE_ENFORCE_EQ(in_var->GetType(), proto::VarType::LOD_TENSOR,
-                      "The %d-th output of Output(%s) must be LoDTensor.", j,
-                      out);
-
     out_var->SetLoDLevel(in_var->GetLoDLevel());
   }
 
@@ -519,20 +515,14 @@ void OpDesc::InferShape(const BlockDesc &block) const {
 }
 
 void OpDesc::InferVarType(BlockDesc *block) const {
+  // There are a few places that var type can be set.
+  // When VarDesc is created, default set to LOD_TENSOR.
+  // When output variable is created, default is defaut set to LOD_TENSOR.
+  // We limit here to be the only place that operator defines its customized
+  // var type inference. Hence, we don't do any "default" setting here.
   auto &info = OpInfoMap::Instance().Get(this->Type());
   if (info.infer_var_type_) {
     info.infer_var_type_(*this, block);
-  } else {
-    // all output type is LoDTensor by default
-    VLOG(10) << this->Type()
-             << " has not registered InferVarType. Set output variables to "
-                "LOD_TENSOR";
-    for (auto &out_pair : this->outputs_) {
-      for (auto &out_var_name : out_pair.second) {
-        block->FindRecursiveOrCreateVar(out_var_name)
-            .SetType(proto::VarType::LOD_TENSOR);
-      }
-    }
   }
 }
 

paddle/fluid/framework/parallel_executor.cc

@@ -156,12 +156,10 @@ ParallelExecutor::ParallelExecutor(
                            params, member_->local_scopes_, member_->use_cuda_);
 #endif
 
-  if (VLOG_IS_ON(5)) {
-    // If the loss_var_name is given, the number of graph should be only one.
-    if (loss_var_name.size()) {
-      PADDLE_ENFORCE_EQ(ir::GraphNum(*graph), 1,
-                        "The number of graph should be only one");
-    }
+  // If the loss_var_name is given, the number of graph should be only one.
+  if (loss_var_name.size()) {
+    PADDLE_ENFORCE_EQ(ir::GraphNum(*graph), 1,
+                      "The number of graph should be only one");
   }
 
   if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
@@ -299,6 +297,12 @@ 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/framework/program_desc.cc

@@ -126,7 +126,7 @@ const std::vector<std::string> ProgramDesc::GetFeedTargetNames() {
   std::vector<std::string> feed_target_names;
   for (auto *op : global_block.AllOps()) {
     if (op->Type() == kFeedOpType) {
-      int col = boost::get<int>(op->GetAttr("col"));
+      size_t col = boost::get<int>(op->GetAttr("col"));
       if (col >= feed_target_names.size()) {
         feed_target_names.resize(col + 1);
       }
@@ -143,7 +143,7 @@ const std::vector<std::string> ProgramDesc::GetFetchTargetNames() {
   std::vector<std::string> fetch_target_names;
   for (auto *op : global_block.AllOps()) {
     if (op->Type() == kFetchOpType) {
-      int col = boost::get<int>(op->GetAttr("col"));
+      size_t col = boost::get<int>(op->GetAttr("col"));
       if (col >= fetch_target_names.size()) {
         fetch_target_names.resize(col + 1);
       }

paddle/fluid/framework/program_desc_test.cc

@@ -103,7 +103,7 @@ TEST(ProgramDesc, copy_ctor) {
       ASSERT_EQ(1, op->GetBlockAttrId("sub_block"));
       found_sub_block = true;
 
-      ASSERT_EQ(2, op->GetBlocksAttrIds("sub_blocks").size());
+      ASSERT_EQ(2UL, op->GetBlocksAttrIds("sub_blocks").size());
       found_sub_blocks = true;
     }
   }

paddle/fluid/framework/reader_test.cc

@@ -39,8 +39,8 @@ TEST(READER, decorate_chain) {
   {
     auto endpoints = root->GetEndPoints();
     ASSERT_EQ(endpoints.size(), 2U);
-    ASSERT_NE(endpoints.count(end_point1.get()), 0);
-    ASSERT_NE(endpoints.count(end_point2.get()), 0);
+    ASSERT_NE(endpoints.count(end_point1.get()), 0UL);
+    ASSERT_NE(endpoints.count(end_point2.get()), 0UL);
   }
 
   {

paddle/fluid/framework/selected_rows_test.cc

@@ -91,7 +91,7 @@ TEST(SelectedRows, SparseTable) {
   ASSERT_TRUE(table.HasKey(10));
   ASSERT_TRUE(table.HasKey(8));
   ASSERT_TRUE(table.HasKey(6));
-  ASSERT_EQ(table.rows().size(), 3);
+  ASSERT_EQ(table.rows().size(), 3UL);
 
   framework::Tensor ids;
   ids.Resize(framework::make_ddim({4}));

paddle/fluid/inference/analysis/analyzer.cc

@@ -101,7 +101,13 @@ Analyzer::Analyzer() { Register("manager1", new DfgPassManagerImpl); }
 
 void Analyzer::Run(Argument* argument) {
   std::vector<std::string> passes;
-  for (auto& pass : all_ir_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);
       passes.push_back("graph_viz_pass");  // add graphviz for debug.
@@ -117,11 +123,26 @@ void Analyzer::Run(Argument* argument) {
   }
 }
 
+Analyzer& Analyzer::IncludeAllIrPasses() {
+  ir_passes_ = all_ir_passes_;
+  return *this;
+}
+
 Analyzer& Analyzer::DisableIrPasses(const std::vector<std::string>& passes) {
   disabled_ir_passes_.insert(passes.begin(), passes.end());
   return *this;
 }
 
+Analyzer& Analyzer::IncludeIrPasses(const std::vector<std::string>& passes) {
+  ir_passes_ = passes;
+  return *this;
+}
+
+Analyzer& Analyzer::SetUseMkldnn(bool use_mkldnn) {
+  use_mkldnn_ = use_mkldnn;
+  return *this;
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/analysis/analyzer.h

@@ -54,6 +54,9 @@ class Analyzer : public OrderedRegistry<PassManager> {
   void Run(Argument* argument);
 
   Analyzer& DisableIrPasses(const std::vector<std::string>& passes);
+  Analyzer& IncludeIrPasses(const std::vector<std::string>& passes);
+  Analyzer& IncludeAllIrPasses();
+  Analyzer& SetUseMkldnn(bool use_mkldnn);
 
   DISABLE_COPY_AND_ASSIGN(Analyzer);
 
@@ -64,23 +67,29 @@ class Analyzer : public OrderedRegistry<PassManager> {
   // larger fusion.
   const std::vector<std::string> all_ir_passes_{{
       // Manual update the passes here.
-      "infer_clean_graph_pass",        //
-      "attention_lstm_fuse_pass",      //
-      "embedding_fc_lstm_fuse_pass",   //
-      "fc_lstm_fuse_pass",             //
-      "mul_lstm_fuse_pass",            //
-      "fc_gru_fuse_pass",              //
-      "mul_gru_fuse_pass",             //
-      "seq_concat_fc_fuse_pass",       //
-      "fc_fuse_pass",                  //
-      "conv_bn_fuse_pass",             //
-      "conv_eltwiseadd_bn_fuse_pass",  //
+      "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",             //
+      "fc_gru_fuse_pass",               //
+      "mul_gru_fuse_pass",              //
+      "seq_concat_fc_fuse_pass",        //
+      "fc_fuse_pass",                   //
+      "conv_bn_fuse_pass",              //
+      "conv_eltwiseadd_bn_fuse_pass",   //
 #ifdef PADDLE_WITH_MKLDNN
-      "conv_relu_mkldnn_fuse_pass",  //
+      "conv_bias_mkldnn_fuse_pass",             //
+      "conv_relu_mkldnn_fuse_pass",             //
+      "conv_elementwise_add_mkldnn_fuse_pass",  //
 #endif
   }};
 
   std::unordered_set<std::string> disabled_ir_passes_;
+  // Ir passes to run
+  std::vector<std::string> ir_passes_;
+  bool use_mkldnn_;
 };
 
 }  // namespace analysis

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);
 
@@ -225,10 +221,24 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
 
   argument_.origin_program_desc.reset(
       new ProgramDesc(*inference_program_->Proto()));
-  PADDLE_ENFORCE(
-      config_.ir_mode == contrib::AnalysisConfig::IrPassMode::kExclude,
-      "Only kExclude is supported yet.");
-  Analyzer().DisableIrPasses(config_.ir_passes).Run(&argument_);
+
+  switch (config_.ir_mode) {
+    case contrib::AnalysisConfig::IrPassMode::kExclude:
+      Analyzer()
+          .IncludeAllIrPasses()
+          .SetUseMkldnn(config_._use_mkldnn)
+          .DisableIrPasses(config_.ir_passes)
+          .Run(&argument_);
+      break;
+    case contrib::AnalysisConfig::IrPassMode::kInclude:
+      Analyzer()
+          .SetUseMkldnn(config_._use_mkldnn)
+          .IncludeIrPasses(config_.ir_passes)
+          .Run(&argument_);
+      break;
+    default:
+      LOG(ERROR) << "Only kExclude and kInclude modes are supoorted yet.";
+  }
 
   CHECK(argument_.transformed_program_desc);
   VLOG(5) << "to prepare executor";

paddle/fluid/inference/api/api_impl_tester.cc

@@ -205,7 +205,7 @@ void MainThreadsWord2Vec(bool use_gpu) {
       float* ref_data = refs[tid].data<float>();
       EXPECT_EQ(refs[tid].numel(), static_cast<int64_t>(len / sizeof(float)));
       for (int i = 0; i < refs[tid].numel(); ++i) {
-        EXPECT_NEAR(ref_data[i], data[i], ACC_DIFF);
+        EXPECT_NEAR(ref_data[i], data[i], 2e-3);
       }
     });
   }

paddle/fluid/inference/api/paddle_inference_api.h

@@ -259,10 +259,17 @@ struct AnalysisConfig : public NativeConfig {
     kExclude   // Specify the disabled passes in `ir_passes`.
   };
 
+  void SetIncludeMode() {
+    ir_mode = IrPassMode::kInclude;
+    // this pass has to be run at the beginning of all fuse passes
+    ir_passes = {"infer_clean_graph_pass"};
+  }
+
   // Determine whether to perform graph optimization.
   bool enable_ir_optim = true;
   // Manually determine the IR passes to run.
   IrPassMode ir_mode{IrPassMode::kExclude};
+  // passes to be excluded/included
   std::vector<std::string> ir_passes{"embedding_fc_lstm_fuse_pass"};
 
   // NOT stable yet.

paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc

@@ -52,9 +52,10 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
 }
 
 // Easy for profiling independently.
-TEST(Analyzer_resnet50, profile) {
+void profile(bool use_mkldnn = false) {
   AnalysisConfig cfg;
   SetConfig(&cfg);
+  cfg._use_mkldnn = use_mkldnn;
   std::vector<PaddleTensor> outputs;
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
@@ -69,6 +70,11 @@ TEST(Analyzer_resnet50, profile) {
   }
 }
 
+TEST(Analyzer_resnet50, profile) { profile(); }
+#ifdef PADDLE_WITH_MKLDNN
+TEST(Analyzer_resnet50, profile_mkldnn) { profile(true /* use_mkldnn */); }
+#endif
+
 // Check the fuse status
 TEST(Analyzer_resnet50, fuse_statis) {
   AnalysisConfig cfg;
@@ -82,15 +88,21 @@ TEST(Analyzer_resnet50, fuse_statis) {
 }
 
 // Compare result of NativeConfig and AnalysisConfig
-TEST(Analyzer_resnet50, compare) {
+void compare(bool use_mkldnn = false) {
   AnalysisConfig cfg;
   SetConfig(&cfg);
+  cfg._use_mkldnn = use_mkldnn;
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
   CompareNativeAndAnalysis(cfg, input_slots_all);
 }
 
+TEST(Analyzer_resnet50, compare) { compare(); }
+#ifdef PADDLE_WITH_MKLDNN
+TEST(Analyzer_resnet50, compare_mkldnn) { compare(true /* use_mkldnn */); }
+#endif
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

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/inference/tests/api/analyzer_vis_tester.cc

@@ -59,9 +59,6 @@ void SetConfig(AnalysisConfig *cfg) {
   cfg->specify_input_name = true;
   // TODO(TJ): fix fusion gru
   cfg->ir_passes.push_back("fc_gru_fuse_pass");
-#ifdef PADDLE_WITH_MKLDNN
-  cfg->_use_mkldnn = true;
-#endif
 }
 
 void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
@@ -84,9 +81,10 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
 
 // Easy for profiling independently.
 //  ocr, mobilenet and se_resnext50
-TEST(Analyzer_vis, profile) {
+void profile(bool use_mkldnn = false) {
   AnalysisConfig cfg;
   SetConfig(&cfg);
+  cfg._use_mkldnn = use_mkldnn;
   std::vector<PaddleTensor> outputs;
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
@@ -108,6 +106,12 @@ TEST(Analyzer_vis, profile) {
   }
 }
 
+TEST(Analyzer_vis, profile) { profile(); }
+
+#ifdef PADDLE_WITH_MKLDNN
+TEST(Analyzer_vis, profile_mkldnn) { profile(true /* use_mkldnn */); }
+#endif
+
 // Check the fuse status
 TEST(Analyzer_vis, fuse_statis) {
   AnalysisConfig cfg;
@@ -118,15 +122,21 @@ TEST(Analyzer_vis, fuse_statis) {
 }
 
 // Compare result of NativeConfig and AnalysisConfig
-TEST(Analyzer_vis, compare) {
+void compare(bool use_mkldnn = false) {
   AnalysisConfig cfg;
   SetConfig(&cfg);
+  cfg._use_mkldnn = use_mkldnn;
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
   CompareNativeAndAnalysis(cfg, input_slots_all);
 }
 
+TEST(Analyzer_vis, compare) { compare(); }
+#ifdef PADDLE_WITH_MKLDNN
+TEST(Analyzer_vis, compare_mkldnn) { compare(true /* use_mkldnn */); }
+#endif
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/tests/api/tester_helper.h

@@ -50,7 +50,7 @@ void CompareResult(const std::vector<PaddleTensor> &outputs,
     auto &ref_out = ref_outputs[i];
     size_t size = VecReduceToInt(out.shape);
     size_t ref_size = VecReduceToInt(ref_out.shape);
-    EXPECT_GT(size, 0);
+    EXPECT_GT(size, 0UL);
     EXPECT_EQ(size, ref_size);
     EXPECT_EQ(out.dtype, ref_out.dtype);
     switch (out.dtype) {
@@ -163,7 +163,8 @@ void TestPrediction(const AnalysisConfig &config,
                     const std::vector<std::vector<PaddleTensor>> &inputs,
                     std::vector<PaddleTensor> *outputs, int num_threads,
                     bool use_analysis = FLAGS_use_analysis) {
-  LOG(INFO) << "use_analysis: " << use_analysis;
+  LOG(INFO) << "use_analysis: " << use_analysis
+            << ", use_mkldnn: " << config._use_mkldnn;
   if (num_threads == 1) {
     TestOneThreadPrediction(config, inputs, outputs, use_analysis);
   } else {
@@ -175,6 +176,7 @@ void TestPrediction(const AnalysisConfig &config,
 void CompareNativeAndAnalysis(
     const AnalysisConfig &config,
     const std::vector<std::vector<PaddleTensor>> &inputs) {
+  LOG(INFO) << "use_mkldnn: " << config._use_mkldnn;
   std::vector<PaddleTensor> native_outputs, analysis_outputs;
   TestOneThreadPrediction(config, inputs, &native_outputs, false);
   TestOneThreadPrediction(config, inputs, &analysis_outputs, true);

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()
@@ -284,10 +284,10 @@ op_library(max_sequence_len_op DEPS lod_rank_table)
 op_library(sequence_conv_op DEPS context_project)
 op_library(sequence_pool_op DEPS sequence_pooling)
 if (NOT WIN32)
-op_library(lstm_op DEPS sequence2batch lstm_compute)
-op_library(hierarchical_sigmoid_op DEPS matrix_bit_code)
-op_library(lstmp_op DEPS sequence2batch lstm_compute)
-op_library(gru_op DEPS sequence2batch gru_compute)
+    op_library(lstm_op DEPS sequence2batch lstm_compute)
+    op_library(hierarchical_sigmoid_op DEPS matrix_bit_code)
+    op_library(lstmp_op DEPS sequence2batch lstm_compute)
+    op_library(gru_op DEPS sequence2batch gru_compute)
 endif(NOT WIN32)
 op_library(recurrent_op DEPS executor)
 op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
@@ -316,7 +316,7 @@ op_library(save_op DEPS lod_tensor)
 op_library(load_op DEPS lod_tensor)
 op_library(save_combine_op DEPS lod_tensor)
 op_library(load_combine_op DEPS lod_tensor)
-op_library(concat_op DEPS concat)
+op_library(concat_op DEPS concat_and_split)
 
 list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
 
@@ -348,6 +348,6 @@ cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor memory)
 cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)
 cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op)
 if(NOT WIN32)
-nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
+    nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
 endif()
 nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor)

paddle/fluid/operators/array_to_lod_tensor_op.cc

@@ -11,7 +11,7 @@ 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/math/concat.h>
+#include <paddle/fluid/operators/math/concat_and_split.h>
 #include <numeric>
 
 #include "paddle/fluid/framework/lod_rank_table.h"

paddle/fluid/operators/concat_op.h

@@ -17,7 +17,7 @@ limitations under the License. */
 #include <utility>
 #include <vector>
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/operators/math/concat.h"
+#include "paddle/fluid/operators/math/concat_and_split.h"
 #include "paddle/fluid/operators/strided_memcpy.h"
 
 namespace paddle {
@@ -89,29 +89,17 @@ class ConcatGradKernel : public framework::OpKernel<T> {
         outputs.push_back(nullptr);
       }
     }
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
 
     // Sometimes direct copies will be faster, this maybe need deeply analysis.
     if (axis == 0 && outs.size() < 10) {
-      size_t input_offset = 0;
-      const auto in_stride = framework::stride_numel(out_grad->dims());
-
-      for (size_t i = 0; i < outs.size(); ++i) {
-        auto out_stride = framework::stride_numel(ins[i]->dims());
-        auto* out = outputs[i];
-        if (out != nullptr) {
-          StridedNumelCopyWithAxis<T>(
-              ctx.device_context(), axis, out->data<T>(), out_stride,
-              out_grad->data<T>() + input_offset, in_stride, out_stride[axis]);
-        }
-        input_offset += out_stride[axis];
-      }
+      std::vector<const framework::Tensor*> ref_shape;
+      ref_shape.insert(ref_shape.begin(), ins.begin(), ins.end());
+      StridedMemcpyWithAxis0<T>(dev_ctx, *out_grad, ref_shape, &outputs);
     } else {
-      auto& dev_ctx = ctx.template device_context<DeviceContext>();
-      paddle::operators::math::ConcatGradFunctor<DeviceContext, T>
-          concat_grad_functor;
-      concat_grad_functor(dev_ctx, *out_grad,
-                          ctx.MultiInput<framework::Tensor>("X"),
-                          static_cast<int>(axis), &outputs);
+      math::SplitFunctor<DeviceContext, T> split_functor;
+      split_functor(dev_ctx, *out_grad, ctx.MultiInput<framework::Tensor>("X"),
+                    static_cast<int>(axis), &outputs);
     }
   }
 };

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 =
-        output->mutable_data<T>(ctx.GetPlace(), handler.GetDstMemorySize());
+    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_proposal_labels_op.cc

@@ -16,7 +16,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detection/bbox_util.h"
 #include "paddle/fluid/operators/gather.h"
-#include "paddle/fluid/operators/math/concat.h"
+#include "paddle/fluid/operators/math/concat_and_split.h"
 #include "paddle/fluid/operators/math/math_function.h"
 
 namespace paddle {

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]
@@ -392,10 +394,9 @@ 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) {
-          return scores_data[i] > scores_data[j];
-        };
+    auto compare = [scores_data](const int64_t &i, const int64_t &j) {
+      return scores_data[i] > scores_data[j];
+    };
 
     if (pre_nms_top_n <= 0 || pre_nms_top_n >= scores_slice.numel()) {
       std::sort(index, index + scores_slice.numel(), compare);
@@ -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);
@@ -191,8 +205,9 @@ __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,
-                          const float *dev_boxes, uint64_t *dev_mask) {
+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;
 
@@ -234,9 +249,9 @@ __global__ void NMSKernel(const int n_boxes, const float nms_overlap_thresh,
 }
 
 template <typename T>
-void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals,
-         const Tensor &sorted_indices, const T nms_threshold,
-         Tensor *keep_out) {
+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];
   PADDLE_ENFORCE_EQ(boxes_num, sorted_indices.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>>>(
-      anchors.data<T>(), bbox_deltas.data<T>(), variances.data<T>(),
-      index_sort.data<int>(), im_info.data<T>(), pre_nms_num,
-      proposals.data<T>());
+
+  {
+    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>(), 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.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, "
-                      "represents the layout of coordinate "
-                      "[xmin, ymin, xmax, ymax]");
+    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] 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/lod_tensor_to_array_op.cc

@@ -17,7 +17,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
-#include "paddle/fluid/operators/math/concat.h"
+#include "paddle/fluid/operators/math/concat_and_split.h"
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/port.h"
 
@@ -79,7 +79,7 @@ struct LoDTensorToArrayFunctor : public boost::static_visitor<void> {
 template <typename DeviceContext>
 template <typename T>
 void LoDTensorToArrayFunctorImpl<DeviceContext>::apply() {
-  math::ConcatGradFunctor<DeviceContext, T> func;
+  math::SplitFunctor<DeviceContext, T> func;
   func(*dev_ctx_, prev_functor_->input_, prev_functor_->ref_inputs_, 0,
        &prev_functor_->outputs_);
 }

paddle/fluid/operators/math/CMakeLists.txt

 if (NOT WIN32)
-add_subdirectory(detail)
+    add_subdirectory(detail)
 endif(NOT WIN32)
 
 function(math_library TARGET)
@@ -35,7 +35,7 @@ function(math_library TARGET)
 endfunction()
 
 # please add new math_library in alphabetical order
-math_library(concat)
+math_library(concat_and_split)
 math_library(context_project DEPS im2col math_function)
 math_library(cross_entropy)
 math_library(cos_sim_functor)
@@ -43,8 +43,8 @@ math_library(depthwise_conv)
 math_library(im2col)
 
 if (NOT WIN32) # windows do not support avx functions yet.
-math_library(gru_compute DEPS activation_functions math_function)
-math_library(lstm_compute DEPS activation_functions)
+    math_library(gru_compute DEPS activation_functions math_function)
+    math_library(lstm_compute DEPS activation_functions)
 endif (NOT WIN32)
 
 cc_library(blas SRCS blas.cc DEPS cblas framework_proto device_context)
@@ -58,7 +58,7 @@ math_library(sequence_pooling DEPS math_function)
 math_library(sequence_scale)
 math_library(softmax DEPS math_function)
 if (NOT WIN32)
-math_library(matrix_bit_code)
+    math_library(matrix_bit_code)
 endif (NOT WIN32)
 math_library(unpooling)
 math_library(vol2col)
@@ -72,9 +72,9 @@ if(WITH_GPU)
     nv_test(math_function_gpu_test SRCS math_function_test.cu DEPS math_function)
     nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor math_function)
 endif()
-cc_test(concat_test SRCS concat_test.cc DEPS concat)
+cc_test(concat_test SRCS concat_test.cc DEPS concat_and_split)
 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/concat.cc → paddle/fluid/operators/math/concat_and_split.cc

@@ -12,7 +12,7 @@ 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/math/concat.h"
+#include "paddle/fluid/operators/math/concat_and_split.h"
 #include <vector>
 
 namespace paddle {
@@ -67,7 +67,7 @@ class ConcatFunctor<platform::CPUDeviceContext, T> {
  * each dimension must be the same, except the axis dimension.
  */
 template <typename T>
-class ConcatGradFunctor<platform::CPUDeviceContext, T> {
+class SplitFunctor<platform::CPUDeviceContext, T> {
  public:
   void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input,
@@ -111,7 +111,7 @@ class ConcatGradFunctor<platform::CPUDeviceContext, T> {
 };
 #define DEFINE_FUNCTOR(type)                                      \
   template class ConcatFunctor<platform::CPUDeviceContext, type>; \
-  template class ConcatGradFunctor<platform::CPUDeviceContext, type>;
+  template class SplitFunctor<platform::CPUDeviceContext, type>;
 
 FOR_ALL_TYPES(DEFINE_FUNCTOR);
 

paddle/fluid/operators/math/concat.cu → paddle/fluid/operators/math/concat_and_split.cu

@@ -15,7 +15,7 @@ limitations under the License. */
 #include <algorithm>
 #include <vector>
 #include "paddle/fluid/framework/mixed_vector.h"
-#include "paddle/fluid/operators/math/concat.h"
+#include "paddle/fluid/operators/math/concat_and_split.h"
 #include "paddle/fluid/platform/cuda_primitives.h"
 #include "paddle/fluid/platform/float16.h"
 
@@ -24,7 +24,7 @@ namespace operators {
 namespace math {
 
 template <typename T>
-__global__ void KernelConcat(T** inputs, const int* input_cols, int col_size,
+__global__ void ConcatKernel(T** inputs, const int* input_cols, int col_size,
                              const int output_rows, const int output_cols,
                              T* output) {
   int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
@@ -50,7 +50,7 @@ __global__ void KernelConcat(T** inputs, const int* input_cols, int col_size,
 }
 
 template <typename T>
-__global__ void KernelConcat(T** inputs_data, const int fixed_in_col,
+__global__ void ConcatKernel(T** inputs_data, const int fixed_in_col,
                              const int out_rows, const int out_cols,
                              T* output_data) {
   int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
@@ -67,9 +67,9 @@ __global__ void KernelConcat(T** inputs_data, const int fixed_in_col,
 }
 
 template <typename T>
-__global__ void KernelConcatGrad(const T* input_data, const int in_row,
-                                 const int in_col, const int* out_cols,
-                                 int out_cols_size, T** outputs_data) {
+__global__ void SplitKernel(const T* input_data, const int in_row,
+                            const int in_col, const int* out_cols,
+                            int out_cols_size, T** outputs_data) {
   int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
   int curr_segment = 0;
   int curr_offset = out_cols[0];
@@ -94,9 +94,9 @@ __global__ void KernelConcatGrad(const T* input_data, const int in_row,
 }
 
 template <typename T>
-__global__ void KernelConcatGrad(const T* input_data, const int in_row,
-                                 const int in_col, const int fixed_out_col,
-                                 T** outputs_data) {
+__global__ void SplitKernel(const T* input_data, const int in_row,
+                            const int in_col, const int fixed_out_col,
+                            T** outputs_data) {
   int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
   for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) {
     int split = tid_x / fixed_out_col;
@@ -170,11 +170,11 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
     dim3 grid_size = dim3(grid_cols, grid_rows, 1);
 
     if (sameShape) {
-      KernelConcat<<<grid_size, block_size, 0, context.stream()>>>(
+      ConcatKernel<<<grid_size, block_size, 0, context.stream()>>>(
           dev_ins_data, in_col, out_row, out_col, output->data<T>());
     } else {
       const int* dev_ins_col_data = inputs_col.CUDAData(context.GetPlace());
-      KernelConcat<<<grid_size, block_size, 0, context.stream()>>>(
+      ConcatKernel<<<grid_size, block_size, 0, context.stream()>>>(
           dev_ins_data, dev_ins_col_data, static_cast<int>(inputs_col.size()),
           out_row, out_col, output->data<T>());
     }
@@ -189,7 +189,7 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
  * each dimension must be the same, except the axis dimension.
  */
 template <typename T>
-class ConcatGradFunctor<platform::CUDADeviceContext, T> {
+class SplitFunctor<platform::CUDADeviceContext, T> {
  public:
   void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input,
@@ -248,11 +248,11 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
     dim3 grid_size = dim3(grid_cols, grid_rows, 1);
 
     if (sameShape) {
-      KernelConcatGrad<<<grid_size, block_size, 0, context.stream()>>>(
+      SplitKernel<<<grid_size, block_size, 0, context.stream()>>>(
           input.data<T>(), in_row, in_col, out0_col, dev_out_gpu_data);
     } else {
       const int* dev_outs_col_data = outputs_cols.CUDAData(context.GetPlace());
-      KernelConcatGrad<<<grid_size, block_size, 0, context.stream()>>>(
+      SplitKernel<<<grid_size, block_size, 0, context.stream()>>>(
           input.data<T>(), in_row, in_col, dev_outs_col_data,
           static_cast<int>(outputs_cols.size()), dev_out_gpu_data);
     }
@@ -264,7 +264,7 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
 
 #define DEFINE_FUNCTOR(type)                                       \
   template class ConcatFunctor<platform::CUDADeviceContext, type>; \
-  template class ConcatGradFunctor<platform::CUDADeviceContext, type>
+  template class SplitFunctor<platform::CUDADeviceContext, type>
 
 FOR_ALL_TYPES(DEFINE_FUNCTOR);
 

paddle/fluid/operators/math/concat.h → paddle/fluid/operators/math/concat_and_split.h

@@ -54,7 +54,7 @@ class ConcatFunctor {
  *     Output[1] = [[5,6]]
  */
 template <typename DeviceContext, typename T>
-class ConcatGradFunctor {
+class SplitFunctor {
  public:
   void operator()(const DeviceContext& context, const framework::Tensor& input,
                   const std::vector<const framework::Tensor*>& ref_inputs,

paddle/fluid/operators/math/concat_test.cc

@@ -12,10 +12,10 @@ 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/math/concat.h"
 #include <gtest/gtest.h>
 #include <vector>
 #include "paddle/fluid/framework/tensor_util.h"
+#include "paddle/fluid/operators/math/concat_and_split.h"
 
 template <typename DeviceContext, typename Place>
 void testConcat() {

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);
+    for (int i = 0; i < M; i++) {
+      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_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,
@@ -175,26 +192,61 @@ class LSTMKernelImpl : public LSTMKernel<T> {
 #endif
 };
 
-#define INTRI8_FLOAT(isa)                                                    \
-  template <>                                                                \
-  void LSTMKernelImpl<float, isa, kEQ8>::ComputeCtHt(                        \
-      float* gates, const float* ct_1, float* ct, float* ht,                 \
-      const float* wp_data, float* checked) const {                          \
-    /* gates: W_ch, W_ih, W_fh, W_oh */                                      \
-    __m256 c, i, f, o;                                                       \
-    c = _mm256_loadu_ps(gates);                                              \
-    i = _mm256_loadu_ps(gates + 8);                                          \
-    f = _mm256_loadu_ps(gates + 16);                                         \
-    o = _mm256_loadu_ps(gates + 24);                                         \
-    /* C_t = C_t-1 * fgated + cand_gated * igated*/                          \
-    c = _mm256_mul_ps(avx_act_cand_->Compute(c), avx_act_gate_->Compute(i)); \
-    i = _mm256_loadu_ps(ct_1);                                               \
-    f = _mm256_mul_ps(i, avx_act_gate_->Compute(f));                         \
-    f = _mm256_add_ps(c, f);                                                 \
-    _mm256_storeu_ps(ct, f);                                                 \
-    /* 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);                                                 \
+#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,                   \
+      const float* wp_data, float* checked) const {                            \
+    /* gates: W_ch, W_ih, W_fh, W_oh */                                        \
+    __m256 c, i, f, o;                                                         \
+    c = _mm256_loadu_ps(gates);                                                \
+    i = _mm256_loadu_ps(gates + 8);                                            \
+    f = _mm256_loadu_ps(gates + 16);                                           \
+    o = _mm256_loadu_ps(gates + 24);                                           \
+    /* C_t = C_t-1 * fgated + cand_gated * igated*/                            \
+    c = _mm256_mul_ps(avx_act_cand_->Compute(c), avx_act_gate_->Compute(i));   \
+    i = _mm256_loadu_ps(ct_1);                                                 \
+    f = _mm256_mul_ps(i, avx_act_gate_->Compute(f));                           \
+    f = _mm256_add_ps(c, f);                                                   \
+    _mm256_storeu_ps(ct, f);                                                   \
+    /* 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_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>);