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

cmake/inference_lib.cmake

@@ -18,7 +18,7 @@ function(copy TARGET)
     set(oneValueArgs "")
     set(multiValueArgs SRCS DSTS DEPS)
     cmake_parse_arguments(copy_lib "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-    set(inference_lib_dist_dep ${TARGET} ${inference_lib_dist_dep} PARENT_SCOPE)
+    set(fluid_lib_dist_dep ${TARGET} ${fluid_lib_dist_dep} PARENT_SCOPE)
 
     list(LENGTH copy_lib_SRCS copy_lib_SRCS_len)
     list(LENGTH copy_lib_DSTS copy_lib_DSTS_len)
@@ -185,7 +185,8 @@ copy(cmake_cache
   SRCS ${CMAKE_CURRENT_BINARY_DIR}/CMakeCache.txt
   DSTS ${FLUID_INSTALL_DIR})
 
-add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep}) 
+# This command generates a complete fluid library for both train and inference
+add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep}) 
 
 # paddle fluid version
 execute_process(

paddle/fluid/API.spec

@@ -75,7 +75,8 @@ paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'outp
 paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
 paddle.fluid.layers.sequence_expand ArgSpec(args=['x', 'y', 'ref_level', 'name'], varargs=None, keywords=None, defaults=(-1, None))
 paddle.fluid.layers.sequence_expand_as ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen', 'name'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.layers.sequence_unpad ArgSpec(args=['x', 'length', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.lstm_unit ArgSpec(args=['x_t', 'hidden_t_prev', 'cell_t_prev', 'forget_bias', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(0.0, None, None, None))
 paddle.fluid.layers.reduce_sum ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
 paddle.fluid.layers.reduce_mean ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
@@ -84,6 +85,7 @@ paddle.fluid.layers.reduce_min ArgSpec(args=['input', 'dim', 'keep_dim', 'name']
 paddle.fluid.layers.reduce_prod ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
 paddle.fluid.layers.sequence_first_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.sequence_last_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.layers.sequence_slice ArgSpec(args=['input', 'offset', 'length', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name'], varargs=None, keywords=None, defaults=(False, None, None))
 paddle.fluid.layers.split ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None))
 paddle.fluid.layers.ctc_greedy_decoder ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,))
@@ -127,6 +129,7 @@ paddle.fluid.layers.relu ArgSpec(args=['x', 'name'], varargs=None, keywords=None
 paddle.fluid.layers.log ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.margin_rank_loss ArgSpec(args=['label', 'left', 'right', 'margin', 'name'], varargs=None, keywords=None, defaults=(0.1, None))
 paddle.fluid.layers.elu ArgSpec(args=['x', 'alpha', 'name'], varargs=None, keywords=None, defaults=(1.0, None))
 paddle.fluid.layers.relu6 ArgSpec(args=['x', 'threshold', 'name'], varargs=None, keywords=None, defaults=(6.0, None))
 paddle.fluid.layers.pow ArgSpec(args=['x', 'factor', 'name'], varargs=None, keywords=None, defaults=(1.0, None))

paddle/fluid/CMakeLists.txt

@@ -12,6 +12,5 @@ endif(NOT WIN32)
 if(WITH_INFERENCE)
   # NOTE: please add subdirectory inference at last.
   add_subdirectory(inference)
+  add_subdirectory(train)
 endif()
-
-add_subdirectory(train)

paddle/fluid/framework/details/op_handle_base.h

@@ -64,7 +64,8 @@ class OpHandleBase {
   virtual bool IsMultiDeviceTransfer() { return false; }
 
   const platform::DeviceContext *DeviceContext(platform::Place place) {
-    return dev_ctxes_[place];
+    auto it = dev_ctxes_.find(place);
+    return it != dev_ctxes_.end() ? it->second : nullptr;
   }
 
   void SetDeviceContext(platform::Place place, platform::DeviceContext *ctx_) {

paddle/fluid/framework/executor.cc

@@ -46,6 +46,41 @@ ExecutorPrepareContext::~ExecutorPrepareContext() {
   VLOG(5) << "destroy ExecutorPrepareContext";
 }
 
+template <typename RefCntMap>
+static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
+                                GarbageCollector<Tensor>* gc,
+                                RefCntMap* ref_cnts) {
+  std::unordered_set<Tensor*> erase_tensors;
+
+  auto handler = [&](const VariableNameMap& name_map) {
+    for (auto& name_pair : name_map) {
+      for (auto& name : name_pair.second) {
+        auto it = ref_cnts->find(name);
+        if (it == ref_cnts->end()) continue;
+        if ((it->second)-- == 1) {
+          auto* var = scope.FindVar(name);
+          if (var != nullptr) {
+            VLOG(10) << "Erase tensor \'" << name << "\'";
+            if (var->IsType<LoDTensor>()) {
+              erase_tensors.insert(var->GetMutable<LoDTensor>());
+            } else if (var->IsType<SelectedRows>()) {
+              erase_tensors.insert(
+                  var->GetMutable<SelectedRows>()->mutable_value());
+            }
+          }
+        }
+      }
+    }
+  };
+
+  handler(op->Inputs());
+  handler(op->Outputs());
+
+  if (!erase_tensors.empty()) {
+    gc->Add(erase_tensors);
+  }
+}
+
 Executor::Executor(const platform::Place& place) : place_(place) {}
 
 void Executor::Close() {
@@ -66,7 +101,7 @@ void InitializeVariable(Variable* var, proto::VarType::Type var_type) {
   } else if (var_type == proto::VarType::FETCH_LIST) {
     var->GetMutable<FeedFetchList>();
   } else if (var_type == proto::VarType::STEP_SCOPES) {
-    var->GetMutable<std::vector<framework::Scope>>();
+    var->GetMutable<std::vector<framework::Scope*>>();
   } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
     var->GetMutable<LoDRankTable>();
   } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {
@@ -331,9 +366,13 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
   }
 
   int64_t max_memory_size = GetEagerDeletionThreshold();
-
   std::unique_ptr<GarbageCollector<Tensor>> gc;
-  if (max_memory_size >= 0) {
+  // WhileOp would set keep_kids to false
+  // WhileGradOp would need the scopes created in WhileOp
+  // Perhaps, we should not perform eager deletion in WhileOp
+  // The scopes and variables created by WhileOp would be deleted
+  // in WhileGradOp.
+  if (max_memory_size >= 0 && !keep_kids) {
     ctx->ResetReferenceCount();
 #ifdef PADDLE_WITH_CUDA
     if (platform::is_gpu_place(place_)) {
@@ -352,45 +391,8 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
     op->Run(*local_scope, place_);
 
     if (gc != nullptr) {
-      std::vector<std::string> erase_vars;
-      for (auto& input : op->Inputs()) {
-        for (auto& input_name : input.second) {
-          auto it = ctx->cur_ref_cnts_.find(input_name);
-          if (it == ctx->cur_ref_cnts_.end()) continue;
-          if (it->second == 1) {  // should delete it
-            erase_vars.emplace_back(input_name);
-            ctx->cur_ref_cnts_.erase(input_name);
-          } else {
-            --(it->second);
-          }
-        }
-      }
-
-      for (auto& output : op->Outputs()) {
-        for (auto& output_name : output.second) {
-          auto it = ctx->cur_ref_cnts_.find(output_name);
-          if (it == ctx->cur_ref_cnts_.end()) continue;
-          if (it->second == 1) {
-            erase_vars.emplace_back(output_name);
-            ctx->cur_ref_cnts_.erase(output_name);
-          } else {
-            --(it->second);
-          }
-        }
-      }
-
-      if (!erase_vars.empty()) {
-        std::vector<framework::LoDTensor*> erase_tensors;
-        for (auto& name : erase_vars) {
-          auto* var = local_scope->FindVar(name);
-          if (var == nullptr) continue;
-          if (var->IsType<framework::LoDTensor>()) {
-            auto* tensor = var->GetMutable<framework::LoDTensor>();
-            erase_tensors.push_back(tensor);
-          }
-        }
-        if (!erase_tensors.empty()) gc->Add(erase_tensors);
-      }
+      DeleteUnusedTensors(*local_scope, op.get(), gc.get(),
+                          &(ctx->cur_ref_cnts_));
     }
 
     if (FLAGS_benchmark) {

paddle/fluid/framework/executor.h

@@ -32,38 +32,32 @@ template <typename T>
 std::unordered_map<std::string, T> GetNonPersistableReferenceCount(
     const ProgramDesc& prog, size_t block_id) {
   auto& block = prog.Block(block_id);
-  std::unordered_set<std::string> ignored_vars;
   std::unordered_map<std::string, T> ref_cnts;
 
-  for (auto var_desc : block.AllVars()) {
-    auto type = var_desc->Proto()->type().type();
-    if (type != proto::VarType::LOD_TENSOR || var_desc->Persistable()) {
-      ignored_vars.insert(var_desc->Name());  // ignore persistable vars
-    }
-  }
-
-  for (auto op_desc : block.AllOps()) {
-    for (auto& input : op_desc->Inputs()) {
-      for (auto& input_name : input.second) {
-        if (!ignored_vars.count(input_name)) {
-          if (ref_cnts.count(input_name))
-            ++ref_cnts[input_name];
-          else
-            ref_cnts[input_name] = 1;
+  auto update_ref_cnts = [&](OpDesc* op_desc, const VariableNameMap& name_map) {
+    for (auto& name_pair : name_map) {
+      for (auto& name : name_pair.second) {
+        auto* var_desc = block.FindVar(name);
+        if (var_desc == nullptr || var_desc->Persistable()) continue;
+        auto type = var_desc->Proto()->type().type();
+        if (type != proto::VarType::LOD_TENSOR &&
+            type != proto::VarType::SELECTED_ROWS) {
+          continue;
         }
-      }
-    }
 
-    for (auto& output : op_desc->Outputs()) {
-      for (auto output_name : output.second) {
-        if (!ignored_vars.count(output_name)) {
-          if (ref_cnts.count(output_name))
-            ++ref_cnts[output_name];
-          else
-            ref_cnts[output_name] = 1;
+        auto it = ref_cnts.find(name);
+        if (it != ref_cnts.end()) {
+          ++it->second;
+        } else {
+          ref_cnts[name] = 1;
         }
       }
     }
+  };
+
+  for (auto op_desc : block.AllOps()) {
+    update_ref_cnts(op_desc, op_desc->Inputs());
+    update_ref_cnts(op_desc, op_desc->Outputs());
   }
   return ref_cnts;
 }

paddle/fluid/framework/feed_fetch_method.cc

@@ -27,8 +27,7 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input,
   // be created.
   VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index;
   Variable* g_feed_value = scope->Var(var_name);
-  auto& feed_inputs =
-      *(g_feed_value->GetMutable<std::vector<paddle::framework::LoDTensor>>());
+  auto& feed_inputs = *(g_feed_value->GetMutable<FeedFetchList>());
   if (index >= feed_inputs.size()) {
     feed_inputs.resize(index + 1);
   }

paddle/fluid/framework/ir/CMakeLists.txt

@@ -38,6 +38,7 @@ pass_library(fc_lstm_fuse_pass inference)
 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)
 
 cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector )
 

paddle/fluid/framework/ir/conv_bn_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_bn_fuse_pass.h"
+#include <functional>
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/operators/math/cpu_vec.h"
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+#define GET_CONV_BN_NODES(pattern_name)                                      \
+  /* OPERATORS */                                                            \
+  GET_IR_NODE_FROM_SUBGRAPH(conv, conv, pattern_name);                       \
+  GET_IR_NODE_FROM_SUBGRAPH(batch_norm, batch_norm, pattern_name);           \
+  /* CONV inputs */                                                          \
+  GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight, pattern_name);         \
+  /* CONV outputs */                                                         \
+  GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, pattern_name);               \
+  /* BN inputs */                                                            \
+  GET_IR_NODE_FROM_SUBGRAPH(bn_scale, bn_scale, pattern_name);               \
+  GET_IR_NODE_FROM_SUBGRAPH(bn_bias, bn_bias, pattern_name);                 \
+  GET_IR_NODE_FROM_SUBGRAPH(bn_mean, bn_mean, pattern_name);                 \
+  GET_IR_NODE_FROM_SUBGRAPH(bn_variance, bn_variance, pattern_name);         \
+  /* BN outputs */                                                           \
+  GET_IR_NODE_FROM_SUBGRAPH(bn_out, bn_out, pattern_name); /* Out */         \
+  GET_IR_NODE_FROM_SUBGRAPH(bn_mean_out, bn_mean_out, pattern_name);         \
+  GET_IR_NODE_FROM_SUBGRAPH(bn_variance_out, bn_variance_out, pattern_name); \
+  GET_IR_NODE_FROM_SUBGRAPH(bn_saved_mean, bn_saved_mean, pattern_name);     \
+  GET_IR_NODE_FROM_SUBGRAPH(bn_saved_variance, bn_saved_variance, pattern_name)
+
+void recompute_bias_and_weights(const Scope* scope,
+                                ir::Node* conv_weight,            //
+                                const ir::Node& bn_scale,         //
+                                const LoDTensor& bn_bias_tensor,  //
+                                const ir::Node& bn_mean,          //
+                                const ir::Node& bn_variance,      //
+                                LoDTensor* eltwise_y_in_tensor,   //
+                                float epsilon) {
+  using EigenVectorArrayMap =
+      Eigen::Map<Eigen::Array<float, Eigen::Dynamic, 1>>;
+  using ConstEigenVectorArrayMap =
+      Eigen::Map<const Eigen::Array<float, Eigen::Dynamic, 1>>;
+  using EigenMatrixArrayMap = Eigen::Map<
+      Eigen::Array<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>>;
+
+  // Re-compute bias of conv2d from BN
+  PADDLE_ENFORCE_EQ(eltwise_y_in_tensor->dims(), bn_bias_tensor.dims());
+
+  auto* scale_tensor = scope->FindVar(bn_scale.Name())->GetMutable<LoDTensor>();
+  auto* variance_tensor =
+      scope->FindVar(bn_variance.Name())->GetMutable<LoDTensor>();
+  auto* mean_tensor = scope->FindVar(bn_mean.Name())->GetMutable<LoDTensor>();
+
+  ConstEigenVectorArrayMap scale_array(scale_tensor->data<float>(),
+                                       scale_tensor->numel(), 1);
+  EigenVectorArrayMap variance_array(
+      variance_tensor->mutable_data<float>(platform::CPUPlace()),
+      variance_tensor->numel(), 1);
+  ConstEigenVectorArrayMap mean_array(mean_tensor->data<float>(),
+                                      mean_tensor->numel(), 1);
+  ConstEigenVectorArrayMap bn_bias_array(bn_bias_tensor.data<float>(),
+                                         bn_bias_tensor.numel(), 1);
+
+  // variance will not be used anymore, so make it std_array and then tmp_array
+  variance_array += epsilon;
+  variance_array = variance_array.sqrt();
+  variance_array = scale_array / variance_array;
+
+  EigenVectorArrayMap eltwise_y_in_array(
+      eltwise_y_in_tensor->mutable_data<float>(platform::CPUPlace()),
+      eltwise_y_in_tensor->numel(), 1);
+
+  eltwise_y_in_array =
+      ((eltwise_y_in_array - mean_array) * variance_array) + bn_bias_array;
+
+  // Re-compute weight of conv2d from BN
+  auto* weights = scope->FindVar(conv_weight->Name())->GetMutable<LoDTensor>();
+  auto weights_shape = weights->dims();
+  auto weights_shape_2d = flatten_to_2d(weights_shape, 1);
+
+  EigenMatrixArrayMap weights_array_2d(
+      weights->mutable_data<float>(platform::CPUPlace()), weights_shape_2d[0],
+      weights_shape_2d[1]);
+
+  weights_array_2d.colwise() *= variance_array;
+}
+
+std::unique_ptr<ir::Graph> ConvBNFusePass::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::ConvBN conv_bn_pattern(gpd.mutable_pattern(), name_scope_);
+  conv_bn_pattern(conv_input, false /*with_eltwise_add*/);
+
+  int found_conv_bn_count = 0;
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    VLOG(4) << "handle ConvBN fuse";
+
+    // 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
+    GET_CONV_BN_NODES(conv_bn_pattern);
+
+    // Create eltwise_y (conv bias) variable
+    VarDesc eltwise_y_in_desc(
+        patterns::PDNodeName(name_scope_, "eltwise_y_in"));
+    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>();
+
+    // Get batch norm bias
+    auto* bn_bias_tensor =
+        scope->FindVar(bn_bias->Name())->GetMutable<LoDTensor>();
+
+    // Initialize eltwise_y
+    eltwise_y_in_tensor->Resize(bn_bias_tensor->dims());
+    std::fill_n(eltwise_y_in_tensor->mutable_data<float>(platform::CPUPlace()),
+                eltwise_y_in_tensor->numel(), 0.0f);
+
+    // update weights and biases
+    float epsilon = boost::get<float>(batch_norm->Op()->GetAttr("epsilon"));
+    recompute_bias_and_weights(scope, conv_weight, *bn_scale, *bn_bias_tensor,
+                               *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++;
+  };
+
+  gpd(graph.get(), handler);
+
+  AddStatis(found_conv_bn_count);
+  return graph;
+}
+
+std::unique_ptr<ir::Graph> ConvEltwiseAddBNFusePass::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::ConvBN conv_bn_pattern(gpd.mutable_pattern(), name_scope_);
+  conv_bn_pattern(conv_input, true /*with_eltwise_add*/);
+
+  int found_conv_bn_count = 0;
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    VLOG(4) << "handle ConvBN fuse";
+
+    // 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
+    GET_CONV_BN_NODES(conv_bn_pattern);
+    // OPERATORS
+    GET_IR_NODE_FROM_SUBGRAPH(eltwise, eltwise, conv_bn_pattern);
+    // BIAS inputs
+    GET_IR_NODE_FROM_SUBGRAPH(eltwise_y_in, eltwise_y_in, conv_bn_pattern);
+    // BIAS outputs
+    GET_IR_NODE_FROM_SUBGRAPH(eltwise_out, eltwise_out, conv_bn_pattern);
+
+    // Get eltwise_y (conv bias) variable
+    auto* eltwise_y_in_tensor =
+        scope->FindVar(eltwise_y_in->Name())->GetMutable<LoDTensor>();
+
+    // Get batch norm bias
+    auto* bn_bias_tensor =
+        scope->FindVar(bn_bias->Name())->GetMutable<LoDTensor>();
+
+    // update weights and biases
+    float epsilon = boost::get<float>(batch_norm->Op()->GetAttr("epsilon"));
+    recompute_bias_and_weights(scope, conv_weight, *bn_scale, *bn_bias_tensor,
+                               *bn_mean, *bn_variance, eltwise_y_in_tensor,
+                               epsilon);
+
+    // Update the elementwise_add node
+    eltwise->Op()->SetAttr("axis", 1);
+    eltwise->Op()->SetOutput("Out", std::vector<std::string>({bn_out->Name()}));
+
+    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, eltwise_out});
+
+    PADDLE_ENFORCE(subgraph.count(conv_input));
+    IR_NODE_LINK_TO(eltwise, bn_out);
+
+    found_conv_bn_count++;
+  };
+
+  gpd(graph.get(), handler);
+
+  AddStatis(found_conv_bn_count);
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(conv_bn_fuse_pass, paddle::framework::ir::ConvBNFusePass);
+REGISTER_PASS(conv_eltwiseadd_bn_fuse_pass,
+              paddle::framework::ir::ConvEltwiseAddBNFusePass);

paddle/fluid/framework/ir/conv_bn_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 {
+
+/*
+ * Fuse the Conv and BatchNorm to a ConvBNMKLDNNOp.
+ */
+class ConvBNFusePass : public FusePassBase {
+ public:
+  virtual ~ConvBNFusePass() {}
+
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
+  const std::string name_scope_{"conv_bn_fuse"};
+};
+
+class ConvEltwiseAddBNFusePass : public FusePassBase {
+ public:
+  virtual ~ConvEltwiseAddBNFusePass() {}
+
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
+  const std::string name_scope_{"conv_eltwiseadd_bn_fuse"};
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -626,6 +626,112 @@ bool VarLinksFromOp(Node *node, const std::string &op_type) {
   return false;
 }
 
+PDNode *patterns::ConvBN::operator()(paddle::framework::ir::PDNode *conv_input,
+                                     bool with_eltwise_add) {
+  // Create Operators
+  conv_input->assert_is_op_input("conv2d", "Input");
+  auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d");
+
+  PDNode *eltwise_op = nullptr;
+  if (with_eltwise_add) {
+    eltwise_op =
+        pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add");
+  }
+  auto *batch_norm_op =
+      pattern->NewNode(batch_norm_repr())->assert_is_op("batch_norm");
+  // Create variables
+  // Conv Filter
+  auto *conv_weight_var = pattern->NewNode(conv_weight_repr())
+                              ->AsInput()
+                              ->assert_is_persistable_var()
+                              ->assert_is_op_input("conv2d", "Filter");
+
+  auto *conv_out_var = pattern->NewNode(conv_out_repr())
+                           ->AsIntermediate()
+                           ->assert_is_only_output_of_op("conv2d");
+
+  PDNode *eltwise_y_in_var = nullptr;
+  PDNode *eltwise_out_var = nullptr;
+  if (with_eltwise_add) {
+    // Conv output as Bias input
+    conv_out_var->assert_is_op_input("elementwise_add", "X");
+    // Bias
+    eltwise_y_in_var = pattern->NewNode(eltwise_y_in_repr())
+                           ->assert_is_op_input("elementwise_add", "Y")
+                           ->AsInput();
+    eltwise_out_var = pattern->NewNode(eltwise_out_repr())
+                          ->AsIntermediate()
+                          ->assert_is_only_output_of_op("elementwise_add");
+  } else {
+    // Conv output as BN input
+    conv_out_var->assert_is_op_input("batch_norm", "X");
+  }
+
+  // BN Scale
+  auto *bn_scale_var = pattern->NewNode(bn_scale_repr())
+                           ->AsInput()
+                           ->assert_is_persistable_var()
+                           ->assert_is_op_input("batch_norm", "Scale");
+  // BN Bias
+  auto *bn_bias_var = pattern->NewNode(bn_bias_repr())
+                          ->AsInput()
+                          ->assert_is_persistable_var()
+                          ->assert_is_op_input("batch_norm", "Bias");
+  // BN Mean
+  auto *bn_mean_var = pattern->NewNode(bn_mean_repr())
+                          ->AsInput()
+                          ->assert_is_persistable_var()
+                          ->assert_is_op_input("batch_norm", "Mean");
+  // BN Variance
+  auto *bn_variance_var = pattern->NewNode(bn_variance_repr())
+                              ->AsInput()
+                              ->assert_is_persistable_var()
+                              ->assert_is_op_input("batch_norm", "Variance");
+
+  // BN output
+  auto *bn_out_var = pattern->NewNode(bn_out_repr())
+                         ->AsOutput()
+                         ->assert_is_op_output("batch_norm");
+
+  auto *bn_mean_out_var = pattern->NewNode(bn_mean_out_repr())
+                              ->AsOutput()
+                              ->assert_is_op_output("batch_norm", "MeanOut");
+
+  auto *bn_variance_out_var =
+      pattern->NewNode(bn_variance_out_repr())
+          ->AsOutput()
+          ->assert_is_op_output("batch_norm", "VarianceOut");
+
+  auto *bn_saved_mean_var =
+      pattern->NewNode(bn_saved_mean_repr())
+          ->AsOutput()
+          ->assert_is_op_output("batch_norm", "SavedMean");
+
+  auto *bn_saved_variance_var =
+      pattern->NewNode(bn_saved_variance_repr())
+          ->AsOutput()
+          ->assert_is_op_output("batch_norm", "SavedVariance");
+
+  conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var});
+
+  if (with_eltwise_add) {
+    eltwise_op->LinksFrom({conv_out_var, eltwise_y_in_var})
+        .LinksTo({eltwise_out_var});
+    batch_norm_op
+        ->LinksFrom({eltwise_out_var, bn_scale_var, bn_bias_var, bn_mean_var,
+                     bn_variance_var})
+        .LinksTo({bn_out_var, bn_mean_out_var, bn_variance_out_var,
+                  bn_saved_mean_var, bn_saved_variance_var});
+  } else {
+    batch_norm_op
+        ->LinksFrom({conv_out_var, bn_scale_var, bn_bias_var, bn_mean_var,
+                     bn_variance_var})
+        .LinksTo({bn_out_var, bn_mean_out_var, bn_variance_out_var,
+                  bn_saved_mean_var, bn_saved_variance_var});
+  }
+  return bn_out_var;
+}
+
 PDNode *patterns::ConvReLU::operator()(
     paddle::framework::ir::PDNode *conv_input) {
   // Create Operators

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -375,6 +375,44 @@ struct PatternBase {
   size_t id_;
 };
 
+// Conv with batch norm
+// op: conv + (elementwise_add +) batch_norm
+// named nodes:
+// conv_weight, conv_out, conv,
+// bn_x, bn_scale, bn_bias, bn_mean,  bn_variance,
+// bn_batch_norm, bn_y, bn_mean_out, bn_variance_out,
+// bn_saved_mean, bn_saved_variance
+struct ConvBN : public PatternBase {
+  ConvBN(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "conv_bn") {}
+
+  PDNode* operator()(PDNode* conv_input, bool with_eltwise_add);
+
+  // declare operator node's name
+  PATTERN_DECL_NODE(conv);
+  PATTERN_DECL_NODE(batch_norm);
+  PATTERN_DECL_NODE(eltwise);  // ELEMENTWISE_ADD
+  // CONV inputs
+  PATTERN_DECL_NODE(conv_weight);  // Filter
+  // CONV outputs
+  PATTERN_DECL_NODE(conv_out);  // tmp
+  // ELTWISE inputs
+  PATTERN_DECL_NODE(eltwise_y_in);
+  // ELTWISE outputs
+  PATTERN_DECL_NODE(eltwise_out);  // tmp
+  // BN inputs
+  PATTERN_DECL_NODE(bn_scale);
+  PATTERN_DECL_NODE(bn_bias);
+  PATTERN_DECL_NODE(bn_mean);
+  PATTERN_DECL_NODE(bn_variance);
+  // BN outputs
+  PATTERN_DECL_NODE(bn_out);  // Out
+  PATTERN_DECL_NODE(bn_mean_out);
+  PATTERN_DECL_NODE(bn_variance_out);
+  PATTERN_DECL_NODE(bn_saved_mean);
+  PATTERN_DECL_NODE(bn_saved_variance);
+};
+
 // CONV with ReLU
 // op: conv + relu
 // named nodes:

paddle/fluid/framework/naive_executor.cc

@@ -37,7 +37,7 @@ static void InitializeVariable(Variable *var, proto::VarType::Type var_type) {
   } else if (var_type == proto::VarType::FETCH_LIST) {
     var->GetMutable<FeedFetchList>();
   } else if (var_type == proto::VarType::STEP_SCOPES) {
-    var->GetMutable<std::vector<framework::Scope>>();
+    var->GetMutable<std::vector<framework::Scope *>>();
   } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
     var->GetMutable<LoDRankTable>();
   } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {

paddle/fluid/framework/operator.cc

@@ -149,9 +149,17 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
     platform::SetDeviceId(dev_id);
 #endif
   }
-  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-  platform::RecordEvent record_event(Type(), pool.Get(place));
-  RunImpl(scope, place);
+
+  // The profile has a process-wide mutex, results in serious performance issue
+  // in concurrency scenerio. Here use an `if` to fix this issue.
+  // Please not remove the `if`, ask @Superjomn if there are any concern.
+  if (platform::IsProfileEnabled()) {
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+    platform::RecordEvent record_event(Type(), pool.Get(place));
+    RunImpl(scope, place);
+  } else {
+    RunImpl(scope, place);
+  }
   VLOG(3) << place << " " << DebugStringEx(&scope);
 }
 

paddle/fluid/framework/parallel_executor.cc

@@ -307,6 +307,10 @@ ParallelExecutor::~ParallelExecutor() {
       }
     }
   }
+
+  // member_ must be destructed before gcs_ since the destructor of
+  // ReferenceCountOpHandle use raw pointers of gcs_ inside.
+  member_.reset();
 }
 
 }  // namespace framework

paddle/fluid/framework/parallel_executor.h

@@ -75,7 +75,7 @@ class ParallelExecutor {
  private:
   void BCastParamsToDevices(const std::unordered_set<std::string> &vars) const;
 
-  ParallelExecutorPrivate *member_;
+  std::unique_ptr<ParallelExecutorPrivate> member_;
 
 #ifdef PADDLE_WITH_CUDA
   // ref_cnts_ is only initialized when ParallelExecutor constructs, and then

paddle/fluid/framework/scope.cc

@@ -49,18 +49,18 @@ int64_t GetEagerDeletionThreshold() {
 Scope::~Scope() { DropKids(); }
 
 Scope& Scope::NewScope() const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   kids_.push_back(new Scope(this));
   return *kids_.back();
 }
 
 Variable* Scope::Var(const std::string& name) {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   return VarInternal(name);
 }
 
 Variable* Scope::Var(std::string* name) {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   auto new_name = string::Sprintf("%p.%d", this, vars_.size());
   if (name != nullptr) {
     *name = new_name;
@@ -69,29 +69,34 @@ Variable* Scope::Var(std::string* name) {
 }
 
 Variable* Scope::FindVar(const std::string& name) const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   return FindVarInternal(name);
 }
 
+Variable* Scope::FindLocalVar(const std::string& name) const {
+  std::lock_guard<std::mutex> lock(mutex_);
+  return FindVarLocally(name);
+}
+
 const Scope* Scope::FindScope(const Variable* var) const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   return FindScopeInternal(var);
 }
 
 void Scope::DropKids() {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   for (Scope* s : kids_) delete s;
   kids_.clear();
 }
 
 bool Scope::HasKid(const Scope* scope) const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
   return it != this->kids_.end();
 }
 
 std::vector<std::string> Scope::LocalVarNames() const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   std::vector<std::string> known_vars;
   known_vars.reserve(this->vars_.size());
   for (auto& p : vars_) {
@@ -101,7 +106,7 @@ std::vector<std::string> Scope::LocalVarNames() const {
 }
 
 void Scope::DeleteScope(Scope* scope) const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
   PADDLE_ENFORCE(it != this->kids_.end(), "Cannot find %p as kid scope", scope);
   this->kids_.erase(it);
@@ -114,7 +119,7 @@ void Scope::DeleteScope(Scope* scope) const {
 }
 
 void Scope::EraseVars(const std::vector<std::string>& var_names) {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   std::set<std::string> var_set(var_names.begin(), var_names.end());
   for (auto it = vars_.begin(); it != vars_.end();) {
     if (var_set.find(it->first) != var_set.end()) {
@@ -127,12 +132,12 @@ void Scope::EraseVars(const std::vector<std::string>& var_names) {
 
 void Scope::Rename(const std::string& origin_name,
                    const std::string& new_name) const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   RenameInternal(origin_name, new_name);
 }
 
 std::string Scope::Rename(const std::string& origin_name) const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   auto new_name = string::Sprintf("%p.%d", this, vars_.size());
   RenameInternal(origin_name, new_name);
   return new_name;

paddle/fluid/framework/scope.h

@@ -63,6 +63,11 @@ class Scope {
   /// Caller doesn't own the returned Variable.
   Variable* FindVar(const std::string& name) const;
 
+  /// Find a variable in the current scope.
+  /// Return nullptr if cannot find.
+  /// Caller doesn't own the returned Variable.
+  Variable* FindLocalVar(const std::string& name) const;
+
   const Scope* parent() const { return parent_; }
 
   /// Find the scope or an ancestor scope that contains the given variable.

paddle/fluid/framework/tensor_util.cc

@@ -36,6 +36,11 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
   auto size = src.numel() * SizeOfType(src.type());
 
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
+    if (src_ptr == dst_ptr) {
+      VLOG(3) << "Skip copy the same data async from " << src_place << " to "
+              << dst_place;
+      return;
+    }
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                  boost::get<platform::CPUPlace>(src_place), src_ptr, size);
   }
@@ -71,6 +76,11 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
     auto stream =
         reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
     if (platform::is_same_place(src_place, dst_place)) {
+      if (src_ptr == dst_ptr) {
+        VLOG(3) << "Skip copy the same data async from " << src_place << " to "
+                << dst_place;
+        return;
+      }
       memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
                    stream);
     } else {
@@ -114,6 +124,11 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
   auto dst_ptr = dst->mutable_data(dst_place, src.type());
   auto size = src.numel() * SizeOfType(src.type());
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
+    if (src_ptr == dst_ptr) {
+      VLOG(3) << "Skip copy the same data from " << src_place << " to "
+              << dst_place;
+      return;
+    }
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                  boost::get<platform::CPUPlace>(src_place), src_ptr, size);
   }
@@ -130,6 +145,11 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
     memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, nullptr);
   } else if (platform::is_gpu_place(src_place) &&
              platform::is_gpu_place(dst_place)) {
+    if (src_ptr == dst_ptr && platform::is_same_place(src_place, dst_place)) {
+      VLOG(3) << "Skip copy the same data from " << src_place << " to "
+              << dst_place;
+      return;
+    }
     auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
     auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
     memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr);

paddle/fluid/framework/tensor_util_test.cc

@@ -41,6 +41,11 @@ TEST(TensorCopy, Tensor) {
     EXPECT_EQ(src_ptr[i], dst_ptr[i]);
   }
 
+  TensorCopy(dst_tensor, *cpu_place, &dst_tensor);
+  for (size_t i = 0; i < 9; ++i) {
+    EXPECT_EQ(src_ptr[i], dst_ptr[i]);
+  }
+
   EXPECT_TRUE(dst_tensor.layout() == src_tensor.layout());
 
   Tensor slice_tensor = src_tensor.Slice(1, 2);
@@ -82,6 +87,15 @@ TEST(TensorCopy, Tensor) {
       EXPECT_EQ(src_ptr[i], dst_ptr[i]);
     }
 
+    // Copy the same tensor
+    TensorCopy(gpu_tensor, *gpu_place, gpu_ctx, &gpu_tensor);
+    gpu_ctx.Wait();
+    const int* dst_ptr_tmp = dst_tensor.data<int>();
+    EXPECT_NE(src_ptr, dst_ptr_tmp);
+    for (size_t i = 0; i < 9; ++i) {
+      EXPECT_EQ(src_ptr[i], dst_ptr_tmp[i]);
+    }
+
     Tensor slice_tensor = src_tensor.Slice(1, 2);
 
     // CPU Slice Tensor to GPU Tensor

paddle/fluid/framework/var_desc.h

@@ -59,6 +59,7 @@ class VarDesc {
  public:
   explicit VarDesc(const std::string &name) {
     desc_.set_name(name);
+    // TODO(paddle-dev): Why default to lodtensor.
     desc_.mutable_type()->set_type(proto::VarType::LOD_TENSOR);
   }
 

paddle/fluid/framework/variable.h

@@ -38,8 +38,12 @@ class Variable {
 
   template <typename T>
   T* GetMutable() {
-    if (!IsType<T>()) {
+    if (!holder_) {
       holder_.reset(new PlaceholderImpl<T>(new T()));
+    } else {
+      PADDLE_ENFORCE(IsType<T>(),
+                     "Variable must be type %s, the holding type is %s",
+                     typeid(T).name(), holder_->Type().name());
     }
     return static_cast<T*>(holder_->Ptr());
   }

paddle/fluid/framework/variable_test.cc

@@ -33,9 +33,10 @@ TEST(Variable, GetMutable) {
   const Tensor& tt = v->Get<Tensor>();
   EXPECT_EQ(1234, tt.content_);
 
-  std::string* s = v->GetMutable<std::string>();
-  *s = "hello";
-
-  const std::string& ss = v->Get<std::string>();
-  EXPECT_EQ("hello", ss);
+  try {
+    v->GetMutable<std::string>();
+  } catch (std::exception& e) {
+    return;
+  }
+  EXPECT_TRUE(false);
 }

paddle/fluid/inference/CMakeLists.txt

@@ -19,9 +19,19 @@ cc_library(paddle_fluid_origin DEPS ${fluid_modules} paddle_fluid_api)
 
 add_subdirectory(api)
 
+set(STATIC_INFERENCE_APIS paddle_fluid_api paddle_inference_api analysis_predictor)
+set(SHARED_INFERENCE_SRCS
+    io.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/api/analysis_predictor.cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/api/details/zero_copy_tensor.cc)
+if (WITH_GPU AND TENSORRT_FOUND)
+  set(STATIC_INFERENCE_APIS ${STATIC_INFERENCE_APIS} paddle_inference_tensorrt_subgraph_engine)
+  set(SHARED_INFERENCE_SRCS ${SHARED_INFERENCE_SRCS} ${CMAKE_CURRENT_SOURCE_DIR}/api/api_tensorrt_subgraph_engine.cc)
+endif()
+
 # Create static library
-cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api paddle_inference_api 
-           analysis_predictor zero_copy_tensor)
+cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor)
+
 if(NOT APPLE)
   # TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac.
   set(LINK_FLAGS "-Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.sym")
@@ -29,10 +39,7 @@ if(NOT APPLE)
 endif()
 
 # Create shared library
-cc_library(paddle_fluid_shared SHARED
-    SRCS io.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/api/analysis_predictor.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/api/details/zero_copy_tensor.cc
+cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS}
     DEPS ${fluid_modules} paddle_fluid_api)
 
 set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)

paddle/fluid/inference/analysis/analyzer.cc

@@ -70,7 +70,7 @@ class DfgPassManagerImpl final : public DfgPassManager {
       auto trt_teller = [&](const Node* node) {
         std::unordered_set<std::string> teller_set(
             {"mul", "conv2d", "pool2d", "relu", "softmax", "sigmoid",
-             "depthwise_conv2d", "batch_norm", "concat", "tanh",
+             "depthwise_conv2d", "batch_norm", "concat", "tanh", "pad",
              "elementwise_add", "dropout"});
         if (!node->IsFunction()) return false;
 

paddle/fluid/inference/analysis/analyzer.h

@@ -64,15 +64,17 @@ 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",                 //
+      "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",  //
 #ifdef PADDLE_WITH_MKLDNN
       "conv_relu_mkldnn_fuse_pass",  //
 #endif

paddle/fluid/inference/api/analysis_predictor.cc

@@ -25,9 +25,11 @@
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/api/paddle_inference_pass.h"
 #include "paddle/fluid/inference/utils/singleton.h"
+#include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 
 DECLARE_bool(profile);
+DECLARE_int32(paddle_num_threads);
 
 namespace paddle {
 
@@ -47,6 +49,9 @@ bool AnalysisPredictor::Init(
   }
 #endif
 
+  // no matter with or without MKLDNN
+  paddle::platform::SetNumThreads(FLAGS_paddle_num_threads);
+
   if (config_.use_gpu) {
     place_ = paddle::platform::CUDAPlace(config_.device);
     LOG(WARNING) << "ir optimize only supports CPU currently, enable_ir_optim "
@@ -335,6 +340,19 @@ bool AnalysisPredictor::LoadProgramDesc() {
   }
   return true;
 }
+
+AnalysisPredictor::~AnalysisPredictor() {
+#if !defined(_WIN32)
+  if (FLAGS_profile) {
+    platform::DisableProfiler(platform::EventSortingKey::kTotal,
+                              "./profile.log");
+  }
+#endif
+  if (sub_scope_) {
+    scope_->DeleteScope(sub_scope_);
+  }
+}
+
 std::unique_ptr<PaddlePredictor> AnalysisPredictor::Clone() {
   auto *x = new AnalysisPredictor(config_);
   x->Init(scope_, inference_program_);

paddle/fluid/inference/api/analysis_predictor.h

@@ -72,6 +72,7 @@ class AnalysisPredictor : public PaddlePredictor {
   template <typename T>
   void GetFetchOne(const framework::LoDTensor &fetchs,
                    PaddleTensor *output_data);
+  ~AnalysisPredictor();
 
  private:
   contrib::AnalysisConfig config_;

paddle/fluid/inference/api/api_impl.cc

@@ -23,9 +23,11 @@ limitations under the License. */
 #include "paddle/fluid/framework/feed_fetch_method.h"
 #include "paddle/fluid/inference/api/api_impl.h"
 #include "paddle/fluid/inference/api/helper.h"
+#include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 
 DEFINE_bool(profile, false, "Turn on profiler for fluid");
+DECLARE_int32(paddle_num_threads);
 
 namespace paddle {
 namespace {
@@ -72,6 +74,9 @@ bool NativePaddlePredictor::Init(
   }
 #endif
 
+  // no matter with or without MKLDNN
+  paddle::platform::SetNumThreads(FLAGS_paddle_num_threads);
+
   if (config_.use_gpu) {
     place_ = paddle::platform::CUDAPlace(config_.device);
   } else {

paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc

@@ -185,3 +185,4 @@ USE_TRT_CONVERTER(softmax);
 USE_TRT_CONVERTER(batch_norm);
 USE_TRT_CONVERTER(concat);
 USE_TRT_CONVERTER(dropout);
+USE_TRT_CONVERTER(pad);

paddle/fluid/inference/api/demo_ci/CMakeLists.txt

@@ -3,6 +3,7 @@ project(cpp_inference_demo CXX C)
 option(WITH_MKL        "Compile demo with MKL/OpenBlas support, default use MKL."       ON)
 option(WITH_GPU        "Compile demo with GPU/CPU, default use CPU."                    OFF)
 option(WITH_STATIC_LIB "Compile demo with static/shared library, default use static."   ON)
+option(USE_TENSORRT "Compile demo with TensorRT."   OFF)
 
 macro(safe_set_static_flag)
     foreach(flag_var
@@ -60,6 +61,13 @@ endif(NOT WIN32)
 include_directories("${PADDLE_LIB}/third_party/boost")
 include_directories("${PADDLE_LIB}/third_party/eigen3")
 
+if (NOT WIN32) 
+  if (USE_TENSORRT AND WITH_GPU) 
+      include_directories("${TENSORRT_INCLUDE_DIR}")
+      link_directories("${TENSORRT_LIB_DIR}")
+  endif()
+endif(NOT WIN32)
+
 if (NOT WIN32)
 link_directories("${PADDLE_LIB}/third_party/install/snappy/lib")
 link_directories("${PADDLE_LIB}/third_party/install/snappystream/lib")
@@ -112,6 +120,10 @@ endif(NOT WIN32)
 
 if(WITH_GPU)
   if(NOT WIN32)
+    if (USE_TENSORRT) 
+      set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer${CMAKE_STATIC_LIBRARY_SUFFIX})
+      set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer_plugin${CMAKE_STATIC_LIBRARY_SUFFIX})
+    endif()
     set(DEPS ${DEPS} ${CUDA_LIB}/libcudart${CMAKE_SHARED_LIBRARY_SUFFIX})
   else()
     set(DEPS ${DEPS} ${CUDA_LIB}/cudart${CMAKE_STATIC_LIBRARY_SUFFIX} )

paddle/fluid/inference/api/demo_ci/run.sh

@@ -3,6 +3,9 @@ PADDLE_ROOT=$1
 TURN_ON_MKL=$2 # use MKL or Openblas
 TEST_GPU_CPU=$3 # test both GPU/CPU mode or only CPU mode
 DATA_DIR=$4 # dataset
+TENSORRT_INCLUDE_DIR=$5 # TensorRT header file dir, defalut to /usr/local/TensorRT/include
+TENSORRT_LIB_DIR=$6 # TensorRT lib file dir, default to /usr/local/TensorRT/lib
+
 cd `dirname $0`
 current_dir=`pwd`
 if [ $2 == ON ]; then
@@ -16,6 +19,11 @@ else
   use_gpu_list='false'
 fi
 
+USE_TENSORRT=OFF
+if [ [-d"$TENSORRT_INCLUDE_DIR"] -a [-d"$TENSORRT_LIB_DIR"] ]; then
+  USE_TENSORRT=ON
+fi
+
 PREFIX=inference-vis-demos%2F
 URL_ROOT=http://paddlemodels.cdn.bcebos.com/${PREFIX}
 
@@ -86,5 +94,23 @@ for WITH_STATIC_LIB in ON OFF; do
       fi
     done
   done
+  
+  # --------tensorrt mobilenet------
+  if [ $USE_TENSORRT == ON -a $TEST_GPU_CPU == ON ]; then
+    rm -rf *
+    cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \
+      -DWITH_MKL=$TURN_ON_MKL \
+      -DDEMO_NAME=trt_mobilenet_demo \
+      -DWITH_GPU=$TEST_GPU_CPU \
+      -DWITH_STATIC_LIB=$WITH_STATIC_LIB \
+      -DUSE_TENSORRT=$USE_TENSORRT \
+      -DTENSORRT_INCLUDE_DIR=$TENSORRT_INCLUDE_DIR \
+      -DTENSORRT_LIB_DIR=$TENSORRT_LIB_DIR
+    make -j 
+    ./trt_mobilenet_demo \
+      --modeldir=$DATA_DIR/mobilenet/model \
+      --data=$DATA_DIR/mobilenet/data.txt \
+      --refer=$DATA_DIR/mobilenet/result.txt 
+  fi
 done
 set +x

paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.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. */
+
+/*
+ * This file contains demo of mobilenet for tensorrt.
+ */
+
+#include <gflags/gflags.h>
+#include <glog/logging.h>  // use glog instead of CHECK to avoid importing other paddle header files.
+#include "paddle/fluid/inference/demo_ci/utils.h"
+
+DECLARE_double(fraction_of_gpu_memory_to_use);
+DEFINE_string(modeldir, "", "Directory of the inference model.");
+DEFINE_string(refer, "", "path to reference result for comparison.");
+DEFINE_string(
+    data, "",
+    "path of data; each line is a record, format is "
+    "'<space splitted floats as data>\t<space splitted ints as shape'");
+
+namespace paddle {
+namespace demo {
+
+/*
+ * Use the tensorrt fluid engine to inference the demo.
+ */
+void Main() {
+  std::unique_ptr<PaddlePredictor> predictor;
+  paddle::contrib::MixedRTConfig config;
+  config.param_file = FLAGS_modeldir + "/__params__";
+  config.prog_file = FLAGS_modeldir + "/__model__";
+  config.use_gpu = true;
+  config.device = 0;
+  config.max_batch_size = 1;
+  config.fraction_of_gpu_memory = 0.1;  // set by yourself
+  predictor = CreatePaddlePredictor<paddle::contrib::MixedRTConfig>(config);
+
+  VLOG(3) << "begin to process data";
+  // Just a single batch of data.
+  std::string line;
+  std::ifstream file(FLAGS_data);
+  std::getline(file, line);
+  auto record = ProcessALine(line);
+  file.close();
+
+  // Inference.
+  PaddleTensor input;
+  input.shape = record.shape;
+  input.data =
+      PaddleBuf(record.data.data(), record.data.size() * sizeof(float));
+  input.dtype = PaddleDType::FLOAT32;
+
+  VLOG(3) << "run executor";
+  std::vector<PaddleTensor> output;
+  predictor->Run({input}, &output, 1);
+
+  VLOG(3) << "output.size " << output.size();
+  auto& tensor = output.front();
+  VLOG(3) << "output: " << SummaryTensor(tensor);
+
+  // compare with reference result
+  CheckOutput(FLAGS_refer, tensor);
+}
+
+}  // namespace demo
+}  // namespace paddle
+
+int main(int argc, char** argv) {
+  google::ParseCommandLineFlags(&argc, &argv, true);
+  paddle::demo::Main();
+  return 0;
+}

paddle/fluid/inference/api/demo_ci/utils.h

@@ -14,6 +14,8 @@
 
 #pragma once
 #include <algorithm>
+#include <fstream>
+#include <iostream>
 #include <string>
 #include <vector>
 #include "paddle/fluid/inference/paddle_inference_api.h"
@@ -21,6 +23,11 @@
 namespace paddle {
 namespace demo {
 
+struct Record {
+  std::vector<float> data;
+  std::vector<int32_t> shape;
+};
+
 static void split(const std::string& str, char sep,
                   std::vector<std::string>* pieces) {
   pieces->clear();
@@ -39,6 +46,58 @@ static void split(const std::string& str, char sep,
   }
 }
 
+Record ProcessALine(const std::string& line) {
+  VLOG(3) << "process a line";
+  std::vector<std::string> columns;
+  split(line, '\t', &columns);
+  CHECK_EQ(columns.size(), 2UL)
+      << "data format error, should be <data>\t<shape>";
+
+  Record record;
+  std::vector<std::string> data_strs;
+  split(columns[0], ' ', &data_strs);
+  for (auto& d : data_strs) {
+    record.data.push_back(std::stof(d));
+  }
+
+  std::vector<std::string> shape_strs;
+  split(columns[1], ' ', &shape_strs);
+  for (auto& s : shape_strs) {
+    record.shape.push_back(std::stoi(s));
+  }
+  VLOG(3) << "data size " << record.data.size();
+  VLOG(3) << "data shape size " << record.shape.size();
+  return record;
+}
+
+void CheckOutput(const std::string& referfile, const PaddleTensor& output) {
+  std::string line;
+  std::ifstream file(referfile);
+  std::getline(file, line);
+  auto refer = ProcessALine(line);
+  file.close();
+
+  size_t numel = output.data.length() / PaddleDtypeSize(output.dtype);
+  VLOG(3) << "predictor output numel " << numel;
+  VLOG(3) << "reference output numel " << refer.data.size();
+  CHECK_EQ(numel, refer.data.size());
+  switch (output.dtype) {
+    case PaddleDType::INT64: {
+      for (size_t i = 0; i < numel; ++i) {
+        CHECK_EQ(static_cast<int64_t*>(output.data.data())[i], refer.data[i]);
+      }
+      break;
+    }
+    case PaddleDType::FLOAT32:
+      for (size_t i = 0; i < numel; ++i) {
+        CHECK_LT(
+            fabs(static_cast<float*>(output.data.data())[i] - refer.data[i]),
+            1e-5);
+      }
+      break;
+  }
+}
+
 /*
  * Get a summary of a PaddleTensor content.
  */

paddle/fluid/inference/api/demo_ci/vis_demo.cc

@@ -18,10 +18,6 @@ limitations under the License. */
 
 #include <gflags/gflags.h>
 #include <glog/logging.h>  // use glog instead of CHECK to avoid importing other paddle header files.
-#include <fstream>
-#include <iostream>
-
-// #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/inference/demo_ci/utils.h"
 
 #ifdef PADDLE_WITH_CUDA
@@ -38,69 +34,11 @@ DEFINE_bool(use_gpu, false, "Whether use gpu.");
 namespace paddle {
 namespace demo {
 
-struct Record {
-  std::vector<float> data;
-  std::vector<int32_t> shape;
-};
-
-void split(const std::string& str, char sep, std::vector<std::string>* pieces);
-
-Record ProcessALine(const std::string& line) {
-  VLOG(3) << "process a line";
-  std::vector<std::string> columns;
-  split(line, '\t', &columns);
-  CHECK_EQ(columns.size(), 2UL)
-      << "data format error, should be <data>\t<shape>";
-
-  Record record;
-  std::vector<std::string> data_strs;
-  split(columns[0], ' ', &data_strs);
-  for (auto& d : data_strs) {
-    record.data.push_back(std::stof(d));
-  }
-
-  std::vector<std::string> shape_strs;
-  split(columns[1], ' ', &shape_strs);
-  for (auto& s : shape_strs) {
-    record.shape.push_back(std::stoi(s));
-  }
-  VLOG(3) << "data size " << record.data.size();
-  VLOG(3) << "data shape size " << record.shape.size();
-  return record;
-}
-
-void CheckOutput(const std::string& referfile, const PaddleTensor& output) {
-  std::string line;
-  std::ifstream file(referfile);
-  std::getline(file, line);
-  auto refer = ProcessALine(line);
-  file.close();
-
-  size_t numel = output.data.length() / PaddleDtypeSize(output.dtype);
-  VLOG(3) << "predictor output numel " << numel;
-  VLOG(3) << "reference output numel " << refer.data.size();
-  CHECK_EQ(numel, refer.data.size());
-  switch (output.dtype) {
-    case PaddleDType::INT64: {
-      for (size_t i = 0; i < numel; ++i) {
-        CHECK_EQ(static_cast<int64_t*>(output.data.data())[i], refer.data[i]);
-      }
-      break;
-    }
-    case PaddleDType::FLOAT32:
-      for (size_t i = 0; i < numel; ++i) {
-        CHECK_LT(
-            fabs(static_cast<float*>(output.data.data())[i] - refer.data[i]),
-            1e-5);
-      }
-      break;
-  }
-}
-
 /*
  * Use the native fluid engine to inference the demo.
  */
 void Main(bool use_gpu) {
+  std::unique_ptr<PaddlePredictor> predictor;
   NativeConfig config;
   config.param_file = FLAGS_modeldir + "/__params__";
   config.prog_file = FLAGS_modeldir + "/__model__";
@@ -111,7 +49,7 @@ void Main(bool use_gpu) {
   }
 
   VLOG(3) << "init predictor";
-  auto predictor =
+  predictor =
       CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
 
   VLOG(3) << "begin to process data";
@@ -131,7 +69,7 @@ void Main(bool use_gpu) {
 
   VLOG(3) << "run executor";
   std::vector<PaddleTensor> output;
-  predictor->Run({input}, &output);
+  predictor->Run({input}, &output, 1);
 
   VLOG(3) << "output.size " << output.size();
   auto& tensor = output.front();
@@ -146,9 +84,10 @@ void Main(bool use_gpu) {
 
 int main(int argc, char** argv) {
   google::ParseCommandLineFlags(&argc, &argv, true);
-  paddle::demo::Main(false /* use_gpu*/);
   if (FLAGS_use_gpu) {
     paddle::demo::Main(true /*use_gpu*/);
+  } else {
+    paddle::demo::Main(false /*use_gpu*/);
   }
   return 0;
 }

paddle/fluid/inference/tensorrt/convert/CMakeLists.txt

 # Add TRT tests
 nv_library(tensorrt_converter
   SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc
-batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc dropout_op.cc
+batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc dropout_op.cc pad_op.cc
   DEPS tensorrt_engine operator scope framework_proto op_registry)
 
 nv_test(test_op_converter SRCS test_op_converter.cc DEPS
@@ -26,6 +26,8 @@ nv_test(test_trt_batch_norm_op SRCS test_batch_norm_op.cc batch_norm_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine batch_norm_op SERIAL)
 nv_test(test_trt_concat_op SRCS test_concat_op.cc concat_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine concat_op SERIAL)
-
 nv_test(test_trt_dropout_op SRCS test_dropout_op.cc dropout_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine dropout_op SERIAL)
+
+nv_test(test_trt_pad_op SRCS test_pad_op.cc pad_op.cc
+        DEPS ${FLUID_CORE_MODULES} tensorrt_engine pad_op SERIAL)

paddle/fluid/inference/tensorrt/convert/pad_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/inference/tensorrt/convert/op_converter.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+/*
+ * PadOp.
+ */
+class PadOpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope, bool test_mode) override {
+    VLOG(4) << "convert a fluid transpose op to tensorrt tranpose layer";
+
+    framework::OpDesc op_desc(op, nullptr);
+    // Declare inputs
+    auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
+
+    const std::vector<int> paddings =
+        boost::get<std::vector<int>>(op_desc.GetAttr("paddings"));
+    const float pad_value = boost::get<float>(op_desc.GetAttr("pad_value"));
+
+    nvinfer1::Dims input_shape = input->getDimensions();
+    int nbDims = input_shape.nbDims;
+    int pad_size = static_cast<int>(paddings.size());
+    PADDLE_ENFORCE_GE(nbDims, 2);
+    PADDLE_ENFORCE_EQ((nbDims + 1) * 2, pad_size);
+    PADDLE_ENFORCE(pad_value == 0.0, "The pad layer of TRT only support zero.");
+
+    nvinfer1::DimsHW pre_pad(paddings[pad_size - 4], paddings[pad_size - 2]);
+    nvinfer1::DimsHW post_pad(paddings[pad_size - 3], paddings[pad_size - 1]);
+
+    auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Padding,
+                                       *const_cast<nvinfer1::ITensor*>(input),
+                                       pre_pad, post_pad);
+
+    PADDLE_ENFORCE(layer != nullptr);
+    auto output_name = op_desc.Output("Out")[0];
+    engine_->SetITensor(output_name, layer->getOutput(0));
+    layer->setName(("scale (Output: " + output_name + ")").c_str());
+    layer->getOutput(0)->setName(output_name.c_str());
+    if (test_mode) {  // the test framework can not determine which is the
+                      // output, so place the declaration inside.
+      engine_->DeclareOutput(output_name);
+    }
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+REGISTER_TRT_OP_CONVERTER(pad, PadOpConverter);

paddle/fluid/inference/tensorrt/convert/test_pad_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 <gtest/gtest.h>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+TEST(PadConverter, main) {
+  framework::Scope scope;
+  std::unordered_set<std::string> parameters;
+  TRTConvertValidation validator(10, parameters, scope, 1000);
+  validator.DeclInputVar("pad-X", nvinfer1::Dims3(3, 2, 2));
+  validator.DeclOutputVar("pad-Out", nvinfer1::Dims3(3, 3, 5));
+
+  // Prepare Op description
+  framework::OpDesc desc;
+  desc.SetType("pad");
+  desc.SetInput("X", {"pad-X"});
+  desc.SetOutput("Out", {"pad-Out"});
+
+  std::vector<int> paddings = {0, 0, 0, 0, 0, 1, 1, 2};
+  float pad_value = 0.0;
+  desc.SetAttr("paddings", paddings);
+  desc.SetAttr("pad_value", pad_value);
+
+  LOG(INFO) << "set OP";
+  validator.SetOp(*desc.Proto());
+  LOG(INFO) << "execute";
+
+  validator.Execute(2);
+}
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+USE_OP(pad);

paddle/fluid/operators/CMakeLists.txt

@@ -230,7 +230,7 @@ if(WITH_DISTRIBUTE)
         op_library(${dist_op} DEPS ${DISTRIBUTE_DEPS})
         set_source_files_properties(${dist_op}.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     endforeach()
-    
+
     #set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     #cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op
     #        listen_and_serv_op sum_op executor SERIAL)
@@ -268,6 +268,7 @@ if (WITH_GPU AND TENSORRT_FOUND)
 else()
     set(DEPS_OPS ${DEPS_OPS} tensorrt_engine_op)
 endif()
+op_library(clip_by_norm_op DEPS selected_rows_functor selected_rows)
 op_library(sum_op DEPS selected_rows_functor)
 op_library(sgd_op DEPS selected_rows_functor)
 op_library(print_op DEPS lod_tensor)
@@ -299,7 +300,7 @@ op_library(flatten_op DEPS reshape_op)
 op_library(sequence_pad_op DEPS sequence_padding)
 op_library(unstack_op DEPS stack_op)
 op_library(fake_quantize_op DEPS memory)
-op_library(fusion_lstm_op DEPS cpu_lstm_compute)
+op_library(fusion_lstm_op DEPS jit_kernel)
 if (WITH_GPU)
     op_library(conv_op DEPS vol2col depthwise_conv im2col)
     op_library(layer_norm_op DEPS cub)

paddle/fluid/operators/adadelta_op.cc

@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
+
 class AdadeltaOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -31,6 +32,16 @@ class AdadeltaOp : public framework::OperatorWithKernel {
                    "Input(AvgSquaredGrad) of AdadeltaOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("AvgSquaredUpdate"),
                    "Input(AvgSquaredUpdate) of AdadeltaOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->GetInputsVarType("Param").front() ==
+            framework::proto::VarType::LOD_TENSOR,
+        "The input var's type should be LoDTensor, but the received is %s",
+        ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front());
+    PADDLE_ENFORCE(
+        ctx->GetInputsVarType("Grad").front() ==
+            framework::proto::VarType::LOD_TENSOR,
+        "The input var's type should be LoDTensor, but the received is %s",
+        ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front());
 
     PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
                    "Output(ParamOut) of AdadeltaOp should not be null.");
@@ -56,6 +67,7 @@ class AdadeltaOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("AvgSquaredGradOut", param_dim);
     ctx->SetOutputDim("AvgSquaredUpdateOut", param_dim);
   }
+
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
     auto input_data_type =

paddle/fluid/operators/adadelta_op.h

@@ -23,6 +23,17 @@ template <typename DeviceContext, typename T>
 class AdadeltaOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    const auto* param_var = ctx.InputVar("Param");
+    PADDLE_ENFORCE(param_var->IsType<framework::LoDTensor>(),
+                   "The Var(%s)'s type should be LoDTensor, "
+                   "but the received is %s",
+                   ctx.Inputs("Param").front(), param_var->Type().name());
+    const auto* grad_var = ctx.InputVar("Grad");
+    PADDLE_ENFORCE(grad_var->IsType<framework::LoDTensor>(),
+                   "The Var(%s)'s type should be LoDTensor, "
+                   "but the received is %s",
+                   ctx.Inputs("Grad").front(), grad_var->Type().name());
+
     auto param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
     auto avg_squared_grad_out_tensor =
         ctx.Output<framework::Tensor>("AvgSquaredGradOut");

paddle/fluid/operators/adagrad_op.h

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 
@@ -21,25 +22,31 @@ namespace operators {
 
 template <typename DeviceContext, typename T>
 struct SparseAdagradFunctor {
-  void operator()(const DeviceContext& context,
-                  const framework::SelectedRows& grad,
-                  const framework::Tensor& learning_rate, T epsilon,
-                  framework::Tensor* moment, framework::Tensor* param);
+  void operator()(const DeviceContext &context,
+                  const framework::SelectedRows &grad,
+                  const framework::Tensor &learning_rate, T epsilon,
+                  framework::Tensor *moment, framework::Tensor *param);
 };
 
 template <typename DeviceContext, typename T>
 class AdagradOpKernel : public framework::OpKernel<T> {
  public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
-    auto* param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
-    auto* moment_out_tensor = ctx.Output<framework::Tensor>("MomentOut");
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    const auto *param_var = ctx.InputVar("Param");
+    PADDLE_ENFORCE(param_var->IsType<framework::LoDTensor>(),
+                   "The Var(%s)'s type should be LoDTensor, "
+                   "but the received is %s",
+                   ctx.Inputs("Param").front(), param_var->Type().name());
+
+    auto *param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
+    auto *moment_out_tensor = ctx.Output<framework::Tensor>("MomentOut");
 
     param_out_tensor->mutable_data<T>(ctx.GetPlace());
     moment_out_tensor->mutable_data<T>(ctx.GetPlace());
 
     T epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
 
-    auto* grad_var = ctx.InputVar("Grad");
+    auto *grad_var = ctx.InputVar("Grad");
     if (grad_var->IsType<framework::LoDTensor>()) {
       auto param = framework::EigenVector<T>::Flatten(
           *ctx.Input<framework::Tensor>("Param"));
@@ -47,16 +54,16 @@ class AdagradOpKernel : public framework::OpKernel<T> {
           *ctx.Input<framework::Tensor>("Grad"));
       auto moment = framework::EigenVector<T>::Flatten(
           *ctx.Input<framework::Tensor>("Moment"));
-      auto* learning_rate = ctx.Input<framework::Tensor>("LearningRate");
+      auto *learning_rate = ctx.Input<framework::Tensor>("LearningRate");
 
       auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
       auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
-      auto* place = ctx.template device_context<DeviceContext>().eigen_device();
+      auto *place = ctx.template device_context<DeviceContext>().eigen_device();
 
       moment_out.device(*place) = moment + grad * grad;
       Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
       if (platform::is_cpu_place(ctx.GetPlace())) {
-        auto* lr = learning_rate->data<T>();
+        auto *lr = learning_rate->data<T>();
         param_out.device(*place) =
             param - lr[0] * grad / (moment_out.sqrt() + epsilon);
       } else {
@@ -66,10 +73,10 @@ class AdagradOpKernel : public framework::OpKernel<T> {
             lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon);
       }
     } else if (grad_var->IsType<framework::SelectedRows>()) {
-      auto* param_tensor = ctx.Input<framework::Tensor>("Param");
+      auto *param_tensor = ctx.Input<framework::Tensor>("Param");
       PADDLE_ENFORCE_EQ(param_tensor, param_out_tensor);
 
-      auto* moment_tensor = ctx.Input<framework::Tensor>("Moment");
+      auto *moment_tensor = ctx.Input<framework::Tensor>("Moment");
       PADDLE_ENFORCE_EQ(moment_tensor, moment_out_tensor);
 
       SparseAdagradFunctor<DeviceContext, T> functor;

paddle/fluid/operators/adam_op.h

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
+#include "paddle/fluid/operators/math/algorithm.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
 #include "paddle/fluid/platform/for_range.h"
 
@@ -199,23 +200,9 @@ struct SparseAdamFunctor {
         row_numel_(row_numel),
         row_count_(row_count) {}
 
-  inline HOSTDEVICE int64_t BinarySearchInRows(int64_t row) const {
-    int64_t beg = 0, end = row_count_ - 1;
-    while (beg <= end) {
-      auto mid = ((beg + end) >> 1);
-      if (rows_[mid] == row)
-        return mid;
-      else if (rows_[mid] < row)
-        beg = mid + 1;
-      else
-        end = mid - 1;
-    }
-    return -1;
-  }
-
   inline HOSTDEVICE void operator()(size_t i) const {
-    int64_t row = i / row_numel_;
-    auto row_idx = BinarySearchInRows(row);
+    auto row_idx =
+        math::BinarySearch<int64_t>(rows_, row_count_, i / row_numel_);
     T g = row_idx >= 0 ? grad_[row_idx * row_numel_ + i % row_numel_] : 0;
 
     // The following code is the same as dense
@@ -244,6 +231,12 @@ template <typename DeviceContext, typename T>
 class AdamOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    const auto* param_var = ctx.InputVar("Param");
+    PADDLE_ENFORCE(param_var->IsType<framework::LoDTensor>(),
+                   "The Var(%s)'s type should be LoDTensor, "
+                   "but the received is %s",
+                   ctx.Inputs("Param").front(), param_var->Type().name());
+
     using paddle::framework::LoDTensor;
     using paddle::operators::detail::Ref;
 

paddle/fluid/operators/adamax_op.cc

@@ -35,6 +35,16 @@ class AdamaxOp : public framework::OperatorWithKernel {
                    "Input(LearningRate) of AdamaxOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Beta1Pow"),
                    "Input(Beta1Pow) of AdamaxOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->GetInputsVarType("Param").front() ==
+            framework::proto::VarType::LOD_TENSOR,
+        "The input var's type should be LoDTensor, but the received is %s",
+        ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front());
+    PADDLE_ENFORCE(
+        ctx->GetInputsVarType("Grad").front() ==
+            framework::proto::VarType::LOD_TENSOR,
+        "The input var's type should be LoDTensor, but the received is %s",
+        ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front());
 
     PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
                    "Output(ParamOut) of AdamaxOp should not be null.");

paddle/fluid/operators/adamax_op.h

@@ -23,6 +23,17 @@ template <typename DeviceContext, typename T>
 class AdamaxOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    const auto* param_var = ctx.InputVar("Param");
+    PADDLE_ENFORCE(param_var->IsType<framework::LoDTensor>(),
+                   "The Var(%s)'s type should be LoDTensor, "
+                   "but the received is %s",
+                   ctx.Inputs("Param").front(), param_var->Type().name());
+    const auto* grad_var = ctx.InputVar("Grad");
+    PADDLE_ENFORCE(grad_var->IsType<framework::LoDTensor>(),
+                   "The Var(%s)'s type should be LoDTensor, "
+                   "but the received is %s",
+                   ctx.Inputs("Grad").front(), grad_var->Type().name());
+
     auto param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
     auto moment_out_tensor = ctx.Output<framework::Tensor>("MomentOut");
     auto inf_norm_out_tensor = ctx.Output<framework::Tensor>("InfNormOut");

paddle/fluid/operators/clip_by_norm_op.h

@@ -16,12 +16,15 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/selected_rows.h"
+#include "paddle/fluid/operators/math/selected_rows_functor.h"
 #include "paddle/fluid/platform/transform.h"
 
 namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
+using SelectedRows = framework::SelectedRows;
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
@@ -31,9 +34,40 @@ class ClipByNormKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto max_norm = context.Attr<T>("max_norm");
-    auto* input = context.Input<Tensor>("X");
-    auto* output = context.Output<Tensor>("Out");
-    output->mutable_data<T>(context.GetPlace());
+    auto in_var = context.InputVar("X");
+
+    Tensor* output = nullptr;
+    const Tensor* input = nullptr;
+    if (in_var->IsType<framework::LoDTensor>()) {
+      input = context.Input<Tensor>("X");
+
+      output = context.Output<Tensor>("Out");
+      output->mutable_data<T>(context.GetPlace());
+    } else if (in_var->IsType<SelectedRows>()) {
+      auto* x = context.Input<SelectedRows>("X");
+
+      // merge ids in selected rows first
+      math::scatter::MergeAdd<DeviceContext, T> merge_func;
+      SelectedRows* merged_input =
+          const_cast<framework::Scope&>(context.scope())
+              .Var()
+              ->GetMutable<SelectedRows>();
+      merge_func(context.template device_context<DeviceContext>(), *x,
+                 merged_input);
+      input = &(merged_input->value());
+
+      SelectedRows* output_selected_rows = context.Output<SelectedRows>("Out");
+      output_selected_rows->set_rows(merged_input->rows());
+      output_selected_rows->set_height(merged_input->height());
+      output = output_selected_rows->mutable_value();
+      output->Resize(merged_input->value().dims());
+      output->mutable_data<T>(context.GetPlace());
+    } else {
+      PADDLE_THROW("Unexpected branch, input variable type is %s",
+                   in_var->Type().name());
+    }
+
+    PADDLE_ENFORCE_NOT_NULL(input);
 
     auto x = EigenVector<T>::Flatten(*input);
     auto out = EigenVector<T>::Flatten(*output);

paddle/fluid/operators/decayed_adagrad_op.cc

@@ -32,6 +32,16 @@ class DecayedAdagradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(
         ctx->HasInput("LearningRate"),
         "Input(LearningRate) of DecayedAdagradOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->GetInputsVarType("Param").front() ==
+            framework::proto::VarType::LOD_TENSOR,
+        "The input var's type should be LoDTensor, but the received is %s",
+        ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front());
+    PADDLE_ENFORCE(
+        ctx->GetInputsVarType("Grad").front() ==
+            framework::proto::VarType::LOD_TENSOR,
+        "The input var's type should be LoDTensor, but the received is %s",
+        ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front());
 
     PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
                    "Output(ParamOut) of DecayedAdagradOp should not be null.");

paddle/fluid/operators/decayed_adagrad_op.h

@@ -23,6 +23,17 @@ template <typename DeviceContext, typename T>
 class DecayedAdagradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    const auto* param_var = ctx.InputVar("Param");
+    PADDLE_ENFORCE(param_var->IsType<framework::LoDTensor>(),
+                   "The Var(%s)'s type should be LoDTensor, "
+                   "but the received is %s",
+                   ctx.Inputs("Param").front(), param_var->Type().name());
+    const auto* grad_var = ctx.InputVar("Grad");
+    PADDLE_ENFORCE(grad_var->IsType<framework::LoDTensor>(),
+                   "The Var(%s)'s type should be LoDTensor, "
+                   "but the received is %s",
+                   ctx.Inputs("Grad").front(), grad_var->Type().name());
+
     auto param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
     auto moment_out_tensor = ctx.Output<framework::Tensor>("MomentOut");
 

paddle/fluid/operators/fill_constant_op.cc

@@ -70,6 +70,12 @@ class FillConstantOp : public framework::OperatorBase {
   }
 };
 
+class FillConstantOpVarTypeInference : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc &op_desc,
+                  framework::BlockDesc *block) const override {}
+};
+
 class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
@@ -102,4 +108,5 @@ Fill up a variable with specified constant value.
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(fill_constant, ops::FillConstantOp,
                   ops::FillConstantInferShape, ops::FillConstantOpMaker,
-                  paddle::framework::EmptyGradOpMaker);
+                  paddle::framework::EmptyGradOpMaker,
+                  ops::FillConstantOpVarTypeInference);

paddle/fluid/operators/ftrl_op.cc

@@ -34,6 +34,16 @@ class FTRLOp : public framework::OperatorWithKernel {
                    "Input(Grad) of FTRL should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("LearningRate"),
                    "Input(LearningRate) of FTRL should not be null.");
+    PADDLE_ENFORCE(
+        ctx->GetInputsVarType("Param").front() ==
+            framework::proto::VarType::LOD_TENSOR,
+        "The input var's type should be LoDTensor, but the received is %s",
+        ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front());
+    PADDLE_ENFORCE(
+        ctx->GetInputsVarType("Grad").front() ==
+            framework::proto::VarType::LOD_TENSOR,
+        "The input var's type should be LoDTensor, but the received is %s",
+        ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front());
 
     PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
                    "Output(ParamOut) of FTRL should not be null.");

paddle/fluid/operators/ftrl_op.h

@@ -28,6 +28,17 @@ template <typename DeviceContext, typename T>
 class FTRLOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    const auto* param_var = ctx.InputVar("Param");
+    PADDLE_ENFORCE(param_var->IsType<framework::LoDTensor>(),
+                   "The Var(%s)'s type should be LoDTensor, "
+                   "but the received is %s",
+                   ctx.Inputs("Param").front(), param_var->Type().name());
+    const auto* grad_var = ctx.InputVar("Grad");
+    PADDLE_ENFORCE(grad_var->IsType<framework::LoDTensor>(),
+                   "The Var(%s)'s type should be LoDTensor, "
+                   "but the received is %s",
+                   ctx.Inputs("Grad").front(), grad_var->Type().name());
+
     auto* param_out = ctx.Output<Tensor>("ParamOut");
     auto* sq_accum_out = ctx.Output<Tensor>("SquaredAccumOut");
     auto* lin_accum_out = ctx.Output<Tensor>("LinearAccumOut");

paddle/fluid/operators/fused_embedding_fc_lstm_op.cc

@@ -93,11 +93,7 @@ void FusedEmbeddingFCLSTMOp::InferShape(
   ctx->SetOutputDim("Cell", out_dims);
   ctx->ShareLoD("Ids", "Hidden");
   ctx->ShareLoD("Ids", "Cell");
-  int xx_width;
-  if (ctx->Attrs().Get<bool>("use_seq")) {
-    xx_width = wh_dims[1];
-  } else {
-    xx_width = x_dims[1] > wh_dims[1] ? wh_dims[1] : x_dims[1];
+  if (!ctx->Attrs().Get<bool>("use_seq")) {
     PADDLE_ENFORCE(ctx->HasOutput("BatchedInput"),
                    "Assert only one Output(BatchedInput) of LSTM.");
     PADDLE_ENFORCE(ctx->HasOutput("BatchedHidden"),
@@ -112,7 +108,7 @@ void FusedEmbeddingFCLSTMOp::InferShape(
     ctx->SetOutputDim("BatchedHidden", out_dims);
     ctx->SetOutputDim("BatchedCell", out_dims);
   }
-  ctx->SetOutputDim("XX", {x_dims[0], xx_width});
+  ctx->SetOutputDim("XX", {x_dims[0], wh_dims[1]});
   ctx->ShareLoD("Ids", "XX");
 }
 
@@ -435,8 +431,6 @@ class FusedEmbeddingFCLSTMKernel : public framework::OpKernel<T> {
     INIT_VEC_FUNC
     INIT_BASE_INPUT_DATAS
 
-    // std::cout << "===> Batch Compute" << std::endl;
-
     auto* reordered_h0 = ctx.Output<Tensor>("ReorderedH0");
     auto* reordered_c0 = ctx.Output<Tensor>("ReorderedC0");
     auto* batched_input = ctx.Output<LoDTensor>("BatchedInput");

paddle/fluid/operators/fusion_lstm_op.cc

@@ -15,11 +15,9 @@ limitations under the License. */
 #include "paddle/fluid/operators/fusion_lstm_op.h"
 #include <string>
 #include "paddle/fluid/operators/math/blas.h"
-#include "paddle/fluid/operators/math/cpu_lstm_compute.h"
-#include "paddle/fluid/operators/math/cpu_vec.h"
 #include "paddle/fluid/operators/math/fc_compute.h"
+#include "paddle/fluid/operators/math/jit_kernel.h"
 #include "paddle/fluid/operators/math/sequence2batch.h"
-#include "paddle/fluid/platform/cpu_info.h"
 
 namespace paddle {
 namespace operators {
@@ -219,121 +217,55 @@ This operator fuse the X into LSTM, more details can refer to LSTM op.
 template <typename T>
 class FuisonLSTMKernel : public framework::OpKernel<T> {
  public:
-#define INIT_VEC_FUNC                                                          \
-  std::function<void(const int, const T *, T *)> act_gate, act_cell, act_cand; \
-  auto& act_gate_str = ctx.Attr<std::string>("gate_activation");               \
-  auto& act_cell_str = ctx.Attr<std::string>("cell_activation");               \
-  auto& act_cand_str = ctx.Attr<std::string>("candidate_activation");          \
-  if (platform::jit::MayIUse(platform::jit::avx)) {                            \
-    math::VecActivations<T, platform::jit::avx> act_functor;                   \
-    act_gate = act_functor(act_gate_str);                                      \
-    act_cell = act_functor(act_cell_str);                                      \
-    act_cand = act_functor(act_cand_str);                                      \
-  } else {                                                                     \
-    math::VecActivations<T, platform::jit::isa_any> act_functor;               \
-    act_gate = act_functor(act_gate_str);                                      \
-    act_cell = act_functor(act_cell_str);                                      \
-    act_cand = act_functor(act_cand_str);                                      \
-  }
-
-#define INIT_BASE_INPUT_OUTPUT                        \
-  auto* x = ctx.Input<LoDTensor>("X");                \
-  auto* h0 = ctx.Input<Tensor>("H0");                 \
-  auto* c0 = ctx.Input<Tensor>("C0");                 \
-  auto* wx = ctx.Input<Tensor>("WeightX");            \
-  auto* wh = ctx.Input<Tensor>("WeightH");            \
-  auto* bias = ctx.Input<Tensor>("Bias");             \
-  auto* xx = ctx.Output<LoDTensor>("XX");             \
-  auto* hidden_out = ctx.Output<LoDTensor>("Hidden"); \
-  auto* cell_out = ctx.Output<LoDTensor>("Cell");     \
-  bool is_reverse = ctx.Attr<bool>("is_reverse");     \
-  bool use_peepholes = ctx.Attr<bool>("use_peepholes");
-
-#define INIT_BASE_SIZES                  \
-  auto x_dims = x->dims();   /* T x M*/  \
-  auto wh_dims = wh->dims(); /* D x 4D*/ \
-  const int M = x_dims[1];               \
-  const int D = wh_dims[0];              \
-  const int D2 = D * 2;                  \
-  const int D3 = D * 3;                  \
-  const int D4 = wh_dims[1];
-
-#define INIT_BASE_INPUT_DATAS                                 \
-  const T* x_data = x->data<T>();                             \
-  const T* wx_data = wx->data<T>();                           \
-  const T* wh_data = wh->data<T>();                           \
-  /* diagonal weight*/                                        \
-  const T* wc_data = bias->data<T>() + D4;                    \
-  /* for peephole only*/                                      \
-  T* checked_cell_data = nullptr;                             \
-  auto place = ctx.GetPlace();                                \
-  if (use_peepholes) {                                        \
-    /* w_ic * Ct-1, w_fc * Ct-1  ; w_oc * Ct => ih*/          \
-    auto* checked_cell = ctx.Output<Tensor>("CheckedCell");   \
-    checked_cell_data = checked_cell->mutable_data<T>(place); \
-  }
-
-/// Compute LSTM
+#define INIT_BASE_DEFINES                                   \
+  using DeviceContext = paddle::platform::CPUDeviceContext; \
+  auto* x = ctx.Input<LoDTensor>("X");                      \
+  auto* h0 = ctx.Input<Tensor>("H0");                       \
+  auto* c0 = ctx.Input<Tensor>("C0");                       \
+  auto* wx = ctx.Input<Tensor>("WeightX");                  \
+  auto* wh = ctx.Input<Tensor>("WeightH");                  \
+  auto* bias = ctx.Input<Tensor>("Bias");                   \
+  auto* xx = ctx.Output<LoDTensor>("XX");                   \
+  auto* hidden_out = ctx.Output<LoDTensor>("Hidden");       \
+  auto* cell_out = ctx.Output<LoDTensor>("Cell");           \
+  bool is_reverse = ctx.Attr<bool>("is_reverse");           \
+  bool use_peepholes = ctx.Attr<bool>("use_peepholes");     \
+  auto x_dims = x->dims();   /* T x M*/                     \
+  auto wh_dims = wh->dims(); /* D x 4D*/                    \
+  const int M = x_dims[1];                                  \
+  const int D = wh_dims[0];                                 \
+  const int D4 = wh_dims[1]
+
+#define INIT_OTHER_DEFINES                                                  \
+  const T* x_data = x->data<T>();                                           \
+  const T* wx_data = wx->data<T>();                                         \
+  const T* wh_data = wh->data<T>();                                         \
+  /* diagonal weight*/                                                      \
+  const T* wp_data = bias->data<T>() + D4;                                  \
+  /* for peephole only*/                                                    \
+  T* checked_cell_data = nullptr;                                           \
+  auto place = ctx.GetPlace();                                              \
+  if (use_peepholes) {                                                      \
+    /* w_ic * Ct-1, w_fc * Ct-1  ; w_oc * Ct => ih*/                        \
+    auto* checked_cell = ctx.Output<Tensor>("CheckedCell");                 \
+    checked_cell_data = checked_cell->mutable_data<T>(place);               \
+  }                                                                         \
+  const auto& ker =                                                         \
+      math::jitkernel::KernelPool::Instance()                               \
+          .template Get<math::jitkernel::LSTMKernel<T>, const std::string&, \
+                        const std::string&, const std::string&>(            \
+              ctx.Attr<std::string>("gate_activation"),                     \
+              ctx.Attr<std::string>("candidate_activation"),                \
+              ctx.Attr<std::string>("cell_activation"), D, use_peepholes)
+
+// Wh GEMM
 #define GEMM_WH_ADDON(bs, prev, out)                                           \
   blas.GEMM(CblasNoTrans, CblasNoTrans, bs, D4, D, static_cast<T>(1), prev, D, \
             wh_data, D4, static_cast<T>(1), out, D4)
 
-#define GET_Ct(ct_1, gates, ct)                   \
-  /* C_t = C_t-1 * fgated + cand_gated * igated*/ \
-  act_cand(D, gates, gates);                      \
-  blas.VMUL(D, gates, gates + D, gates + D);      \
-  blas.VMUL(D, ct_1, gates + D2, gates + D2);     \
-  blas.VADD(D, gates + D, gates + D2, ct)
-
-#define GET_Ht(ct, gates, ht)        \
-  /* H_t = act_cell(C_t) * ogated */ \
-  act_cell(D, ct, gates + D2);       \
-  blas.VMUL(D, gates + D2, gates + D3, ht)
-
-#define GET_Ct_NOH0C0(gates, ct)     \
-  /* C_t = igated * cgated*/         \
-  act_gate(D, gates + D, gates + D); \
-  act_cand(D, gates, gates);         \
-  blas.VMUL(D, gates, gates + D, ct)
-
-#define COMPUTE_CtHt_NOH0C0(gates, ct, ht) \
-  GET_Ct_NOH0C0(gates, ct);                \
-  act_gate(D, gates + D3, gates + D3);     \
-  GET_Ht(ct, gates, ht)
-
-#define COMPUTE_CtHt_PEEPHOLE_NOH0C0(gates, ct, ht) \
-  GET_Ct_NOH0C0(gates, ct);                         \
-  /* get outgated, put W_oc * C_t on igated */      \
-  blas.VMUL(D, wc_data + D2, ct, gates + D);        \
-  blas.VADD(D, gates + D, gates + D3, gates + D3);  \
-  act_gate(D, gates + D3, gates + D3);              \
-  GET_Ht(ct, gates, ht)
-
-#define COMPUTE_CtHt(gates, ct_1, ct, ht) \
-  act_gate(D3, gates + D, gates + D);     \
-  GET_Ct(ct_1, gates, ct);                \
-  GET_Ht(ct, gates, ht)
-
-#define COMPUTE_CtHt_PEEPHOLE(gates, ct_1, ct, ht)        \
-  /* get fgated and igated*/                              \
-  blas.VMUL(D, wc_data, ct_1, checked_cell_data);         \
-  blas.VMUL(D, wc_data + D, ct_1, checked_cell_data + D); \
-  blas.VADD(D2, checked_cell_data, gates + D, gates + D); \
-  act_gate(D2, gates + D, gates + D);                     \
-  GET_Ct(ct_1, gates, ct);                                \
-  /* get ogated*/                                         \
-  blas.VMUL(D, wc_data + D2, ct, gates + D);              \
-  blas.VADD(D, gates + D, gates + D3, gates + D3);        \
-  act_gate(D, gates + D3, gates + D3);                    \
-  GET_Ht(ct, gates, ht)
-
   void SeqCompute(const framework::ExecutionContext& ctx) const {
-    using DeviceContext = paddle::platform::CPUDeviceContext;
-    INIT_BASE_INPUT_OUTPUT
-    INIT_BASE_SIZES
-    INIT_VEC_FUNC
-    INIT_BASE_INPUT_DATAS
-
+    INIT_BASE_DEFINES;
+    INIT_OTHER_DEFINES;
     auto x_lod = x->lod();
     const int total_T = x_dims[0];
     const int N = x_lod[0].size() - 1;
@@ -357,89 +289,47 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
       gate_offset = -D;
     }
 
-#define MOVE_ONE_STEP                    \
-  prev_h_data = h_out_data;              \
-  prev_c_data = c_out_data;              \
-  xx_data = xx_data + xx_offset;         \
-  h_out_data = h_out_data + gate_offset; \
-  c_out_data = c_out_data + gate_offset
-
-#define PROCESS_H0C0_DEFINES                       \
-  int bid = is_reverse ? N - 1 - i : i;            \
-  int seq_len = x_lod[0][bid + 1] - x_lod[0][bid]; \
-  const T* prev_c_data = nullptr;                  \
-  const T* prev_h_data = nullptr;                  \
-  int tstart = 0
-
-#define PROCESS_H0C0_PEEPHOLE                                      \
-  PROCESS_H0C0_DEFINES;                                            \
-  if (h0_data) {                                                   \
-    prev_h_data = h0_data + bid * D;                               \
-    prev_c_data = c0_data + bid * D;                               \
-  } else {                                                         \
-    COMPUTE_CtHt_PEEPHOLE_NOH0C0(xx_data, c_out_data, h_out_data); \
-    MOVE_ONE_STEP;                                                 \
-    tstart = 1;                                                    \
-  }
-
-#define PROCESS_H0C0                                      \
-  PROCESS_H0C0_DEFINES;                                   \
-  if (h0_data) {                                          \
-    prev_h_data = h0_data + bid * D;                      \
-    prev_c_data = c0_data + bid * D;                      \
-  } else {                                                \
-    COMPUTE_CtHt_NOH0C0(xx_data, c_out_data, h_out_data); \
-    MOVE_ONE_STEP;                                        \
-    tstart = 1;                                           \
-  }
-
-    if (use_peepholes) {
-      for (int i = 0; i < N; ++i) {
-        PROCESS_H0C0_PEEPHOLE
-        for (int step = tstart; step < seq_len; ++step) {
-          GEMM_WH_ADDON(1, prev_h_data, xx_data);
-          COMPUTE_CtHt_PEEPHOLE(xx_data, prev_c_data, c_out_data, h_out_data);
-          MOVE_ONE_STEP;
-        }
-      }
-    } else {
-      // TODO(TJ): unly workaround, clean me
-      std::function<void(T*, const T*, T*, T*)> compute_ctht;
-      if (platform::jit::MayIUse(platform::jit::avx) &&
-          act_gate_str == "sigmoid" && act_cand_str == "tanh" &&
-          act_cell_str == "tanh" && D == 8) {
-        compute_ctht = math::lstm_compute_ctht<T>;
+    for (int i = 0; i < N; ++i) {
+      int bid = is_reverse ? N - 1 - i : i;
+      int seq_len = x_lod[0][bid + 1] - x_lod[0][bid];
+      const T* prev_c_data = nullptr;
+      const T* prev_h_data = nullptr;
+      int tstart = 0;
+      if (h0_data) {
+        prev_h_data = h0_data + bid * D;
+        prev_c_data = c0_data + bid * D;
       } else {
-        compute_ctht = [&](T* gates, const T* ct_1, T* ct, T* ht) {
-          COMPUTE_CtHt(gates, ct_1, ct, ht);
-        };
+        ker->ComputeC1H1(xx_data, c_out_data, h_out_data, wp_data);
+        tstart = 1;
+        // move one step
+        prev_h_data = h_out_data;
+        prev_c_data = c_out_data;
+        xx_data = xx_data + xx_offset;
+        h_out_data = h_out_data + gate_offset;
+        c_out_data = c_out_data + gate_offset;
       }
-      for (int i = 0; i < N; ++i) {
-        PROCESS_H0C0
-        for (int step = tstart; step < seq_len; ++step) {
-          GEMM_WH_ADDON(1, prev_h_data, xx_data);
-          compute_ctht(xx_data, prev_c_data, c_out_data, h_out_data);
-          MOVE_ONE_STEP;
-        }
+      for (int step = tstart; step < seq_len; ++step) {
+        GEMM_WH_ADDON(1, prev_h_data, xx_data);
+        ker->ComputeCtHt(xx_data, prev_c_data, c_out_data, h_out_data, wp_data,
+                         checked_cell_data);
+        // move one step
+        prev_h_data = h_out_data;
+        prev_c_data = c_out_data;
+        xx_data = xx_data + xx_offset;
+        h_out_data = h_out_data + gate_offset;
+        c_out_data = c_out_data + gate_offset;
       }
     }
-#undef PROCESS_H0C0_DEFINES
-#undef PROCESS_H0C0_PEEPHOLE
-#undef PROCESS_H0C0
-#undef MOVE_ONE_STEP
   }
 
   void BatchCompute(const framework::ExecutionContext& ctx) const {
-    using DeviceContext = platform::CPUDeviceContext;
-    INIT_BASE_INPUT_OUTPUT
-    INIT_BASE_SIZES
+    INIT_BASE_DEFINES;
     if (x->lod()[0].size() == 2) {
       xx->Resize({x_dims[0], D4});
       SeqCompute(ctx);
       return;
     }
-    INIT_VEC_FUNC
-    INIT_BASE_INPUT_DATAS
+    INIT_OTHER_DEFINES;
 
     auto* reordered_h0 = ctx.Output<Tensor>("ReorderedH0");
     auto* reordered_c0 = ctx.Output<Tensor>("ReorderedC0");
@@ -487,8 +377,8 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
       prev_c_data = reordered_c0_data;
       size_t sz = sizeof(T) * D;
       for (int i = 0; i < max_bs; ++i) {
-        std::memcpy(reordered_h0_data, h0_data + seq_order[i] * D, sz);
-        std::memcpy(reordered_c0_data, c0_data + seq_order[i] * D, sz);
+        blas.VCOPY(sz, h0_data + seq_order[i] * D, reordered_h0_data);
+        blas.VCOPY(sz, c0_data + seq_order[i] * D, reordered_c0_data);
         reordered_h0_data += D;
         reordered_c0_data += D;
       }
@@ -498,13 +388,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
       T* cur_h_out_data = batched_h_out_data;
       T* cur_c_out_data = batched_c_out_data;
       for (int i = 0; i < max_bs; ++i) {
-        GET_Ct_NOH0C0(cur_in_data, cur_c_out_data);
-        if (use_peepholes) {
-          blas.VMUL(D, wc_data + D2, cur_c_out_data, cur_in_data + D);
-          blas.VADD(D, cur_in_data + D, cur_in_data + D3, cur_in_data + D3);
-        }
-        act_gate(D, cur_in_data + D3, cur_in_data + D3);
-        GET_Ht(cur_c_out_data, cur_in_data, cur_h_out_data);
+        ker->ComputeC1H1(cur_in_data, cur_c_out_data, cur_h_out_data, wp_data);
         cur_in_data += D4;
         cur_c_out_data += D;
         cur_h_out_data += D;
@@ -513,71 +397,37 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
       prev_h_data = batched_h_out_data;
       prev_c_data = batched_c_out_data;
     }
+
+    // compute kernel part
     const auto& batch_starts = batched_lod[0];
     const int max_seq_len = batch_starts.size() - 1;
     const int offset = tstart * max_bs * D;
     batched_input_data = batched_input_data + offset * 4;
     batched_h_out_data = batched_h_out_data + offset;
     batched_c_out_data = batched_c_out_data + offset;
-
-#define DEFINE_CUR                        \
-  T* cur_in_data = batched_input_data;    \
-  T* cur_prev_c_data = prev_c_data;       \
-  T* cur_c_out_data = batched_c_out_data; \
-  T* cur_h_out_data = batched_h_out_data
-
-#define MOVE_ONE_BATCH  \
-  cur_in_data += D4;    \
-  cur_prev_c_data += D; \
-  cur_c_out_data += D;  \
-  cur_h_out_data += D
-
-#define MOVE_ONE_STEP                  \
-  prev_c_data = batched_c_out_data;    \
-  prev_h_data = batched_h_out_data;    \
-  batched_c_out_data = cur_c_out_data; \
-  batched_h_out_data = cur_h_out_data; \
-  batched_input_data = cur_in_data
-
-    if (use_peepholes) {
-      for (int step = tstart; step < max_seq_len; ++step) {
-        const int cur_bs = batch_starts[step + 1] - batch_starts[step];
-        GEMM_WH_ADDON(cur_bs, prev_h_data, batched_input_data);
-        DEFINE_CUR;
-        for (int i = 0; i < cur_bs; ++i) {
-          COMPUTE_CtHt_PEEPHOLE(cur_in_data, cur_prev_c_data, cur_c_out_data,
-                                cur_h_out_data);
-          MOVE_ONE_BATCH;
-        }
-        MOVE_ONE_STEP;
-      }
-    } else {
-      // TODO(TJ): unly workaround, clean me
-      std::function<void(T*, const T*, T*, T*)> compute_ctht;
-      if (platform::jit::MayIUse(platform::jit::avx) &&
-          act_gate_str == "sigmoid" && act_cand_str == "tanh" &&
-          act_cell_str == "tanh" && D == 8) {
-        compute_ctht = math::lstm_compute_ctht<T>;
-      } else {
-        compute_ctht = [&](T* gates, const T* ct_1, T* ct, T* ht) {
-          COMPUTE_CtHt(gates, ct_1, ct, ht);
-        };
-      }
-      for (int step = tstart; step < max_seq_len; ++step) {
-        const int cur_bs = batch_starts[step + 1] - batch_starts[step];
-        GEMM_WH_ADDON(cur_bs, prev_h_data, batched_input_data);
-        DEFINE_CUR;
-        for (int i = 0; i < cur_bs; ++i) {
-          compute_ctht(cur_in_data, cur_prev_c_data, cur_c_out_data,
-                       cur_h_out_data);
-          MOVE_ONE_BATCH;
-        }
-        MOVE_ONE_STEP;
+    for (int step = tstart; step < max_seq_len; ++step) {
+      const int cur_bs = batch_starts[step + 1] - batch_starts[step];
+      GEMM_WH_ADDON(cur_bs, prev_h_data, batched_input_data);
+      T* cur_in_data = batched_input_data;
+      T* cur_prev_c_data = prev_c_data;
+      T* cur_c_out_data = batched_c_out_data;
+      T* cur_h_out_data = batched_h_out_data;
+      for (int i = 0; i < cur_bs; ++i) {
+        ker->ComputeCtHt(cur_in_data, cur_prev_c_data, cur_c_out_data,
+                         cur_h_out_data, wp_data, checked_cell_data);
+        // move one batch
+        cur_in_data += D4;
+        cur_prev_c_data += D;
+        cur_c_out_data += D;
+        cur_h_out_data += D;
       }
+      // move one step
+      prev_c_data = batched_c_out_data;
+      prev_h_data = batched_h_out_data;
+      batched_c_out_data = cur_c_out_data;
+      batched_h_out_data = cur_h_out_data;
+      batched_input_data = cur_in_data;
     }
-#undef MOVE_ONE_STEP
-#undef MOVE_ONE_BATCH
-#undef DEFINE_CUR
 
     math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
     batched_h_out->set_lod(batched_lod);
@@ -594,18 +444,9 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
     }
   }
 
-#undef COMPUTE_CtHt_PEEPHOLE
-#undef COMPUTE_CtHt
-#undef GET_Ct_NOH0C0
-#undef COMPUTE_CtHt_NOH0C0
-#undef COMPUTE_CtHt_PEEPHOLE_NOH0C0
-#undef GET_Ht
-#undef GET_Ct
 #undef GEMM_WH_ADDON
-#undef INIT_BASE_INPUT_DATAS
-#undef INIT_BASE_SIZES
-#undef INIT_BASE_INPUT_OUTPUT
-#undef INIT_VEC_FUNC
+#undef INIT_OTHER_DEFINES
+#undef INIT_BASE_DEFINES
 };
 
 }  // namespace operators

paddle/fluid/operators/isfinite_op.cc

@@ -60,7 +60,7 @@ class OverflowOpMaker : public framework::OpProtoAndCheckerMaker {
               "(Tensor) 1-dim tensor, contains a bool scalar. The output "
               "tensor of overflow operator.");
     AddComment(string::Sprintf(R"DOC(
-Overflow operator.
+Overflow %s operator.
 
 $$Out = any(X)$$
 
@@ -69,6 +69,8 @@ Out = Inf if any X contains Inf,
 Out = Nan if any X contains Nan,
 Out = 0 if no Inf/Nan detected.
 If X contains both Inf/Nan, it will return the first indicator it meeted.
+
+%s
 )DOC",
                                GetName(), GetComments()));
   }

paddle/fluid/operators/math/CMakeLists.txt

@@ -3,8 +3,8 @@ add_subdirectory(detail)
 endif(NOT WIN32)
 
 function(math_library TARGET)
-    # math_library is a function to create math library. 
-    # The interface is the same as cc_library. 
+    # math_library is a function to create math library.
+    # The interface is the same as cc_library.
     # But it handle split GPU/CPU code and link some common library.
     set(cc_srcs)
     set(cu_srcs)
@@ -45,15 +45,13 @@ 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)
-# TODO(TJ): ugly workaround, clean me
-cc_library(cpu_lstm_compute SRCS cpu_lstm_compute.cc DEPS activation_functions cblas cpu_info)
 endif (NOT WIN32)
 
 cc_library(blas SRCS blas.cc DEPS cblas framework_proto device_context)
 math_library(math_function DEPS blas)
 math_library(maxouting)
 math_library(pooling)
-math_library(selected_rows_functor DEPS selected_rows math_function)
+math_library(selected_rows_functor DEPS selected_rows math_function blas)
 math_library(sequence2batch)
 math_library(sequence_padding)
 math_library(sequence_pooling DEPS math_function)
@@ -76,3 +74,7 @@ if(WITH_GPU)
 endif()
 cc_test(concat_test SRCS concat_test.cc DEPS concat)
 cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info)
+cc_library(jit_kernel 
+    SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_lstm.cc
+    DEPS cpu_info cblas activation_functions)
+cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel)

paddle/fluid/operators/math/algorithm.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <algorithm>
+#include <cstdint>  // for int64_t
+#include <numeric>
+
+#include "paddle/fluid/platform/hostdevice.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+template <typename T>
+HOSTDEVICE inline int64_t BinarySearch(const T *x, int64_t num, const T &val) {
+  int64_t beg = 0, end = num - 1;
+  while (beg <= end) {
+    auto mid = ((beg + end) >> 1);
+    if (x[mid] == val)
+      return mid;
+    else if (x[mid] < val)
+      beg = mid + 1;
+    else
+      end = mid - 1;
+  }
+  return -1;
+}
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/cpu_lstm_compute.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/operators/math/cpu_vec.h"
-#include "paddle/fluid/platform/cpu_info.h"
-#ifdef __AVX__
-#include <immintrin.h>
-#endif
-
-namespace paddle {
-namespace operators {
-namespace math {
-
-// TODO(TJ): ugly workaround, clean me
-template <typename T>
-void lstm_compute_ctht(T* gates, const T* ct_1, T* ct, T* ht) {
-  // gates: W_ch, W_ih, W_fh, W_oh
-  vec_sigmoid<T, platform::jit::avx>(24, gates + 8, gates + 8);
-  vec_tanh<T, platform::jit::avx>(8, gates, gates);
-  const T *i = gates + 8, *f = gates + 16, *o = gates + 24;
-  const T min = SIGMOID_THRESHOLD_MIN;
-  const T max = SIGMOID_THRESHOLD_MAX;
-  for (int d = 0; d < 8; ++d) {
-    // C_t = C_t-1 * fgated + cand_gated * igated
-    ct[d] = ct_1[d] * f[d] + gates[d] * i[d];
-    // H_t = act_cell(C_t) * ogated
-    T tmp = ct[d] * 2;
-    tmp = static_cast<T>(0) - ((tmp < min) ? min : ((tmp > max) ? max : tmp));
-    vec_exp<T>(1, &tmp, &tmp);
-    tmp = static_cast<T>(2) / (static_cast<T>(1) + tmp) - static_cast<T>(1);
-    ht[d] = tmp * o[d];
-  }
-}
-
-#ifdef __AVX__
-namespace detail {
-namespace forward {
-namespace avx {
-__m256 Sigmoid(const __m256 a);
-__m256 Tanh(const __m256 a);
-
-}  // namespace avx
-}  // namespace forward
-}  // namespace detail
-
-template <>
-void lstm_compute_ctht<float>(float* gates, const float* ct_1, float* ct,
-                              float* ht);
-
-#endif
-
-}  // namespace math
-}  // namespace operators
-}  // namespace paddle

paddle/fluid/operators/math/cpu_vec.h

@@ -125,10 +125,8 @@ inline void vec_scal<float, platform::jit::avx2>(const int n, const float a,
 }
 
 template <>
-inline void vec_scal<float, platform::jit::avx512_common>(const int n,
-                                                          const float a,
-                                                          const float* x,
-                                                          float* y) {
+inline void vec_scal<float, platform::jit::avx512f>(const int n, const float a,
+                                                    const float* x, float* y) {
   // TODO(TJ): enable me
   vec_scal<float, platform::jit::avx2>(n, a, x, y);
 }
@@ -181,10 +179,10 @@ inline void vec_bias_sub<float, platform::jit::avx2>(const int n, const float a,
 }
 
 template <>
-inline void vec_bias_sub<float, platform::jit::avx512_common>(const int n,
-                                                              const float a,
-                                                              const float* x,
-                                                              float* y) {
+inline void vec_bias_sub<float, platform::jit::avx512f>(const int n,
+                                                        const float a,
+                                                        const float* x,
+                                                        float* y) {
   // TODO(TJ): enable me
   vec_bias_sub<float, platform::jit::avx2>(n, a, x, y);
 }
@@ -242,7 +240,7 @@ inline void vec_cross<float, platform::jit::avx2>(const int n, const float* x,
 }
 
 template <>
-inline void vec_cross<float, platform::jit::avx512_common>(
+inline void vec_cross<float, platform::jit::avx512f>(
     const int n, const float* x, const float* y, const float* z, float* out) {
   // TODO(TJ): enable me
   vec_cross<float, platform::jit::avx>(n, x, y, z, out);
@@ -296,10 +294,10 @@ inline void vec_add_bias<float, platform::jit::avx2>(const int n, const float a,
 }
 
 template <>
-inline void vec_add_bias<float, platform::jit::avx512_common>(const int n,
-                                                              const float a,
-                                                              const float* x,
-                                                              float* y) {
+inline void vec_add_bias<float, platform::jit::avx512f>(const int n,
+                                                        const float a,
+                                                        const float* x,
+                                                        float* y) {
   // TODO(TJ): enable me
   vec_add_bias<float, platform::jit::avx2>(n, a, x, y);
 }
@@ -390,9 +388,9 @@ inline void vec_sigmoid<float, platform::jit::avx2>(const int n, const float* x,
 }
 
 template <>
-inline void vec_sigmoid<float, platform::jit::avx512_common>(const int n,
-                                                             const float* x,
-                                                             float* y) {
+inline void vec_sigmoid<float, platform::jit::avx512f>(const int n,
+                                                       const float* x,
+                                                       float* y) {
   // TODO(TJ): enable me
   vec_sigmoid<float, platform::jit::avx2>(n, x, y);
 }
@@ -454,9 +452,8 @@ inline void vec_relu<float, platform::jit::avx2>(const int n, const float* x,
 }
 
 template <>
-inline void vec_relu<float, platform::jit::avx512_common>(const int n,
-                                                          const float* x,
-                                                          float* y) {
+inline void vec_relu<float, platform::jit::avx512f>(const int n, const float* x,
+                                                    float* y) {
   // TODO(TJ): enable me
   vec_relu<float, platform::jit::avx2>(n, x, y);
 }

paddle/fluid/operators/math/cpu_vec_test.cc

@@ -110,7 +110,7 @@ TEST(CpuVecTest, sigmoid) {
     TestAndBench<float>(sz, vec_sigmoid<float>, ref_sigmoid<float>);
     TestAndBench<float>(sz, vec_sigmoid<float, jit::avx>, ref_sigmoid<float>);
     TestAndBench<float>(sz, vec_sigmoid<float, jit::avx2>, ref_sigmoid<float>);
-    TestAndBench<float>(sz, vec_sigmoid<float, jit::avx512_common>,
+    TestAndBench<float>(sz, vec_sigmoid<float, jit::avx512f>,
                         ref_sigmoid<float>);
   }
   TestAndBench<double>(30, vec_sigmoid<double>, ref_sigmoid<double>);
@@ -123,8 +123,7 @@ TEST(CpuVecTest, tanh) {
     TestAndBench<float>(sz, vec_tanh<float>, ref_tanh<float>);
     TestAndBench<float>(sz, vec_tanh<float, jit::avx>, ref_tanh<float>);
     TestAndBench<float>(sz, vec_tanh<float, jit::avx2>, ref_tanh<float>);
-    TestAndBench<float>(sz, vec_tanh<float, jit::avx512_common>,
-                        ref_tanh<float>);
+    TestAndBench<float>(sz, vec_tanh<float, jit::avx512f>, ref_tanh<float>);
   }
   TestAndBench<double>(30, vec_tanh<double>, ref_tanh<double>);
 }
@@ -136,8 +135,7 @@ TEST(CpuVecTest, relu) {
     TestAndBench<float>(sz, vec_relu<float>, ref_relu<float>);
     TestAndBench<float>(sz, vec_relu<float, jit::avx>, ref_relu<float>);
     TestAndBench<float>(sz, vec_relu<float, jit::avx2>, ref_relu<float>);
-    TestAndBench<float>(sz, vec_relu<float, jit::avx512_common>,
-                        ref_relu<float>);
+    TestAndBench<float>(sz, vec_relu<float, jit::avx512f>, ref_relu<float>);
   }
   TestAndBench<double>(30, vec_relu<double>, ref_relu<double>);
 }
@@ -170,7 +168,7 @@ TEST(CpuVecTest, inplace_sigmoid) {
     TestInplace<float>(sz, vec_sigmoid<float>, ref_sigmoid<float>);
     TestInplace<float>(sz, vec_sigmoid<float, jit::avx>, ref_sigmoid<float>);
     TestInplace<float>(sz, vec_sigmoid<float, jit::avx2>, ref_sigmoid<float>);
-    TestInplace<float>(sz, vec_sigmoid<float, jit::avx512_common>,
+    TestInplace<float>(sz, vec_sigmoid<float, jit::avx512f>,
                        ref_sigmoid<float>);
   }
   TestInplace<double>(30, vec_sigmoid<double>, ref_sigmoid<double>);
@@ -183,8 +181,7 @@ TEST(CpuVecTest, inplace_tanh) {
     TestInplace<float>(sz, vec_tanh<float>, ref_tanh<float>);
     TestInplace<float>(sz, vec_tanh<float, jit::avx>, ref_tanh<float>);
     TestInplace<float>(sz, vec_tanh<float, jit::avx2>, ref_tanh<float>);
-    TestInplace<float>(sz, vec_tanh<float, jit::avx512_common>,
-                       ref_tanh<float>);
+    TestInplace<float>(sz, vec_tanh<float, jit::avx512f>, ref_tanh<float>);
   }
   TestInplace<double>(30, vec_tanh<double>, ref_tanh<double>);
 }
@@ -196,8 +193,7 @@ TEST(CpuVecTest, inplace_relu) {
     TestInplace<float>(sz, vec_relu<float>, ref_relu<float>);
     TestInplace<float>(sz, vec_relu<float, jit::avx>, ref_relu<float>);
     TestInplace<float>(sz, vec_relu<float, jit::avx2>, ref_relu<float>);
-    TestInplace<float>(sz, vec_relu<float, jit::avx512_common>,
-                       ref_relu<float>);
+    TestInplace<float>(sz, vec_relu<float, jit::avx512f>, ref_relu<float>);
   }
   TestInplace<double>(30, vec_relu<double>, ref_relu<double>);
 }

paddle/fluid/operators/math/depthwise_conv.cu

@@ -46,17 +46,20 @@ __forceinline__ __device__ unsigned warp_id() {
   return ret;
 }
 
+#define ARG_DEFINE_KernelDepthwiseConv                                         \
+  const T *const input_data, const T *const filter_data, const int batch_size, \
+      const int output_channels, const int output_height,                      \
+      const int output_width, const int input_channels,                        \
+      const int input_height, const int input_width,                           \
+      const int filter_multiplier, const int filter_height,                    \
+      const int filter_width, const int stride_height, const int stride_width, \
+      const int padding_height, const int padding_width,                       \
+      const int dilate_height, const int dilate_width, T *const output_data
+
 // A Cuda kernel to compute the depthwise convolution forward pass
 // in NCHW format.
 template <typename T>
-__device__ __inline__ void KernelDepthwiseConv(
-    const T* const input_data, const T* const filter_data, const int batch_size,
-    const int output_channels, const int output_height, const int output_width,
-    const int input_channels, const int input_height, const int input_width,
-    const int filter_multiplier, const int filter_height,
-    const int filter_width, const int stride_height, const int stride_width,
-    const int padding_height, const int padding_width, const int dilate_height,
-    const int dilate_width, T* const output_data) {
+__device__ __inline__ void KernelDepthwiseConv(ARG_DEFINE_KernelDepthwiseConv) {
   for (int w_out = threadIdx.x; w_out < output_width; w_out += blockDim.x) {
     for (int h_out = threadIdx.y; h_out < output_height; h_out += blockDim.y) {
       const int batch = blockIdx.y;
@@ -97,42 +100,105 @@ __device__ __inline__ void KernelDepthwiseConv(
   }
 }
 
-template <typename T, int c_filter_multiplier, int c_stride>
-__global__ void KernelDepthwiseConvSp(
-    const T* const input_data, const T* const filter_data, const int batch_size,
-    const int output_channels, const int output_height, const int output_width,
-    const int input_channels, const int input_height, const int input_width,
-    const int filter_multiplier, const int filter_height,
-    const int filter_width, const int stride_height, const int stride_width,
-    const int padding_height, const int padding_width, const int dilate_height,
-    const int dilate_width, T* const output_data) {
-  if (c_filter_multiplier == 0)
-    KernelDepthwiseConv<T>(input_data, filter_data, batch_size, output_channels,
-                           output_height, output_width, input_channels,
-                           input_height, input_width, filter_multiplier,
-                           filter_height, filter_width, stride_height,
-                           stride_width, padding_height, padding_width,
-                           dilate_height, dilate_width, output_data);
+template <typename T, int c_filter>
+__device__ __inline__ void KernelDepthwiseConvCFilter(
+    ARG_DEFINE_KernelDepthwiseConv) {
+  const int kWeghtSize = c_filter * c_filter;
+  T r_weight[kWeghtSize];
+  const int batch = blockIdx.y;
+  const int c_out = blockIdx.x;
+  const T* weight = filter_data + c_out * c_filter * c_filter;
+  for (int i = 0; i < c_filter * c_filter; i++) r_weight[i] = weight[i];
 
-  else
-    KernelDepthwiseConv<T>(input_data, filter_data, batch_size, output_channels,
-                           output_height, output_width, input_channels,
-                           input_height, input_width, c_filter_multiplier,
-                           filter_height, filter_height, c_stride, c_stride,
-                           padding_height, padding_width, dilate_height,
-                           dilate_width, output_data);
+  for (int w_out = threadIdx.x; w_out < output_width; w_out += blockDim.x) {
+    for (int h_out = threadIdx.y; h_out < output_height; h_out += blockDim.y) {
+      const int batch = blockIdx.y;
+      const int c_out = blockIdx.x;
+
+      const int c_in = c_out / filter_multiplier;
+      T value = 0;
+      const int h_in_start = -padding_height + h_out * stride_height;
+      const int w_in_start = -padding_width + w_out * stride_width;
+      const int h_in_end = h_in_start + c_filter * dilate_height;
+      const int w_in_end = w_in_start + c_filter * dilate_width;
+
+      const int in_offset =
+          ((batch * input_channels + c_in) * input_height) * input_width;
+
+      const int h_end = h_in_end < input_height ? h_in_end : input_height;
+      const int w_end = w_in_end < input_width ? w_in_end : input_width;
+      const int h_start = h_in_start > 0 ? h_in_start : 0;
+      const int w_start = w_in_start > 0 ? w_in_start : 0;
+
+      for (int h_in = h_in_start, h_f = 0; h_f < c_filter;
+           h_in += dilate_height, h_f++) {
+        for (int w_in = w_in_start, w_f = 0; w_f < c_filter;
+             w_in += dilate_width, w_f++) {
+          if (h_in >= 0 && h_in < input_height && w_in >= 0 &&
+              w_in < input_width) {
+            const int offset = in_offset + h_in * input_width + w_in;
+            value += r_weight[h_f * c_filter + w_f] * input_data[offset];
+          }
+        }
+      }
+      int index =
+          ((batch * gridDim.x + c_out) * output_height + h_out) * output_width +
+          w_out;
+      output_data[index] = value;
+    }
+  }
+}
+
+template <typename T, int c_filter_multiplier, int c_stride, int c_filter>
+__global__ void KernelDepthwiseConvSp(ARG_DEFINE_KernelDepthwiseConv) {
+  if (c_filter_multiplier == 0) {
+    if (c_filter == -1)
+      KernelDepthwiseConv<T>(
+          input_data, filter_data, batch_size, output_channels, output_height,
+          output_width, input_channels, input_height, input_width,
+          filter_multiplier, filter_height, filter_width, stride_height,
+          stride_width, padding_height, padding_width, dilate_height,
+          dilate_width, output_data);
+    else
+      KernelDepthwiseConvCFilter<T, c_filter>(
+          input_data, filter_data, batch_size, output_channels, output_height,
+          output_width, input_channels, input_height, input_width,
+          filter_multiplier, filter_height, filter_width, stride_height,
+          stride_width, padding_height, padding_width, dilate_height,
+          dilate_width, output_data);
+  } else {
+    if (c_filter == -1)
+      KernelDepthwiseConv<T>(input_data, filter_data, batch_size,
+                             output_channels, output_height, output_width,
+                             input_channels, input_height, input_width,
+                             c_filter_multiplier, filter_height, filter_height,
+                             c_stride, c_stride, padding_height, padding_width,
+                             dilate_height, dilate_width, output_data);
+    else
+      KernelDepthwiseConvCFilter<T, c_filter>(
+          input_data, filter_data, batch_size, output_channels, output_height,
+          output_width, input_channels, input_height, input_width,
+          c_filter_multiplier, filter_height, filter_height, c_stride, c_stride,
+          padding_height, padding_width, dilate_height, dilate_width,
+          output_data);
+  }
 }
 
 // CUDA kernel to compute the depthwise convolution backprop w.r.t input.
+#define ARG_DEFINE_KernelDepthwiseConvInputGrad                                \
+  const T *const output_grad_data, const T *const filter_data,                 \
+      const int batch_size, const int output_channels,                         \
+      const int output_height, const int output_width,                         \
+      const int input_channels, const int input_height, const int input_width, \
+      const int filter_multiplier, const int filter_height,                    \
+      const int filter_width, const int stride_height, const int stride_width, \
+      const int padding_height, const int padding_width,                       \
+      const int dilate_height, const int dilate_width,                         \
+      T *const input_grad_data
+
 template <typename T>
 __device__ __inline__ void KernelDepthwiseConvInputGrad(
-    const T* const output_grad_data, const T* const filter_data,
-    const int batch_size, const int output_channels, const int output_height,
-    const int output_width, const int input_channels, const int input_height,
-    const int input_width, const int filter_multiplier, const int filter_height,
-    const int filter_width, const int stride_height, const int stride_width,
-    const int padding_height, const int padding_width, const int dilate_height,
-    const int dilate_width, T* const input_grad_data) {
+    ARG_DEFINE_KernelDepthwiseConvInputGrad) {
   for (int w_in = threadIdx.x; w_in < input_width; w_in += blockDim.x) {
     for (int h_in = threadIdx.y; h_in < input_height; h_in += blockDim.y) {
       const int batch = blockIdx.y;
@@ -184,15 +250,67 @@ __device__ __inline__ void KernelDepthwiseConvInputGrad(
   }
 }
 
-template <typename T, int c_filter_multiplier, int c_stride>
+template <typename T, int c_filter, int c_filter_multiplier>
+__device__ __inline__ void KernelDepthwiseConvInputGradCFilter(
+    ARG_DEFINE_KernelDepthwiseConvInputGrad) {
+  const int kWeghtSize = c_filter * c_filter * c_filter_multiplier + 1;
+  T r_weight[kWeghtSize];
+  const int batch = blockIdx.y;
+  const int c_in = blockIdx.x;
+
+  for (int c_i = 0; c_i < filter_multiplier; c_i++) {
+    int c_out = c_in * filter_multiplier + c_i;
+    const T* weight = filter_data + c_out * c_filter * c_filter;
+    for (int i = 0; i < c_filter * c_filter; i++)
+      r_weight[i + c_i * c_filter * c_filter] =
+          weight[c_filter * c_filter - i - 1];
+  }
+
+  for (int w_in = threadIdx.x; w_in < input_width; w_in += blockDim.x) {
+    for (int h_in = threadIdx.y; h_in < input_height; h_in += blockDim.y) {
+      const int batch = blockIdx.y;
+      const int c_in = blockIdx.x;
+
+      int h_out_start = h_in - (c_filter - 1) * dilate_height + padding_height;
+
+      int w_out_start = w_in - (c_filter - 1) * dilate_width + padding_width;
+
+      T value = 0;
+
+      for (int c_i = 0; c_i < filter_multiplier; c_i++) {
+        int c_out = c_in * filter_multiplier + c_i;
+        for (int h_out = h_out_start, h_f = 0; h_f < c_filter;
+             h_out += dilate_height, h_f++) {
+          for (int w_out = w_out_start, w_f = 0; w_f < c_filter;
+               w_out += dilate_width, w_f++) {
+            int s_h_out = h_out / stride_height;
+            int s_w_out = w_out / stride_width;
+            if (h_out % stride_height == 0 && w_out % stride_width == 0 &&
+                s_h_out >= 0 && s_h_out < output_height && s_w_out >= 0 &&
+                s_w_out < output_width) {
+              const int output_grad_offset =
+                  ((batch * output_channels + c_out) * output_height +
+                   s_h_out) *
+                      output_width +
+                  s_w_out;
+              value +=
+                  output_grad_data[output_grad_offset] *
+                  r_weight[h_f * c_filter + w_f + c_i * c_filter * c_filter];
+            }
+          }
+        }
+      }
+      int index =
+          ((batch * gridDim.x + c_in) * input_height + h_in) * input_width +
+          w_in;
+      input_grad_data[index] = value;
+    }
+  }
+}
+
+template <typename T, int c_filter_multiplier, int c_stride, int c_filter>
 __global__ void KernelDepthwiseConvInputGradSp(
-    const T* const output_grad_data, const T* const filter_data,
-    const int batch_size, const int output_channels, const int output_height,
-    const int output_width, const int input_channels, const int input_height,
-    const int input_width, const int filter_multiplier, const int filter_height,
-    const int filter_width, const int stride_height, const int stride_width,
-    const int padding_height, const int padding_width, const int dilate_height,
-    const int dilate_width, T* const input_grad_data) {
+    ARG_DEFINE_KernelDepthwiseConvInputGrad) {
   if (c_filter_multiplier == 0)
     KernelDepthwiseConvInputGrad<T>(
         output_grad_data, filter_data, batch_size, output_channels,
@@ -200,13 +318,20 @@ __global__ void KernelDepthwiseConvInputGradSp(
         filter_multiplier, filter_height, filter_width, stride_height,
         stride_width, padding_height, padding_width, dilate_height,
         dilate_width, input_grad_data);
-  else
+  else if (c_filter == -1)
     KernelDepthwiseConvInputGrad<T>(
         output_grad_data, filter_data, batch_size, output_channels,
         output_height, output_width, input_channels, input_height, input_width,
         c_filter_multiplier, filter_height, filter_width, c_stride, c_stride,
         padding_height, padding_width, dilate_height, dilate_width,
         input_grad_data);
+  else
+    KernelDepthwiseConvInputGradCFilter<T, c_filter, c_filter_multiplier>(
+        output_grad_data, filter_data, batch_size, output_channels,
+        output_height, output_width, input_channels, input_height, input_width,
+        c_filter_multiplier, filter_height, filter_width, c_stride, c_stride,
+        padding_height, padding_width, dilate_height, dilate_width,
+        input_grad_data);
 }
 
 // Cuda kernel to compute the depthwise convolution backprop w.r.t. filter.
@@ -325,12 +450,14 @@ class DepthwiseConvFunctor<platform::CUDADeviceContext, T> {
     dim3 threads(std::min(output_width, thread), blocks, 1);
     dim3 grid(output_channels, batch_size, 1);
     int filter_multiplier = output_channels / input_channels;
-#define check_case(c_filter_multiplier, c_stride)                            \
+#define check_case(c_filter_multiplier, c_stride, c_filter)                  \
   if (c_filter_multiplier == 0 ||                                            \
       filter_multiplier == c_filter_multiplier &&                            \
-          stride_height == stride_width && stride_height == c_stride) {      \
-    KernelDepthwiseConvSp<T, c_filter_multiplier,                            \
-                          c_stride><<<grid, threads, 0, context.stream()>>>( \
+          stride_height == stride_width && stride_height == c_stride &&      \
+          (ksize_height == ksize_width && ksize_height == c_filter ||        \
+           c_filter == -1)) {                                                \
+    KernelDepthwiseConvSp<T, c_filter_multiplier, c_stride,                  \
+                          c_filter><<<grid, threads, 0, context.stream()>>>( \
         input_data, filter_data, batch_size, output_channels, output_height, \
         output_width, input_channels, input_height, input_width,             \
         filter_multiplier, ksize_height, ksize_width, stride_height,         \
@@ -338,11 +465,17 @@ class DepthwiseConvFunctor<platform::CUDADeviceContext, T> {
         dilate_width, output_data);                                          \
     return;                                                                  \
   }
-    check_case(1, 1);
-    check_case(1, 2);
-    // NOTE(liangdun): 0,0 for other case
-    // add other case if needed, e.g. check_case(2^n,1)
-    check_case(0, 0);
+    check_case(1, 1, 3);
+    check_case(1, 1, 5);
+    check_case(1, 1, -1);
+    check_case(1, 2, 3);
+    check_case(1, 2, 5);
+    check_case(1, 2, -1);
+    check_case(0, 0, 3);
+    check_case(0, 0, 5);
+    check_case(0, 0, -1);
+// NOTE(liangdun): 0,0 for other case
+// add other case if needed, e.g. check_case(2^n,1)
 #undef check_case
   }
 };
@@ -384,13 +517,15 @@ class DepthwiseConvInputGradFunctor<platform::CUDADeviceContext, T> {
     dim3 grid(input_channels, batch_size, 1);
     int filter_multiplier = output_channels / input_channels;
 
-#define check_case(c_filter_multiplier, c_stride)                       \
+#define check_case(c_filter_multiplier, c_stride, c_filter)             \
   if (c_filter_multiplier == 0 ||                                       \
       filter_multiplier == c_filter_multiplier &&                       \
-          stride_height == stride_width && stride_height == c_stride) { \
+          stride_height == stride_width && stride_height == c_stride && \
+          (ksize_height == ksize_width && ksize_height == c_filter ||   \
+           c_filter == -1)) {                                           \
     KernelDepthwiseConvInputGradSp<                                     \
-        T, c_filter_multiplier,                                         \
-        c_stride><<<grid, threads, 0, context.stream()>>>(              \
+        T, c_filter_multiplier, c_stride,                               \
+        c_filter><<<grid, threads, 0, context.stream()>>>(              \
         output_grad_data, filter_data, batch_size, output_channels,     \
         output_height, output_width, input_channels, input_height,      \
         input_width, filter_multiplier, ksize_height, ksize_width,      \
@@ -398,11 +533,21 @@ class DepthwiseConvInputGradFunctor<platform::CUDADeviceContext, T> {
         dilate_height, dilate_width, input_grad_data);                  \
     return;                                                             \
   }
-    check_case(1, 1);
-    check_case(1, 2);
-    // NOTE(liangdun): 0,0 for other case
-    // add other case if needed, e.g. check_case(2^n,1)
-    check_case(0, 0);
+    check_case(1, 1, 3);
+    check_case(1, 1, 5);
+    check_case(1, 1, -1);
+    check_case(1, 2, 3);
+    check_case(1, 2, 5);
+    check_case(1, 2, -1);
+    check_case(2, 1, 3);
+    check_case(2, 1, 5);
+    check_case(2, 1, -1);
+    check_case(2, 2, 3);
+    check_case(2, 2, 5);
+    check_case(2, 2, -1);
+    check_case(0, 0, -1);
+// NOTE(liangdun): 0,0 for other case
+// add other case if needed, e.g. check_case(2^n,1)
 #undef check_case
   }
 };

paddle/fluid/operators/math/cpu_lstm_compute.cc → paddle/fluid/operators/math/jit_kernel.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/math/cpu_lstm_compute.h"
+#include "paddle/fluid/operators/math/jit_kernel.h"
+#include <iostream>
+#include <string>
 
 namespace paddle {
 namespace operators {
 namespace math {
-#ifdef __AVX__
-template <>
-void lstm_compute_ctht<float>(float* gates, const float* ct_1, float* ct,
-                              float* ht) {
-  namespace act = detail::forward::avx;
-  // 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(act::Tanh(c), act::Sigmoid(i));
-  i = _mm256_loadu_ps(ct_1);
-  f = _mm256_mul_ps(i, act::Sigmoid(f));
-  f = _mm256_add_ps(c, f);
-  _mm256_storeu_ps(ct, f);
-
-  /* H_t = act_cell(C_t) * ogated */
-  o = _mm256_mul_ps(act::Tanh(f), act::Sigmoid(o));
-  _mm256_storeu_ps(ht, o);
+namespace jitkernel {
+
+namespace jit = platform::jit;
+
+KernelPool& KernelPool::Instance() {
+  static thread_local KernelPool g_jit_kernels;
+  return g_jit_kernels;
+}
+
+std::shared_ptr<const Kernel> KernelPool::Get(const std::string& key) const {
+  if (kers_.find(key) == kers_.end()) {
+    return nullptr;
+  }
+  return kers_.at(key);
 }
-#endif
+
+}  // namespace jitkernel
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/math/jit_kernel.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 <functional>
+#include <memory>  // for shared_ptr
+#include <string>
+#include <unordered_map>
+#include "paddle/fluid/platform/cpu_info.h"
+#include "paddle/fluid/platform/macros.h"
+
+// Note: Only support on CPU yet.
+namespace paddle {
+namespace operators {
+namespace math {
+namespace jitkernel {
+
+#define SIGMOID_THRESHOLD_MIN -40.0
+#define SIGMOID_THRESHOLD_MAX 13.0
+#define EXP_MAX_INPUT 40.0
+#define AVX_FLOAT_BLOCK 8
+#define AVX2_FLOAT_BLOCK 8
+#define AVX512_FLOAT_BLOCK 16
+
+typedef enum { kLT8, kEQ8, kGT8LT16, kEQ16, kGT16 } jit_block;
+
+class Kernel {
+ public:
+  Kernel() = default;
+  virtual ~Kernel() = default;
+  int num_{0};
+  int end_{0};
+  int rest_{0};
+  DISABLE_COPY_AND_ASSIGN(Kernel);
+};
+
+class KernelPool {
+ public:
+  static KernelPool &Instance();
+
+  template <typename Ker, typename... ARGS>
+  std::shared_ptr<const Ker> Get(ARGS... args);
+
+  std::shared_ptr<const Kernel> Get(const std::string &key) const;
+
+ private:
+  KernelPool() = default;
+  std::unordered_map<std::string, std::shared_ptr<const Kernel>> kers_;
+
+  DISABLE_COPY_AND_ASSIGN(KernelPool);
+};
+
+template <typename T>
+class VMulKernel : public Kernel {
+ public:
+  virtual void Compute(const T *x, const T *y, T *z) const = 0;
+};
+
+template <typename T>
+class VAddKernel : public Kernel {
+ public:
+  virtual void Compute(const T *x, const T *y, T *z) const = 0;
+};
+
+template <typename T>
+class VScalKernel : public Kernel {
+ public:
+  virtual void Compute(const T a, const T *x, T *y) const = 0;
+  virtual void Compute(const T a, T *x) const = 0;
+};
+
+template <typename T>
+class VAddBiasKernel : public Kernel {
+ public:
+  virtual void Compute(const T a, const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class VActKernel : public Kernel {
+ public:
+  virtual void Compute(const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class VReluKernel : public VActKernel<T> {
+ public:
+  virtual void Compute(const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class VIdentityKernel : public VActKernel<T> {
+ public:
+  virtual void Compute(const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class VExpKernel : public VActKernel<T> {
+ public:
+  virtual void Compute(const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class VSigmoidKernel : public VActKernel<T> {
+ public:
+  virtual void Compute(const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class VTanhKernel : public VActKernel<T> {
+ public:
+  virtual void Compute(const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class LSTMKernel : public Kernel {
+ public:
+  virtual void ComputeCtHt(T *gates, const T *ct_1, T *ct, T *ht,
+                           /* below only used in peephole*/
+                           const T *wp_data = nullptr,
+                           T *checked = nullptr) const = 0;
+
+  // compute c1 and h1 without c0 or h0
+  virtual void ComputeC1H1(T *gates, T *ct, T *ht,
+                           /* below only used in peephole*/
+                           const T *wp_data = nullptr) const = 0;
+};
+
+}  // namespace jitkernel
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/jit_kernel_blas.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/math/jit_kernel.h"
+#include <string>
+#include "paddle/fluid/operators/math/jit_kernel_macro.h"
+#ifdef PADDLE_WITH_MKLML
+#include "paddle/fluid/platform/dynload/mklml.h"
+#endif
+
+#ifdef __AVX__
+#include <immintrin.h>
+#endif
+
+namespace paddle {
+namespace operators {
+namespace math {
+namespace jitkernel {
+
+namespace jit = platform::jit;
+
+/* VMUL JitKernel */
+template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+class VMulKernelImpl : public VMulKernel<T> {
+ public:
+  explicit VMulKernelImpl(int d) : VMulKernel<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];
+    }
+  }
+};
+
+#ifdef PADDLE_WITH_MKLML
+#define MKL_FLOAT(isa, block)                           \
+  template <>                                           \
+  void VMulKernelImpl<float, isa, block>::Compute(      \
+      const float* x, const float* y, float* z) const { \
+    platform::dynload::vsMul(this->num_, x, y, z);      \
+  }
+
+#define MKL_DOUBLE(isa, block)                             \
+  template <>                                              \
+  void VMulKernelImpl<double, isa, block>::Compute(        \
+      const double* x, const double* y, double* z) const { \
+    platform::dynload::vdMul(this->num_, x, y, z);         \
+  }
+
+FOR_EACH_ISA(MKL_FLOAT, kGT16);
+FOR_EACH_ISA_BLOCK(MKL_DOUBLE);
+#endif
+
+#define INTRI8_FLOAT(isa)                               \
+  template <>                                           \
+  void VMulKernelImpl<float, isa, kEQ8>::Compute(       \
+      const float* x, const float* y, float* z) const { \
+    __m256 tmpx, tmpy;                                  \
+    tmpx = _mm256_loadu_ps(x);                          \
+    tmpy = _mm256_loadu_ps(y);                          \
+    tmpx = _mm256_mul_ps(tmpx, tmpy);                   \
+    _mm256_storeu_ps(z, tmpx);                          \
+  }
+
+// avx > for > mkl
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+#endif
+// TODO(TJ): eq16 test and complete avx512
+#undef INTRI8_FLOAT
+#undef MKL_FLOAT
+#undef MKL_DOUBLE
+
+/* VADD JitKernel */
+template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+class VAddKernelImpl : public VAddKernel<T> {
+ public:
+  explicit VAddKernelImpl(int d) : VAddKernel<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];
+    }
+  }
+};
+
+#ifdef PADDLE_WITH_MKLML
+#define MKL_FLOAT(isa, block)                           \
+  template <>                                           \
+  void VAddKernelImpl<float, isa, block>::Compute(      \
+      const float* x, const float* y, float* z) const { \
+    platform::dynload::vsAdd(this->num_, x, y, z);      \
+  }
+
+#define MKL_DOUBLE(isa, block)                             \
+  template <>                                              \
+  void VAddKernelImpl<double, isa, block>::Compute(        \
+      const double* x, const double* y, double* z) const { \
+    platform::dynload::vdAdd(this->num_, x, y, z);         \
+  }
+
+FOR_EACH_ISA(MKL_FLOAT, kGT16);
+FOR_EACH_ISA_BLOCK(MKL_DOUBLE);
+#endif
+
+#define INTRI8_FLOAT(isa)                               \
+  template <>                                           \
+  void VAddKernelImpl<float, isa, kEQ8>::Compute(       \
+      const float* x, const float* y, float* z) const { \
+    __m256 tmpx, tmpy;                                  \
+    tmpx = _mm256_loadu_ps(x);                          \
+    tmpy = _mm256_loadu_ps(y);                          \
+    tmpx = _mm256_add_ps(tmpx, tmpy);                   \
+    _mm256_storeu_ps(z, tmpx);                          \
+  }
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+#endif
+// TODO(TJ): eq16 test and complete avx512
+
+#undef INTRI8_FLOAT
+#undef MKL_FLOAT
+#undef MKL_DOUBLE
+
+/* VSCAL JitKernel */
+template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+class VScalKernelImpl : public VScalKernel<T> {
+ public:
+  explicit VScalKernelImpl(int d) : VScalKernel<T>() { this->num_ = d; }
+  void Compute(const T a, const T* x, T* y) const override {
+    for (int i = 0; i < this->num_; ++i) {
+      y[i] = a * x[i];
+    }
+  }
+  void Compute(const T a, T* x) const override {
+    for (int i = 0; i < this->num_; ++i) {
+      x[i] = a * x[i];
+    }
+  }
+};
+
+#ifdef PADDLE_WITH_MKLML
+#define MKL_FLOAT(isa, block)                                               \
+  template <>                                                               \
+  void VScalKernelImpl<float, isa, block>::Compute(const float a, float* x) \
+      const {                                                               \
+    platform::dynload::cblas_sscal(this->num_, a, x, 1);                    \
+  }
+
+#define MKL_DOUBLE(isa, block)                                                 \
+  template <>                                                                  \
+  void VScalKernelImpl<double, isa, block>::Compute(const double a, double* x) \
+      const {                                                                  \
+    platform::dynload::cblas_dscal(this->num_, a, x, 1);                       \
+  }
+
+FOR_EACH_ISA(MKL_FLOAT, kGT16);
+FOR_EACH_ISA_BLOCK(MKL_DOUBLE);
+#endif
+
+#define INTRI8_FLOAT(isa)                              \
+  template <>                                          \
+  void VScalKernelImpl<float, isa, kEQ8>::Compute(     \
+      const float a, const float* x, float* y) const { \
+    __m256 tmp;                                        \
+    __m256 scalar = _mm256_set1_ps(a);                 \
+    tmp = _mm256_loadu_ps(x);                          \
+    tmp = _mm256_mul_ps(tmp, scalar);                  \
+    _mm256_storeu_ps(y, tmp);                          \
+  }
+#define INTRI8_INPLACE_FLOAT(isa)                                          \
+  template <>                                                              \
+  void VScalKernelImpl<float, isa, kEQ8>::Compute(const float a, float* x) \
+      const {                                                              \
+    __m256 tmp;                                                            \
+    __m256 scalar = _mm256_set1_ps(a);                                     \
+    tmp = _mm256_loadu_ps(x);                                              \
+    tmp = _mm256_mul_ps(tmp, scalar);                                      \
+    _mm256_storeu_ps(x, tmp);                                              \
+  }
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+INTRI8_INPLACE_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+INTRI8_INPLACE_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+INTRI8_INPLACE_FLOAT(jit::avx512f);
+#endif
+// TODO(TJ): eq16 test and complete avx512
+
+#undef INTRI8_FLOAT
+#undef INTRI8_INPLACE_FLOAT
+#undef MKL_FLOAT
+#undef MKL_DOUBLE
+
+/* VAddBias JitKernel */
+template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+class VAddBiasKernelImpl : public VAddBiasKernel<T> {
+ public:
+  explicit VAddBiasKernelImpl(int d) : VAddBiasKernel<T>() { this->num_ = d; }
+  void Compute(const T a, const T* x, T* y) const override {
+    for (int i = 0; i < this->num_; ++i) {
+      y[i] = x[i] + a;
+    }
+  }
+};
+
+#define INTRI8_FLOAT(isa)                              \
+  template <>                                          \
+  void VAddBiasKernelImpl<float, isa, kEQ8>::Compute(  \
+      const float a, const float* x, float* y) const { \
+    __m256 tmp = _mm256_loadu_ps(x);                   \
+    tmp = _mm256_add_ps(tmp, _mm256_set1_ps(a));       \
+    _mm256_storeu_ps(y, tmp);                          \
+  }
+
+#define INTRI16_FLOAT(isa)                             \
+  template <>                                          \
+  void VAddBiasKernelImpl<float, isa, kEQ16>::Compute( \
+      const float a, const float* x, float* y) const { \
+    __m256 tmp0 = _mm256_loadu_ps(x);                  \
+    __m256 tmp1 = _mm256_loadu_ps(x + 8);              \
+    tmp0 = _mm256_add_ps(tmp0, _mm256_set1_ps(a));     \
+    tmp1 = _mm256_add_ps(tmp1, _mm256_set1_ps(a));     \
+    _mm256_storeu_ps(y, tmp0);                         \
+    _mm256_storeu_ps(y + 8, tmp1);                     \
+  }
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+INTRI16_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+INTRI16_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+INTRI16_FLOAT(jit::avx512f);
+#endif
+// TODO(TJ): eq16 test and complete avx512
+
+#undef INTRI8_FLOAT
+#undef INTRI16_FLOAT
+
+/* VRelu JitKernel */
+template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+class VReluKernelImpl : public VReluKernel<T> {
+ public:
+  explicit VReluKernelImpl(int d) : VReluKernel<T>() { this->num_ = d; }
+  void Compute(const T* x, T* y) const override {
+    for (int i = 0; i < this->num_; ++i) {
+      y[i] = x[i] > 0 ? x[i] : 0;
+    }
+  }
+};
+
+#define INTRI8_FLOAT(isa)                                                   \
+  template <>                                                               \
+  void VReluKernelImpl<float, isa, kEQ8>::Compute(const float* x, float* y) \
+      const {                                                               \
+    __m256 tmp = _mm256_loadu_ps(x);                                        \
+    tmp = _mm256_max_ps(tmp, _mm256_setzero_ps());                          \
+    _mm256_storeu_ps(y, tmp);                                               \
+  }
+
+#define INTRI16_FLOAT(isa)                                                   \
+  template <>                                                                \
+  void VReluKernelImpl<float, isa, kEQ16>::Compute(const float* x, float* y) \
+      const {                                                                \
+    __m256 zeros = _mm256_setzero_ps();                                      \
+    __m256 tmp0 = _mm256_loadu_ps(x);                                        \
+    __m256 tmp1 = _mm256_loadu_ps(x + 8);                                    \
+    tmp0 = _mm256_max_ps(tmp0, zeros);                                       \
+    tmp1 = _mm256_max_ps(tmp1, zeros);                                       \
+    _mm256_storeu_ps(y, tmp0);                                               \
+    _mm256_storeu_ps(y + 8, tmp1);                                           \
+  }
+
+#define INTRI_GT8LT16_FLOAT(isa)                                        \
+  template <>                                                           \
+  VReluKernelImpl<float, isa, kGT8LT16>::VReluKernelImpl(int d)         \
+      : VReluKernel<float>() {                                          \
+    this->num_ = d;                                                     \
+    this->end_ = AVX_FLOAT_BLOCK;                                       \
+    this->rest_ = d - AVX_FLOAT_BLOCK;                                  \
+  }                                                                     \
+  template <>                                                           \
+  void VReluKernelImpl<float, isa, kGT8LT16>::Compute(const float* x,   \
+                                                      float* y) const { \
+    __m256 zeros = _mm256_setzero_ps();                                 \
+    __m256 tmp0 = _mm256_loadu_ps(x);                                   \
+    __m256 tmp1 = _mm256_loadu_ps(x + this->rest_);                     \
+    tmp0 = _mm256_max_ps(tmp0, zeros);                                  \
+    tmp1 = _mm256_max_ps(tmp1, zeros);                                  \
+    _mm256_storeu_ps(y, tmp0);                                          \
+    _mm256_storeu_ps(y + this->rest_, tmp1);                            \
+  }
+
+#define INTRI_GT16_FLOAT(isa)                                                \
+  template <>                                                                \
+  VReluKernelImpl<float, isa, kGT16>::VReluKernelImpl(int d)                 \
+      : VReluKernel<float>() {                                               \
+    this->num_ = d;                                                          \
+    this->end_ = d - d % AVX_FLOAT_BLOCK;                                    \
+    this->rest_ = d - AVX_FLOAT_BLOCK;                                       \
+  }                                                                          \
+  template <>                                                                \
+  void VReluKernelImpl<float, isa, kGT16>::Compute(const float* x, float* y) \
+      const {                                                                \
+    __m256 zeros = _mm256_setzero_ps();                                      \
+    for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) {                  \
+      __m256 tmp = _mm256_loadu_ps(x + i);                                   \
+      tmp = _mm256_max_ps(tmp, zeros);                                       \
+      _mm256_storeu_ps(y + i, tmp);                                          \
+    }                                                                        \
+    __m256 tmp = _mm256_loadu_ps(x + this->rest_);                           \
+    tmp = _mm256_max_ps(tmp, zeros);                                         \
+    _mm256_storeu_ps(y + this->rest_, tmp);                                  \
+  }
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+INTRI16_FLOAT(jit::avx);
+INTRI_GT8LT16_FLOAT(jit::avx);
+INTRI_GT16_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+INTRI16_FLOAT(jit::avx2);
+INTRI_GT8LT16_FLOAT(jit::avx2);
+INTRI_GT16_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+// TODO(TJ): refine avx512
+INTRI8_FLOAT(jit::avx512f);
+INTRI16_FLOAT(jit::avx512f);
+INTRI_GT8LT16_FLOAT(jit::avx512f);
+INTRI_GT16_FLOAT(jit::avx512f);
+#endif
+
+#undef INTRI8_FLOAT
+#undef INTRI16_FLOAT
+#undef INTRI_GT8LT16_FLOAT
+#undef INTRI_GT16_FLOAT
+
+/* An empty JitKernel */
+template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+class VIdentityKernelImpl : public VIdentityKernel<T> {
+ public:
+  explicit VIdentityKernelImpl(int d) : VIdentityKernel<T>() { this->num_ = d; }
+  void Compute(const T* x, T* y) const override {}
+};
+
+REGISTER_JITKERNEL(vmul, VMulKernel);
+REGISTER_JITKERNEL(vadd, VAddKernel);
+REGISTER_JITKERNEL(vscal, VScalKernel);
+REGISTER_JITKERNEL(vaddb, VAddBiasKernel);
+REGISTER_JITKERNEL(vrelu, VReluKernel);
+REGISTER_JITKERNEL(videntity, VIdentityKernel);
+
+}  // namespace jitkernel
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/jit_kernel_exp.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/math/jit_kernel.h"
+#include <cmath>  // for exp
+#include <string>
+#include "paddle/fluid/operators/math/jit_kernel_macro.h"
+#ifdef PADDLE_WITH_MKLML
+#include "paddle/fluid/platform/dynload/mklml.h"
+#endif
+
+#ifdef __AVX__
+#include <immintrin.h>
+#endif
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+#ifdef __AVX__
+namespace detail {
+__m256 Exp(__m256 a);
+}  // namespace detail
+#endif
+
+namespace jitkernel {
+namespace jit = platform::jit;
+
+/* VExp JitKernel */
+template <typename T, jit::cpu_isa_t isa, jit_block>
+class VExpKernelImpl : public VExpKernel<T> {
+ public:
+  explicit VExpKernelImpl(int d) : VExpKernel<T>() { this->num_ = d; }
+  void Compute(const T* x, T* y) const override {
+    for (int i = 0; i < this->num_; ++i) {
+      y[i] = std::exp(x[i]);
+    }
+  }
+};
+
+#ifdef PADDLE_WITH_MKLML
+#define MKL_FLOAT(isa, block)                                               \
+  template <>                                                               \
+  void VExpKernelImpl<float, isa, block>::Compute(const float* x, float* y) \
+      const {                                                               \
+    platform::dynload::vsExp(this->num_, x, y);                             \
+  }
+
+#define MKL_DOUBLE(isa, block)                                                 \
+  template <>                                                                  \
+  void VExpKernelImpl<double, isa, block>::Compute(const double* x, double* y) \
+      const {                                                                  \
+    platform::dynload::vdExp(this->num_, x, y);                                \
+  }
+FOR_EACH_ISA(MKL_FLOAT, kLT8);
+FOR_EACH_ISA(MKL_FLOAT, kGT8LT16);
+FOR_EACH_ISA(MKL_FLOAT, kGT16);
+FOR_EACH_ISA_BLOCK(MKL_DOUBLE);
+#endif
+
+#define INTRI8_FLOAT(isa)                                                  \
+  template <>                                                              \
+  void VExpKernelImpl<float, isa, kEQ8>::Compute(const float* x, float* y) \
+      const {                                                              \
+    __m256 tmp = _mm256_loadu_ps(x);                                       \
+    _mm256_storeu_ps(y, detail::Exp(tmp));                                 \
+  }
+
+#define INTRI16_FLOAT(isa)                                                  \
+  template <>                                                               \
+  void VExpKernelImpl<float, isa, kEQ16>::Compute(const float* x, float* y) \
+      const {                                                               \
+    __m256 tmp0 = _mm256_loadu_ps(x);                                       \
+    __m256 tmp1 = _mm256_loadu_ps(x + 8);                                   \
+    tmp0 = detail::Exp(tmp0);                                               \
+    tmp1 = detail::Exp(tmp1);                                               \
+    _mm256_storeu_ps(y, tmp0);                                              \
+    _mm256_storeu_ps(y + 8, tmp1);                                          \
+  }
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+INTRI16_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+INTRI16_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+INTRI16_FLOAT(jit::avx512f);
+#endif
+// TODO(TJ): eq16 test and complete avx512
+
+#undef INTRI8_FLOAT
+#undef INTRI16_FLOAT
+#undef MKL_FLOAT
+#undef MKL_DOUBLE
+
+REGISTER_JITKERNEL(vexp, VExpKernel);
+
+/* VSigmoid JitKernel */
+template <typename T, jit::cpu_isa_t isa, jit_block>
+class VSigmoidKernelImpl : public VSigmoidKernel<T> {
+ public:
+  explicit VSigmoidKernelImpl(int d) : VSigmoidKernel<T>() {
+    this->num_ = d;
+    vexp_ = KernelPool::Instance().template Get<VExpKernel<T>>(d);
+  }
+  void Compute(const T* x, T* y) const override {
+    const T min = SIGMOID_THRESHOLD_MIN;
+    const T max = SIGMOID_THRESHOLD_MAX;
+    for (int i = 0; i < this->num_; ++i) {
+      y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]);
+      y[i] = static_cast<T>(0) - y[i];
+    }
+    vexp_->Compute(y, y);
+    for (int i = 0; i < this->num_; ++i) {
+      y[i] = static_cast<T>(1) / (static_cast<T>(1) + y[i]);
+    }
+  }
+
+ private:
+  std::shared_ptr<const VExpKernel<T>> vexp_;
+};
+
+#define INTRI_SIGMOID(tmp, min, max)              \
+  tmp = _mm256_max_ps(tmp, min);                  \
+  tmp = _mm256_min_ps(tmp, max);                  \
+  tmp = _mm256_sub_ps(_mm256_set1_ps(0.0f), tmp); \
+  tmp = detail::Exp(tmp);                         \
+  tmp = _mm256_add_ps(_mm256_set1_ps(1.0f), tmp); \
+  tmp = _mm256_div_ps(_mm256_set1_ps(1.0f), tmp)
+
+#define INTRI8_FLOAT(isa)                                                      \
+  template <>                                                                  \
+  void VSigmoidKernelImpl<float, isa, kEQ8>::Compute(const float* x, float* y) \
+      const {                                                                  \
+    __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);                        \
+    __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);                        \
+    __m256 tmp = _mm256_loadu_ps(x);                                           \
+    INTRI_SIGMOID(tmp, min, max);                                              \
+    _mm256_storeu_ps(y, tmp);                                                  \
+  }
+
+#define INTRI16_FLOAT(isa)                                              \
+  template <>                                                           \
+  void VSigmoidKernelImpl<float, isa, kEQ16>::Compute(const float* x,   \
+                                                      float* y) const { \
+    __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);                 \
+    __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);                 \
+    __m256 tmp0 = _mm256_loadu_ps(x);                                   \
+    __m256 tmp1 = _mm256_loadu_ps(x + 8);                               \
+    INTRI_SIGMOID(tmp0, min, max);                                      \
+    INTRI_SIGMOID(tmp1, min, max);                                      \
+    _mm256_storeu_ps(y, tmp0);                                          \
+    _mm256_storeu_ps(y + 8, tmp1);                                      \
+  }
+
+#define INTRI_GT8LT16_FLOAT(isa)                                             \
+  template <>                                                                \
+  VSigmoidKernelImpl<float, isa, kGT8LT16>::VSigmoidKernelImpl(int d)        \
+      : VSigmoidKernel<float>() {                                            \
+    this->num_ = d;                                                          \
+    this->end_ = AVX_FLOAT_BLOCK;                                            \
+    this->rest_ = d - this->end_;                                            \
+    vexp_ =                                                                  \
+        KernelPool::Instance().template Get<VExpKernel<float>>(this->rest_); \
+  }                                                                          \
+  template <>                                                                \
+  void VSigmoidKernelImpl<float, isa, kGT8LT16>::Compute(const float* x,     \
+                                                         float* y) const {   \
+    __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);                      \
+    __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);                      \
+    __m256 tmp = _mm256_loadu_ps(x);                                         \
+    INTRI_SIGMOID(tmp, min, max);                                            \
+    _mm256_storeu_ps(y, tmp);                                                \
+    const float min_ = SIGMOID_THRESHOLD_MIN;                                \
+    const float max_ = SIGMOID_THRESHOLD_MAX;                                \
+    for (int i = this->end_; i < this->num_; ++i) {                          \
+      y[i] = (x[i] < min_) ? min_ : ((x[i] > max_) ? max_ : x[i]);           \
+      y[i] = 0.f - y[i];                                                     \
+    }                                                                        \
+    vexp_->Compute(y + this->end_, y + this->end_);                          \
+    for (int i = this->end_; i < this->num_; ++i) {                          \
+      y[i] = 1.f / (1.f + y[i]);                                             \
+    }                                                                        \
+  }
+
+#define INTRI_GT16_FLOAT(isa)                                                \
+  template <>                                                                \
+  VSigmoidKernelImpl<float, isa, kGT16>::VSigmoidKernelImpl(int d)           \
+      : VSigmoidKernel<float>() {                                            \
+    this->num_ = d;                                                          \
+    this->rest_ = d % AVX_FLOAT_BLOCK;                                       \
+    this->end_ = d - this->rest_;                                            \
+    vexp_ =                                                                  \
+        KernelPool::Instance().template Get<VExpKernel<float>>(this->rest_); \
+  }                                                                          \
+  template <>                                                                \
+  void VSigmoidKernelImpl<float, isa, kGT16>::Compute(const float* x,        \
+                                                      float* y) const {      \
+    __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);                      \
+    __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);                      \
+    for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) {                  \
+      __m256 tmp = _mm256_loadu_ps(x + i);                                   \
+      INTRI_SIGMOID(tmp, min, max);                                          \
+      _mm256_storeu_ps(y + i, tmp);                                          \
+    }                                                                        \
+    const float min_ = SIGMOID_THRESHOLD_MIN;                                \
+    const float max_ = SIGMOID_THRESHOLD_MAX;                                \
+    for (int i = this->end_; i < this->num_; ++i) {                          \
+      y[i] = (x[i] < min_) ? min_ : ((x[i] > max_) ? max_ : x[i]);           \
+      y[i] = 0.f - y[i];                                                     \
+    }                                                                        \
+    vexp_->Compute(y + this->end_, y + this->end_);                          \
+    for (int i = this->end_; i < this->num_; ++i) {                          \
+      y[i] = 1.f / (1.f + y[i]);                                             \
+    }                                                                        \
+  }
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+INTRI16_FLOAT(jit::avx);
+INTRI_GT8LT16_FLOAT(jit::avx);
+INTRI_GT16_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+INTRI16_FLOAT(jit::avx2);
+// INTRI_GT8LT16_FLOAT(jit::avx2);
+// INTRI_GT16_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+INTRI16_FLOAT(jit::avx512f);
+// INTRI_GT8LT16_FLOAT(jit::avx512f);
+// INTRI_GT16_FLOAT(jit::avx512f);
+#endif
+
+#undef INTRI8_FLOAT
+#undef INTRI16_FLOAT
+#undef INTRI_GT8LT16_FLOAT
+#undef INTRI_GT16_FLOAT
+#undef INTRI_VSIGMOID
+
+REGISTER_JITKERNEL(vsigmoid, VSigmoidKernel);
+
+/* VTanh JitKernel */
+template <typename T, jit::cpu_isa_t isa, jit_block>
+class VTanhKernelImpl : public VTanhKernel<T> {
+ public:
+  explicit VTanhKernelImpl(int d) : VTanhKernel<T>() {
+    this->num_ = d;
+    vscal_ = KernelPool::Instance().template Get<VScalKernel<T>>(d);
+    vsigmoid_ = KernelPool::Instance().template Get<VSigmoidKernel<T>>(d);
+    vaddbias_ = KernelPool::Instance().template Get<VAddBiasKernel<T>>(d);
+  }
+  void Compute(const T* x, T* y) const override {
+    vscal_->Compute(static_cast<T>(2), x, y);
+    vsigmoid_->Compute(y, y);
+    vscal_->Compute(static_cast<T>(2), y);
+    vaddbias_->Compute(static_cast<T>(-1), y, y);
+  }
+
+ private:
+  std::shared_ptr<const VScalKernel<T>> vscal_;
+  std::shared_ptr<const VSigmoidKernel<T>> vsigmoid_;
+  std::shared_ptr<const VAddBiasKernel<T>> vaddbias_;
+};
+
+#define INTRI_VTANH(tmp)                                   \
+  tmp = _mm256_mul_ps(_mm256_set1_ps(-2.0f), tmp);         \
+  tmp = _mm256_min_ps(tmp, _mm256_set1_ps(EXP_MAX_INPUT)); \
+  tmp = detail::Exp(tmp);                                  \
+  tmp = _mm256_add_ps(_mm256_set1_ps(1.0f), tmp);          \
+  tmp = _mm256_div_ps(_mm256_set1_ps(2.0f), tmp);          \
+  tmp = _mm256_sub_ps(tmp, _mm256_set1_ps(1.0f))
+
+#define INTRI8_FLOAT(isa)                                                   \
+  template <>                                                               \
+  void VTanhKernelImpl<float, isa, kEQ8>::Compute(const float* x, float* y) \
+      const {                                                               \
+    __m256 tmp = _mm256_loadu_ps(x);                                        \
+    INTRI_VTANH(tmp);                                                       \
+    _mm256_storeu_ps(y, tmp);                                               \
+  }
+
+#define INTRI16_FLOAT(isa)                                                   \
+  template <>                                                                \
+  void VTanhKernelImpl<float, isa, kEQ16>::Compute(const float* x, float* y) \
+      const {                                                                \
+    __m256 tmp0 = _mm256_loadu_ps(x);                                        \
+    __m256 tmp1 = _mm256_loadu_ps(x + 8);                                    \
+    INTRI_VTANH(tmp0);                                                       \
+    INTRI_VTANH(tmp1);                                                       \
+    _mm256_storeu_ps(y, tmp0);                                               \
+    _mm256_storeu_ps(y + 8, tmp1);                                           \
+  }
+
+#define INTRI_GT8LT16_FLOAT(isa)                                              \
+  template <>                                                                 \
+  VTanhKernelImpl<float, isa, kGT8LT16>::VTanhKernelImpl(int d)               \
+      : VTanhKernel<float>() {                                                \
+    this->num_ = d;                                                           \
+    this->end_ = AVX_FLOAT_BLOCK;                                             \
+    this->rest_ = d - this->end_;                                             \
+    vscal_ =                                                                  \
+        KernelPool::Instance().template Get<VScalKernel<float>>(this->rest_); \
+    vsigmoid_ = KernelPool::Instance().template Get<VSigmoidKernel<float>>(   \
+        this->rest_);                                                         \
+    vaddbias_ = KernelPool::Instance().template Get<VAddBiasKernel<float>>(   \
+        this->rest_);                                                         \
+  }                                                                           \
+  template <>                                                                 \
+  void VTanhKernelImpl<float, isa, kGT8LT16>::Compute(const float* x,         \
+                                                      float* y) const {       \
+    __m256 tmp = _mm256_loadu_ps(x);                                          \
+    INTRI_VTANH(tmp);                                                         \
+    _mm256_storeu_ps(y, tmp);                                                 \
+    x += AVX_FLOAT_BLOCK;                                                     \
+    y += AVX_FLOAT_BLOCK;                                                     \
+    vscal_->Compute(2.f, x, y);                                               \
+    vsigmoid_->Compute(y, y);                                                 \
+    vscal_->Compute(2.f, y);                                                  \
+    vaddbias_->Compute(-1.f, y, y);                                           \
+  }
+
+#define INTRI_GT16_FLOAT(isa)                                                 \
+  template <>                                                                 \
+  VTanhKernelImpl<float, isa, kGT16>::VTanhKernelImpl(int d)                  \
+      : VTanhKernel<float>() {                                                \
+    this->num_ = d;                                                           \
+    this->rest_ = d % AVX_FLOAT_BLOCK;                                        \
+    this->end_ = d - this->rest_;                                             \
+    vscal_ =                                                                  \
+        KernelPool::Instance().template Get<VScalKernel<float>>(this->rest_); \
+    vsigmoid_ = KernelPool::Instance().template Get<VSigmoidKernel<float>>(   \
+        this->rest_);                                                         \
+    vaddbias_ = KernelPool::Instance().template Get<VAddBiasKernel<float>>(   \
+        this->rest_);                                                         \
+  }                                                                           \
+  template <>                                                                 \
+  void VTanhKernelImpl<float, isa, kGT16>::Compute(const float* x, float* y)  \
+      const {                                                                 \
+    for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) {                   \
+      __m256 tmp = _mm256_loadu_ps(x + i);                                    \
+      INTRI_VTANH(tmp);                                                       \
+      _mm256_storeu_ps(y + i, tmp);                                           \
+    }                                                                         \
+    x += this->end_;                                                          \
+    y += this->end_;                                                          \
+    vscal_->Compute(2.f, x, y);                                               \
+    vsigmoid_->Compute(y, y);                                                 \
+    vscal_->Compute(2.f, y);                                                  \
+    vaddbias_->Compute(-1.f, y, y);                                           \
+  }
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+INTRI16_FLOAT(jit::avx);
+INTRI_GT8LT16_FLOAT(jit::avx);
+INTRI_GT16_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+INTRI16_FLOAT(jit::avx2);
+// maybe use avx at gt8lt16 and gt16
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+INTRI16_FLOAT(jit::avx512f);
+// maybe use avx at gt8lt16 and gt16
+#endif
+
+#undef INTRI8_FLOAT
+#undef INTRI16_FLOAT
+#undef INTRI_GT8LT16_FLOAT
+#undef INTRI_GT16_FLOAT
+#undef INTRI_VTANH
+
+REGISTER_JITKERNEL(vtanh, VTanhKernel);
+
+#undef JITKERNEL_NEW_ACT_IMPL
+
+}  // namespace jitkernel
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/jit_kernel_lstm.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/math/jit_kernel.h"
+#include <string>
+#include "paddle/fluid/operators/math/jit_kernel_macro.h"
+#include "paddle/fluid/platform/enforce.h"
+#include "paddle/fluid/platform/macros.h"
+
+#ifdef __AVX__
+#include <immintrin.h>
+#endif
+
+namespace paddle {
+namespace operators {
+namespace math {
+#ifdef __AVX__
+namespace detail {
+__m256 Exp(__m256 a);
+}  // namespace detail
+#endif
+
+namespace jitkernel {
+namespace jit = platform::jit;
+
+#ifdef __AVX__
+typedef enum { kSigmoid, kRelu, kTanh, kIdentity } act_type;
+
+class AVXAct {
+ public:
+  virtual ~AVXAct() = default;
+  virtual __m256 Compute(__m256 x) const = 0;
+};
+
+template <act_type type>
+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);
+}
+
+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);
+}
+
+template <>
+__m256 AVXActImpl<kRelu>::Compute(__m256 x) const {
+  return _mm256_max_ps(x, _mm256_setzero_ps());
+}
+
+template <>
+__m256 AVXActImpl<kIdentity>::Compute(__m256 x) const {
+  return x;
+}
+#endif
+
+template <typename T>
+static std::shared_ptr<const VActKernel<T>> GetActKernel(
+    const std::string& type, int n) {
+  if (type == "sigmoid") {
+    return std::dynamic_pointer_cast<const VActKernel<T>>(
+        KernelPool::Instance().template Get<VSigmoidKernel<T>>(n));
+  } else if (type == "relu") {
+    return std::dynamic_pointer_cast<const VActKernel<T>>(
+        KernelPool::Instance().template Get<VReluKernel<T>>(n));
+  } else if (type == "tanh") {
+    return std::dynamic_pointer_cast<const VActKernel<T>>(
+        KernelPool::Instance().template Get<VTanhKernel<T>>(n));
+  } else if (type == "identity" || type == "") {
+    return std::dynamic_pointer_cast<const VActKernel<T>>(
+        KernelPool::Instance().template Get<VIdentityKernel<T>>(n));
+  }
+  PADDLE_THROW("Not support type: %s", type);
+  return nullptr;
+}
+
+/* LSTM JitKernel */
+template <typename T, jit::cpu_isa_t isa, jit_block>
+class LSTMKernelImpl : public LSTMKernel<T> {
+ public:
+  explicit LSTMKernelImpl(const std::string& act_gate,
+                          const std::string& act_cand,
+                          const std::string& act_cell, int d)
+      : LSTMKernel<T>() {
+    d_ = d;
+    d2_ = d * 2;
+    d3_ = d * 3;
+    act_gate_d3_ = GetActKernel<T>(act_gate, d3_);
+    act_gate_d_ = GetActKernel<T>(act_gate, d);
+    act_cand_d_ = GetActKernel<T>(act_cand, d);
+    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,
+                   T* checked) const override {
+    // gates: W_ch, W_ih, W_fh, W_oh
+    act_gate_d3_->Compute(gates + d_, gates + d_);
+
+    /* C_t = C_t-1 * fgated + cand_gated * igated */
+    act_cand_d_->Compute(gates, gates);
+    vmul_d_->Compute(gates, gates + d_, gates + d_);
+    vmul_d_->Compute(ct_1, gates + d2_, gates + d2_);
+    vadd_d_->Compute(gates + d_, gates + d2_, ct);
+
+    /* H_t = act_cell(C_t) * ogated */
+    act_cell_d_->Compute(ct, gates + d2_);
+    vmul_d_->Compute(gates + d2_, gates + d3_, ht);
+  }
+  void ComputeC1H1(T* gates, T* ct, T* ht, const T* wp_data) const override {
+    /* C_t = igated * cgated*/
+    act_gate_d_->Compute(gates + d_, gates + d_);
+    act_cand_d_->Compute(gates, gates);
+    vmul_d_->Compute(gates, gates + d_, ct);
+    /* H_t = act_cell(C_t) * ogated */
+    act_gate_d_->Compute(gates + d3_, gates + d3_);
+    act_cell_d_->Compute(ct, gates + d2_);
+    vmul_d_->Compute(gates + d2_, gates + d3_, ht);
+  }
+
+ private:
+  int d_, d2_, d3_;
+  std::shared_ptr<const VActKernel<T>> act_gate_d3_, act_gate_d_, act_cand_d_,
+      act_cell_d_;
+  std::shared_ptr<const VMulKernel<T>> vmul_d_;
+  std::shared_ptr<const VAddKernel<T>> vadd_d_;
+#ifdef __AVX__
+  std::unique_ptr<const AVXAct> avx_act_gate_, avx_act_cand_, avx_act_cell_;
+#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);                                                 \
+  }
+
+// TODO(TJ): optimize keq16
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+#endif
+
+/* Peephole JitKernel */
+template <typename T, jit::cpu_isa_t isa, jit_block>
+class PeepholeKernelImpl : public LSTMKernel<T> {
+ public:
+  explicit PeepholeKernelImpl(const std::string& act_gate,
+                              const std::string& act_cand,
+                              const std::string& act_cell, int d)
+      : LSTMKernel<T>() {
+    d_ = d;
+    d2_ = d * 2;
+    d3_ = d * 3;
+    act_gate_d_ = GetActKernel<T>(act_gate, d);
+    act_cand_d_ = GetActKernel<T>(act_cand, d);
+    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);
+    vadd_d2_ = KernelPool::Instance().template Get<VAddKernel<T>>(d2_);
+    act_gate_d2_ = GetActKernel<T>(act_gate, d2_);
+  }
+
+  void ComputeCtHt(T* gates, const T* ct_1, T* ct, T* ht, const T* wp_data,
+                   T* checked) const override {
+    /* get fgated and igated*/
+    vmul_d_->Compute(wp_data, ct_1, checked);
+    vmul_d_->Compute(wp_data + d_, ct_1, checked + d_);
+    vadd_d2_->Compute(checked, gates + d_, gates + d_);
+    act_gate_d2_->Compute(gates + d_, gates + d_);
+    /* C_t = C_t-1 * fgated + cand_gated * igated*/
+    act_cand_d_->Compute(gates, gates);
+    vmul_d_->Compute(gates, gates + d_, gates + d_);
+    vmul_d_->Compute(ct_1, gates + d2_, gates + d2_);
+    vadd_d_->Compute(gates + d_, gates + d2_, ct);
+    /* get ogated*/
+    vmul_d_->Compute(wp_data + d2_, ct, gates + d_);
+    vadd_d_->Compute(gates + d_, gates + d3_, gates + d3_);
+    act_gate_d_->Compute(gates + d3_, gates + d3_);
+    /* H_t = act_cell(C_t) * ogated */
+    act_cell_d_->Compute(ct, gates + d2_);
+    vmul_d_->Compute(gates + d2_, gates + d3_, ht);
+  }
+
+  void ComputeC1H1(T* gates, T* ct, T* ht, const T* wp_data) const override {
+    /* C_t = igated * cgated*/
+    act_gate_d_->Compute(gates + d_, gates + d_);
+    act_cand_d_->Compute(gates, gates);
+    vmul_d_->Compute(gates, gates + d_, ct);
+    /* get outgated, put W_oc * C_t on igated */
+    vmul_d_->Compute(wp_data + d2_, ct, gates + d_);
+    vadd_d_->Compute(gates + d_, gates + d3_, gates + d3_);
+    /* H_t = act_cell(C_t) * ogated */
+    act_gate_d_->Compute(gates + d3_, gates + d3_);
+    act_cell_d_->Compute(ct, gates + d2_);
+    vmul_d_->Compute(gates + d2_, gates + d3_, ht);
+  }
+
+ private:
+  int d_, d2_, d3_;
+  std::shared_ptr<const VActKernel<T>> act_gate_d2_, act_gate_d_, act_cand_d_,
+      act_cell_d_;
+  std::shared_ptr<const VMulKernel<T>> vmul_d_;
+  std::shared_ptr<const VAddKernel<T>> vadd_d_, vadd_d2_;
+};
+
+#define JITKERNEL_DECLARE_LSTM(ker_class, ker_dtype)                  \
+  template <>                                                         \
+  std::shared_ptr<const LSTMKernel<ker_dtype>>                        \
+  KernelPool::Get<LSTMKernel<ker_dtype>, const std::string&,          \
+                  const std::string&, const std::string&, int, bool>( \
+      const std::string& act_gate, const std::string& act_cand,       \
+      const std::string& act_cell, int d, bool use_peephole)
+
+#define JITKERNEL_KEY_LSTM(ker_key, dtype_key)                               \
+  #ker_key #dtype_key + std::to_string(d) + act_gate + act_cand + act_cell + \
+                                       (use_peephole ? "p" : "n")
+
+#define JITKERNEL_NEW_LSTM_IMPL(ker, dtype, isa, k)                    \
+  if (use_peephole) {                                                  \
+    p = std::dynamic_pointer_cast<ker<dtype>>(                         \
+        std::make_shared<PeepholeKernelImpl<dtype, isa, k>>(           \
+            act_gate, act_cand, act_cell, d));                         \
+  } else {                                                             \
+    p = std::dynamic_pointer_cast<ker<dtype>>(                         \
+        std::make_shared<ker##Impl<dtype, isa, k>>(act_gate, act_cand, \
+                                                   act_cell, d));      \
+  }
+
+REGISTER_JITKERNEL_ARGS(lstm, LSTMKernel, JITKERNEL_DECLARE_LSTM,
+                        JITKERNEL_KEY_LSTM, JITKERNEL_NEW_LSTM_IMPL);
+
+#undef INTRI8_FLOAT
+#undef JITKERNEL_DECLARE_LSTM
+#undef JITKERNEL_KEY_LSTM
+#undef JITKERNEL_NEW_LSTM_IMPL
+}  // namespace jitkernel
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/jit_kernel_macro.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/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+namespace jitkernel {
+
+namespace jit = platform::jit;
+
+#define SEARCH_BLOCK(macro_, ker, dtype, isa)                 \
+  if (d < AVX_FLOAT_BLOCK) {                                  \
+    macro_(ker, dtype, isa, kLT8);                            \
+  } else if (d == AVX_FLOAT_BLOCK) {                          \
+    macro_(ker, dtype, isa, kEQ8);                            \
+  } else if (d > AVX_FLOAT_BLOCK && d < AVX512_FLOAT_BLOCK) { \
+    macro_(ker, dtype, isa, kGT8LT16);                        \
+  } else if (d == AVX512_FLOAT_BLOCK) {                       \
+    macro_(ker, dtype, isa, kEQ16);                           \
+  } else {                                                    \
+    macro_(ker, dtype, isa, kGT16);                           \
+  }
+
+#define SEARCH_ISA_BLOCK(macro_, ker, dtype)        \
+  if (jit::MayIUse(jit::avx512f)) {                 \
+    SEARCH_BLOCK(macro_, ker, dtype, jit::avx512f); \
+  } else if (jit::MayIUse(jit::avx2)) {             \
+    SEARCH_BLOCK(macro_, ker, dtype, jit::avx2);    \
+  } else if (jit::MayIUse(jit::avx)) {              \
+    SEARCH_BLOCK(macro_, ker, dtype, jit::avx);     \
+  } else {                                          \
+    SEARCH_BLOCK(macro_, ker, dtype, jit::isa_any); \
+  }
+
+#define JITKERNEL_DECLARE(ker_class, ker_dtype) \
+  template <>                                   \
+  std::shared_ptr<const ker_class<ker_dtype>>   \
+  KernelPool::Get<ker_class<ker_dtype>, int>(int d)
+
+#define JITKERNEL_KEY(ker_key, dtype_key) \
+  #ker_key #dtype_key + std::to_string(d)
+
+#define JITKERNEL_NEW_IMPL(ker, dtype, isa, k) \
+  p = std::dynamic_pointer_cast<ker<dtype>>(   \
+      std::make_shared<ker##Impl<dtype, isa, k>>(d))
+
+#define JITKERNEL_WITH_DTYPE(ker_key, ker_class, ker_dtype, dtype_key, \
+                             marco_declare, macro_key, macro_impl)     \
+  marco_declare(ker_class, ker_dtype) {                                \
+    std::string key = macro_key(ker_key, dtype_key);                   \
+    if (kers_.find(key) == kers_.end()) {                              \
+      std::shared_ptr<ker_class<ker_dtype>> p;                         \
+      SEARCH_ISA_BLOCK(macro_impl, ker_class, ker_dtype);              \
+      kers_.insert({key, std::dynamic_pointer_cast<Kernel>(p)});       \
+      return p;                                                        \
+    }                                                                  \
+    return std::dynamic_pointer_cast<const ker_class<ker_dtype>>(      \
+        kers_.at(key));                                                \
+  }
+
+#define REGISTER_JITKERNEL(ker_key, ker_class)                           \
+  JITKERNEL_WITH_DTYPE(ker_key, ker_class, float, f, JITKERNEL_DECLARE,  \
+                       JITKERNEL_KEY, JITKERNEL_NEW_IMPL);               \
+  JITKERNEL_WITH_DTYPE(ker_key, ker_class, double, d, JITKERNEL_DECLARE, \
+                       JITKERNEL_KEY, JITKERNEL_NEW_IMPL)
+
+#define REGISTER_JITKERNEL_ARGS(ker_key, ker_class, marco_declare, macro_key,  \
+                                macro_impl)                                    \
+  JITKERNEL_WITH_DTYPE(ker_key, ker_class, float, f, marco_declare, macro_key, \
+                       macro_impl);                                            \
+  JITKERNEL_WITH_DTYPE(ker_key, ker_class, double, d, marco_declare,           \
+                       macro_key, macro_impl)
+
+#define FOR_EACH_ISA(macro_, block) \
+  macro_(jit::avx512f, block);      \
+  macro_(jit::avx2, block);         \
+  macro_(jit::avx, block);          \
+  macro_(jit::isa_any, block)
+
+#define FOR_EACH_BLOCK(macro_, isa) \
+  macro_(isa, kLT8);                \
+  macro_(isa, kEQ8);                \
+  macro_(isa, kGT8LT16);            \
+  macro_(isa, kEQ16);               \
+  macro_(isa, kGT16)
+
+#define FOR_EACH_ISA_BLOCK(macro_)      \
+  FOR_EACH_BLOCK(macro_, jit::avx512f); \
+  FOR_EACH_BLOCK(macro_, jit::avx2);    \
+  FOR_EACH_BLOCK(macro_, jit::avx);     \
+  FOR_EACH_BLOCK(macro_, jit::isa_any)
+
+}  // namespace jitkernel
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/selected_rows_functor.cc

@@ -12,10 +12,11 @@ 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 <map>
 #include <set>
 #include <vector>
 
-#include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
 
 namespace paddle {
@@ -150,6 +151,45 @@ template struct SelectedRowsAddTo<platform::CPUDeviceContext, double>;
 template struct SelectedRowsAddTo<platform::CPUDeviceContext, int>;
 template struct SelectedRowsAddTo<platform::CPUDeviceContext, int64_t>;
 
+template <typename T>
+struct SelectedRowsSumTo<platform::CPUDeviceContext, T> {
+  void operator()(const platform::CPUDeviceContext& context,
+                  const std::vector<framework::SelectedRows*>& input1,
+                  const std::vector<int64_t>& input2_offsets,
+                  framework::SelectedRows* input2) {
+    // Ensure all selected rows have the same height
+    size_t size = 0u;
+    for (auto iter = input1.begin(); iter != input1.end(); ++iter) {
+      auto& in_rows = (*iter)->rows();
+      size += in_rows.end() - in_rows.begin();
+      auto in1_height = (*iter)->height();
+      PADDLE_ENFORCE_EQ(in1_height, input2->height());
+    }
+    // concat rows
+    std::vector<int64_t> in2_rows;
+    in2_rows.reserve(in2_rows.size() + size);
+    for (auto iter = input1.begin(); iter != input1.end(); ++iter) {
+      const framework::Vector<int64_t>& in_rows = (*iter)->rows();
+      in2_rows.insert(in2_rows.end(), in_rows.begin(), in_rows.end());
+    }
+    input2->set_rows(in2_rows);
+
+    auto* in2_value = input2->mutable_value();
+    auto* in2_data = in2_value->data<T>();
+    auto blas = math::GetBlas<platform::CPUDeviceContext, T>(context);
+    size_t offset = 0u;
+    for (size_t i = 0u; i != input1.size(); ++i) {
+      auto& in_value = input1[i]->value();
+      const auto* in_data = in_value.data<T>();
+      offset += input2_offsets[i];
+      blas.VCOPY(in_value.numel(), in_data, in2_data + offset);
+    }
+  }
+};
+
+template struct SelectedRowsSumTo<platform::CPUDeviceContext, float>;
+template struct SelectedRowsSumTo<platform::CPUDeviceContext, double>;
+
 template <typename T>
 struct SelectedRowsAddToTensor<platform::CPUDeviceContext, T> {
   void operator()(const platform::CPUDeviceContext& context,
@@ -207,35 +247,45 @@ struct MergeAdd<platform::CPUDeviceContext, T> {
                   const framework::SelectedRows& input,
                   framework::SelectedRows* output) {
     framework::SelectedRows& out = *output;
-    auto input_rows = input.rows();
-    std::set<int64_t> row_set(input_rows.begin(), input_rows.end());
-    std::vector<int64_t> merge_rows(row_set.begin(), row_set.end());
+    std::vector<int64_t> input_rows(input.rows());
 
-    auto input_width = input.value().dims()[1];
-    out.set_rows(merge_rows);
+    std::map<int64_t, std::vector<int64_t>> merge_row_map;
+    for (size_t i = 0; i < input_rows.size(); ++i) {
+      merge_row_map[input_rows[i]].push_back(i);
+    }
+
+    std::vector<int64_t> merge_rows(merge_row_map.size());
+    size_t idx = 0;
+    int64_t input_width = input.value().dims()[1];
     out.set_height(input.height());
-    out.mutable_value()->mutable_data<T>(
+
+    T* out_data = out.mutable_value()->mutable_data<T>(
         framework::make_ddim(
             {static_cast<int64_t>(merge_rows.size()), input_width}),
         context.GetPlace());
-
-    math::SetConstant<platform::CPUDeviceContext, T> constant_functor;
-    constant_functor(context, out.mutable_value(), 0.0);
-
-    auto* out_data = out.mutable_value()->data<T>();
-    auto* input_data = input.value().data<T>();
-
-    for (size_t i = 0; i < input_rows.size(); i++) {
-      size_t out_i = FindPos(merge_rows, input_rows[i]);
-      for (int64_t j = 0; j < input_width; j++) {
-        out_data[out_i * input_width + j] += input_data[i * input_width + j];
+    const T* in_data = input.value().data<T>();
+
+    for (auto& row_pair : merge_row_map) {
+      auto* out_ptr = out_data + idx * input_width;
+      auto& rows = row_pair.second;
+      merge_rows[idx] = row_pair.first;
+      ++idx;
+      // rows.size() is always larger than 0
+      std::memcpy(out_ptr, in_data + rows[0] * input_width,
+                  sizeof(T) * input_width);
+
+      for (size_t i = 1; i < rows.size(); ++i) {
+        auto* in_ptr = in_data + rows[i] * input_width;
+        for (int64_t j = 0; j < input_width; ++j) {
+          out_ptr[j] += in_ptr[j];
+        }
       }
     }
+
+    out.set_rows(merge_rows);
   }
 };
 
-template struct MergeAdd<platform::CPUDeviceContext, float>;
-template struct MergeAdd<platform::CPUDeviceContext, double>;
 template struct MergeAdd<platform::CPUDeviceContext, int>;
 template struct MergeAdd<platform::CPUDeviceContext, int64_t>;
 

paddle/fluid/operators/math/selected_rows_functor.h

@@ -12,8 +12,14 @@ 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 <map>
+#include <vector>
+
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/selected_rows.h"
+#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/platform/device_context.h"
 
 #define INLINE_FOR2(sizei, sizej)     \
@@ -49,6 +55,15 @@ struct SelectedRowsAddTo {
                   const int64_t input2_offset, framework::SelectedRows* input2);
 };
 
+// input2 = [all input in input1] + input2
+template <typename DeviceContext, typename T>
+struct SelectedRowsSumTo {
+  void operator()(const DeviceContext& context,
+                  const std::vector<framework::SelectedRows*>& input1,
+                  const std::vector<int64_t>& input2_offsets,
+                  framework::SelectedRows* input2);
+};
+
 // input2 = input1 + input2
 template <typename DeviceContext, typename T>
 struct SelectedRowsAddToTensor {
@@ -70,6 +85,104 @@ struct MergeAdd {
                   framework::SelectedRows* output);
 };
 
+template <>
+struct MergeAdd<platform::CPUDeviceContext, float> {
+  framework::SelectedRows operator()(const platform::CPUDeviceContext& context,
+                                     const framework::SelectedRows& input) {
+    framework::SelectedRows out;
+    (*this)(context, input, &out);
+    return out;
+  }
+
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::SelectedRows& input,
+                  framework::SelectedRows* output) {
+    framework::SelectedRows& out = *output;
+    std::vector<int64_t> input_rows(input.rows());
+
+    std::map<int64_t, std::vector<int64_t>> merge_row_map;
+    for (size_t i = 0; i < input_rows.size(); ++i) {
+      merge_row_map[input_rows[i]].push_back(i);
+    }
+
+    std::vector<int64_t> merge_rows(merge_row_map.size());
+    size_t idx = 0;
+    int64_t input_width = input.value().dims()[1];
+    out.set_height(input.height());
+
+    auto* out_data = out.mutable_value()->mutable_data<float>(
+        framework::make_ddim(
+            {static_cast<int64_t>(merge_rows.size()), input_width}),
+        context.GetPlace());
+    auto* in_data = input.value().data<float>();
+
+    auto blas = GetBlas<platform::CPUDeviceContext, float>(context);
+    for (auto& row_pair : merge_row_map) {
+      auto* out_ptr = out_data + idx * input_width;
+      auto& rows = row_pair.second;
+      merge_rows[idx] = row_pair.first;
+      ++idx;
+      // rows.size() is always larger than 0
+      blas.VCOPY(input_width, in_data + rows[0] * input_width, out_ptr);
+
+      for (size_t i = 1; i < rows.size(); ++i) {
+        blas.AXPY(input_width, 1., in_data + rows[i] * input_width, out_ptr);
+      }
+    }
+
+    out.set_rows(merge_rows);
+  }
+};
+
+template <>
+struct MergeAdd<platform::CPUDeviceContext, double> {
+  framework::SelectedRows operator()(const platform::CPUDeviceContext& context,
+                                     const framework::SelectedRows& input) {
+    framework::SelectedRows out;
+    (*this)(context, input, &out);
+    return out;
+  }
+
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::SelectedRows& input,
+                  framework::SelectedRows* output) {
+    framework::SelectedRows& out = *output;
+    std::vector<int64_t> input_rows(input.rows());
+
+    std::map<int64_t, std::vector<int64_t>> merge_row_map;
+    for (size_t i = 0; i < input_rows.size(); ++i) {
+      merge_row_map[input_rows[i]].push_back(i);
+    }
+
+    std::vector<int64_t> merge_rows(merge_row_map.size());
+    size_t idx = 0;
+    int64_t input_width = input.value().dims()[1];
+    out.set_height(input.height());
+
+    auto* out_data = out.mutable_value()->mutable_data<double>(
+        framework::make_ddim(
+            {static_cast<int64_t>(merge_rows.size()), input_width}),
+        context.GetPlace());
+    auto* in_data = input.value().data<double>();
+
+    auto blas = GetBlas<platform::CPUDeviceContext, double>(context);
+    for (auto& row_pair : merge_row_map) {
+      auto* out_ptr = out_data + idx * input_width;
+      auto& rows = row_pair.second;
+      merge_rows[idx] = row_pair.first;
+      ++idx;
+      // rows.size() is always larger than 0
+      blas.VCOPY(input_width, in_data + rows[0] * input_width, out_ptr);
+
+      for (size_t i = 1; i < rows.size(); ++i) {
+        blas.AXPY(input_width, 1., in_data + rows[i] * input_width, out_ptr);
+      }
+    }
+
+    out.set_rows(merge_rows);
+  }
+};
+
 template <typename DeviceContext, typename T>
 struct Add {
   framework::SelectedRows operator()(const DeviceContext& context,

paddle/fluid/operators/math/selected_rows_functor_test.cc

@@ -219,3 +219,174 @@ TEST(selected_rows_functor, cpu_add_to) {
   // row9: 2.0 + 3.0
   EXPECT_EQ(tensor1_data[9 * row_numel + 6], 5.0);
 }
+
+TEST(selected_rows_functor, cpu_merge_add_float) {
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CPUDeviceContext ctx(cpu_place);
+  paddle::operators::math::SetConstant<paddle::platform::CPUDeviceContext,
+                                       float>
+      functor;
+  int64_t height = 10;
+  int64_t row_numel = 10;
+
+  std::vector<int64_t> rows{0, 4, 4, 7};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows{
+      new paddle::framework::SelectedRows(rows, height)};
+  auto* in_value = selected_rows->mutable_value();
+  in_value->mutable_data<float>(
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows.size()), row_numel}),
+      cpu_place);
+  functor(ctx, in_value, 1.0);
+
+  std::unique_ptr<paddle::framework::SelectedRows> output{
+      new paddle::framework::SelectedRows()};
+
+  paddle::operators::math::scatter::MergeAdd<paddle::platform::CPUDeviceContext,
+                                             float>
+      merge_add_functor;
+  merge_add_functor(ctx, *selected_rows, output.get());
+
+  auto out_height = output->height();
+  EXPECT_EQ(out_height, height);
+
+  auto& out_rows = output->rows();
+  EXPECT_EQ(out_rows[0], 0);
+  EXPECT_EQ(out_rows[1], 4);
+  EXPECT_EQ(out_rows[2], 7);
+
+  auto* out_data = output->value().data<float>();
+
+  EXPECT_EQ(out_data[0 * row_numel], 1.0);
+  EXPECT_EQ(out_data[1 * row_numel], 2.0);
+  EXPECT_EQ(out_data[2 * row_numel], 1.0);
+}
+
+TEST(selected_rows_functor, cpu_merge_add_int) {
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CPUDeviceContext ctx(cpu_place);
+  paddle::operators::math::SetConstant<paddle::platform::CPUDeviceContext, int>
+      functor;
+  int64_t height = 10;
+  int64_t row_numel = 10;
+
+  std::vector<int64_t> rows{0, 4, 4, 7};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows{
+      new paddle::framework::SelectedRows(rows, height)};
+  auto* in_value = selected_rows->mutable_value();
+  in_value->mutable_data<int>(
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows.size()), row_numel}),
+      cpu_place);
+  functor(ctx, in_value, 1);
+
+  std::unique_ptr<paddle::framework::SelectedRows> output{
+      new paddle::framework::SelectedRows()};
+
+  paddle::operators::math::scatter::MergeAdd<paddle::platform::CPUDeviceContext,
+                                             int>
+      merge_add_functor;
+  merge_add_functor(ctx, *selected_rows, output.get());
+
+  auto out_height = output->height();
+  EXPECT_EQ(out_height, height);
+
+  auto& out_rows = output->rows();
+  EXPECT_EQ(out_rows[0], 0);
+  EXPECT_EQ(out_rows[1], 4);
+  EXPECT_EQ(out_rows[2], 7);
+
+  auto* out_data = output->value().data<int>();
+
+  EXPECT_EQ(out_data[0 * row_numel], 1);
+  EXPECT_EQ(out_data[1 * row_numel], 2);
+  EXPECT_EQ(out_data[2 * row_numel], 1);
+}
+TEST(selected_rows_functor, cpu_sum_to) {
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CPUDeviceContext ctx(cpu_place);
+  paddle::operators::math::SetConstant<paddle::platform::CPUDeviceContext,
+                                       float>
+      functor;
+  int64_t height = 10;
+  int64_t row_numel = 10;
+  std::vector<int64_t> rows1{0, 4, 7};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows1{
+      new paddle::framework::SelectedRows(rows1, height)};
+  auto* in1_value = selected_rows1->mutable_value();
+  in1_value->mutable_data<float>(
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows1.size()), row_numel}),
+      cpu_place);
+  functor(ctx, in1_value, 1.0);
+  std::vector<int64_t> rows2{0, 5, 7, 9};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows2{
+      new paddle::framework::SelectedRows(rows2, height)};
+  auto* in2_value = selected_rows2->mutable_value();
+  in2_value->mutable_data<float>(
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows2.size()), row_numel}),
+      cpu_place);
+  functor(ctx, in2_value, 2.0);
+  std::unique_ptr<paddle::framework::SelectedRows> output{
+      new paddle::framework::SelectedRows()};
+  output->set_height(height);
+  auto* out_value = output->mutable_value();
+  // simplely concat two SelectedRows
+  out_value->mutable_data<float>(paddle::framework::make_ddim({7, 10}),
+                                 cpu_place);
+  paddle::operators::math::SelectedRowsSumTo<paddle::platform::CPUDeviceContext,
+                                             float>
+      sum_to_functor;
+  sum_to_functor(ctx, std::vector<paddle::framework::SelectedRows*>(
+                          {selected_rows1.get(), selected_rows2.get()}),
+                 std::vector<int64_t>({0, in1_value->numel()}), output.get());
+  auto out_height = output->height();
+  EXPECT_EQ(out_height, height);
+  auto& out_rows = output->rows();
+  // input1 rows
+  EXPECT_EQ(out_rows[0], 0);
+  EXPECT_EQ(out_rows[1], 4);
+  EXPECT_EQ(out_rows[2], 7);
+  // input2 rows
+  EXPECT_EQ(out_rows[3], 0);
+  EXPECT_EQ(out_rows[4], 5);
+  EXPECT_EQ(out_rows[5], 7);
+  EXPECT_EQ(out_rows[6], 9);
+  auto* out_data = output->value().data<float>();
+  // input1 value
+  EXPECT_EQ(out_data[0 * row_numel + 0], 1.0);
+  EXPECT_EQ(out_data[0 * row_numel + 8], 1.0);
+  EXPECT_EQ(out_data[1 * row_numel + 1], 1.0);
+  EXPECT_EQ(out_data[2 * row_numel + 6], 1.0);
+  // input2 value
+  EXPECT_EQ(out_data[3 * row_numel + 3], 2.0);
+  EXPECT_EQ(out_data[3 * row_numel + 8], 2.0);
+  EXPECT_EQ(out_data[4 * row_numel + 4], 2.0);
+  EXPECT_EQ(out_data[5 * row_numel + 7], 2.0);
+  EXPECT_EQ(out_data[6 * row_numel + 9], 2.0);
+  std::unique_ptr<paddle::framework::Tensor> tensor1{
+      new paddle::framework::Tensor()};
+  tensor1->mutable_data<float>(
+      paddle::framework::make_ddim({height, row_numel}), cpu_place);
+  functor(ctx, tensor1.get(), 3.0);
+  paddle::operators::math::SelectedRowsAddToTensor<
+      paddle::platform::CPUDeviceContext, float>
+      add_to_tensor_functor;
+  add_to_tensor_functor(ctx, *output, tensor1.get());
+  auto* tensor1_data = tensor1->data<float>();
+  // row0: 1.0 + 2.0 + 3.0
+  EXPECT_EQ(tensor1_data[0 * row_numel + 0], 6.0);
+  // row1: 3.0
+  EXPECT_EQ(tensor1_data[1 * row_numel + 1], 3.0);
+  // row4 : 1.0 + 3.0
+  EXPECT_EQ(tensor1_data[4 * row_numel + 6], 4.0);
+  // row5: 2.0 + 3.0
+  EXPECT_EQ(tensor1_data[5 * row_numel + 7], 5.0);
+  // row6: 3.0
+  EXPECT_EQ(tensor1_data[6 * row_numel + 1], 3.0);
+  // row7: 1.0 + 2.0 + 3.0
+  EXPECT_EQ(tensor1_data[7 * row_numel + 3], 6.0);
+  // row9: 2.0 + 3.0
+  EXPECT_EQ(tensor1_data[9 * row_numel + 6], 5.0);
+}

paddle/fluid/operators/math/sequence_pooling.cc

@@ -12,9 +12,11 @@ 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/sequence_pooling.h"
 #include <string>
+
+#include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/operators/math/sequence_pooling.h"
 
 namespace paddle {
 namespace operators {
@@ -180,6 +182,7 @@ class SequencePoolFunctor<platform::CPUDeviceContext, T> {
     }
     auto lod = input.lod()[0];
     auto& place = *context.eigen_device();
+    auto blas = math::GetBlas<platform::CPUDeviceContext, T>(context);
     for (int i = 0; i < static_cast<int>(lod.size()) - 1; ++i) {
       Tensor in_t =
           input.Slice(static_cast<int>(lod[i]), static_cast<int>(lod[i + 1]));
@@ -191,7 +194,14 @@ class SequencePoolFunctor<platform::CPUDeviceContext, T> {
       if (pooltype == "AVERAGE") {
         out_e.device(place) = in_e.mean(Eigen::array<int, 1>({{0}}));
       } else if (pooltype == "SUM") {
-        out_e.device(place) = in_e.sum(Eigen::array<int, 1>({{0}}));
+        if (h > 0) {
+          const T* in_data = in_t.data<T>();
+          T* out_data = out_t.mutable_data<T>(context.GetPlace());
+          blas.VCOPY(w, in_data, out_data);
+          for (int64_t r = 1; r != h; ++r) {
+            blas.AXPY(w, 1., in_data + r * w, out_data);
+          }
+        }
       } else if (pooltype == "SQRT") {
         out_e.device(place) = in_e.sum(Eigen::array<int, 1>({{0}})) /
                               std::sqrt(static_cast<T>(h));
@@ -223,6 +233,7 @@ class SequencePoolGradFunctor<platform::CPUDeviceContext, T> {
     }
     auto lod = in_grad->lod()[0];
     auto& place = *context.eigen_device();
+    auto blas = math::GetBlas<platform::CPUDeviceContext, T>(context);
     for (int i = 0; i < static_cast<int>(lod.size()) - 1; ++i) {
       auto in_g_t = in_grad->Slice(static_cast<int>(lod[i]),
                                    static_cast<int>(lod[i + 1]));
@@ -237,7 +248,11 @@ class SequencePoolGradFunctor<platform::CPUDeviceContext, T> {
       if (pooltype == "AVERAGE") {
         in_g_e.device(place) = (out_g_e / static_cast<T>(h)).broadcast(bcast);
       } else if (pooltype == "SUM") {
-        in_g_e.device(place) = (out_g_e).broadcast(bcast);
+        const T* out_g_data = out_g_t.data<T>();
+        T* in_g_data = in_g_t.mutable_data<T>(context.GetPlace());
+        for (int r = 0; r != h; ++r) {
+          blas.VCOPY(w, out_g_data, in_g_data + r * w);
+        }
       } else if (pooltype == "SQRT") {
         in_g_e.device(place) =
             (out_g_e / std::sqrt(static_cast<T>(h))).broadcast(bcast);

paddle/fluid/operators/momentum_op.cc

@@ -33,6 +33,11 @@ class MomentumOp : public framework::OperatorWithKernel {
                    "Input(velocity) of Momentum should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("LearningRate"),
                    "Input(LearningRate) of Momentum should not be null.");
+    PADDLE_ENFORCE(
+        ctx->GetInputsVarType("Param").front() ==
+            framework::proto::VarType::LOD_TENSOR,
+        "The input var's type should be LoDTensor, but the received is %s",
+        ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front());
 
     PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
                    "Output(ParamOut) of Momentum should not be null.");

paddle/fluid/operators/momentum_op.cu

@@ -46,6 +46,17 @@ template <typename T>
 class MomentumOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    const auto* param_var = ctx.InputVar("Param");
+    PADDLE_ENFORCE(param_var->IsType<framework::LoDTensor>(),
+                   "The Var(%s)'s type should be LoDTensor, "
+                   "but the received is %s",
+                   ctx.Inputs("Param").front(), param_var->Type().name());
+    const auto* grad_var = ctx.InputVar("Grad");
+    PADDLE_ENFORCE(grad_var->IsType<framework::LoDTensor>(),
+                   "The Var(%s)'s type should be LoDTensor, "
+                   "but the received is %s",
+                   ctx.Inputs("Grad").front(), grad_var->Type().name());
+
     auto param_out = ctx.Output<framework::Tensor>("ParamOut");
     auto velocity_out = ctx.Output<framework::Tensor>("VelocityOut");
     auto param = ctx.Input<framework::Tensor>("Param");

paddle/fluid/operators/momentum_op.h

@@ -23,6 +23,12 @@ template <typename T>
 class MomentumOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    const auto* param_var = ctx.InputVar("Param");
+    PADDLE_ENFORCE(param_var->IsType<framework::LoDTensor>(),
+                   "The Var(%s)'s type should be LoDTensor, "
+                   "but the received is %s",
+                   ctx.Inputs("Param").front(), param_var->Type().name());
+
     auto param_out = ctx.Output<framework::Tensor>("ParamOut");
     auto velocity_out = ctx.Output<framework::Tensor>("VelocityOut");
     auto param = ctx.Input<framework::Tensor>("Param");

paddle/fluid/operators/parallel_do_op.cc

@@ -397,6 +397,24 @@ class ParallelDoGradOpShapeInference : public framework::InferShapeBase {
   }
 };
 
+class ParallelDoGradOpVarTypeInference : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc &op_desc,
+                  framework::BlockDesc *block) const override {
+    framework::BlockDesc *sub_block =
+        boost::get<framework::BlockDesc *>(op_desc.GetAttr(kParallelBlock));
+    for (auto &out_vars : op_desc.Outputs()) {
+      for (auto &out_var : out_vars.second) {
+        auto &var = block->FindRecursiveOrCreateVar(out_var);
+        auto sub_var = sub_block->FindRecursiveOrCreateVar(out_var);
+        if (sub_var.GetType() != var.GetType()) {
+          var.SetType(sub_var.GetType());
+        }
+      }
+    }
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -404,4 +422,5 @@ REGISTER_OPERATOR(parallel_do, paddle::operators::ParallelDoOp,
                   paddle::operators::ParallelDoOpProtoMaker,
                   paddle::operators::ParallelDoGradOpDescMaker);
 REGISTER_OPERATOR(parallel_do_grad, paddle::operators::ParallelDoGradOp,
-                  paddle::operators::ParallelDoGradOpShapeInference);
+                  paddle::operators::ParallelDoGradOpShapeInference,
+                  paddle::operators::ParallelDoGradOpVarTypeInference);

paddle/fluid/operators/rmsprop_op.cc

@@ -32,6 +32,11 @@ class RmspropOp : public framework::OperatorWithKernel {
                    "Input(Grad) of RmspropOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Moment"),
                    "Input(Moment) of RmspropOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->GetInputsVarType("Param").front() ==
+            framework::proto::VarType::LOD_TENSOR,
+        "The input var's type should be LoDTensor, but the received is %s",
+        ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front());
 
     PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
                    "Output(param_out) of RmspropOp should not be null.");