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/jit_kernel_test.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 <sys/time.h>
+#include <cmath>    // for exp
+#include <cstring>  // for memcpy
+#include <string>
+#include <vector>
+#include "gflags/gflags.h"
+#include "glog/logging.h"
+#include "gtest/gtest.h"
+
+#ifdef PADDLE_WITH_MKLML
+#include "paddle/fluid/platform/dynload/mklml.h"
+#endif
+
+#ifdef __AVX__
+#include <immintrin.h>
+#endif
+
+constexpr int repeat = 20000;
+
+inline double GetCurrentUS() {
+  struct timeval time;
+  gettimeofday(&time, NULL);
+  return 1e+6 * time.tv_sec + time.tv_usec;
+}
+
+template <typename T>
+void RandomVec(const int n, T* a, const T lower = static_cast<T>(-20.f),
+               const T upper = static_cast<T>(20.f)) {
+  static unsigned int seed = 100;
+  std::mt19937 rng(seed++);
+  std::uniform_real_distribution<double> uniform_dist(0, 1);
+  for (int i = 0; i < n; ++i) {
+    a[i] = static_cast<T>(uniform_dist(rng) * (upper - lower) + lower);
+  }
+}
+
+void vrelu_ref(const int n, const float* x, float* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i] > 0.f ? x[i] : 0.f;
+  }
+}
+
+#if defined __AVX__ || defined __AVX2__
+void vrelu_intri8(const int n, const float* x, float* y) {
+  __m256 tmp = _mm256_loadu_ps(x);
+  tmp = _mm256_max_ps(tmp, _mm256_setzero_ps());
+  _mm256_storeu_ps(y, tmp);
+}
+#endif
+
+TEST(JitKernel, vrelu) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
+    std::vector<float> x(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data(), -10.f, 1.f);
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VReluKernel<float>>(d);
+    const float* x_data = x.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vrelu_ref(d, x_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+#if defined __AVX__ || defined __AVX2__
+    if (d == 8) {
+      auto si0 = GetCurrentUS();
+      for (int i = 0; i < repeat; ++i) {
+        vrelu_intri8(d, x_data, zref_data);
+      }
+      auto si1 = GetCurrentUS();
+      VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat;
+    }
+#endif
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+void vaddbias_ref(const int n, const float a, const float* x, float* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i] + a;
+  }
+}
+
+TEST(JitKernel, vaddbias) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 64, 100, 128, 256}) {
+    std::vector<float> x(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data(), -2.f, 2.f);
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VAddBiasKernel<float>>(d);
+    const float a = 2.f;
+    const float* x_data = x.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vaddbias_ref(d, a, x_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(a, x_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+void vexp_ref(const int n, const float* x, float* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = std::exp(x[i]);
+  }
+}
+
+#ifdef PADDLE_WITH_MKLML
+void vexp_mkl(const int n, const float* x, float* y) {
+  paddle::platform::dynload::vsExp(n, x, y);
+}
+#endif
+
+TEST(JitKernel, vexp) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 128, 256}) {
+    std::vector<float> x(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data(), -2.f, 2.f);
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VExpKernel<float>>(d);
+    const float* x_data = x.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vexp_ref(d, x_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+
+#ifdef PADDLE_WITH_MKLML
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vexp_mkl(d, x_data, zref_data);
+    }
+    auto tmkle = GetCurrentUS();
+#endif
+
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+#ifdef PADDLE_WITH_MKLML
+            << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, "
+#else
+            << " us, "
+#endif
+            << "tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+inline float _sigmoid(float x) {
+  const float min = SIGMOID_THRESHOLD_MIN;
+  const float max = SIGMOID_THRESHOLD_MAX;
+  float tmp = (x < min) ? min : ((x > max) ? max : x);
+  return 1.f / (1.f + std::exp(-tmp));
+}
+
+void vsigmoid_ref(const int n, const float* x, float* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = _sigmoid(x[i]);
+  }
+}
+
+void vsigmoid_better(
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VExpKernel<float>>& vexp,
+    const int n, const float* x, float* y) {
+  const float min = SIGMOID_THRESHOLD_MIN;
+  const float max = SIGMOID_THRESHOLD_MAX;
+  for (int i = 0; i < n; ++i) {
+    y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]);
+    y[i] = 0.f - y[i];
+  }
+  vexp->Compute(y, y);
+  for (int i = 0; i < n; ++i) {
+    y[i] = 1.f / (1.f + y[i]);
+  }
+}
+
+TEST(JitKernel, vsigmoid) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) {
+    std::vector<float> x(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data(), -2.f, 2.f);
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VSigmoidKernel<float>>(d);
+    const auto& vexp =
+        jit::KernelPool::Instance().template Get<jit::VExpKernel<float>>(d);
+    const float* x_data = x.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vsigmoid_better(vexp, d, x_data, zref_data);
+    }
+    auto tmkle = GetCurrentUS();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vsigmoid_ref(d, x_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, better(jit exp) takes: " << (tmkle - tmkls) / repeat
+            << " us, tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+inline float _tanh(float x) { return 2.f * _sigmoid(2.f * x) - 1.f; }
+
+void vtanh_ref(const int n, const float* x, float* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = _tanh(x[i]);
+  }
+}
+
+void vtanh_better(
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VScalKernel<float>>& vscal,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VSigmoidKernel<float>>&
+        vsigmoid,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VAddBiasKernel<float>>&
+        vaddbias,
+    const int n, const float* x, float* y) {
+  vscal->Compute(2.f, x, y);
+  vsigmoid->Compute(y, y);
+  vscal->Compute(2.f, y);
+  vaddbias->Compute(-1.f, y, y);
+}
+
+TEST(JitKernel, vtanh) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) {
+    std::vector<float> x(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data(), -2.f, 2.f);
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VTanhKernel<float>>(d);
+    const auto& vscal =
+        jit::KernelPool::Instance().template Get<jit::VScalKernel<float>>(d);
+    const auto& vsigmoid =
+        jit::KernelPool::Instance().template Get<jit::VSigmoidKernel<float>>(d);
+    const auto& vaddbias =
+        jit::KernelPool::Instance().template Get<jit::VAddBiasKernel<float>>(d);
+    const float* x_data = x.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vtanh_better(vscal, vsigmoid, vaddbias, d, x_data, zref_data);
+    }
+    auto tmkle = GetCurrentUS();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vtanh_ref(d, x_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, better(jit exp) takes: " << (tmkle - tmkls) / repeat
+            << " us, tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+void lstm_ctht_ref(
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VSigmoidKernel<float>>&
+        vsigmoid_3d,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VTanhKernel<float>>& vtanh_d,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VExpKernel<float>>& vexp_1,
+    const int d, float* gates, const float* ct_1, float* ct, float* ht) {
+  vsigmoid_3d->Compute(gates + d, gates + d);
+  vtanh_d->Compute(gates, gates);
+  const float *i = gates + d, *f = gates + d * 2, *o = gates + d * 3;
+  const float min = SIGMOID_THRESHOLD_MIN;
+  const float max = SIGMOID_THRESHOLD_MAX;
+  for (int k = 0; k < d; ++k) {
+    // C_t = C_t-1 * fgated + cand_gated * igated
+    ct[k] = ct_1[k] * f[k] + gates[k] * i[k];
+    // H_t = act_cell(C_t) * ogated
+    float tmp = ct[k] * 2;
+    tmp = 0.f - ((tmp < min) ? min : ((tmp > max) ? max : tmp));
+    vexp_1->Compute(&tmp, &tmp);
+    tmp = 2.f / (1.f + tmp) - 1.f;
+    ht[k] = tmp * o[k];
+  }
+}
+
+void lstm_ctht_better(
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VSigmoidKernel<float>>&
+        vsigmoid_3d,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VTanhKernel<float>>& vtanh_d,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VMulKernel<float>>& vmul_d,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VAddKernel<float>>& vadd_d,
+    const int d, float* gates, const float* ct_1, float* ct, float* ht) {
+  int d2 = d * 2;
+  vsigmoid_3d->Compute(gates + d, gates + d);
+  vtanh_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 */
+  vtanh_d->Compute(ct, gates + d2);
+  vmul_d->Compute(gates + d2, gates + d * 3, ht);
+}
+
+TEST(JitKernel, lstm) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 32, 64, 100}) {
+    int d4 = d * 4;
+    int d3 = d * 3;
+    std::vector<float> x(d4), xref(d4);
+    std::vector<float> ct_1(d), ct_tgt(d), ht_tgt(d);
+    std::vector<float> ct_ref(d), ht_ref(d);
+    RandomVec<float>(d4, x.data(), -2.f, 2.f);
+    RandomVec<float>(d, ct_1.data(), -2.f, 2.f);
+    memcpy(xref.data(), x.data(), sizeof(float) * d4);
+    std::string act_gate = "sigmoid", act_cand = "tanh", act_cell = "tanh";
+    const auto& ker =
+        jit::KernelPool::Instance()
+            .template Get<jit::LSTMKernel<float>, const std::string&,
+                          const std::string&, const std::string&>(
+                act_gate, act_cand, act_cell, d, false);
+    // below kernels are used to compute refer
+    const auto& vsigmoid_3d =
+        jit::KernelPool::Instance().template Get<jit::VSigmoidKernel<float>>(
+            d3);
+    const auto& vtanh_d =
+        jit::KernelPool::Instance().template Get<jit::VTanhKernel<float>>(d);
+    const auto& vexp_1 =
+        jit::KernelPool::Instance().template Get<jit::VExpKernel<float>>(1);
+    const auto& vmul_d =
+        jit::KernelPool::Instance().template Get<jit::VMulKernel<float>>(d);
+    const auto& vadd_d =
+        jit::KernelPool::Instance().template Get<jit::VAddKernel<float>>(d);
+
+    float* x_data = x.data();
+    float* xref_data = xref.data();
+    const float* ct_1_data = ct_1.data();
+    float* ct_tgt_data = ct_tgt.data();
+    float* ht_tgt_data = ht_tgt.data();
+    float* ct_ref_data = ct_ref.data();
+    float* ht_ref_data = ht_ref.data();
+    // compute once to check correctness
+    lstm_ctht_ref(vsigmoid_3d, vtanh_d, vexp_1, d, xref_data, ct_1_data,
+                  ct_ref_data, ht_ref_data);
+    ker->ComputeCtHt(x_data, ct_1_data, ct_tgt_data, ht_tgt_data);
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ct_tgt_data[i], ct_ref_data[i], 1e-3);
+      EXPECT_NEAR(ht_tgt_data[i], ht_ref_data[i], 1e-3);
+    }
+
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      lstm_ctht_better(vsigmoid_3d, vtanh_d, vmul_d, vadd_d, d, xref_data,
+                       ct_1_data, ct_ref_data, ht_ref_data);
+    }
+    auto tmkle = GetCurrentUS();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      lstm_ctht_ref(vsigmoid_3d, vtanh_d, vexp_1, d, xref_data, ct_1_data,
+                    ct_ref_data, ht_ref_data);
+    }
+    auto trefe = GetCurrentUS();
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->ComputeCtHt(x_data, ct_1_data, ct_tgt_data, ht_tgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, better(jit) takes: " << (tmkle - tmkls) / repeat
+            << " us, tgt takes: " << (ttgte - ttgts) / repeat;
+  }
+}
+
+void vscal_ref(const int n, const float a, const float* x, float* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = a * x[i];
+  }
+}
+void vscal_inp_ref(const int n, const float a, float* x) {
+  for (int i = 0; i < n; ++i) {
+    x[i] = a * x[i];
+  }
+}
+#if defined __AVX__ || defined __AVX2__
+void vscal_intri8(const int n, const float a, const float* x, float* y) {
+  __m256 tmp;
+  __m256 scalar = _mm256_set1_ps(a);
+  tmp = _mm256_loadu_ps(x);
+  tmp = _mm256_mul_ps(tmp, scalar);
+  _mm256_storeu_ps(y, tmp);
+}
+void vscal_inp_intri8(const int n, const float a, float* x) {
+  __m256 tmp;
+  __m256 scalar = _mm256_set1_ps(a);
+  tmp = _mm256_loadu_ps(x);
+  tmp = _mm256_mul_ps(tmp, scalar);
+  _mm256_storeu_ps(x, tmp);
+}
+#endif
+
+#ifdef PADDLE_WITH_MKLML
+void vscal_inp_mkl(const int n, const float a, float* x) {
+  paddle::platform::dynload::cblas_sscal(n, a, x, 1);
+}
+#endif
+
+TEST(JitKernel, vscal) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
+    std::vector<float> x(d), y(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data());
+    std::memcpy(y.data(), x.data(), sizeof(float) * d);
+    float a = 2.f;
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VScalKernel<float>>(d);
+    const float* x_data = x.data();
+    float* y_data = y.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vscal_ref(d, a, x_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+    auto trefs1 = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vscal_inp_ref(d, a, y_data);
+    }
+    auto trefe1 = GetCurrentUS();
+
+#ifdef PADDLE_WITH_MKLML
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vscal_inp_mkl(d, a, y_data);
+    }
+    auto tmkle = GetCurrentUS();
+#endif
+
+#if defined __AVX__ || defined __AVX2__
+    if (d == 8) {
+      auto si0 = GetCurrentUS();
+      for (int i = 0; i < repeat; ++i) {
+        vscal_intri8(d, a, x_data, zref_data);
+      }
+      auto si1 = GetCurrentUS();
+      auto si2 = GetCurrentUS();
+      for (int i = 0; i < repeat; ++i) {
+        vscal_inp_intri8(d, a, y_data);
+      }
+      auto si3 = GetCurrentUS();
+      VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat
+              << " us, inplace: " << (si3 - si2) / repeat;
+    }
+#endif
+
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(a, x_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+    auto ttgts1 = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(a, y_data);
+    }
+    auto ttgte1 = GetCurrentUS();
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, inplace takes: " << (trefe1 - trefs1) / repeat
+#ifdef PADDLE_WITH_MKLML
+            << " us, mkl inplace takes: " << (tmkle - tmkls) / repeat << " us, "
+#else
+            << " us, "
+#endif
+            << "tgt takes: " << (ttgte - ttgts) / repeat
+            << "us, tgt inplace takes: " << (ttgte1 - ttgts1) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+void vmul_ref(const int n, const float* x, const float* y, float* z) {
+  for (int i = 0; i < n; ++i) {
+    z[i] = x[i] * y[i];
+  }
+}
+
+#if defined __AVX__ || defined __AVX2__
+void vmul_intri8(const int n, const float* x, const float* y, float* z) {
+  __m256 tmpx, tmpy;
+  tmpx = _mm256_loadu_ps(x);
+  tmpy = _mm256_loadu_ps(y);
+  tmpx = _mm256_mul_ps(tmpx, tmpy);
+  _mm256_storeu_ps(z, tmpx);
+}
+#endif
+
+#ifdef PADDLE_WITH_MKLML
+void vmul_mkl(const int n, const float* x, const float* y, float* z) {
+  paddle::platform::dynload::vsMul(n, x, y, z);
+}
+#endif
+
+TEST(JitKernel, vmul) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
+    std::vector<float> x(d), y(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data());
+    RandomVec<float>(d, y.data());
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VMulKernel<float>>(d);
+    const float* x_data = x.data();
+    const float* y_data = y.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vmul_ref(d, x_data, y_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+
+#ifdef PADDLE_WITH_MKLML
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vmul_mkl(d, x_data, y_data, zref_data);
+    }
+    auto tmkle = GetCurrentUS();
+#endif
+
+#if defined __AVX__ || defined __AVX2__
+    if (d == 8) {
+      auto si0 = GetCurrentUS();
+      for (int i = 0; i < repeat; ++i) {
+        vmul_intri8(d, x_data, y_data, zref_data);
+      }
+      auto si1 = GetCurrentUS();
+      VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat;
+    }
+#endif
+
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, y_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+#ifdef PADDLE_WITH_MKLML
+            << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, "
+#else
+            << " us, "
+#endif
+            << "tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+void vadd_ref(const int n, const float* x, const float* y, float* z) {
+  for (int i = 0; i < n; ++i) {
+    z[i] = x[i] + y[i];
+  }
+}
+
+#if defined __AVX__ || defined __AVX2__
+void vadd_intri8(const int n, const float* x, const float* y, float* z) {
+  __m256 tmpx, tmpy;
+  tmpx = _mm256_loadu_ps(x);
+  tmpy = _mm256_loadu_ps(y);
+  tmpx = _mm256_add_ps(tmpx, tmpy);
+  _mm256_storeu_ps(z, tmpx);
+}
+#endif
+
+#ifdef PADDLE_WITH_MKLML
+void vadd_mkl(const int n, const float* x, const float* y, float* z) {
+  paddle::platform::dynload::vsAdd(n, x, y, z);
+}
+#endif
+
+TEST(JitKernel, vadd) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
+    std::vector<float> x(d), y(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data());
+    RandomVec<float>(d, y.data());
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VAddKernel<float>>(d);
+    const float* x_data = x.data();
+    const float* y_data = y.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vadd_ref(d, x_data, y_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+
+#ifdef PADDLE_WITH_MKLML
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vadd_mkl(d, x_data, y_data, zref_data);
+    }
+    auto tmkle = GetCurrentUS();
+#endif
+
+#if defined __AVX__ || defined __AVX2__
+    if (d == 8) {
+      auto si0 = GetCurrentUS();
+      for (int i = 0; i < repeat; ++i) {
+        vadd_intri8(d, x_data, y_data, zref_data);
+      }
+      auto si1 = GetCurrentUS();
+      VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat;
+    }
+#endif
+
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, y_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+#ifdef PADDLE_WITH_MKLML
+            << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, "
+#else
+            << " us, "
+#endif
+            << "tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+TEST(JitKernel, pool) {
+  namespace jit = paddle::operators::math::jitkernel;
+  const int frame_size = 4;
+  std::string act_gate = "sigmoid", act_cand = "tanh", act_cell = "tanh";
+  const auto& plstm1 =
+      jit::KernelPool::Instance()
+          .template Get<jit::LSTMKernel<float>, const std::string&,
+                        const std::string&, const std::string&>(
+              act_gate, act_cand, act_cell, frame_size, false);
+  const auto& plstm2 =
+      jit::KernelPool::Instance()
+          .template Get<jit::LSTMKernel<float>, const std::string&,
+                        const std::string&, const std::string&>(
+              act_gate, act_cand, act_cell, frame_size, false);
+  const auto& peephole =
+      jit::KernelPool::Instance()
+          .template Get<jit::LSTMKernel<float>, const std::string&,
+                        const std::string&, const std::string&>(
+              act_gate, act_cand, act_cell, frame_size, true);
+  EXPECT_TRUE(plstm1 != peephole);
+
+  const auto& pvmul_f =
+      jit::KernelPool::Instance().template Get<jit::VMulKernel<float>>(4);
+  EXPECT_TRUE(std::dynamic_pointer_cast<const jit::Kernel>(plstm2) !=
+              std::dynamic_pointer_cast<const jit::Kernel>(pvmul_f));
+
+  const auto& pvmul_d =
+      jit::KernelPool::Instance().template Get<jit::VMulKernel<double>>(4);
+  EXPECT_TRUE(std::dynamic_pointer_cast<const jit::Kernel>(pvmul_f) !=
+              std::dynamic_pointer_cast<const jit::Kernel>(pvmul_d));
+
+  const auto& pvmul_from_key = jit::KernelPool::Instance().Get("vmulf4");
+  EXPECT_EQ(pvmul_f, pvmul_from_key);
+  const auto& pvmul_from_key2 = jit::KernelPool::Instance().Get("vmulf5");
+  EXPECT_TRUE(pvmul_from_key2 == nullptr);
+}

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/reader/blocking_queue.h

@@ -31,8 +31,8 @@ class BlockingQueue {
   // is a workaround and a simplified version of framework::Channel as it
   // doesn't support GPU and it implements on buffered blocking queue.
  public:
-  explicit BlockingQueue(size_t capacity)
-      : capacity_(capacity), closed_(false) {
+  explicit BlockingQueue(size_t capacity, bool speed_test_mode = false)
+      : capacity_(capacity), speed_test_mode_(speed_test_mode), closed_(false) {
     PADDLE_ENFORCE_GT(
         capacity_, 0,
         "The capacity of a reader::BlockingQueue must be greater than 0.");
@@ -72,7 +72,9 @@ class BlockingQueue {
     if (!queue_.empty()) {
       PADDLE_ENFORCE_NOT_NULL(elem);
       *elem = queue_.front();
-      queue_.pop_front();
+      if (LIKELY(!speed_test_mode_)) {
+        queue_.pop_front();
+      }
       send_cv_.notify_one();
       return true;
     } else {
@@ -114,6 +116,7 @@ class BlockingQueue {
 
  private:
   size_t capacity_;
+  bool speed_test_mode_;
   bool closed_;
   std::deque<T> queue_;
 

paddle/fluid/operators/reader/lod_tensor_blocking_queue.h

@@ -33,8 +33,9 @@ class LoDTensorBlockingQueue {
 
  private:
   LoDTensorBlockingQueue(size_t capacity,
-                         const std::vector<framework::DDim>& dims)
-      : queue_(capacity), dims_(dims) {}
+                         const std::vector<framework::DDim>& dims,
+                         bool speed_test_mode = false)
+      : queue_(capacity, speed_test_mode), dims_(dims) {}
 
  public:
   bool Push(const std::vector<framework::LoDTensor>& lod_tensor_vec) {
@@ -69,11 +70,12 @@ class LoDTensorBlockingQueue {
 
 class LoDTensorBlockingQueueHolder {
  public:
-  void InitOnce(size_t capacity, const std::vector<framework::DDim>& dims) {
+  void InitOnce(size_t capacity, const std::vector<framework::DDim>& dims,
+                bool speed_test_mode = false) {
     PADDLE_ENFORCE(
         queue_ == nullptr,
         "LoDTensorBlockingQueueHolder::InitOnce() can only be called once");
-    queue_.reset(new LoDTensorBlockingQueue(capacity, dims));
+    queue_.reset(new LoDTensorBlockingQueue(capacity, dims, speed_test_mode));
   }
 
   inline const std::shared_ptr<LoDTensorBlockingQueue>& GetQueue() const {

paddle/fluid/operators/reader/reader_blocking_queue_test.cc

@@ -217,3 +217,27 @@ TEST(BlockingQueue, MyClassTest) {
   q.Receive(&b);
   EXPECT_EQ(a.val_, b.val_);
 }
+
+TEST(BlockingQueue, speed_test_mode) {
+  size_t queue_size = 10;
+  BlockingQueue<size_t> q1(queue_size, false);
+  for (size_t i = 0; i < queue_size; ++i) {
+    q1.Send(i);
+  }
+  size_t b;
+  for (size_t i = 0; i < queue_size; ++i) {
+    q1.Receive(&b);
+    EXPECT_EQ(b, i);
+  }
+  EXPECT_EQ(q1.Size(), 0);
+
+  BlockingQueue<size_t> q2(queue_size, true);
+  for (size_t i = 0; i < queue_size; ++i) {
+    q2.Send(i);
+  }
+  for (size_t i = 0; i < queue_size; ++i) {
+    q2.Receive(&b);
+    EXPECT_EQ(b, 0);
+  }
+  EXPECT_EQ(q2.Size(), queue_size);
+}

paddle/fluid/operators/reshape_op.cc

@@ -164,7 +164,7 @@ dimension value will be copied from Input(X) at runtime. Note that the index of
 [2, 3, 4], Attr(shape) = [2, 3, 2, 0] is an invalid input.
 
 3. Input(Shape) has a higher priority than Attr(shape) if it is provided, while
-Attr(shape) still should be set correctly to gurantee shape inference in 
+Attr(shape) still should be set correctly to gurantee shape inference in
 compile-time.
 
 )DOC");
@@ -259,7 +259,6 @@ class Reshape2Op : public ReshapeOp {
       : ReshapeOp(type, inputs, outputs, attrs) {}
 
   void InferShape(framework::InferShapeContext *ctx) const override {
-    ReshapeOp::InferShape(ctx);
     PADDLE_ENFORCE(ctx->HasOutput("XShape"),
                    "Output(XShape) of ReshapeOp should not be null.");
     const auto &x_dims = ctx->GetInputDim("X");
@@ -270,6 +269,8 @@ class Reshape2Op : public ReshapeOp {
     }
     ctx->SetOutputDim("XShape", framework::make_ddim(xshape_dims));
     ctx->ShareLoD("X", /*->*/ "XShape");
+
+    ReshapeOp::InferShape(ctx);
   }
 };
 

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.");

paddle/fluid/operators/rmsprop_op.h

@@ -13,66 +13,254 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <math.h>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/algorithm.h"
+#include "paddle/fluid/operators/math/selected_rows_functor.h"
+#include "paddle/fluid/platform/for_range.h"
 
 namespace paddle {
 namespace operators {
 
-using Tensor = framework::Tensor;
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
+template <typename T>
+struct DenseRmspropGradFunctor {
+  inline explicit DenseRmspropGradFunctor(const T *grad) : grad_(grad) {}
+
+  HOSTDEVICE inline T operator()(int64_t idx) const { return grad_[idx]; }
+
+  const T *grad_;
+};
+
+template <typename T>
+struct SparseRmspropGradFunctor {
+  inline SparseRmspropGradFunctor(const T *grad, const int64_t *rows,
+                                  int64_t row_numel, int64_t row_count)
+      : grad_(grad),
+        rows_(rows),
+        row_numel_(row_numel),
+        row_count_(row_count) {}
+
+  HOSTDEVICE inline T operator()(int64_t idx) const {
+    auto row_idx = math::BinarySearch(rows_, row_count_, idx / row_numel_);
+    return row_idx >= 0 ? grad_[row_idx * row_numel_ + idx % row_numel_] : 0;
+  }
+
+  const T *grad_;
+  const int64_t *rows_;
+  int64_t row_numel_;
+  int64_t row_count_;
+};
+
+template <typename T, typename GradFunctor>
+struct UncenteredRmspropFunctor {
+  UncenteredRmspropFunctor(T *param, T *ms, T *mom, const T *lr, T rho,
+                           T epsilon, T momentum,
+                           const GradFunctor &grad_functor)
+      : param_(param),
+        ms_(ms),
+        mom_(mom),
+        lr_(lr),
+        rho_(rho),
+        epsilon_(epsilon),
+        momentum_(momentum),
+        grad_functor_(grad_functor) {}
+
+  HOSTDEVICE inline void operator()(int64_t idx) const {
+    T g = grad_functor_(idx);
+    T ms_out = rho_ * ms_[idx] + (1 - rho_) * g * g;
+    T mom_out = momentum_ * mom_[idx] + lr_[0] * g / sqrt(ms_out + epsilon_);
+    param_[idx] -= mom_out;
+    ms_[idx] = ms_out;
+    mom_[idx] = mom_out;
+  }
+
+  T *param_;
+  T *ms_;
+  T *mom_;
+  const T *lr_;
+  T rho_;
+  T epsilon_;
+  T momentum_;
+  GradFunctor grad_functor_;
+};
+
+template <typename T, typename GradFunctor>
+struct CenteredRmspropFunctor {
+  CenteredRmspropFunctor(T *param, T *ms, T *mom, T *mean_grad, const T *lr,
+                         T rho, T epsilon, T momentum,
+                         const GradFunctor &grad_functor)
+      : param_(param),
+        ms_(ms),
+        mom_(mom),
+        mean_grad_(mean_grad),
+        lr_(lr),
+        rho_(rho),
+        epsilon_(epsilon),
+        momentum_(momentum),
+        grad_functor_(grad_functor) {}
+
+  HOSTDEVICE inline void operator()(int64_t idx) const {
+    T g = grad_functor_(idx);
+    T ms_out = rho_ * ms_[idx] + (1 - rho_) * g * g;
+    T mg_out = rho_ * mean_grad_[idx] + (1 - rho_) * g;
+    T mom_out = momentum_ * mom_[idx] +
+                lr_[0] * g / sqrt(ms_out - mg_out * mg_out + epsilon_);
+    param_[idx] -= mom_out;
+    ms_[idx] = ms_out;
+    mom_[idx] = mom_out;
+    mean_grad_[idx] = mg_out;
+  }
+
+  T *param_;
+  T *ms_;
+  T *mom_;
+  T *mean_grad_;
+  const T *lr_;
+  T rho_;
+  T epsilon_;
+  T momentum_;
+  GradFunctor grad_functor_;
+};
+
 template <typename DeviceContext, typename T>
 class RmspropOpKernel : public framework::OpKernel<T> {
  public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
-    auto* param_out = ctx.Output<Tensor>("ParamOut");
-    auto* moment_out = ctx.Output<Tensor>("MomentOut");
-    auto* mean_square_out = ctx.Output<Tensor>("MeanSquareOut");
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    using LoDTensor = framework::LoDTensor;
+    auto *grad_var = ctx.InputVar("Grad");
+    auto *param_out = ctx.Output<LoDTensor>("ParamOut");
+    auto *moment_out = ctx.Output<LoDTensor>("MomentOut");
+    auto *mean_square_out = ctx.Output<LoDTensor>("MeanSquareOut");
 
-    auto grad = ctx.Input<Tensor>("Grad");
+    auto epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
+    auto rho = static_cast<T>(ctx.Attr<float>("decay"));
+    auto momentum = static_cast<T>(ctx.Attr<float>("momentum"));
+    bool centered = ctx.Attr<bool>("centered");
 
-    param_out->mutable_data<T>(ctx.GetPlace());
-    moment_out->mutable_data<T>(ctx.GetPlace());
-    mean_square_out->mutable_data<T>(ctx.GetPlace());
+    auto &p_tensor = *ctx.Input<LoDTensor>("Param");
+    auto &ms_tensor = *ctx.Input<LoDTensor>("MeanSquare");
+    auto &lr_tensor = *ctx.Input<LoDTensor>("LearningRate");
+    auto &mom_tensor = *ctx.Input<LoDTensor>("Moment");
 
-    float epsilon = ctx.Attr<float>("epsilon");
-    float rho = ctx.Attr<float>("decay");
-    float momentum = ctx.Attr<float>("momentum");
-    bool centered = ctx.Attr<bool>("centered");
+    PADDLE_ENFORCE_EQ(&p_tensor, param_out,
+                      "Param and ParamOut must be the same Tensor");
+    PADDLE_ENFORCE_EQ(&mom_tensor, moment_out,
+                      "Moment and MomentOut must be the same Tensor");
+    PADDLE_ENFORCE_EQ(&ms_tensor, mean_square_out,
+                      "MeanSquare and MeanSquareOut must be the same Tensor");
+
+    auto &dev_ctx = ctx.template device_context<DeviceContext>();
+    size_t limit = static_cast<size_t>(ms_tensor.numel());
+
+    if (grad_var->IsType<LoDTensor>()) {
+      auto &grad_tensor = grad_var->Get<LoDTensor>();
+
+      if (std::is_same<DeviceContext, platform::CPUDeviceContext>::value) {
+        auto &place =
+            *ctx.template device_context<DeviceContext>().eigen_device();
+        auto lr_value = lr_tensor.data<T>()[0];
+
+        auto p = EigenVector<T>::Flatten(p_tensor);
+        auto ms = EigenVector<T>::Flatten(ms_tensor);
+        auto g = EigenVector<T>::Flatten(grad_tensor);
+        auto mom = EigenVector<T>::Flatten(mom_tensor);
+
+        auto p_out = EigenVector<T>::Flatten(*param_out);
+        auto mom_out = EigenVector<T>::Flatten(*moment_out);
+        auto ms_out = EigenVector<T>::Flatten(*mean_square_out);
+
+        ms_out.device(place) = rho * ms + (1 - rho) * g * g;
+        if (centered) {
+          auto &mg_tensor = *ctx.Input<LoDTensor>("MeanGrad");
+          auto mg = EigenVector<T>::Flatten(mg_tensor);
+          auto *mean_grad_out = ctx.Output<LoDTensor>("MeanGradOut");
+          PADDLE_ENFORCE(&mg_tensor, mean_grad_out,
+                         "MeanGrad and MeanGradOut must be the same Tensor");
+          auto mg_out = EigenVector<T>::Flatten(*mean_grad_out);
+
+          mg_out.device(place) = rho * mg + (1 - rho) * g;
+          mom_out.device(place) =
+              momentum * mom +
+              lr_value * g / (ms_out - mg_out.square() + epsilon).sqrt();
+        } else {
+          mom_out.device(place) =
+              momentum * mom + lr_value * g / (ms_out + epsilon).sqrt();
+        }
+        p_out.device(place) = p - mom_out;
+      } else {
+        DenseRmspropGradFunctor<T> grad_func(grad_tensor.data<T>());
+        platform::ForRange<DeviceContext> for_range(dev_ctx, limit);
+        if (centered) {
+          auto &mg_tensor = *ctx.Input<LoDTensor>("MeanGrad");
+          auto *mean_grad_out = ctx.Output<LoDTensor>("MeanGradOut");
+          PADDLE_ENFORCE(&mg_tensor, mean_grad_out,
+                         "MeanGrad and MeanGradOut must be the same Tensor");
+          for_range(CenteredRmspropFunctor<T, DenseRmspropGradFunctor<T>>(
+              param_out->mutable_data<T>(ctx.GetPlace()),
+              mean_square_out->mutable_data<T>(ctx.GetPlace()),
+              moment_out->mutable_data<T>(ctx.GetPlace()),
+              mean_grad_out->mutable_data<T>(ctx.GetPlace()),
+              lr_tensor.data<T>(), rho, epsilon, momentum, grad_func));
+        } else {
+          for_range(UncenteredRmspropFunctor<T, DenseRmspropGradFunctor<T>>(
+              param_out->mutable_data<T>(ctx.GetPlace()),
+              mean_square_out->mutable_data<T>(ctx.GetPlace()),
+              moment_out->mutable_data<T>(ctx.GetPlace()), lr_tensor.data<T>(),
+              rho, epsilon, momentum, grad_func));
+        }
+      }
+    } else if (grad_var->IsType<framework::SelectedRows>()) {
+      auto &grad = grad_var->Get<framework::SelectedRows>();
+      auto *merged_grad = const_cast<framework::Scope &>(ctx.scope())
+                              .Var()
+                              ->GetMutable<framework::SelectedRows>();
+
+      math::scatter::MergeAdd<DeviceContext, T> merge_func;
+      merge_func(dev_ctx, grad, merged_grad);
+
+      platform::ForRange<DeviceContext> for_range(dev_ctx, limit);
+      const int64_t *rows;
+#ifdef PADDLE_WITH_CUDA
+      if (platform::is_gpu_place(ctx.GetPlace())) {
+        rows = merged_grad->rows().CUDAData(ctx.GetPlace());
+      } else {
+#endif
+        rows = merged_grad->rows().data();
+#ifdef PADDLE_WITH_CUDA
+      }
+#endif
+      auto &merged_tensor = merged_grad->value();
+      int64_t row_count = merged_grad->rows().size();
+      int64_t row_numel = merged_tensor.numel() / row_count;
+      SparseRmspropGradFunctor<T> grad_func(merged_tensor.data<T>(), rows,
+                                            row_numel, row_count);
 
-    auto p = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Param"));
-    auto ms = EigenVector<T>::Flatten(*ctx.Input<Tensor>("MeanSquare"));
-    auto lr = EigenVector<T>::Flatten(*ctx.Input<Tensor>("LearningRate"));
-    auto g = EigenVector<T>::Flatten(*grad);
-    auto mom = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Moment"));
-
-    auto p_out = EigenVector<T>::Flatten(*param_out);
-    auto mom_out = EigenVector<T>::Flatten(*moment_out);
-    auto ms_out = EigenVector<T>::Flatten(*mean_square_out);
-    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
-
-    Eigen::DSizes<int, 1> grad_dsize(static_cast<int>(grad->numel()));
-
-    ms_out.device(place) = rho * ms + (1 - rho) * g * g;
-    if (centered) {
-      auto mg = EigenVector<T>::Flatten(*ctx.Input<Tensor>("MeanGrad"));
-      auto* mean_grad_out = ctx.Output<Tensor>("MeanGradOut");
-      mean_grad_out->mutable_data<T>(ctx.GetPlace());
-      auto mg_out = EigenVector<T>::Flatten(*mean_grad_out);
-
-      mg_out.device(place) = rho * mg + (1 - rho) * g;
-      mom_out.device(place) = momentum * mom +
-                              lr.broadcast(grad_dsize) * g /
-                                  (ms_out - mg_out.square() + epsilon).sqrt();
+      if (centered) {
+        auto &mg_tensor = *ctx.Input<LoDTensor>("MeanGrad");
+        auto *mean_grad_out = ctx.Output<LoDTensor>("MeanGradOut");
+        PADDLE_ENFORCE(&mg_tensor, mean_grad_out,
+                       "MeanGrad and MeanGradOut must be the same Tensor");
+        for_range(CenteredRmspropFunctor<T, SparseRmspropGradFunctor<T>>(
+            param_out->mutable_data<T>(ctx.GetPlace()),
+            mean_square_out->mutable_data<T>(ctx.GetPlace()),
+            moment_out->mutable_data<T>(ctx.GetPlace()),
+            mean_grad_out->mutable_data<T>(ctx.GetPlace()), lr_tensor.data<T>(),
+            rho, epsilon, momentum, grad_func));
+      } else {
+        for_range(UncenteredRmspropFunctor<T, SparseRmspropGradFunctor<T>>(
+            param_out->mutable_data<T>(ctx.GetPlace()),
+            mean_square_out->mutable_data<T>(ctx.GetPlace()),
+            moment_out->mutable_data<T>(ctx.GetPlace()), lr_tensor.data<T>(),
+            rho, epsilon, momentum, grad_func));
+      }
     } else {
-      mom_out.device(place) =
-          momentum * mom +
-          lr.broadcast(grad_dsize) * g / (ms_out + epsilon).sqrt();
+      PADDLE_THROW("RMSProp only supports LoDTensor or SelectedRows gradient");
     }
-    p_out.device(place) = p - mom_out;
   }
 };
 

paddle/fluid/operators/sequence_concat_op.cc

@@ -90,11 +90,13 @@ REGISTER_OPERATOR(sequence_concat, paddle::framework::OperatorWithKernel,
                   paddle::framework::DefaultGradOpDescMaker<false>);
 template <typename T>
 using Kernel = op::SeqConcatKernel<paddle::platform::CPUDeviceContext, T>;
-REGISTER_OP_CPU_KERNEL(sequence_concat, Kernel<float>, Kernel<double>);
+REGISTER_OP_CPU_KERNEL(sequence_concat, Kernel<float>, Kernel<double>,
+                       Kernel<int64_t>);
+
 REGISTER_OPERATOR(sequence_concat_grad, paddle::framework::OperatorWithKernel,
                   op::SeqConcatGradShapeInferer);
 template <typename T>
 using GradKernel =
     op::SeqConcatGradKernel<paddle::platform::CPUDeviceContext, T>;
 REGISTER_OP_CPU_KERNEL(sequence_concat_grad, GradKernel<float>,
-                       GradKernel<double>);
+                       GradKernel<double>, GradKernel<int64_t>);

paddle/fluid/operators/sequence_unpad_op.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/sequence_unpad_op.h"
+
+namespace paddle {
+namespace operators {
+
+class SequenceUnpadOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of SequenceUnpadOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Length"),
+                   "Input(Length) of SequenceUnpadOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of SequenceUnpadOp should not be null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    PADDLE_ENFORCE_GE(x_dims.size(), 2,
+                      "The rank of Input(X) can't be less than 2.");
+
+    auto len_dims = ctx->GetInputDim("Length");
+    PADDLE_ENFORCE(len_dims.size() == 2 && len_dims[1] == 1,
+                   "The shape of Input(Length) should be [batch_size, 1].");
+    PADDLE_ENFORCE(
+        len_dims[0] == x_dims[0],
+        "Input(X) and Input(Length) should have the same first dimension.");
+
+    int64_t out_dim_0 = -1;
+    if (ctx->IsRuntime()) {
+      out_dim_0 = x_dims[0] * x_dims[1];
+    }
+
+    std::vector<int64_t> out_dims_vec{out_dim_0};
+    if (x_dims.size() == 2) {
+      out_dims_vec.push_back(1);
+    } else {
+      for (size_t i = 2; i < x_dims.size(); ++i) {
+        out_dims_vec.push_back(x_dims[i]);
+      }
+    }
+    ctx->SetOutputDim("Out", framework::make_ddim(out_dims_vec));
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("X"));
+    return framework::OpKernelType(data_type, ctx.device_context());
+  }
+};
+
+class SequenceUnpadOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(LoDTensor, default LoDTensor<float>) Input tensor which "
+             "contains the padded sequences with equal length.");
+    AddInput("Length",
+             "(LoDTensor) The input tensor which specifies the actual ength of "
+             "sequences after unpadding.");
+    AddOutput(
+        "Out",
+        "(LoDTensor) The output tensor which contains unpadded sequences.");
+    AddComment(R"DOC(
+      Sequence Unpad Operator
+
+      This operator removes the padding data in the input sequences and convert 
+      them into sequences with actual length as output, identitied by lod 
+      information.
+
+      Example:
+
+      Given input tensor Input(X):
+          X.data = [[ 1.0,  2.0,  3.0,  4.0,  5.0],
+                    [ 6.0,  7.0,  8.0,  9.0, 10.0],
+                    [11.0, 12.0, 13.0, 14.0, 15.0]], 
+`     
+      in which there are 3 sequences padded to length 5, and the acutal length 
+      specified by Input(Length):
+
+          Length.data = [[2], [3], [4]],
+
+      after unpadding, Output(Out) will be:
+
+          Out.data = [[1.0, 2.0, 6.0, 7.0, 8.0, 11.0, 12.0, 13.0, 14.0]]
+          Out.lod = [[0, 2, 5, 9]]      
+
+    )DOC");
+  }
+};
+
+class SequenceUnpadGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of SequenceUnpadGradOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasInput(framework::GradVarName("Out")),
+        "Input(Out@GRAD) of SequenceUnpadGradOp should not be null.");
+
+    if (ctx->HasOutput(framework::GradVarName("X"))) {
+      ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
+      ctx->ShareLoD("X", /*->*/ framework::GradVarName("X"));
+    }
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("X"));
+    return framework::OpKernelType(data_type, ctx.device_context());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(sequence_unpad, ops::SequenceUnpadOp,
+                  ops::SequenceUnpadOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+REGISTER_OPERATOR(sequence_unpad_grad, ops::SequenceUnpadGradOp);
+REGISTER_OP_CPU_KERNEL(
+    sequence_unpad,
+    ops::SequenceUnpadOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SequenceUnpadOpKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::SequenceUnpadOpKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::SequenceUnpadOpKernel<paddle::platform::CPUDeviceContext, int64_t>);
+REGISTER_OP_CPU_KERNEL(
+    sequence_unpad_grad,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CPUDeviceContext,
+                                   int64_t>);

paddle/fluid/operators/sequence_unpad_op.cu

+/* 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/sequence_unpad_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    sequence_unpad,
+    ops::SequenceUnpadOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SequenceUnpadOpKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::SequenceUnpadOpKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::SequenceUnpadOpKernel<paddle::platform::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
+    sequence_unpad_grad,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CUDADeviceContext,
+                                   int64_t>);

paddle/fluid/operators/sequence_unpad_op.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <vector>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/memory/memcpy.h"
+#include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/operators/math/sequence_padding.h"
+
+namespace paddle {
+namespace operators {
+
+using LoDTensor = framework::LoDTensor;
+using LoD = framework::LoD;
+
+template <typename DeviceContext, typename T>
+class SequenceUnpadOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x_t = ctx.Input<LoDTensor>("X");
+    auto* len_t = ctx.Input<LoDTensor>("Length");
+    auto* out_t = ctx.Output<LoDTensor>("Out");
+    out_t->mutable_data<T>(ctx.GetPlace());
+
+    const int64_t* seq_len_ptr = nullptr;
+    if (platform::is_gpu_place(ctx.GetPlace())) {
+      LoDTensor seq_len_cpu;
+      seq_len_cpu.Resize(len_t->dims());
+      seq_len_ptr = seq_len_cpu.mutable_data<int64_t>(platform::CPUPlace());
+      framework::TensorCopy(*len_t, platform::CPUPlace(),
+                            ctx.template device_context<DeviceContext>(),
+                            &seq_len_cpu);
+    } else {
+      seq_len_ptr = len_t->data<int64_t>();
+    }
+
+    size_t batch_size = x_t->dims()[0];
+    std::vector<size_t> out_lod0(batch_size + 1, 0);
+    for (size_t i = 0; i < batch_size; ++i) {
+      out_lod0[i + 1] = out_lod0[i] + seq_len_ptr[i];
+    }
+
+    framework::LoD out_lod;
+    out_lod.push_back(out_lod0);
+    out_t->set_lod(out_lod);
+
+    std::vector<int64_t> out_dims_vec{static_cast<int64_t>(out_lod0.back())};
+    if (x_t->dims().size() == 2) {
+      out_dims_vec.push_back(1);
+    } else {
+      for (size_t i = 2; i < x_t->dims().size(); ++i) {
+        out_dims_vec.push_back(x_t->dims()[i]);
+      }
+    }
+    out_t->Resize(framework::make_ddim(out_dims_vec));
+
+    int64_t padded_length = x_t->dims()[1];
+    math::UnpaddingLoDTensorFunctor<DeviceContext, T>()(
+        ctx.template device_context<DeviceContext>(), *x_t, out_t,
+        padded_length, 0, false, math::kBatchLengthWidth);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class SequenceUnpadGradOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* d_x = ctx.Output<LoDTensor>(framework::GradVarName("X"));
+    if (d_x) {
+      const auto* d_out = ctx.Input<LoDTensor>(framework::GradVarName("Out"));
+      const auto* x_t = ctx.Input<LoDTensor>("X");
+      d_x->mutable_data<T>(ctx.GetPlace());
+
+      int padded_length = x_t->dims()[1];
+
+      LoDTensor zero_pads;
+      zero_pads.Resize({1, 1});
+      zero_pads.mutable_data<T>(ctx.GetPlace());
+      math::SetConstant<DeviceContext, T> set_zero;
+      auto& dev_ctx = ctx.template device_context<DeviceContext>();
+      set_zero(dev_ctx, &zero_pads, static_cast<T>(0));
+
+      math::PaddingLoDTensorFunctor<DeviceContext, T>()(
+          ctx.template device_context<DeviceContext>(), *d_out, d_x, zero_pads,
+          padded_length, 0, false, math::kBatchLengthWidth);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/sgd_op.cc

@@ -21,7 +21,7 @@ class SGDOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
-  void InferShape(framework::InferShapeContext* ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Param"),
                    "Input(Param) of SGDOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Grad"),
@@ -42,7 +42,7 @@ class SGDOp : public framework::OperatorWithKernel {
 
  protected:
   framework::OpKernelType GetExpectedKernelType(
-      const framework::ExecutionContext& ctx) const override {
+      const framework::ExecutionContext &ctx) const override {
     auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("Param"));
     return framework::OpKernelType(data_type, ctx.device_context());
   }
@@ -50,17 +50,20 @@ class SGDOp : public framework::OperatorWithKernel {
 
 class SGDOpInferVarType : public framework::VarTypeInference {
  public:
-  void operator()(const framework::OpDesc& op_desc,
-                  framework::BlockDesc* block) const override {
-    auto input_var = op_desc.Input("Param")[0];
-    for (auto& out_var : op_desc.Output("ParamOut")) {
-      if (block->FindRecursiveOrCreateVar(input_var).GetType() ==
-          framework::proto::VarType::SELECTED_ROWS) {
-        block->FindRecursiveOrCreateVar(out_var).SetType(
-            framework::proto::VarType::SELECTED_ROWS);
-      } else {
-        block->FindRecursiveOrCreateVar(out_var).SetType(
-            framework::proto::VarType::LOD_TENSOR);
+  void operator()(const framework::OpDesc &op_desc,
+                  framework::BlockDesc *block) const override {
+    auto input_var_n = op_desc.Input("Param")[0];
+    auto in_var_type = block->FindRecursiveOrCreateVar(input_var_n).GetType();
+    PADDLE_ENFORCE(in_var_type == framework::proto::VarType::SELECTED_ROWS ||
+                       in_var_type == framework::proto::VarType::LOD_TENSOR,
+                   "The input Var's type should be LoDtensor or SelectedRows,"
+                   " but the received var(%s)'s type is %s",
+                   input_var_n, in_var_type);
+
+    for (auto &out_var_n : op_desc.Output("ParamOut")) {
+      auto &out_var = block->FindRecursiveOrCreateVar(out_var_n);
+      if (out_var.GetType() != in_var_type) {
+        out_var.SetType(in_var_type);
       }
     }
   }

paddle/fluid/operators/sgd_op.cu

@@ -56,6 +56,12 @@ template <typename T>
 class SGDOpCUDAKernel : 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 = ctx.Input<framework::Tensor>("Param");
     auto* param_out = ctx.Output<framework::Tensor>("ParamOut");
     auto* learning_rate = ctx.Input<framework::Tensor>("LearningRate");

paddle/fluid/operators/truncated_gaussian_random_op.cc

@@ -148,7 +148,7 @@ struct TruncatedNormal {
 
   T operator()(T value) const {
     auto p = a_normal_cdf + (b_normal_cdf - a_normal_cdf) * value;
-    return (std::sqrt(2.0) * Erfinv(2 * p - 1) + mean) * std;
+    return std::sqrt(2.0) * Erfinv(2 * p - 1) * std + mean;
   }
 };
 

paddle/fluid/operators/truncated_gaussian_random_op.cu

@@ -42,7 +42,7 @@ struct TruncatedNormal {
     rng.discard(n);
     T value = dist(rng);
     auto p = a_normal_cdf + (b_normal_cdf - a_normal_cdf) * value;
-    return (std::sqrt(2.0) * erfinvf(2 * p - 1) + mean) * std;
+    return std::sqrt(2.0) * erfinvf(2 * p - 1) * std + mean;
   }
 };
 
@@ -52,6 +52,7 @@ class GPUTruncatedGaussianRandomKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& context) const override {
     auto* tensor = context.Output<framework::Tensor>("Out");
     T* data = tensor->mutable_data<T>(context.GetPlace());
+
     unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
     if (seed == 0) {
       std::random_device rd;

paddle/fluid/operators/uniform_random_op.cc

@@ -23,14 +23,14 @@ namespace operators {
 template <typename T>
 class CPUUniformRandomKernel : public framework::OpKernel<T> {
  public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
-    framework::Tensor* tensor = nullptr;
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    framework::Tensor *tensor = nullptr;
     auto out_var = ctx.OutputVar("Out");
     if (out_var->IsType<framework::LoDTensor>()) {
       tensor = out_var->GetMutable<framework::LoDTensor>();
     } else if (out_var->IsType<framework::SelectedRows>()) {
       auto shape = ctx.Attr<std::vector<int>>("shape");
-      auto* selected_rows = out_var->GetMutable<framework::SelectedRows>();
+      auto *selected_rows = out_var->GetMutable<framework::SelectedRows>();
       tensor = selected_rows->mutable_value();
       tensor->Resize(framework::make_ddim(shape));
       selected_rows->mutable_rows()->reserve(shape[0]);
@@ -39,7 +39,7 @@ class CPUUniformRandomKernel : public framework::OpKernel<T> {
           "uniform_random_op's output only"
           "supports SelectedRows and LoDTensor");
     }
-    T* data = tensor->mutable_data<T>(ctx.GetPlace());
+    T *data = tensor->mutable_data<T>(ctx.GetPlace());
     unsigned int seed = static_cast<unsigned int>(ctx.Attr<int>("seed"));
     std::minstd_rand engine;
     if (seed == 0) {
@@ -60,14 +60,14 @@ class UniformRandomOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
-  void InferShape(framework::InferShapeContext* ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of UniformRandomOp should not be null.");
 
     PADDLE_ENFORCE(
         ctx->Attrs().Get<float>("min") < ctx->Attrs().Get<float>("max"),
         "uniform_random's min must less then max");
-    auto& shape = ctx->Attrs().Get<std::vector<int>>("shape");
+    auto &shape = ctx->Attrs().Get<std::vector<int>>("shape");
     std::vector<int64_t> temp;
     temp.reserve(shape.size());
     for (auto dim : shape) {
@@ -78,7 +78,7 @@ class UniformRandomOp : public framework::OperatorWithKernel {
 
  protected:
   framework::OpKernelType GetExpectedKernelType(
-      const framework::ExecutionContext& ctx) const override {
+      const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         static_cast<framework::proto::VarType::Type>(ctx.Attr<int>("dtype")),
         ctx.GetPlace());
@@ -112,17 +112,17 @@ uniform distribution. The random result is in set [min, max].
 
 class UniformRandomOpVarTypeInference : public framework::VarTypeInference {
  public:
-  void operator()(const framework::OpDesc& op_desc,
-                  framework::BlockDesc* block) const override {
+  void operator()(const framework::OpDesc &op_desc,
+                  framework::BlockDesc *block) const override {
     auto out_var_name = op_desc.Output("Out").front();
-    if (block->FindRecursiveOrCreateVar(out_var_name).GetType() ==
-        framework::proto::VarType::SELECTED_ROWS) {
-      block->FindRecursiveOrCreateVar(out_var_name)
-          .SetType(framework::proto::VarType::SELECTED_ROWS);
-    } else {
-      block->FindRecursiveOrCreateVar(out_var_name)
-          .SetType(framework::proto::VarType::LOD_TENSOR);
+    auto var_data_type = static_cast<framework::proto::VarType::Type>(
+        boost::get<int>(op_desc.GetAttr("dtype")));
+
+    auto out_var = block->FindRecursiveOrCreateVar(out_var_name);
+    if (out_var.GetType() != framework::proto::VarType::SELECTED_ROWS) {
+      out_var.SetType(framework::proto::VarType::LOD_TENSOR);
     }
+    out_var.SetDataType(var_data_type);
   }
 };
 

paddle/fluid/platform/cpu_info.cc

@@ -128,7 +128,7 @@ bool MayIUse(const cpu_isa_t cpu_isa) {
       return cpu.has(Cpu::tAVX);
     case avx2:
       return cpu.has(Cpu::tAVX2);
-    case avx512_common:
+    case avx512f:
       return cpu.has(Cpu::tAVX512F);
     case avx512_core:
       return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512BW) &&

paddle/fluid/platform/cpu_info.h

@@ -43,7 +43,7 @@ typedef enum {
   sse42,
   avx,
   avx2,
-  avx512_common,
+  avx512f,
   avx512_core,
   avx512_core_vnni,
   avx512_mic,

paddle/fluid/platform/device_context.cc

@@ -198,9 +198,9 @@ class CudnnHolder {
 CUDADeviceContext::CUDADeviceContext(CUDAPlace place)
     : place_(place), cudnn_holder_(nullptr) {
   SetDeviceId(place_.device);
-  compute_capability = GetCUDAComputeCapability(place_.device);
-  multi_process = GetCUDAMultiProcessors(place_.device);
-  max_threads_per_mp = GetCUDAMaxThreadsPerMultiProcessor(place_.device);
+  compute_capability_ = GetCUDAComputeCapability(place_.device);
+  multi_process_ = GetCUDAMultiProcessors(place_.device);
+  max_threads_per_mp_ = GetCUDAMaxThreadsPerMultiProcessor(place_.device);
   PADDLE_ENFORCE(cudaStreamCreate(&stream_));
   eigen_stream_.reset(new EigenCudaStreamDevice());
   eigen_stream_->Reinitialize(&stream_, place);
@@ -211,6 +211,16 @@ CUDADeviceContext::CUDADeviceContext(CUDAPlace place)
     cudnn_holder_.reset(new CudnnHolder(&stream_, place));
   }
 
+  driver_version_ = GetCUDADriverVersion(place_.device);
+  runtime_version_ = GetCUDARuntimeVersion(place_.device);
+
+  LOG(INFO) << "device: " << place_.device
+            << ", CUDA Capability: " << compute_capability_
+            << ", Driver Version: " << driver_version_ / 1000 << "."
+            << (driver_version_ % 100) / 10
+            << ", Runtime Version: " << runtime_version_ / 1000 << "."
+            << (runtime_version_ % 100) / 10;
+
   callback_manager_.reset(new StreamCallbackManager(stream_));
 }
 
@@ -232,11 +242,11 @@ void CUDADeviceContext::Wait() const {
 }
 
 int CUDADeviceContext::GetComputeCapability() const {
-  return compute_capability;
+  return compute_capability_;
 }
 
 int CUDADeviceContext::GetMaxPhysicalThreadCount() const {
-  return multi_process * max_threads_per_mp;
+  return multi_process_ * max_threads_per_mp_;
 }
 
 Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {

paddle/fluid/platform/device_context.h

@@ -135,9 +135,11 @@ class CUDADeviceContext : public DeviceContext {
   cudaStream_t stream_;
   cublasHandle_t cublas_handle_;
 
-  int compute_capability;
-  int multi_process;
-  int max_threads_per_mp;
+  int compute_capability_;
+  int runtime_version_;
+  int driver_version_;
+  int multi_process_;
+  int max_threads_per_mp_;
 
   mutable std::mutex mtx_;
 

paddle/fluid/platform/enforce.h

@@ -130,6 +130,13 @@ struct EOFException : public std::exception {
 #define UNLIKELY(condition) (condition == 0)
 #endif
 
+#if !defined(_WIN32)
+#define LIKELY(condition) __builtin_expect(static_cast<bool>(condition), 1)
+#else
+// there is no equivalent intrinsics in msvc.
+#define LIKELY(condition) (condition != 0)
+#endif
+
 template <typename... Args>
 inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
     bool stat, const Args&... args) {

paddle/fluid/platform/gpu_info.cc

@@ -46,6 +46,24 @@ int GetCUDAComputeCapability(int id) {
   return device_prop.major * 10 + device_prop.minor;
 }
 
+int GetCUDARuntimeVersion(int id) {
+  PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
+  int runtime_version = 0;
+  PADDLE_ENFORCE(cudaRuntimeGetVersion(&runtime_version),
+                 "cudaRuntimeGetVersion failed in "
+                 "paddle::platform::cudaRuntimeGetVersion");
+  return runtime_version;
+}
+
+int GetCUDADriverVersion(int id) {
+  PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
+  int driver_version = 0;
+  PADDLE_ENFORCE(cudaDriverGetVersion(&driver_version),
+                 "cudaDriverGetVersion failed in "
+                 "paddle::platform::GetCUDADriverVersion");
+  return driver_version;
+}
+
 int GetCUDAMultiProcessors(int id) {
   PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
   int count;

paddle/fluid/platform/gpu_info.h

@@ -29,6 +29,12 @@ int GetCUDADeviceCount();
 //! Get the compute capability of the ith GPU (format: major * 10 + minor)
 int GetCUDAComputeCapability(int i);
 
+//! Get the runtime version of the ith GPU
+int GetCUDARuntimeVersion(int id);
+
+//! Get the driver version of the ith GPU
+int GetCUDADriverVersion(int id);
+
 //! Get the MultiProcessors of the ith GPU.
 int GetCUDAMultiProcessors(int i);
 

paddle/fluid/platform/init.cc

@@ -116,7 +116,7 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
   platform::SetNumThreads(FLAGS_paddle_num_threads);
 #endif
 
-  if (platform::jit::MayIUse(platform::jit::avx512_common)) {
+  if (platform::jit::MayIUse(platform::jit::avx512f)) {
 #ifndef __AVX512F__
     LOG(WARNING) << "AVX512F is available, Please re-compile on local machine";
 #endif

paddle/fluid/platform/profiler.cc

@@ -276,7 +276,7 @@ struct EventItem {
 // Print results
 void PrintProfiler(const std::vector<std::vector<EventItem>>& events_table,
                    const std::string& sorted_domain, const size_t name_width,
-                   const size_t data_width, double total) {
+                   const size_t data_width, bool merge_thread) {
   // Output header information
   std::cout << "\n------------------------->"
             << "     Profiling Report     "
@@ -292,6 +292,10 @@ void PrintProfiler(const std::vector<std::vector<EventItem>>& events_table,
     PADDLE_THROW("Invalid profiler state", g_state);
   }
 
+  if (merge_thread) {
+    std::cout << "Note! This Report merge all thread info into one."
+              << std::endl;
+  }
   std::cout << "Place: " << place << std::endl;
   std::cout << "Time unit: ms" << std::endl;
   std::cout << "Sorted by " << sorted_domain
@@ -312,8 +316,7 @@ void PrintProfiler(const std::vector<std::vector<EventItem>>& events_table,
                 << std::setw(data_width) << event_item.min_time
                 << std::setw(data_width) << event_item.max_time
                 << std::setw(data_width) << event_item.ave_time
-                << std::setw(data_width) << event_item.total_time / total
-                << std::endl;
+                << std::setw(data_width) << event_item.ratio << std::endl;
     }
   }
   std::cout << std::endl;
@@ -321,8 +324,10 @@ void PrintProfiler(const std::vector<std::vector<EventItem>>& events_table,
 
 // Parse the event list and output the profiling report
 void ParseEvents(const std::vector<std::vector<Event>>& events,
+                 bool merge_thread,
                  EventSortingKey sorted_by = EventSortingKey::kDefault) {
   if (g_state == ProfilerState::kDisabled) return;
+  if (merge_thread && events.size() < 2) return;
 
   std::string sorted_domain;
   std::function<bool(const EventItem&, const EventItem&)> sorted_func;
@@ -361,34 +366,55 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
       sorted_domain = "event first end time";
   }
 
+  const std::vector<std::vector<Event>>* analyze_events;
+  std::vector<std::vector<Event>> merged_events_list;
+  if (merge_thread) {
+    std::vector<Event> merged_events;
+    for (size_t i = 0; i < events.size(); ++i) {
+      for (size_t j = 0; j < events[i].size(); ++j) {
+        merged_events.push_back(events[i][j]);
+      }
+    }
+    merged_events_list.push_back(merged_events);
+    analyze_events = &merged_events_list;
+  } else {
+    analyze_events = &events;
+  }
+
   std::vector<std::vector<EventItem>> events_table;
   size_t max_name_width = 0;
-  double total = 0.;  // the total time
-  for (size_t i = 0; i < events.size(); i++) {
+  for (size_t i = 0; i < (*analyze_events).size(); i++) {
+    double total = 0.;  // the total time in one thread
     std::list<Event> pushed_events;
     std::vector<EventItem> event_items;
     std::unordered_map<std::string, int> event_idx;
 
-    for (size_t j = 0; j < events[i].size(); j++) {
-      if (events[i][j].type() == EventType::kPushRange) {
-        pushed_events.push_back(events[i][j]);
-      } else if (events[i][j].type() == EventType::kPopRange) {
+    for (size_t j = 0; j < (*analyze_events)[i].size(); j++) {
+      if ((*analyze_events)[i][j].type() == EventType::kPushRange) {
+        pushed_events.push_back((*analyze_events)[i][j]);
+      } else if ((*analyze_events)[i][j].type() == EventType::kPopRange) {
         std::list<Event>::reverse_iterator rit = pushed_events.rbegin();
         while (rit != pushed_events.rend() &&
-               rit->name() != events[i][j].name()) {
+               rit->name() != (*analyze_events)[i][j].name()) {
           ++rit;
         }
 
         if (rit != pushed_events.rend()) {
           double event_time = (g_state == ProfilerState::kCUDA ||
                                g_state == ProfilerState::kAll)
-                                  ? rit->CudaElapsedMs(events[i][j])
-                                  : rit->CpuElapsedMs(events[i][j]);
+                                  ? rit->CudaElapsedMs((*analyze_events)[i][j])
+                                  : rit->CpuElapsedMs((*analyze_events)[i][j]);
           total += event_time;
 
-          std::string event_name =
-              "thread" + std::to_string(rit->thread_id()) + "::" + rit->name();
-          max_name_width = std::max(max_name_width, event_name.size());
+          std::string event_name;
+          if (merge_thread) {
+            event_name = rit->name();
+            max_name_width = std::max(max_name_width, event_name.size());
+          } else {
+            event_name = "thread" + std::to_string(rit->thread_id()) + "::" +
+                         rit->name();
+            max_name_width = std::max(max_name_width, event_name.size());
+          }
 
           if (event_idx.find(event_name) == event_idx.end()) {
             event_idx[event_name] = event_items.size();
@@ -413,7 +439,7 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
           pushed_events.erase((++rit).base());
         } else {
           LOG(WARNING) << "Cannot find the push marker of event \'"
-                       << events[i][j].name()
+                       << (*analyze_events)[i][j].name()
                        << "\', which will be ignored in profiling report.";
         }
       }
@@ -421,6 +447,7 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
     // average time
     for (auto& item : event_items) {
       item.ave_time = item.total_time / item.calls;
+      item.ratio = item.total_time / total;
     }
     // sort
     if (sorted_by != EventSortingKey::kDefault) {
@@ -438,7 +465,8 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
   }
 
   // Print report
-  PrintProfiler(events_table, sorted_domain, max_name_width + 4, 12, total);
+  PrintProfiler(events_table, sorted_domain, max_name_width + 4, 12,
+                merge_thread);
 }
 
 void DisableProfiler(EventSortingKey sorted_key,
@@ -449,7 +477,8 @@ void DisableProfiler(EventSortingKey sorted_key,
   Mark("_stop_profiler_", nullptr);
 
   std::vector<std::vector<Event>> all_events = GetAllEvents();
-  ParseEvents(all_events, sorted_key);
+  ParseEvents(all_events, true, sorted_key);
+  ParseEvents(all_events, false, sorted_key);
   ResetProfiler();
   DeviceTracer* tracer = GetDeviceTracer();
   if (tracer->IsEnabled()) {

paddle/fluid/pybind/pybind.cc

@@ -57,6 +57,10 @@ limitations under the License. */
 
 #include "pybind11/stl.h"
 
+DEFINE_bool(reader_queue_speed_test_mode, false,
+            "If set true, the queue.pop will only get data from queue but not "
+            "remove the data from queue for speed testing");
+
 // disable auto conversion to list in Python
 PYBIND11_MAKE_OPAQUE(paddle::framework::LoDTensorArray);
 
@@ -157,7 +161,50 @@ PYBIND11_PLUGIN(core) {
       .def("_get_double_element", TensorGetElement<double>)
       .def("_dtype", [](Tensor &self) { return ToDataType(self.type()); });
 
-  py::class_<LoDTensor, Tensor>(m, "LoDTensor")
+  py::class_<LoDTensor, Tensor>(m, "LoDTensor", R"DOC(
+    LoDTensor is a Tensor with optional LoD information.
+
+    np.array(lod_tensor) can convert LoDTensor to numpy array.
+    lod_tensor.lod() can retrieve the LoD information.
+
+    LoD is short for Level of Details and is usually used for varied sequence
+    length. You can skip the following comment if you don't need optional LoD.
+
+  For example:
+     A LoDTensor X can look like the example below. It contains 2 sequences.
+     The first has length 2 and the second has length 3, as described by x.lod.
+
+     The first tensor dimension 5=2+3 is calculated from LoD if it's available.
+     It means the total number of sequence element. In X, each element has 2
+     columns, hence [5, 2].
+
+      x.lod  = [[2, 3]]
+      x.data = [[1, 2], [3, 4],
+                [5, 6], [7, 8], [9, 10]]
+      x.shape = [5, 2]
+
+      LoD can have multiple levels (for example, a paragraph can have multiple
+      sentences and a sentence can have multiple words). In the following
+      LodTensor Y, the lod_level is 2. It means there are 2 sequence, the
+      first sequence length is 2 (has 2 sub-sequences), the second one's
+      length is 1. The first sequence's 2 sub-sequences have length 2 and 2,
+      respectively. And the second sequence's 1 sub-sequence has length 3.
+
+      y.lod = [[2 1], [2 2 3]]
+      y.shape = [2+2+3, ...]
+
+  Note:
+      In above description, LoD is length-based. In Paddle internal
+      implementation, lod is offset-based. Hence, internally,
+      y.lod is represented as [[0, 2, 3], [0, 2, 4, 7]] (length-based
+      equivlent would be [[2-0, 3-2], [2-0, 4-2, 7-4]]).
+
+      Sometimes LoD is called recursive_sequence_length to be more
+      self-explanatory. In this case, it must be length-based. Due to history
+      reasons. when LoD is called lod in public API, it might be offset-based.
+      Users should be careful about it.
+
+        )DOC")
       .def_buffer(
           [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
       .def("__init__",
@@ -337,7 +384,8 @@ All parameter, weight, gradient are variables in Paddle.
                                return make_ddim(shape);
                              });
               auto *holder = var.GetMutable<LoDTensorBlockingQueueHolder>();
-              holder->InitOnce(capacity, dims);
+              holder->InitOnce(capacity, dims,
+                               FLAGS_reader_queue_speed_test_mode);
               return holder->GetQueue();
             },
         py::return_value_policy::copy);
@@ -624,16 +672,17 @@ All parameter, weight, gradient are variables in Paddle.
     ExecutionStrategy allows the user to more preciously control how to run
     the program in ParallelExecutor by setting the property.
 
-    The available properties include:
-        use_cuda (bool): Whether to use CUDA or not. Default True.
-        num_threads (int): The number of threads that used to run the
-            operators in ParallelExecutor. If it is not set, it will be
-            set in ParallelExecutor according to the device count.
-            Default 0.
-        allow_op_delay (bool): Whether to delay the communication operators
-            to run. Default False.
-        num_iteration_per_drop_scope (int): how many iterations between
-            the two dropping local scopes. Default 100.
+    Examples:
+        .. code-block:: python
+
+          exec_strategy = fluid.ExecutionStrategy()
+          exec_strategy.num_threads = 4
+
+          train_exe = fluid.ParallelExecutor(use_cuda=True,
+                                             loss_name=loss.name,
+                                             exec_strategy=exec_strategy)
+
+          train_loss, = train_exe.run([loss.name], feed=feed_dict)
 
         )DOC");
 
@@ -643,19 +692,34 @@ All parameter, weight, gradient are variables in Paddle.
           [](const ExecutionStrategy &self) { return self.num_threads_; },
           [](ExecutionStrategy &self, size_t num_threads) {
             self.num_threads_ = num_threads;
-          })
+          },
+          R"DOC(The type is INT, num_threads represents the size of thread pool that
+            used to run the operators of the current program in ParallelExecutor.
+            If :math:`num\_threads=1`, all the operators will execute one by one,
+            but the order maybe difference between iterations.
+            If it is not set, it will be set in ParallelExecutor according to the
+            device type and device count, for GPU, :math:`num\_threads=device\_count*4`, for CPU,
+            :math:`num\_threads=CPU\_NUM*4`, the explanation of:math:`CPU\_NUM` is in ParallelExecutor.
+            if it is not set, ParallelExecutor will get the cpu count by calling
+            `multiprocessing.cpu_count()`. Default 0.)DOC")
       .def_property(
           "use_cuda",
           [](const ExecutionStrategy &self) { return self.use_cuda_; },
           [](ExecutionStrategy &self, bool use_cuda) {
             self.use_cuda_ = use_cuda;
-          })
+          })  // FIXME(chengduo): Doesn't add doc for 'use_cuda', use_cuda may
+      // make user confuse, because ParallelExecutor has a parameter named
+      // 'use_cuda' too, in current implementation, ParallelExecutor's
+      // 'use_cuda' will rewrite ExecutionStrategy's 'use_cuda'.
       .def_property(
           "allow_op_delay",
           [](const ExecutionStrategy &self) { return self.allow_op_delay_; },
           [](ExecutionStrategy &self, bool allow_op_delay) {
             self.allow_op_delay_ = allow_op_delay;
-          })
+          },
+          R"DOC(The type is BOOL, allow_op_delay represents whether to delay the
+                communication operators to run, it may make the execution faster.
+                Note that in some models, allow_op_delay may cause program hang. Default False.)DOC")
       .def_property(
           "num_iteration_per_drop_scope",
           [](const ExecutionStrategy &self) {
@@ -663,7 +727,19 @@ All parameter, weight, gradient are variables in Paddle.
           },
           [](ExecutionStrategy &self, size_t num_iteration_per_drop_scope) {
             self.num_iteration_per_drop_scope_ = num_iteration_per_drop_scope;
-          });
+          },
+          R"DOC(The type is INT, num_iteration_per_drop_scope indicates how
+                many iterations to clean up the temp variables which
+                is generated during execution. It may make the execution faster,
+                because the temp variable's shape maybe the same between two iterations. Default 100.
+
+                NOTES:
+                    1. If you fetch data when calling the 'run', the ParallelExecutor
+                       will clean up the temp variables at the end of the current iteration.
+                    2. In some NLP model, it may cause the GPU memory is insufficient,
+                       in this case, you should reduce `num_iteration_per_drop_scope`.
+              )DOC");
+
   exec_strategy.def_property(
       "use_experimental_executor",
       [](const ExecutionStrategy &self) {
@@ -678,20 +754,17 @@ All parameter, weight, gradient are variables in Paddle.
     BuildStrategy allows the user to more preciously control how to
     build the SSA Graph in ParallelExecutor by setting the property.
 
-    The available properties include:
-        reduce_strategy (str): There are two reduce strategies, 'AllReduce'
-            and 'Reduce'. If you want that all parameters will be optimized
-            on all devices, you can choose 'AllReduce'; if you choose
-            'Reduce', all parameters will be evenly allocated to different
-            devices for optimization, and then broadcast the optimized
-            parameter to other devices. Default 'AllReduce'.
-        gradient_scale_strategy (str): There are two ways of defining loss@grad,
-            'CoeffNumDevice' and 'Customized'. By default, ParallelExecutor
-            sets the loss@grad according to the number of devices. If you want
-            to customize loss@grad, you can choose 'Customized'.
-            Default 'CoeffNumDevice'.
-        debug_graphviz_path (str): Whether to write the SSA Graph to file in the
-            form of graphviz. It is useful for debugging. Default "".
+    Examples:
+        .. code-block:: python
+
+          build_strategy = fluid.BuildStrategy()
+          build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
+
+          train_exe = fluid.ParallelExecutor(use_cuda=True,
+                                             loss_name=loss.name,
+                                             build_strategy=build_strategy)
+
+          train_loss, = train_exe.run([loss.name], feed=feed_dict)
 )DOC");
 
   py::enum_<BuildStrategy::ReduceStrategy>(build_strategy, "ReduceStrategy")
@@ -710,31 +783,51 @@ All parameter, weight, gradient are variables in Paddle.
           [](const BuildStrategy &self) { return self.reduce_; },
           [](BuildStrategy &self, BuildStrategy::ReduceStrategy strategy) {
             self.reduce_ = strategy;
-          })
+          },
+          R"DOC(The type is STR, there are two reduce strategies in ParallelExecutor,
+                  'AllReduce' and 'Reduce'. If you want that all the parameters'
+                  optimization are done on all devices independently, you should choose 'AllReduce';
+                  if you choose 'Reduce', all the parameters' optimization will be evenly distributed
+                  to different devices, and then broadcast the optimized parameter to other devices.
+                  In some models, `Reduce` is faster. Default 'AllReduce'. )DOC")
       .def_property(
           "gradient_scale_strategy",
           [](const BuildStrategy &self) { return self.gradient_scale_; },
           [](BuildStrategy &self,
              BuildStrategy::GradientScaleStrategy strategy) {
             self.gradient_scale_ = strategy;
-          })
+          },
+          R"DOC(The type is STR, there are three ways of defining :math:`loss@grad` in
+                   ParallelExecutor, 'CoeffNumDevice', 'One' and 'Customized'. By default,
+                   ParallelExecutor sets the :math:`loss@grad` according to the number of devices.
+                   If you want to customize :math:`loss@grad`, you can choose 'Customized'.
+                   Default 'CoeffNumDevice'.)DOC")
       .def_property(
           "debug_graphviz_path",
           [](const BuildStrategy &self) { return self.debug_graphviz_path_; },
           [](BuildStrategy &self, const std::string &path) {
             self.debug_graphviz_path_ = path;
-          })
+          },
+          R"DOC(The type is STR, debug_graphviz_path indicate the path that
+                    writing the SSA Graph to file in the form of graphviz, you.
+                    It is useful for debugging. Default "")DOC")
       .def_property(
           "enable_data_balance",
           [](const BuildStrategy &self) { return self.enable_data_balance_; },
-          [](BuildStrategy &self, bool b) { self.enable_data_balance_ = b; })
-      .def_property("fuse_elewise_add_act_ops",
-                    [](const BuildStrategy &self) {
-                      return self.fuse_elewise_add_act_ops_;
-                    },
-                    [](BuildStrategy &self, bool b) {
-                      self.fuse_elewise_add_act_ops_ = b;
-                    })
+          [](BuildStrategy &self, bool b) {
+            self.enable_data_balance_ = b;
+          })  // FIXME(chengudo): enable_data_balance seems not important
+      .def_property(
+          "fuse_elewise_add_act_ops",
+          [](const BuildStrategy &self) {
+            return self.fuse_elewise_add_act_ops_;
+          },
+          [](BuildStrategy &self, bool b) {
+            self.fuse_elewise_add_act_ops_ = b;
+          },
+          R"DOC(The type is BOOL, fuse_elewise_add_act_ops indicate whether
+                     to fuse elementwise_add_op and activation_op,
+                     it may make the execution faster. Default False)DOC")
       .def("_create_passes_from_strategy",
            [](BuildStrategy &self) -> std::shared_ptr<ir::PassBuilder> {
              return self.CreatePassesFromStrategy();

paddle/fluid/train/demo/README.md

@@ -15,7 +15,7 @@ cmake .. -DFLUID_INSTALL_DIR=$PADDLE_LIB \
          -DWITH_MKL=OFF \
          -DWITH_MKLDNN=OFF
 make -j8
-make -j8 inference_lib_dist
+make -j8 fluid_lib_dist
 ```
 
 ### step 2. generate program desc

paddle/scripts/paddle_build.sh

@@ -648,25 +648,25 @@ function gen_capi_package() {
     fi
 }
 
-function gen_fluid_inference_lib() {
+function gen_fluid_lib() {
     mkdir -p ${PADDLE_ROOT}/build
     cd ${PADDLE_ROOT}/build
     if [[ ${WITH_C_API:-OFF} == "OFF" && ${WITH_INFERENCE:-ON} == "ON" ]] ; then
         cat <<EOF
     ========================================
-    Generating fluid inference library ...
+    Generating fluid library for train and inference ...
     ========================================
 EOF
         cmake .. -DWITH_DISTRIBUTE=OFF
-        make -j `nproc` inference_lib_dist
+        make -j `nproc` fluid_lib_dist
       fi
 }
 
-function tar_fluid_inference_lib() {
+function tar_fluid_lib() {
     if [[ ${WITH_C_API:-OFF} == "OFF" && ${WITH_INFERENCE:-ON} == "ON" ]] ; then
         cat <<EOF
     ========================================
-    Taring fluid inference library ...
+    Taring fluid library for train and inference ...
     ========================================
 EOF
         cd ${PADDLE_ROOT}/build
@@ -675,15 +675,15 @@ EOF
       fi
 }
 
-function test_fluid_inference_lib() {
+function test_fluid_lib() {
     if [[ ${WITH_C_API:-OFF} == "OFF" && ${WITH_INFERENCE:-ON} == "ON" ]] ; then
         cat <<EOF
     ========================================
-    Testing fluid inference library ...
+    Testing fluid library for inference ...
     ========================================
 EOF
         cd ${PADDLE_ROOT}/paddle/fluid/inference/api/demo_ci
-        ./run.sh ${PADDLE_ROOT} ${WITH_MKL:-ON} ${WITH_GPU:-OFF} ${INFERENCE_DEMO_INSTALL_DIR}
+        ./run.sh ${PADDLE_ROOT} ${WITH_MKL:-ON} ${WITH_GPU:-OFF} ${INFERENCE_DEMO_INSTALL_DIR} ${TENSORRT_INCLUDE_DIR:-/usr/local/TensorRT/include} ${TENSORRT_LIB_DIR:-/usr/local/TensorRT/lib}
         ./clean.sh
       fi
 }
@@ -731,9 +731,9 @@ function main() {
         ;;
       fluid_inference_lib)
         cmake_gen ${PYTHON_ABI:-""}
-        gen_fluid_inference_lib
-        tar_fluid_inference_lib
-        test_fluid_inference_lib
+        gen_fluid_lib
+        tar_fluid_lib
+        test_fluid_lib
         ;;
       check_style)
         check_style
@@ -744,8 +744,8 @@ function main() {
         assert_api_not_changed ${PYTHON_ABI:-""}
         run_test
         gen_capi_package
-        gen_fluid_inference_lib
-        test_fluid_inference_lib
+        gen_fluid_lib
+        test_fluid_lib
         assert_api_spec_approvals
         ;;
       maccheck)

python/paddle/fluid/__init__.py

@@ -113,7 +113,8 @@ def __bootstrap__():
         'use_pinned_memory', 'check_nan_inf', 'benchmark', 'warpctc_dir',
         'eager_delete_scope', 'use_mkldnn', 'initial_cpu_memory_in_mb',
         'init_allocated_mem', 'free_idle_memory', 'paddle_num_threads',
-        "dist_threadpool_size", 'cpu_deterministic', 'eager_delete_tensor_gb'
+        'dist_threadpool_size', 'cpu_deterministic', 'eager_delete_tensor_gb',
+        'reader_queue_speed_test_mode'
     ]
     if core.is_compiled_with_dist():
         read_env_flags.append('rpc_deadline')

python/paddle/fluid/layers/io.py

@@ -55,7 +55,11 @@ def data(name,
     Args:
        name(str): The name/alias of the function
        shape(list): Tuple declaring the shape.
-       append_batch_size(bool): Whether or not to append the data as a batch.
+       append_batch_size(bool):
+          1. If true, it prepends -1 to the shape.
+            For example if shape=[1], the resulting shape is [-1, 1].
+          2. If shape contains -1, such as shape=[1, -1],
+            append_batch_size will be enforced to be be False (ineffective).
        dtype(int|float): The type of data : float32, float_16, int etc
        type(VarType): The output type. By default it is LOD_TENSOR.
        lod_level(int): The LoD Level. 0 means the input data is not a sequence.

python/paddle/fluid/layers/nn.py

@@ -56,6 +56,7 @@ __all__ = [
     'sequence_expand',
     'sequence_expand_as',
     'sequence_pad',
+    'sequence_unpad',
     'lstm_unit',
     'reduce_sum',
     'reduce_mean',
@@ -64,6 +65,7 @@ __all__ = [
     'reduce_prod',
     'sequence_first_step',
     'sequence_last_step',
+    'sequence_slice',
     'dropout',
     'split',
     'ctc_greedy_decoder',
@@ -107,6 +109,7 @@ __all__ = [
     'log',
     'crop',
     'rank_loss',
+    'margin_rank_loss',
     'elu',
     'relu6',
     'pow',
@@ -1901,6 +1904,76 @@ def sequence_last_step(input):
     return sequence_pool(input=input, pool_type="last")
 
 
+def sequence_slice(input, offset, length, name=None):
+    """
+    **Sequence Slice Layer**
+
+    The layer crops a subsequence from given sequence with given start 
+    offset and subsequence length.
+
+    It only supports sequence data (LoDTensor with lod_level equal to 1).
+
+    .. code-block:: text
+    
+	- Case:
+
+            Given the input Variable **input**:
+                
+                input.data = [[a1, a2], [b1, b2], [c1, c2], [d1, d2], [e1, e2]],
+                input.lod = [[3, 2]],
+                input.dims = (5, 2),
+
+            with offset.data = [[0], [1]] and length.data = [[2], [1]],
+
+            the output Variable will be
+                
+                out.data = [[a1, a2], [b1, b2], [e1, e2]],
+                out.lod = [[2, 1]],
+                out.dims = (3, 2).
+	
+    NOTE: The first dimension size of **input**, **offset** and **length** 
+          should be equal. The **offset** should start from 0.
+    
+    Args:
+        input(Variable): The input Variable which consists of the complete 
+                         sequences.
+        offset(Variable): The offset to slice each sequence.
+        length(Variable): The length of each subsequence.
+        name(str|None): A name for this layer(optional). If set None, the
+                        layer will be named automatically.
+
+    Returns:
+        Variable: The output subsequences.
+
+    Examples:
+
+        .. code-block:: python
+
+             import numpy as np
+             seqs = fluid.layers.data(name='x', shape=[10, 5],
+                              dtype='float32', lod_level=1)
+             offset = fluid.layers.assign(input=np.array([[0, 1]]).astype("int32"))
+             length = fluid.layers.assign(input=np.array([[2, 1]]).astype("int32"))
+             subseqs = fluid.layers.sequence_slice(input=seqs, offset=offset, 
+                                                   length=length)
+    """
+    helper = LayerHelper("sequence_slice", **locals())
+    dtype = helper.input_dtype()
+    out = helper.create_tmp_variable(dtype)
+
+    offset.stop_gradient = True
+    length.stop_gradient = True
+
+    helper.append_op(
+        type="sequence_slice",
+        inputs={"X": input,
+                "Offset": offset,
+                "Length": length},
+        outputs={"Out": out})
+
+    return out
+
+
 @templatedoc()
 def pool2d(input,
            pool_size=-1,
@@ -2792,7 +2865,7 @@ def sequence_expand_as(x, y, name=None):
 
 
 @templatedoc()
-def sequence_pad(x, pad_value, maxlen=None):
+def sequence_pad(x, pad_value, maxlen=None, name=None):
     """
     ${comment}
 
@@ -2806,7 +2879,9 @@ def sequence_pad(x, pad_value, maxlen=None):
             None or any positive int. When it is None, all sequences will be
             padded up to the length of the longest one among them; when it a
             certain positive value, it must be greater than the length of the
-            longest original sequence."
+            longest original sequence.
+        name(str|None): A name for this layer(optional). If set None, the layer
+            will be named automatically.
 
     Returns:
         Variable: The padded sequence batch and the original lengths before
@@ -2843,6 +2918,66 @@ def sequence_pad(x, pad_value, maxlen=None):
     return out, length
 
 
+def sequence_unpad(x, length, name=None):
+    """
+    **Sequence Unpad Layer**
+
+    This layer removes the padding data in the input sequences and convert 
+    them into sequences with actual length as output, identitied by lod 
+    information.
+
+    .. code-block:: text
+
+	Example:
+
+	Given input Variable **x**:
+	    x.data = [[ 1.0,  2.0,  3.0,  4.0,  5.0],
+		      [ 6.0,  7.0,  8.0,  9.0, 10.0],
+		      [11.0, 12.0, 13.0, 14.0, 15.0]], 
+     
+	in which there are 3 sequences padded to length 5, and the acutal length 
+	specified by input Variable **length**:
+
+	    length.data = [[2], [3], [4]],
+
+	after unpadding, the output Variable will be:
+
+	    out.data = [[1.0, 2.0, 6.0, 7.0, 8.0, 11.0, 12.0, 13.0, 14.0]]
+	    out.lod = [[2, 3, 4]]      
+
+    Args:
+        x(Variable): Input Variable which contains the padded sequences with
+            equal length.
+        length(Variable): The Variable that specifies the actual ength of
+            sequences after unpadding.
+        name(str|None): A name for this layer(optional). If set None, the layer
+            will be named automatically.
+
+    Returns:
+        Variable: The Variable contains the unpadded sequences.
+
+    Examples:
+        .. code-block:: python
+
+            x = fluid.layers.data(name='x', shape=[10, 5], dtype='float32')
+            len = fluid.layers.data(name='length', shape=[1], dtype='int64')
+            out = fluid.layers.sequence_unpad(x=x, length=len)
+    """
+
+    helper = LayerHelper('sequence_unpad', input=x, **locals())
+    dtype = helper.input_dtype()
+    out = helper.create_tmp_variable(dtype)
+
+    length.stop_gradient = True
+
+    helper.append_op(
+        type='sequence_unpad',
+        inputs={'X': x,
+                'Length': length},
+        outputs={'Out': out})
+    return out
+
+
 def beam_search(pre_ids,
                 pre_scores,
                 ids,
@@ -5827,6 +5962,54 @@ def rank_loss(label, left, right, name=None):
     return out
 
 
+def margin_rank_loss(label, left, right, margin=0.1, name=None):
+    """
+    Margin Ranking Loss Layer for ranking problem,
+    which compares left score and right score passed in.
+    The ranking loss can be defined as following equation:
+
+    .. math::
+
+        rank\_loss &= max(0, -label * (left - right) + margin)
+
+    Args:
+       label (Variable): Indicates whether the left is ranked higher than the right or not.
+       left (Variable): Ranking score for left.
+       right (Variable): Ranking score for right.
+       margin (float): Indicates the given margin.
+       name (str|None): A name for this layer (optional). If set None, the layer
+                       will be named automatically.
+    Returns:
+       Variable: The ranking loss.
+    Raises:
+       ValueError: Any of label, left, and right is not a Variable.
+    Examples:
+        .. code-block:: python
+           label = fluid.layers.data(name="label", shape=[4, 1], dtype="float32")
+           left = fluid.layers.data(name="left", shape=[4, 1], dtype="float32")
+           right = fluid.layers.data(name="right", shape=[4, 1], dtype="float32")
+           out = fluid.layers.margin_rank_loss(label, left, right)
+    """
+    helper = LayerHelper('margin_rank_loss', **locals())
+    if not isinstance(label, Variable):
+        raise ValueError("The label should be a Variable.")
+    if not isinstance(left, Variable):
+        raise ValueError("The left should be a Variable.")
+    if not isinstance(right, Variable):
+        raise ValueError("The right should be a Variable.")
+    out = helper.create_tmp_variable(left.dtype)
+    act = helper.create_tmp_variable(left.dtype)
+    helper.append_op(
+        type='margin_rank_loss',
+        inputs={"Label": label,
+                "X1": left,
+                "X2": right},
+        outputs={'Out': out,
+                 'Activated': act},
+        attrs={'margin': margin})
+    return out
+
+
 def pad2d(input,
           paddings=[0, 0, 0, 0],
           mode='constant',
@@ -6290,6 +6473,7 @@ def sequence_enumerate(input, win_size, pad_value=0, name=None):
         outputs={'Out': out},
         attrs={'win_size': win_size,
                'pad_value': pad_value})
+    return out
 
 
 def sequence_mask(x, maxlen=None, dtype='int64', name=None):

python/paddle/fluid/layers/ops.py

@@ -14,6 +14,8 @@
 
 from __future__ import print_function
 from .layer_function_generator import generate_layer_fn, generate_layer_fn_noattr
+from .. import core
+from ..framework import convert_np_dtype_to_dtype_
 
 __activations_noattr__ = [
     'sigmoid',
@@ -58,8 +60,11 @@ _uniform_random_ = generate_layer_fn('uniform_random')
 
 
 def uniform_random(shape, dtype=None, min=None, max=None, seed=None):
+    locals_var = locals().keys()
+    if not isinstance(dtype, core.VarDesc.VarType):
+        dtype = convert_np_dtype_to_dtype_(dtype)
     kwargs = dict()
-    for name in locals():
+    for name in locals_var:
         val = locals()[name]
         if val is not None:
             kwargs[name] = val
@@ -78,8 +83,9 @@ _hard_shrink_ = generate_layer_fn('hard_shrink')
 
 
 def hard_shrink(x, threshold=None):
+    locals_var = locals().keys()
     kwargs = dict()
-    for name in locals():
+    for name in locals_var:
         val = locals()[name]
         if val is not None:
             kwargs[name] = val
@@ -99,12 +105,12 @@ _cum_sum_ = generate_layer_fn('cumsum')
 
 
 def cumsum(x, axis=None, exclusive=None, reverse=None):
+    locals_var = locals().keys()
     kwargs = dict()
-    for name in locals():
+    for name in locals_var:
         val = locals()[name]
         if val is not None:
             kwargs[name] = val
-
     return _cum_sum_(**kwargs)
 
 
@@ -121,8 +127,9 @@ _thresholded_relu_ = generate_layer_fn('thresholded_relu')
 
 
 def thresholded_relu(x, threshold=None):
+    locals_var = locals().keys()
     kwargs = dict()
-    for name in locals():
+    for name in locals_var:
         val = locals()[name]
         if val is not None:
             kwargs[name] = val

python/paddle/fluid/layers/tensor.py

@@ -100,7 +100,7 @@ def create_global_var(shape,
                       force_cpu=False,
                       name=None):
     """
-    Create a new variable in the global block(block 0).
+    Create a new tensor variable with value in the global block(block 0).
 
     Args:
         shape(list[int]): shape of the variable

python/paddle/fluid/optimizer.py

@@ -659,6 +659,9 @@ class AdamaxOptimizer(Optimizer):
 
             optimizer = fluid.optimizer.Adamax(learning_rate=0.2)
             optimizer.minimize(cost)
+
+    Notes:
+       Currently, AdamaxOptimizer doesn't support sparse parameter optimization.
     """
     _moment_acc_str = "moment"
     _inf_norm_acc_str = "inf_norm"
@@ -778,6 +781,9 @@ class DecayedAdagradOptimizer(Optimizer):
 
             optimizer = fluid.optimizer.DecayedAdagrad(learning_rate=0.2)
             optimizer.minimize(cost)
+
+    Notes:
+       Currently, DecayedAdagradOptimizer doesn't support sparse parameter optimization.
     """
     _moment_acc_str = "moment"
 
@@ -858,6 +864,9 @@ class AdadeltaOptimizer(Optimizer):
             optimizer = fluid.optimizer.Adadelta(
                 learning_rate=0.0003, epsilon=1.0e-6, rho=0.95)
             _, params_grads = optimizer.minimize(cost)
+
+    Notes:
+       Currently, AdadeltaOptimizer doesn't support sparse parameter optimization.
     """
 
     _avg_squared_grad_acc_str = "_avg_squared_grad"
@@ -1126,6 +1135,9 @@ class FtrlOptimizer(Optimizer):
 
               optimizer = fluid.optimizer.Ftrl(0.0001)
               _, params_grads = optimizer.minimize(cost)
+
+    Notes:
+       Currently, FtrlOptimizer doesn't support sparse parameter optimization.
     """
 
     _squared_acc_str = "squared"

python/paddle/fluid/parallel_executor.py

@@ -31,15 +31,32 @@ BuildStrategy = core.ParallelExecutor.BuildStrategy
 
 class ParallelExecutor(object):
     """
-    ParallelExecutor can run program in parallel.
+    ParallelExecutor is designed for data parallelism, which focuses on distributing
+    the data across different nodes and every node operates on the data in parallel.
+    If you use ParallelExecutor to run the current program on GPU, the node means GPU
+    device, and ParallelExecutor will get the available GPU device automatically on
+    the current machine. If you use ParallelExecutor to run the current program on CPU,
+    the node means the CPU device, and you can specify the CPU device number by adding
+    'CPU_NUM' environment variable, for example 'CPU_NUM=4', if the environment variable
+    is not found, ParallelExecutor will call `multiprocessing.cpu_count` to get the number
+    of CPUs in the system.
 
     Args:
         use_cuda (bool): Whether to use CUDA or not.
         loss_name (str): The loss name must set in training. Default None.
         main_program (Program): The program that need to run, if not provided,
             then default_main_program will be used. Default None.
-        share_vars_from(ParallelExecutor): If provied, it will share variables
+        share_vars_from(ParallelExecutor): If provide, it will share variables
             from the specified ParallelExecutor. Default None.
+        exec_strategy(ExecutionStrategy): exec_strategy is used to control how to run
+            the program in ParallelExecutor, for example how many threads are used to
+            execute the program, how many iterations to clean up the temp variables
+            which is generated during execution. For more information, please refer
+            to fluid.ExecutionStrategy. Default None.
+        build_strategy(BuildStrategy): build_strategy is used to control how to
+            build the SSA Graph in ParallelExecutor by setting the property,
+            for example reduce_strategy, gradient_scale_strategy. For more information,
+            please refer to fluid.BuildStrategy. Default None.
         num_trainers(int): If greater than 1, NCCL will be initialized with
             multiple rank of nodes, each node should have same number of GPUs.
             Distributed training will be enabled then. Default 1.

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

@@ -81,7 +81,10 @@ def get_optimizer():
     return optimizer
 
 
-def train_network(batch_size, is_distributed=False, is_sparse=False):
+def train_network(batch_size,
+                  is_distributed=False,
+                  is_sparse=False,
+                  is_self_contained_lr=False):
     # query
     q = fluid.layers.data(
         name="query_ids", shape=[1], dtype="int64", lod_level=1)
@@ -93,7 +96,9 @@ def train_network(batch_size, is_distributed=False, is_sparse=False):
         param_attr=fluid.ParamAttr(
             initializer=fluid.initializer.Constant(value=0.01),
             name="__emb__",
-            learning_rate=emb_lr),
+            learning_rate=emb_lr) if is_self_contained_lr else fluid.ParamAttr(
+                initializer=fluid.initializer.Constant(value=0.01),
+                name="__emb__"),
         is_sparse=is_sparse)
     ## vsum
     q_sum = fluid.layers.sequence_pool(input=q_emb, pool_type='sum')
@@ -119,7 +124,9 @@ def train_network(batch_size, is_distributed=False, is_sparse=False):
         param_attr=fluid.ParamAttr(
             initializer=fluid.initializer.Constant(value=0.01),
             name="__emb__",
-            learning_rate=emb_lr),
+            learning_rate=emb_lr) if is_self_contained_lr else fluid.ParamAttr(
+                initializer=fluid.initializer.Constant(value=0.01),
+                name="__emb__"),
         is_sparse=is_sparse)
     ## vsum
     pt_sum = fluid.layers.sequence_pool(input=pt_emb, pool_type='sum')
@@ -144,7 +151,9 @@ def train_network(batch_size, is_distributed=False, is_sparse=False):
         param_attr=fluid.ParamAttr(
             initializer=fluid.initializer.Constant(value=0.01),
             name="__emb__",
-            learning_rate=emb_lr),
+            learning_rate=emb_lr) if is_self_contained_lr else fluid.ParamAttr(
+                initializer=fluid.initializer.Constant(value=0.01),
+                name="__emb__"),
         is_sparse=is_sparse)
     ## vsum
     nt_sum = fluid.layers.sequence_pool(input=nt_emb, pool_type='sum')
@@ -220,7 +229,10 @@ class TestDistSimnetBow2x2(TestDistRunnerBase):
     def get_model(self, batch_size=2):
         # Train program
         avg_cost, acc, predict = \
-            train_network(batch_size, bool(int(os.environ["IS_DISTRIBUTED"])), bool(int(os.environ["IS_SPARSE"])))
+            train_network(batch_size,
+                          bool(int(os.environ["IS_DISTRIBUTED"])),
+                          bool(int(os.environ["IS_SPARSE"])),
+                          bool(int(os.environ["IS_SELF_CONTAINED_LR"])))
 
         inference_program = fluid.default_main_program().clone()
 

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

@@ -18,6 +18,9 @@ import unittest
 import numpy as np
 from op_test import OpTest
 
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+
 
 class TestClipByNormOp(OpTest):
     def setUp(self):
@@ -62,5 +65,59 @@ class TestCase3(TestClipByNormOp):
         self.max_norm = 1.0
 
 
+class TestClipByNormOpWithSelectedRows(OpTest):
+    def check_with_place(self, place):
+        self.config_test_case()
+        scope = core.Scope()
+
+        # set input
+        x_selected_rows = scope.var('X').get_selected_rows()
+        x_selected_rows.set_rows(self.grad_rows)
+        x_tensor = x_selected_rows.get_tensor()
+        x_np = np.random.random(self.grad_shape).astype("float32")
+        x_np[np.abs(x_np) < self.max_relative_error] = 0.5
+        x_tensor.set(x_np, place)
+
+        # set output
+        out_selected_rows = scope.var('Out').get_selected_rows()
+
+        # run clip_by_norm_op
+        clip_by_norm_op = fluid.op.Operator(
+            "clip_by_norm", max_norm=self.max_norm, X='X', Out='Out')
+        clip_by_norm_op.run(scope, place)
+
+        # check output
+        self.assertEqual(out_selected_rows.rows(), self.grad_clipped_rows)
+        out_tensor = out_selected_rows.get_tensor()
+        y_np = np.zeros(self.grad_clipped_shape)
+        y_np[0] = np.sum(x_np[0:2])
+        y_np[1] = x_np[2]
+        y_np[2] = x_np[3]
+        norm = np.sqrt(np.sum(np.square(y_np)))
+        if norm > self.max_norm:
+            output = self.max_norm * y_np / norm
+        else:
+            output = y_np
+        self.assertTrue(
+            np.allclose(
+                np.array(out_tensor), output, atol=1e-5, equal_nan=False))
+
+    def test_clip_by_norm_with_selected_ros(self):
+        places = [core.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(core.CUDAPlace(0))
+
+        for place in places:
+            self.check_with_place(place)
+
+    def config_test_case(self):
+        self.max_norm = 1.0
+        self.max_relative_error = 0.006
+        self.grad_shape = (4, 1)
+        self.grad_clipped_shape = (3, 1)
+        self.grad_rows = [0, 0, 1, 2]
+        self.grad_clipped_rows = [0, 1, 2]
+
+
 if __name__ == '__main__':
     unittest.main()

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

@@ -25,7 +25,11 @@ class TestDistSimnetBowDense2x2(TestDistBase):
         self._enforce_place = "CPU"
 
     def test_simnet_bow(self):
-        need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '0'}
+        need_envs = {
+            "IS_DISTRIBUTED": '0',
+            "IS_SPARSE": '0',
+            'IS_SELF_CONTAINED_LR': '1'
+        }
         self.check_with_place(
             "dist_simnet_bow.py",
             delta=1e-5,
@@ -39,7 +43,11 @@ class TestDistSimnetBow2x2DenseAsync(TestDistBase):
         self._enforce_place = "CPU"
 
     def test_simnet_bow(self):
-        need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '0'}
+        need_envs = {
+            "IS_DISTRIBUTED": '0',
+            "IS_SPARSE": '0',
+            'IS_SELF_CONTAINED_LR': '1'
+        }
         self.check_with_place(
             "dist_simnet_bow.py",
             delta=100,
@@ -53,7 +61,11 @@ class TestDistSimnetBowSparse2x2(TestDistBase):
         self._enforce_place = "CPU"
 
     def test_simnet_bow(self):
-        need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '1'}
+        need_envs = {
+            "IS_DISTRIBUTED": '0',
+            "IS_SPARSE": '1',
+            'IS_SELF_CONTAINED_LR': '1'
+        }
         self.check_with_place(
             "dist_simnet_bow.py",
             delta=1e-5,
@@ -67,7 +79,11 @@ class TestDistSimnetBow2x2SparseAsync(TestDistBase):
         self._enforce_place = "CPU"
 
     def test_simnet_bow(self):
-        need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '1'}
+        need_envs = {
+            "IS_DISTRIBUTED": '0',
+            "IS_SPARSE": '1',
+            'IS_SELF_CONTAINED_LR': '1'
+        }
         self.check_with_place(
             "dist_simnet_bow.py",
             delta=100,
@@ -75,5 +91,59 @@ class TestDistSimnetBow2x2SparseAsync(TestDistBase):
             need_envs=need_envs)
 
 
+class TestDistSimnetBow2x2LookupTableSync(TestDistBase):
+    def _setup_config(self):
+        self._sync_mode = True
+        self._enforce_place = "CPU"
+
+    def test_simnet_bow(self):
+        need_envs = {
+            "IS_DISTRIBUTED": '1',
+            "IS_SPARSE": '1',
+            'IS_SELF_CONTAINED_LR': '1'
+        }
+        self.check_with_place(
+            "dist_simnet_bow.py",
+            delta=1e-5,
+            check_error_log=False,
+            need_envs=need_envs)
+
+
+class TestDistSimnetBow2x2LookupTableAsync(TestDistBase):
+    def _setup_config(self):
+        self._sync_mode = False
+        self._enforce_place = "CPU"
+
+    def test_simnet_bow(self):
+        need_envs = {
+            "IS_DISTRIBUTED": '1',
+            "IS_SPARSE": '1',
+            'IS_SELF_CONTAINED_LR': '1'
+        }
+        self.check_with_place(
+            "dist_simnet_bow.py",
+            delta=100,
+            check_error_log=False,
+            need_envs=need_envs)
+
+
+class TestDistSimnetBow2x2LookupTableNotContainLRSync(TestDistBase):
+    def _setup_config(self):
+        self._sync_mode = True
+        self._enforce_place = "CPU"
+
+    def test_simnet_bow(self):
+        need_envs = {
+            "IS_DISTRIBUTED": '1',
+            "IS_SPARSE": '1',
+            'IS_SELF_CONTAINED_LR': '0'
+        }
+        self.check_with_place(
+            "dist_simnet_bow.py",
+            delta=1e-5,
+            check_error_log=False,
+            need_envs=need_envs)
+
+
 if __name__ == "__main__":
     unittest.main()

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

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+from op_test import OpTest
+from test_lstm_op import lstm, ACTIVATION
+
+
+def fc(x, w, b):
+    return np.dot(x, w) + b
+
+
+def fused_embedded_fc_lstm(
+        ids,  # T x 1
+        lod,  # 1 x N
+        embeddings=None,  # Dict_size x M
+        wx=None,  # M x 4D
+        bx=None,  # 1 x 4D
+        h0=None,  # N x D
+        c0=None,  # N x D
+        w_h=None,  # D x 4D
+        w_b=None,  # 1 x 4D
+        w_c=None,  # 1 x 3D
+        is_reverse=False,
+        act_gate=None,
+        act_cell=None,
+        act_cand=None):
+    # Make a lookup for embeddings and pass result into lstm reference
+    T = ids.shape[0]
+    M = embeddings.shape[1]
+    x = embeddings[ids].reshape([T, M])
+    return lstm(
+        fc(x, wx, bx), lod, h0, c0, w_h, w_b, w_c, is_reverse, act_gate,
+        act_cell, act_cand)
+
+
+class TestFusionLSTMOp(OpTest):
+    def set_conf(self):
+        pass
+
+    def setUp(self):
+        self.op_type = 'fused_embedding_fc_lstm'
+        self.lod = [[2, 3, 5, 4]]
+        self.M = 8  # Embedding size
+        self.D = 16  # Hidden size 
+        self.dict_size = 18
+        self.has_initial_state = False
+        self.use_peepholes = False
+        self.is_reverse = False
+        self.act_gate = 'sigmoid'
+        self.act_cell = 'tanh'
+        self.act_cand = 'tanh'
+        self.set_conf()
+
+        T = sum(self.lod[0])
+        bs = len(self.lod[0])
+
+        # this is the weight of fc
+        wx = np.random.normal(size=(self.M, 4 * self.D)).astype('float32')
+        # this is the bias of fc
+        bx = np.random.normal(size=(1, 4 * self.D)).astype('float32')
+
+        if self.use_peepholes:
+            b = np.random.normal(size=(1, 7 * self.D)).astype('float32')
+        else:
+            b = np.random.normal(size=(1, 4 * self.D)).astype('float32')
+        w_b = np.copy(b[:, 0:4 * self.D])
+        w_c = b[:, 4 * self.D:] if self.use_peepholes else None
+
+        # low is 0 , high is voc_size - 1
+        ids = np.random.randint(
+            low=0, high=self.dict_size - 1, size=(T, 1)).astype("int64")
+        # embeddings as they were trained , so each entry is of M size
+        embeddings = np.random.random(
+            (self.dict_size, self.M)).astype("float32")
+
+        # multiply embeddings via Weights
+        fc_embeddings = np.dot(embeddings, wx)
+
+        # bias should be manually added into the bias of this fused embedding fc LSTM
+        b[0, 0:4 * self.D] += bx[0, :]
+        combined_biases = b[:, 0:4 * self.D]
+        # So let broadcast it , so they can be added
+        ones = np.ones([self.dict_size, 1])
+        broadcasted_biases = np.dot(ones, combined_biases)
+        # Sum biases with Wx*embeddings
+        fc_embeddings += broadcasted_biases
+
+        if self.has_initial_state:
+            h0 = np.random.normal(size=(bs, self.D)).astype('float32')
+            c0 = np.random.normal(size=(bs, self.D)).astype('float32')
+        else:
+            h0 = np.zeros((bs, self.D)).astype('float32')
+            c0 = np.zeros((bs, self.D)).astype('float32')
+
+        wh = np.random.normal(size=(self.D, 4 * self.D)).astype('float32')
+
+        h, c = fused_embedded_fc_lstm(
+            ids, self.lod, embeddings, wx, bx, h0, c0, wh, w_b, w_c,
+            self.is_reverse, ACTIVATION[self.act_gate],
+            ACTIVATION[self.act_cell], ACTIVATION[self.act_cand])
+
+        self.inputs = {
+            'Ids': (ids, self.lod),
+            'Embeddings': fc_embeddings,
+            'WeightH': wh,
+            'Bias': b
+        }
+
+        if self.has_initial_state:
+            self.inputs['H0'] = h0
+            self.inputs['C0'] = c0
+
+        self.outputs = {
+            'Hidden': (h, self.lod),
+            'Cell': (c, self.lod),
+        }
+        self.attrs = {
+            'use_peepholes': self.use_peepholes,
+            'is_reverse': self.is_reverse,
+            'gate_activation': self.act_gate,
+            'cell_activation': self.act_cell,
+            'candidate_activation': self.act_cand
+        }
+
+    def test_check_output(self):
+        for use_seq in {True, False}:
+            self.attrs['use_seq'] = use_seq
+            self.check_output()
+
+
+class TestFusionLSTMOpInit(TestFusionLSTMOp):
+    def set_conf(self):
+        self.has_initial_state = True
+
+
+class TestFusionLSTMOpReverse(TestFusionLSTMOp):
+    def set_conf(self):
+        self.is_reverse = True
+
+
+class TestFusionLSTMOpInitReverse(TestFusionLSTMOp):
+    def set_conf(self):
+        self.has_initial_state = True
+        self.is_reverse = True
+
+
+class TestFusionLSTMOpMD1(TestFusionLSTMOp):
+    def set_conf(self):
+        self.M = 36
+        self.D = 8
+
+
+class TestFusionLSTMOpMD2(TestFusionLSTMOp):
+    def set_conf(self):
+        self.M = 8
+        self.D = 8
+
+
+class TestFusionLSTMOpMD3(TestFusionLSTMOp):
+    def set_conf(self):
+        self.M = 15
+        self.D = 3
+
+
+class TestFusionLSTMOpBS1(TestFusionLSTMOp):
+    def set_conf(self):
+        self.lod = [[3]]
+        self.D = 16
+
+
+class TestFusionLSTMOpPeepholes(TestFusionLSTMOp):
+    def set_conf(self):
+        self.use_peepholes = True
+
+
+class TestFusionLSTMOpPeepholesInit(TestFusionLSTMOp):
+    def set_conf(self):
+        self.use_peepholes = True
+        self.has_initial_state = True
+
+
+class TestFusionLSTMOpPeepholesReverse(TestFusionLSTMOp):
+    def set_conf(self):
+        self.use_peepholes = True
+        self.is_reverse = True
+
+
+class TestFusionLSTMOpPeepholesInitReverse(TestFusionLSTMOp):
+    def set_conf(self):
+        self.use_peepholes = True
+        self.has_initial_state = True
+        self.is_reverse = True
+
+
+class TestFusionLSTMOpPeepholesBS1(TestFusionLSTMOp):
+    def set_conf(self):
+        self.use_peepholes = True
+        self.lod = [[2]]
+        self.D = 8
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -194,6 +194,14 @@ class TestBook(unittest.TestCase):
             self.assertIsNotNone(layers.sequence_expand(x=x, y=y, ref_level=1))
         print(str(program))
 
+    def test_sequence_unpad(self):
+        program = Program()
+        with program_guard(program):
+            x = layers.data(name='x', shape=[10, 5], dtype='float32')
+            length = layers.data(name='length', shape=[1], dtype='int64')
+            self.assertIsNotNone(layers.sequence_unpad(x=x, length=length))
+        print(str(program))
+
     def test_lstm_unit(self):
         program = Program()
         with program_guard(program):
@@ -406,6 +414,19 @@ class TestBook(unittest.TestCase):
             self.assertIsNotNone(out)
         print(str(program))
 
+    def test_sequence_slice(self):
+        program = Program()
+        with program_guard(program):
+            import numpy as np
+            seqs = layers.data(
+                name='x', shape=[10, 5], dtype='float32', lod_level=1)
+            offset = layers.assign(input=np.array([[0, 1]]).astype('int32'))
+            length = layers.assign(input=np.array([[2, 1]]).astype('int32'))
+            out = layers.sequence_slice(
+                input=seqs, offset=offset, length=length)
+            self.assertIsNotNone(out)
+        print(str(program))
+
     def test_lod_reset(self):
         program = Program()
         with program_guard(program):

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

@@ -19,33 +19,76 @@ import unittest
 import numpy as np
 import paddle.fluid.core as core
 from paddle.fluid.op import Operator
+import paddle.fluid as fluid
+
+
+def create_selected_rows_and_tensor(scope, place, height, row_num,
+                                    embedding_size):
+    sr = scope.var("@selected_rows@").get_selected_rows()
+    tensor = scope.var("grad").get_tensor()
+
+    rows = np.random.random_integers(
+        low=0, high=height - 1, size=[row_num, ]).astype('int64')
+    sr_val = np.random.random(size=[row_num, embedding_size]).astype('float32')
+
+    sr.set_height(height)
+    sr.set_rows(rows)
+    sr.get_tensor().set(sr_val, place)
+
+    tensor_val = np.zeros(shape=[height, embedding_size], dtype='float32')
+    for i in range(row_num):
+        row = rows[i]
+        tensor_val[row, :] = tensor_val[row, :] + sr_val[i, :]
+
+    tensor.set(tensor_val, place)
+    return tensor_val, sr_val
 
 
 class TestBase(unittest.TestCase):
-    def setup(self, centered, epsilon=1e-6):
+    def setup(self,
+              place,
+              is_sparse,
+              centered,
+              size,
+              row_num=None,
+              epsilon=1e-6):
         np.random.seed(5)  # fix seed
 
+        self.scope = fluid.global_scope()
+        self.place = place
+
         self.param_name = "param"
-        self.param = np.random.random((123, 321)).astype("float32")
+        self.param = np.random.random(size).astype("float32")
 
         self.mean_square_name = "mean_square"
-        self.mean_square = np.random.random((123, 321)).astype("float32")
+        self.mean_square = np.random.uniform(
+            low=1, high=2, size=size).astype("float32")
 
         self.mean_grad_name = "mean_grad"
-        self.mean_grad = np.random.random((123, 321)).astype("float32")
+        self.mean_grad = np.random.random(size).astype("float32")
 
         self.lr_name = "lr"
         self.learning_rate = np.array([0.01]).astype("float32")
 
         self.grad_name = "grad"
-        self.grad = np.random.random((123, 321)).astype("float32")
+
+        self.is_sparse = is_sparse
+        if self.is_sparse:
+            self.grad_sr_name = "@selected_rows@"
+            self.grad, self.grad_sr = create_selected_rows_and_tensor(
+                self.scope, place, size[0], row_num, size[1])
+        else:
+            self.grad = np.random.random(size).astype("float32")
+            grad_tensor = self.scope.var(self.grad_name).get_tensor()
+            grad_tensor.set(self.grad, place)
 
         self.moment_name = "moment"
-        self.moment = np.zeros((123, 321)).astype("float32")
+        self.moment = np.random.uniform(
+            low=0, high=1, size=size).astype("float32")
 
         self.epsilon = epsilon
         self.decay = 0.9
-        self.momentum = 0.0
+        self.momentum = 0.1
         self.centered = centered
 
         self.ms_out = self.decay * self.mean_square + (1 - self.decay
@@ -61,118 +104,122 @@ class TestBase(unittest.TestCase):
 
         self.param_out = self.param - self.moment_out
 
-    def check(self,
-              actual_t,
-              expect_t,
-              place,
-              out_name,
-              atol=1e-5,
-              equal_nan=False):
-        self.assertTrue(
-            np.allclose(
-                actual_t, expect_t, atol=atol, equal_nan=equal_nan),
-            "Output (" + out_name + ") has diff at " + str(place) + "\nExpect "
-            + str(expect_t) + "\n" + "But Got" + str(actual_t))
-
-
-class TestRmspropOp(TestBase):
-    def check_with_place(self, place, centered, epsilon):
-        self.setup(centered, epsilon)
-        scope = core.Scope()
-
         # create and initialize Param Variable
-        param = scope.var(self.param_name).get_tensor()
-        param.set(self.param, place)
+        self.param_tensor = self.scope.var(self.param_name).get_tensor()
+        self.param_tensor.set(self.param, place)
 
-        mean_square = scope.var(self.mean_square_name).get_tensor()
-        mean_square.set(self.mean_square, place)
+        self.mean_square_tensor = self.scope.var(
+            self.mean_square_name).get_tensor()
+        self.mean_square_tensor.set(self.mean_square, place)
 
-        lr = scope.var(self.lr_name).get_tensor()
+        lr = self.scope.var(self.lr_name).get_tensor()
         lr.set(self.learning_rate, place)
 
-        grad = scope.var(self.grad_name).get_tensor()
-        grad.set(self.grad, place)
+        self.moment_tensor = self.scope.var(self.moment_name).get_tensor()
+        self.moment_tensor.set(self.moment, place)
 
-        moment = scope.var(self.moment_name).get_tensor()
-        moment.set(self.moment, place)
+        if self.centered:
+            self.mean_grad_tensor = self.scope.var(
+                self.mean_grad_name).get_tensor()
+            self.mean_grad_tensor.set(self.mean_grad, place)
 
-        # create and run sgd operator
+    def check(self, actual_t, expect_t, place, out_name, atol=1e-5):
+        self.assertTrue(
+            np.allclose(
+                actual_t, expect_t, atol=atol),
+            "Output (" + out_name + ") has diff at " + str(place) + "\nExpect "
+            + str(expect_t) + "\n" + "But Got" + str(actual_t))
 
-        if self.centered:
-            mean_grad = scope.var(self.mean_grad_name).get_tensor()
-            mean_grad.set(self.mean_grad, place)
-
-            rmsprop_op = Operator(
-                "rmsprop",
-                Param=self.param_name,
-                Grad=self.grad_name,
-                MeanSquare=self.mean_square_name,
-                MeanGrad=self.mean_grad_name,
-                Moment=self.moment_name,
-                LearningRate=self.lr_name,
-                ParamOut=self.param_name,
-                MeanSquareOut=self.mean_square_name,
-                MomentOut=self.moment_name,
-                MeanGradOut=self.mean_grad_name,
-                epsilon=self.epsilon,
-                decay=self.decay,
-                momentum=self.momentum,
-                centered=True)
-        else:
-            rmsprop_op = Operator(
-                "rmsprop",
-                Param=self.param_name,
-                Grad=self.grad_name,
-                MeanSquare=self.mean_square_name,
-                Moment=self.moment_name,
-                LearningRate=self.lr_name,
-                ParamOut=self.param_name,
-                MeanSquareOut=self.mean_square_name,
-                MomentOut=self.moment_name,
-                epsilon=self.epsilon,
-                decay=self.decay,
-                momentum=self.momentum,
-                centered=False)
-
-        rmsprop_op.run(scope, place)
-
-        atol = 1e-5
-        equal_nan = False
+
+class TestRmspropOp(TestBase):
+    def check_with_place(self,
+                         place,
+                         is_sparse,
+                         centered,
+                         size,
+                         row_num=None,
+                         epsilon=1e-6):
+        self.setup(place, is_sparse, centered, size, row_num, epsilon)
+        self.run_and_check()
+
+    def run_and_check(self):
+        grad_name = self.grad_sr_name if self.is_sparse else self.grad_name
+
+        kwargs = {
+            'Param': self.param_name,
+            'Grad': grad_name,
+            'MeanSquare': self.mean_square_name,
+            'Moment': self.moment_name,
+            'LearningRate': self.lr_name,
+            'ParamOut': self.param_name,
+            'MeanSquareOut': self.mean_square_name,
+            'MomentOut': self.moment_name,
+            'epsilon': self.epsilon,
+            'decay': self.decay,
+            'momentum': self.momentum,
+            'centered': self.centered
+        }
 
         if self.centered:
-            atol = 1e-3
-            equal_nan = True
+            kwargs['MeanGrad'] = self.mean_grad_name
+            kwargs['MeanGradOut'] = self.mean_grad_name
+
+        rmsprop_op = Operator('rmsprop', **kwargs)
+        atol = 1e-6
+
+        rmsprop_op.run(self.scope, self.place)
 
         self.check(
-            np.array(mean_square), self.ms_out, place, self.mean_square_name)
+            np.array(self.mean_square_tensor),
+            self.ms_out,
+            self.place,
+            self.mean_square_name,
+            atol=atol)
         self.check(
-            np.array(moment),
+            np.array(self.moment_tensor),
             self.moment_out,
-            place,
+            self.place,
             self.moment_name,
-            atol=atol,
-            equal_nan=equal_nan)
+            atol=atol)
         self.check(
-            np.array(param),
+            np.array(self.param_tensor),
             self.param_out,
-            place,
+            self.place,
             self.param_name,
-            atol=atol,
-            equal_nan=equal_nan)
+            atol=atol)
 
         if self.centered:
             self.check(
-                np.array(mean_grad), self.mg_out, place, self.mean_grad_name)
+                np.array(self.mean_grad_tensor), self.mg_out, self.place,
+                self.mean_grad_name)
 
     def test_rmsprop(self):
         places = [core.CPUPlace()]
         if core.is_compiled_with_cuda():
             places.append(core.CUDAPlace(0))
+
+        size = (128, 320)
         for place in places:
-            self.check_with_place(place, False, 1e-6)
-            self.check_with_place(place, False, 1e-10)
-            self.check_with_place(place, True, 1e-6)
-            self.check_with_place(place, True, 1e-10)
+            for centered in [False, True]:
+                with fluid.scope_guard(core.Scope()):
+                    self.check_with_place(
+                        place, is_sparse=False, centered=centered, size=size)
+
+                with fluid.scope_guard(core.Scope()):
+                    self.check_with_place(
+                        place,
+                        is_sparse=True,
+                        centered=centered,
+                        row_num=512,
+                        size=size)
+
+                with fluid.scope_guard(core.Scope()):
+                    self.check_with_place(
+                        place,
+                        is_sparse=True,
+                        centered=centered,
+                        row_num=60,
+                        size=size)
 
 
 if __name__ == "__main__":

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

+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+import six
+import numpy as np
+from op_test import OpTest
+
+
+class TestSequenceUnpadOp(OpTest):
+    def init(self):
+        self.length = [2, 3, 4]
+        self.x_shape = (3, 5)
+        self.dtype = "float32"
+
+    def compute(self):
+        assert len(self.length) == self.x_shape[0]
+        x = np.random.random(self.x_shape).astype(self.dtype)
+        out_lod = [self.length]
+
+        out = x[0, 0:self.length[0]]
+        for i in six.moves.xrange(1, x.shape[0]):
+            out = np.append(out, x[i, 0:self.length[i]], axis=0)
+
+        out_shape = (sum(self.length), )
+        if len(self.x_shape) == 2:
+            out_shape = out_shape + (1, )
+        else:
+            out_shape = out_shape + self.x_shape[2:]
+
+        self.inputs = {
+            'X': x,
+            'Length': np.array(self.length).astype('int64').reshape(-1, 1)
+        }
+        self.outputs = {'Out': (out.reshape(out_shape), out_lod)}
+
+    def setUp(self):
+        self.op_type = 'sequence_unpad'
+        self.init()
+        self.compute()
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+
+class TestSequenceUnpadOp2(TestSequenceUnpadOp):
+    def init(self):
+        self.length = [2, 3, 4]
+        self.x_shape = (3, 5, 4, 3)
+        self.dtype = "float32"
+
+
+class TestSequenceUnpadOp3(TestSequenceUnpadOp):
+    def init(self):
+        self.length = [5, 2, 3, 4]
+        self.x_shape = (4, 5, 3, 3, 6)
+        self.dtype = "float64"
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -788,7 +788,8 @@ in a single call.")
             tuple: (main_program, startup_program), of type "Program"
         """
         pserver_prog = self.get_pserver_program(endpoint)
-        pserver_startup = self.get_startup_program(endpoint)
+        pserver_startup = self.get_startup_program(
+            endpoint, pserver_program=pserver_prog)
         return pserver_prog, pserver_startup
 
     def get_startup_program(self,
@@ -1118,6 +1119,7 @@ to transpile() call.")
 
     def _split_table_grad_and_add_send_vars(self, program, pserver_endpoints):
         # 2. add split_ids_op and send_op to send gradient to pservers
+
         # there should only be one table_name
         all_ops = program.global_block().ops
         table_grad_name = grad_var_name(self.table_name)
@@ -1142,7 +1144,7 @@ to transpile() call.")
                         if self.sync_mode else []
                     },
                     attrs={
-                        "sync_mode": self.sync_mode,
+                        "sync_mode": not self.sync_mode,
                         "epmap": pserver_endpoints,
                         RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE,
                         OP_ROLE_VAR_ATTR_NAME: [
@@ -1188,7 +1190,15 @@ to transpile() call.")
     def _create_table_optimize_block(self, pserver_index, pserver_program,
                                      pre_block_idx, grad_to_block_id):
         # STEP: create table optimize block
+        table_opt_block = pserver_program._create_block(pre_block_idx)
         # create table param and grad var in pserver program
+        # create table optimize block in pserver program
+        table_opt_op = [
+            op for op in self.optimize_ops
+            if 'Param' in op.input_names and op.input("Param")[0] ==
+            self.table_name
+        ][0]
+
         origin_param_var = self.origin_program.global_block().vars[
             self.table_name]
 
@@ -1204,19 +1214,16 @@ to transpile() call.")
             dtype=origin_param_var.dtype,
             type=core.VarDesc.VarType.SELECTED_ROWS,
             persistable=True)
+
         # parameter must be selected rows
         param_var.desc.set_type(core.VarDesc.VarType.SELECTED_ROWS)
         grad_var = pserver_program.global_block()._clone_variable(
             self.origin_program.global_block().vars[grad_var_name(
                 self.table_name)])
 
-        # create table optimize block in pserver program
-        table_opt_op = [
-            op for op in self.optimize_ops
-            if 'Param' in op.input_names and op.input("Param")[0] ==
-            self.table_name
-        ][0]
-        table_opt_block = pserver_program._create_block(pre_block_idx)
+        lr_var = pserver_program.global_block()._clone_variable(
+            self.origin_program.global_block().vars[table_opt_op.input(
+                "LearningRate")[0]])
 
         if self.sync_mode:
             # create grad vars in pserver program
@@ -1248,8 +1255,6 @@ to transpile() call.")
             grad_var = pserver_program.global_block()._rename_var(
                 origin_grad_name, splited_grad_name)
 
-        lr_var = pserver_program.global_block().vars[table_opt_op.input(
-            "LearningRate")[0]]
         inputs = {
             "Param": [param_var],
             "Grad": [grad_var],