test=develop

CMakeLists.txt

@@ -127,6 +127,9 @@ set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
 set(FLUID_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_install_dir" CACHE STRING
   "A path setting fluid shared and static libraries")
 
+set(FLUID_INFERENCE_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_inference_install_dir" CACHE STRING
+  "A path setting fluid inference shared and static libraries")
+
 if (WITH_C_API AND WITH_PYTHON)
   message(WARNING "It is suggest not embedded a python interpreter in Paddle "
     "when using C-API. It will give an unpredictable behavior when using a "

README.md

@@ -19,7 +19,7 @@ Our vision is to enable deep learning for everyone via PaddlePaddle.
 Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
 
 
-### Latest PaddlePaddle Release: [Fluid 0.15.0](https://github.com/PaddlePaddle/Paddle/tree/v0.15.0)
+### Latest PaddlePaddle Release: [Fluid 1.0.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.0.0)
 ### Install Latest Stable Release:
 ```
 # Linux CPU
@@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==0.15.0.post85
 
 ## Installation
 
-It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/install/install_doc.html) on our website.
+It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) on our website.
 
 ## Documentation
 
-We provide [English](http://paddlepaddle.org/documentation/docs/en/0.15.0/getstarted/index_en.html) and
-[Chinese](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/index.html) documentation.
+We provide [English](http://paddlepaddle.org/documentation/docs/en/1.0.0/getstarted/index_en.html) and
+[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) documentation.
 
 - [Deep Learning 101](https://github.com/PaddlePaddle/book)
 
   You might want to start from this online interactive book that can run in a Jupyter Notebook.
 
-- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/user_guides/howto/training/cluster_howto.html)
+- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.0/user_guides/howto/training/cluster_howto.html)
 
   You can run distributed training jobs on MPI clusters.
 
-- [Python API](http://paddlepaddle.org/documentation/api/zh/0.15.0/fluid.html)
+- [Python API](http://paddlepaddle.org/documentation/api/zh/1.0/fluid.html)
 
    Our new API enables much shorter programs.
 
-- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/advanced_usage/development/contribute_to_paddle.html)
+- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.0/advanced_usage/development/contribute_to_paddle.html)
 
    We appreciate your contributions!
 

cmake/inference_lib.cmake

@@ -150,16 +150,16 @@ if (WITH_ANAKIN AND WITH_MKL)
         SRCS
         ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libinference_anakin_api* # compiled anakin api
         ${ANAKIN_INSTALL_DIR} # anakin release
-        DSTS ${dst_dir}/inference/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin)
+        DSTS ${FLUID_INSTALL_DIR}/third_party/install/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin)
      list(APPEND inference_deps anakin_inference_lib)
 endif()
 
 set(module "inference")
 copy(inference_lib DEPS ${inference_deps}
   SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
-       ${src_dir}/${module}/api/paddle_inference_api.h ${src_dir}/${module}/api/demo_ci
+       ${src_dir}/${module}/api/paddle_inference_api.h
        ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h
-  DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
+  DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
 )
 
 set(module "platform")
@@ -188,18 +188,38 @@ copy(cmake_cache
 # This command generates a complete fluid library for both train and inference
 add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep}) 
 
+# Following commands generate a inference-only fluid library
+# third_party, version.txt and CMakeCache.txt are the same position with ${FLUID_INSTALL_DIR}
+copy(third_party DEPS fluid_lib_dist
+  SRCS ${FLUID_INSTALL_DIR}/third_party ${FLUID_INSTALL_DIR}/CMakeCache.txt
+  DSTS ${FLUID_INFERENCE_INSTALL_DIR} ${FLUID_INFERENCE_INSTALL_DIR}
+)
+
+# only need libpaddle_fluid.so/a and paddle_inference_api.h for inference-only library
+copy(inference_api_lib DEPS fluid_lib_dist
+  SRCS ${FLUID_INSTALL_DIR}/paddle/fluid/inference/libpaddle_fluid.*
+       ${FLUID_INSTALL_DIR}/paddle/fluid/inference/paddle_inference_api.h
+  DSTS ${FLUID_INFERENCE_INSTALL_DIR}/paddle/lib ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include
+)
+
+add_custom_target(inference_lib_dist DEPENDS third_party inference_api_lib)
+
 # paddle fluid version
-execute_process(
-  COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1
-  WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
-  OUTPUT_VARIABLE PADDLE_GIT_COMMIT)
-set(version_file ${FLUID_INSTALL_DIR}/version.txt)
-file(WRITE ${version_file}
-  "GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n"
-  "WITH_MKL: ${WITH_MKL}\n"
-  "WITH_GPU: ${WITH_GPU}\n")
-if(WITH_GPU)
-  file(APPEND ${version_file}
-    "CUDA version: ${CUDA_VERSION}\n"
-    "CUDNN version: v${CUDNN_MAJOR_VERSION}\n")
-endif()
+function(version version_file)
+  execute_process(
+    COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
+    OUTPUT_VARIABLE PADDLE_GIT_COMMIT)
+  file(WRITE ${version_file}
+    "GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n"
+    "WITH_MKL: ${WITH_MKL}\n"
+    "WITH_MKLDNN: ${WITH_MKLDNN}\n"
+    "WITH_GPU: ${WITH_GPU}\n")
+  if(WITH_GPU)
+    file(APPEND ${version_file}
+      "CUDA version: ${CUDA_VERSION}\n"
+      "CUDNN version: v${CUDNN_MAJOR_VERSION}\n")
+  endif()
+endfunction()
+version(${FLUID_INSTALL_DIR}/version.txt)
+version(${FLUID_INFERENCE_INSTALL_DIR}/version.txt)

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,))
@@ -172,6 +174,7 @@ paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dim
 paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.sequence_reverse ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.affine_channel ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
 paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)

paddle/fluid/framework/executor.cc

@@ -101,7 +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) {

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/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/op_desc.h

@@ -100,16 +100,6 @@ class OpDesc {
   std::vector<std::string> InputNames() const { return MapKeys(inputs_); }
   std::vector<std::string> OutputNames() const { return MapKeys(outputs_); }
 
-  void SetInputMap(const VariableNameMap &input) {
-    this->inputs_ = input;
-    this->need_update_ = true;
-  }
-
-  void SetOutputMap(const VariableNameMap &output) {
-    this->outputs_ = output;
-    this->need_update_ = true;
-  }
-
   const VariableNameMap &Inputs() const { return inputs_; }
 
   const VariableNameMap &Outputs() const { return outputs_; }

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/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/analysis/analyzer_tester.cc

@@ -51,9 +51,7 @@ void TestWord2vecPrediction(const std::string& model_path) {
   config.model_dir = model_path;
   config.use_gpu = false;
   config.device = 0;
-  auto predictor =
-      ::paddle::CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(
-          config);
+  auto predictor = ::paddle::CreatePaddlePredictor<NativeConfig>(config);
 
   // One single batch
 

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/analysis_predictor_tester.cc

@@ -27,9 +27,7 @@ TEST(AnalysisPredictor, ZeroCopy) {
   config.model_dir = FLAGS_dirname + "/word2vec.inference.model";
   config.use_feed_fetch_ops = false;
 
-  auto predictor =
-      CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
-          config);
+  auto predictor = CreatePaddlePredictor<AnalysisConfig>(config);
 
   auto w0 = predictor->GetInputTensor("firstw");
   auto w1 = predictor->GetInputTensor("secondw");

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_impl_tester.cc

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

paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc

@@ -41,11 +41,8 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) {
   config1.device = 0;
   config1.max_batch_size = 10;
 
-  auto predictor0 =
-      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config0);
-  auto predictor1 =
-      CreatePaddlePredictor<MixedRTConfig,
-                            PaddleEngineKind::kAutoMixedTensorRT>(config1);
+  auto predictor0 = CreatePaddlePredictor<NativeConfig>(config0);
+  auto predictor1 = CreatePaddlePredictor<MixedRTConfig>(config1);
 
   for (int batch_id = 0; batch_id < 1; batch_id++) {
     //# 2. Prepare input.

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

@@ -77,7 +77,7 @@ endif(NOT WIN32)
 link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib")
 link_directories("${PADDLE_LIB}/third_party/install/glog/lib")
 link_directories("${PADDLE_LIB}/third_party/install/gflags/lib")
-link_directories("${PADDLE_LIB}/paddle/fluid/inference")
+link_directories("${PADDLE_LIB}/paddle/lib")
 
 add_executable(${DEMO_NAME} ${DEMO_NAME}.cc)
 
@@ -97,10 +97,10 @@ endif()
 # Note: libpaddle_inference_api.so/a must put before libpaddle_fluid.so/a
 if(WITH_STATIC_LIB)
   set(DEPS
-      ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX})
+      ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX})
 else()
   set(DEPS
-      ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX})
+      ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX})
 endif()
 
 if (NOT WIN32)

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

@@ -5,12 +5,13 @@ 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=$6 # TensorRT lib file dir, default to /usr/local/TensorRT/lib
+inference_install_dir=${PADDLE_ROOT}/build/fluid_inference_install_dir
 
 cd `dirname $0`
 current_dir=`pwd`
 if [ $2 == ON ]; then
   # You can export yourself if move the install path
-  MKL_LIB=${PADDLE_ROOT}/build/fluid_install_dir/third_party/install/mklml/lib
+  MKL_LIB=${inference_install_dir}/third_party/install/mklml/lib
   export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${MKL_LIB}
 fi
 if [ $3 == ON ]; then
@@ -55,7 +56,7 @@ cd build
 for WITH_STATIC_LIB in ON OFF; do
   # -----simple_on_word2vec-----
   rm -rf *
-  cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \
+  cmake .. -DPADDLE_LIB=${inference_install_dir} \
     -DWITH_MKL=$TURN_ON_MKL \
     -DDEMO_NAME=simple_on_word2vec \
     -DWITH_GPU=$TEST_GPU_CPU \
@@ -75,7 +76,7 @@ for WITH_STATIC_LIB in ON OFF; do
   fi
   # ---------vis_demo---------
   rm -rf *
-  cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \
+  cmake .. -DPADDLE_LIB=${inference_install_dir} \
     -DWITH_MKL=$TURN_ON_MKL \
     -DDEMO_NAME=vis_demo \
     -DWITH_GPU=$TEST_GPU_CPU \
@@ -98,7 +99,7 @@ for WITH_STATIC_LIB in ON OFF; do
   # --------tensorrt mobilenet------
   if [ $USE_TENSORRT == ON -a $TEST_GPU_CPU == ON ]; then
     rm -rf *
-    cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \
+    cmake .. -DPADDLE_LIB=${inference_install_dir} \
       -DWITH_MKL=$TURN_ON_MKL \
       -DDEMO_NAME=trt_mobilenet_demo \
       -DWITH_GPU=$TEST_GPU_CPU \

paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc

@@ -23,7 +23,7 @@ limitations under the License. */
 #include <memory>
 #include <thread>  //NOLINT
 
-#include "paddle/fluid/inference/paddle_inference_api.h"
+#include "paddle/include/paddle_inference_api.h"
 
 DEFINE_string(dirname, "", "Directory of the inference model.");
 DEFINE_bool(use_gpu, false, "Whether use gpu.");
@@ -42,8 +42,7 @@ void Main(bool use_gpu) {
   config.use_gpu = use_gpu;
   config.fraction_of_gpu_memory = 0.15;
   config.device = 0;
-  auto predictor =
-      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
+  auto predictor = CreatePaddlePredictor<NativeConfig>(config);
 
   for (int batch_id = 0; batch_id < 3; batch_id++) {
     //# 2. Prepare input.
@@ -85,8 +84,7 @@ void MainThreads(int num_threads, bool use_gpu) {
   config.use_gpu = use_gpu;
   config.fraction_of_gpu_memory = 0.15;
   config.device = 0;
-  auto main_predictor =
-      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
+  auto main_predictor = CreatePaddlePredictor<NativeConfig>(config);
 
   std::vector<std::thread> threads;
   for (int tid = 0; tid < num_threads; ++tid) {

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

@@ -18,7 +18,7 @@ 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 "paddle/fluid/inference/demo_ci/utils.h"
+#include "utils.h"  // NOLINT
 
 DECLARE_double(fraction_of_gpu_memory_to_use);
 DEFINE_string(modeldir, "", "Directory of the inference model.");

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

@@ -18,7 +18,7 @@
 #include <iostream>
 #include <string>
 #include <vector>
-#include "paddle/fluid/inference/paddle_inference_api.h"
+#include "paddle/include/paddle_inference_api.h"
 
 namespace paddle {
 namespace demo {

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

@@ -18,7 +18,7 @@ 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 "paddle/fluid/inference/demo_ci/utils.h"
+#include "utils.h"  // NOLINT
 
 #ifdef PADDLE_WITH_CUDA
 DECLARE_double(fraction_of_gpu_memory_to_use);
@@ -34,12 +34,13 @@ DEFINE_bool(use_gpu, false, "Whether use gpu.");
 namespace paddle {
 namespace demo {
 
+using contrib::AnalysisConfig;
 /*
- * Use the native fluid engine to inference the demo.
+ * Use the native and analysis fluid engine to inference the demo.
  */
 void Main(bool use_gpu) {
-  std::unique_ptr<PaddlePredictor> predictor;
-  NativeConfig config;
+  std::unique_ptr<PaddlePredictor> predictor, analysis_predictor;
+  AnalysisConfig config;
   config.param_file = FLAGS_modeldir + "/__params__";
   config.prog_file = FLAGS_modeldir + "/__model__";
   config.use_gpu = use_gpu;
@@ -49,8 +50,8 @@ void Main(bool use_gpu) {
   }
 
   VLOG(3) << "init predictor";
-  predictor =
-      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
+  predictor = CreatePaddlePredictor<NativeConfig>(config);
+  analysis_predictor = CreatePaddlePredictor<AnalysisConfig>(config);
 
   VLOG(3) << "begin to process data";
   // Just a single batch of data.
@@ -68,7 +69,7 @@ void Main(bool use_gpu) {
   input.dtype = PaddleDType::FLOAT32;
 
   VLOG(3) << "run executor";
-  std::vector<PaddleTensor> output;
+  std::vector<PaddleTensor> output, analysis_output;
   predictor->Run({input}, &output, 1);
 
   VLOG(3) << "output.size " << output.size();
@@ -77,6 +78,10 @@ void Main(bool use_gpu) {
 
   // compare with reference result
   CheckOutput(FLAGS_refer, tensor);
+
+  // the analysis_output has some diff with native_output,
+  // TODO(luotao): add CheckOutput for analysis_output later.
+  analysis_predictor->Run({input}, &analysis_output, 1);
 }
 
 }  // namespace demo

paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc

@@ -308,18 +308,13 @@ TEST(Analyzer_rnn1, ZeroCopy) {
   PaddlePlace place;
   int output_size{0};
 
-  auto predictor =
-      CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
-          config);
+  auto predictor = CreatePaddlePredictor<AnalysisConfig>(config);
 
   config.use_feed_fetch_ops = true;
-  auto native_predictor =
-      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
+  auto native_predictor = CreatePaddlePredictor<NativeConfig>(config);
 
   config.use_feed_fetch_ops = true;  // the analysis predictor needs feed/fetch.
-  auto analysis_predictor =
-      CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
-          config);
+  auto analysis_predictor = CreatePaddlePredictor<AnalysisConfig>(config);
 
 #define NEW_TENSOR(name__) \
   auto name__##_tensor = predictor->GetInputTensor(#name__);

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

@@ -77,11 +77,9 @@ void CompareResult(const std::vector<PaddleTensor> &outputs,
 std::unique_ptr<PaddlePredictor> CreateTestPredictor(
     const AnalysisConfig &config, bool use_analysis = true) {
   if (use_analysis) {
-    return CreatePaddlePredictor<contrib::AnalysisConfig,
-                                 PaddleEngineKind::kAnalysis>(config);
+    return CreatePaddlePredictor<contrib::AnalysisConfig>(config);
   } else {
-    return CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(
-        config);
+    return CreatePaddlePredictor<NativeConfig>(config);
   }
 }
 

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

@@ -51,11 +51,8 @@ void CompareTensorRTWithFluid(int batch_size, std::string model_dirname) {
   config1.model_dir = model_dirname;
   config1.max_batch_size = batch_size;
 
-  auto predictor0 =
-      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config0);
-  auto predictor1 =
-      CreatePaddlePredictor<MixedRTConfig,
-                            PaddleEngineKind::kAutoMixedTensorRT>(config1);
+  auto predictor0 = CreatePaddlePredictor<NativeConfig>(config0);
+  auto predictor1 = CreatePaddlePredictor<MixedRTConfig>(config1);
   // Prepare inputs
   int height = 224;
   int width = 224;

paddle/fluid/operators/CMakeLists.txt

@@ -300,11 +300,12 @@ 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)
     op_library(reduce_mean_op DEPS cub)
+    op_library(affine_channel_op DEPS cub)
 else()
     op_library(conv_op DEPS vol2col im2col)
 endif()

paddle/fluid/operators/affine_channel_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+Indicesou 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/data_layout.h"
+#include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+class AffineChannelOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(Tensor) Feature map input can be a 4D tensor with order NCHW "
+             "or NHWC. It also can be a 2D tensor and C is the second "
+             "dimension.");
+    AddInput("Scale",
+             "(Tensor) 1D input of shape (C), the c-th element "
+             "is the scale factor of the affine transformation "
+             "for the c-th channel of the input.");
+    AddInput("Bias",
+             "(Tensor) 1D input of shape (C), the c-th element "
+             "is the bias of the affine transformation for the "
+             "c-th channel of the input.");
+    AddAttr<std::string>(
+        "data_layout",
+        "(string, default NCHW) Only used in "
+        "An optional string from: \"NHWC\", \"NCHW\". "
+        "Defaults to \"NHWC\". Specify the data format of the output data, "
+        "the input will be transformed automatically. ")
+        .SetDefault("AnyLayout");
+    AddOutput("Out", "(Tensor) A tensor of the same shape and order with X.");
+    AddComment(R"DOC(
+
+Applies a separate affine transformation to each channel of the input. Useful
+for replacing spatial batch norm with its equivalent fixed transformation.
+The input also can be 2D tensor and applies a affine transformation in second
+dimension.
+
+$$Out = Scale*X + Bias$$
+
+)DOC");
+  }
+};
+
+class AffineChannelOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of AffineChannelOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Scale"),
+                   "Input(Scale) of AffineChannelOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Bias"),
+                   "Input(Bias) of AffineChannelOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of AffineChannelOp should not be null.");
+    ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
+    ctx->ShareLoD("X", "Out");
+  }
+};
+
+class AffineChannelOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Input(Out@GRAD) should not be null.");
+    if (ctx->HasOutput(framework::GradVarName("X"))) {
+      PADDLE_ENFORCE(ctx->HasInput("Scale"),
+                     "Input(Scale) should not be null.");
+      ctx->SetOutputDim(framework::GradVarName("X"),
+                        ctx->GetInputDim(framework::GradVarName("Out")));
+    }
+    if (ctx->HasOutput(framework::GradVarName("Scale"))) {
+      // Scale@GRAD and Bias@GRAD must exist at the same time.
+      PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Bias")),
+                     "Output(Scale@GRAD) should not be null.");
+      PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null.");
+      ctx->SetOutputDim(framework::GradVarName("Scale"),
+                        ctx->GetInputDim("Scale"));
+      ctx->SetOutputDim(framework::GradVarName("Bias"),
+                        ctx->GetInputDim("Scale"));
+    }
+  }
+};
+
+template <typename T>
+using EigenArrayMap =
+    Eigen::Map<Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
+template <typename T>
+using ConstEigenArrayMap =
+    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
+template <typename T>
+using EigenVectorArrayMap = Eigen::Map<Eigen::Array<T, Eigen::Dynamic, 1>>;
+template <typename T>
+using ConstEigenVectorArrayMap =
+    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, 1>>;
+
+template <typename DeviceContext, typename T>
+class AffineChannelKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<framework::Tensor>("X");
+    auto* scale = ctx.Input<framework::Tensor>("Scale");
+    auto* bias = ctx.Input<framework::Tensor>("Bias");
+
+    auto* y = ctx.Output<framework::Tensor>("Out");
+    y->mutable_data<T>(ctx.GetPlace());
+
+    const framework::DataLayout layout =
+        framework::StringToDataLayout(ctx.Attr<std::string>("data_layout"));
+
+    auto dims = x->dims();
+    int N = dims[0];
+    int C = layout == framework::DataLayout::kNCHW ? dims[1]
+                                                   : dims[dims.size() - 1];
+    int HxW = x->numel() / N / C;
+
+    auto* scale_d = scale->data<T>();
+    auto* bias_d = bias->data<T>();
+    ConstEigenVectorArrayMap<T> a_e(scale_d, C);
+    ConstEigenVectorArrayMap<T> b_e(bias_d, C);
+
+    auto* x_d = x->data<T>();
+    auto* y_d = y->data<T>();
+    if (layout == framework::DataLayout::kNCHW) {
+      int stride = C * HxW;
+      for (int i = 0; i < N; i++) {
+        ConstEigenArrayMap<T> x_e(x_d, HxW, C);
+        EigenArrayMap<T> y_e(y_d, HxW, C);
+        y_e = (x_e.rowwise() * a_e.transpose()).rowwise() + b_e.transpose();
+        x_d += stride;
+        y_d += stride;
+      }
+    } else {
+      int num = N * HxW;
+      ConstEigenArrayMap<T> x_e(x_d, C, num);
+      EigenArrayMap<T> y_e(y_d, C, num);
+      y_e = (x_e.colwise() * a_e).colwise() + b_e;
+    }
+  }
+};
+
+template <typename DeviceContext, typename T>
+class AffineChannelGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<framework::Tensor>("X");
+    auto* scale = ctx.Input<framework::Tensor>("Scale");
+    auto* dy = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+
+    auto* dx = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto* dscale =
+        ctx.Output<framework::Tensor>(framework::GradVarName("Scale"));
+    auto* dbias = ctx.Output<framework::Tensor>(framework::GradVarName("Bias"));
+
+    const framework::DataLayout layout =
+        framework::StringToDataLayout(ctx.Attr<std::string>("data_layout"));
+
+    auto dims = x->dims();
+    int N = dims[0];
+    int C = layout == framework::DataLayout::kNCHW ? dims[1]
+                                                   : dims[dims.size() - 1];
+    int HxW = x->numel() / N / C;
+
+    auto* x_d = x->data<T>();
+    auto* dy_d = dy->data<T>();
+    auto* scale_d = scale->data<T>();
+    ConstEigenVectorArrayMap<T> scale_e(scale_d, C);
+
+    T* dx_d = dx ? dx->mutable_data<T>(ctx.GetPlace()) : nullptr;
+    T* dscale_d = dscale ? dscale->mutable_data<T>(ctx.GetPlace()) : nullptr;
+    T* dbias_d = dbias ? dbias->mutable_data<T>(ctx.GetPlace()) : nullptr;
+    EigenVectorArrayMap<T> dscale_e(dscale_d, C);
+    EigenVectorArrayMap<T> dbias_e(dbias_d, C);
+
+    if (layout == framework::DataLayout::kNCHW) {
+      // compute dx
+      int stride = C * HxW;
+      if (dx) {
+        for (int i = 0; i < N; i++) {
+          ConstEigenArrayMap<T> dy_e(dy_d, HxW, C);
+          EigenArrayMap<T> dx_e(dx_d, HxW, C);
+          dx_e = dy_e.rowwise() * scale_e.transpose();
+          dy_d += stride;
+          dx_d += stride;
+        }
+      }
+      // compute dscale and dbias
+      if (dscale && dbias) {
+        dy_d = dy->data<T>();
+        for (int i = 0; i < N; i++) {
+          ConstEigenArrayMap<T> x_e(x_d, HxW, C);
+          ConstEigenArrayMap<T> dy_e(dy_d, HxW, C);
+          if (i == 0) {
+            dscale_e = (x_e * dy_e).colwise().sum();
+          } else {
+            dscale_e += (x_e * dy_e).colwise().sum();
+          }
+          if (i == 0) {
+            dbias_e = dy_e.colwise().sum();
+          } else {
+            dbias_e += dy_e.colwise().sum();
+          }
+          x_d += stride;
+          dy_d += stride;
+        }
+      }
+    } else {
+      int num = N * HxW;
+      ConstEigenArrayMap<T> dy_e(dy_d, C, num);
+      // compute dx
+      if (dx) {
+        EigenArrayMap<T> dx_e(dx_d, C, num);
+        dx_e = dy_e.colwise() * scale_e;
+      }
+      // compute dscale and dbias
+      if (dscale && dbias) {
+        ConstEigenArrayMap<T> x_e(x_d, C, num);
+        dscale_e = (x_e * dy_e).rowwise().sum();
+        dbias_e = dy_e.rowwise().sum();
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+using CPU = paddle::platform::CPUDeviceContext;
+
+REGISTER_OPERATOR(affine_channel, ops::AffineChannelOp,
+                  ops::AffineChannelOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+REGISTER_OPERATOR(affine_channel_grad, ops::AffineChannelOpGrad);
+
+REGISTER_OP_CPU_KERNEL(affine_channel, ops::AffineChannelKernel<CPU, float>,
+                       ops::AffineChannelKernel<CPU, double>);
+REGISTER_OP_CPU_KERNEL(affine_channel_grad,
+                       ops::AffineChannelGradKernel<CPU, float>,
+                       ops::AffineChannelGradKernel<CPU, double>);

paddle/fluid/operators/affine_channel_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.
+Indicesou 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 "cub/cub.cuh"
+#include "paddle/fluid/framework/data_layout.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T, framework::DataLayout layout, bool HasBias>
+__global__ void KeAffineChannelCUDA(const T* x, const T* scale, const T* bias,
+                                    const int C, const int HxW, const int num,
+                                    T* y) {
+  int gid = blockIdx.x * blockDim.x + threadIdx.x;
+  int stride = blockDim.x * gridDim.x;
+  for (int i = gid; i < num; i += stride) {
+    const int c = layout == framework::DataLayout::kNCHW ? i / HxW % C : i % C;
+    if (HasBias) {
+      y[i] = scale[c] * x[i] + bias[c];
+    } else {
+      y[i] = scale[c] * x[i];
+    }
+  }
+}
+
+template <typename DeviceContext, typename T>
+class AffineChannelCUDAKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<framework::Tensor>("X");
+    auto* scale = ctx.Input<framework::Tensor>("Scale");
+    auto* bias = ctx.Input<framework::Tensor>("Bias");
+
+    auto* y = ctx.Output<framework::Tensor>("Out");
+    y->mutable_data<T>(ctx.GetPlace());
+
+    const framework::DataLayout layout =
+        framework::StringToDataLayout(ctx.Attr<std::string>("data_layout"));
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
+
+    auto dims = x->dims();
+    const int num = x->numel();
+    int N = dims[0];
+    int C = layout == framework::DataLayout::kNCHW ? dims[1]
+                                                   : dims[dims.size() - 1];
+    int HxW = num / N / C;
+
+    const T* x_d = x->data<T>();
+    const T* scale_d = scale->data<T>();
+    const T* bias_d = bias->data<T>();
+    T* y_d = y->data<T>();
+
+    int block = 1024;
+    int grid = (num + block - 1) / block;
+    if (layout == framework::DataLayout::kNCHW) {
+      KeAffineChannelCUDA<T, framework::DataLayout::kNCHW,
+                          true><<<grid, block, 0, dev_ctx.stream()>>>(
+          x_d, scale_d, bias_d, C, HxW, num, y_d);
+    } else {
+      KeAffineChannelCUDA<T, framework::DataLayout::kNHWC,
+                          true><<<grid, block, 0, dev_ctx.stream()>>>(
+          x_d, scale_d, bias_d, C, HxW, num, y_d);
+    }
+  }
+};
+
+template <typename T, int BlockDim, framework::DataLayout layout>
+__global__ void AffineChannelScaleBiasGradientCUDAKernel(
+    const T* dy, const T* x, const int N, const int C, const int HxW, T* dscale,
+    T* dbias) {
+  const int outer_size = C;
+  const int inner_size = N * HxW;
+  typedef cub::BlockReduce<T, BlockDim> BlockReduce;
+  __shared__ typename BlockReduce::TempStorage ds_storage;
+  __shared__ typename BlockReduce::TempStorage db_storage;
+
+  for (int i = blockIdx.x; i < outer_size; i += gridDim.x) {
+    T ds_sum = 0;
+    T db_sum = 0;
+    for (int j = threadIdx.x; j < inner_size; j += blockDim.x) {
+      const int index = layout == framework::DataLayout::kNCHW
+                            ? (j / HxW * C + i) * HxW + j % HxW
+                            : j * outer_size + i;
+      ds_sum += dy[index] * x[index];
+      db_sum += dy[index];
+    }
+    ds_sum = BlockReduce(ds_storage).Reduce(ds_sum, cub::Sum());
+    db_sum = BlockReduce(db_storage).Reduce(db_sum, cub::Sum());
+    if (threadIdx.x == 0) {
+      dscale[i] = ds_sum;
+      dbias[i] = db_sum;
+    }
+    __syncthreads();
+  }
+}
+
+template <typename DeviceContext, typename T>
+class AffineChannelGradCUDAKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<framework::Tensor>("X");
+    auto* scale = ctx.Input<framework::Tensor>("Scale");
+    auto* bias = ctx.Input<framework::Tensor>("Bias");
+    auto* dy = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+
+    auto* dx = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto* dscale =
+        ctx.Output<framework::Tensor>(framework::GradVarName("Scale"));
+    auto* dbias = ctx.Output<framework::Tensor>(framework::GradVarName("Bias"));
+
+    const framework::DataLayout layout =
+        framework::StringToDataLayout(ctx.Attr<std::string>("data_layout"));
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
+
+    auto dims = x->dims();
+    const int num = x->numel();
+    int N = dims[0];
+    int C = layout == framework::DataLayout::kNCHW ? dims[1]
+                                                   : dims[dims.size() - 1];
+    int HxW = num / N / C;
+
+    const T* x_d = x->data<T>();
+    const T* dy_d = dy->data<T>();
+    const T* s_d = scale->data<T>();
+
+    T* dx_d = dx ? dx->mutable_data<T>(ctx.GetPlace()) : nullptr;
+    T* ds_d = dscale ? dscale->mutable_data<T>(ctx.GetPlace()) : nullptr;
+    T* db_d = dbias ? dbias->mutable_data<T>(ctx.GetPlace()) : nullptr;
+
+    const int block = 1024;
+    int max_threads = dev_ctx.GetMaxPhysicalThreadCount();
+    const int max_blocks = std::max(max_threads / block, 1);
+    int grid1 = (num + block - 1) / block;
+    int grid2 = std::min(C, max_blocks);
+    if (layout == framework::DataLayout::kNCHW) {
+      if (dx) {
+        KeAffineChannelCUDA<T, framework::DataLayout::kNCHW,
+                            false><<<grid1, block, 0, dev_ctx.stream()>>>(
+            dy_d, s_d, nullptr, C, HxW, num, dx_d);
+      }
+      if (dscale && dbias) {
+        AffineChannelScaleBiasGradientCUDAKernel<
+            T, block, framework::DataLayout::kNCHW><<<grid2, block, 0,
+                                                      dev_ctx.stream()>>>(
+            dy_d, x_d, N, C, HxW, ds_d, db_d);
+      }
+    } else {
+      if (dx) {
+        KeAffineChannelCUDA<T, framework::DataLayout::kNCHW,
+                            false><<<grid1, block, 0, dev_ctx.stream()>>>(
+            dy_d, s_d, nullptr, C, HxW, num, dx_d);
+      }
+      if (dscale && dbias) {
+        AffineChannelScaleBiasGradientCUDAKernel<
+            T, block, framework::DataLayout::kNHWC><<<grid2, block, 0,
+                                                      dev_ctx.stream()>>>(
+            dy_d, x_d, N, C, HxW, ds_d, db_d);
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+using CUDA = paddle::platform::CUDADeviceContext;
+
+REGISTER_OP_CUDA_KERNEL(affine_channel,
+                        ops::AffineChannelCUDAKernel<CUDA, float>,
+                        ops::AffineChannelCUDAKernel<CUDA, double>);
+REGISTER_OP_CUDA_KERNEL(affine_channel_grad,
+                        ops::AffineChannelGradCUDAKernel<CUDA, float>,
+                        ops::AffineChannelGradCUDAKernel<CUDA, double>);

paddle/fluid/operators/distributed/CMakeLists.txt

@@ -20,7 +20,7 @@ if(WITH_GRPC)
     DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc scope profiler math_function SERIAL)
   cc_test(rpc_server_test SRCS rpc_server_test.cc
     DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor  proto_desc lookup_sparse_table_op SERIAL)
-  cc_test(varhandle_test SRCS varhandle_test.cc)
+  cc_test(varhandle_test SRCS varhandle_test.cc DEPS profiler)
   return()
 endif()
 

paddle/fluid/operators/distributed/grpc_client.cc

@@ -12,14 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/operators/distributed/grpc_client.h"
-
 #include <sys/time.h>
-
 #include <limits>
 
 #include "glog/logging.h"  // For VLOG
 #include "paddle/fluid/framework/threadpool.h"
+#include "paddle/fluid/operators/distributed/grpc_client.h"
 #include "paddle/fluid/operators/distributed/grpc_serde.h"
 #include "paddle/fluid/operators/distributed/request_handler.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -73,10 +71,11 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep,
   const framework::Scope* p_scope = &scope;
   const auto ch = GetChannel(ep_val);
   SendProcessor* s = new SendProcessor(ch);
-  VarHandlePtr h(new VarHandle(ep, "Send", var_name_val, p_ctx, p_scope));
+  const std::string method = "SendRPC";
+  VarHandlePtr h(new VarHandle(ep, method, var_name_val, p_ctx, p_scope));
   s->Prepare(h, time_out);
 
-  framework::AsyncIO([var_name_val, p_scope, p_ctx, s, this] {
+  framework::AsyncIO([var_name_val, p_scope, p_ctx, s, method, h, this] {
     auto* var = p_scope->FindVar(var_name_val);
 
     ::grpc::ByteBuffer req;
@@ -87,10 +86,16 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep,
     // stub context
     s->response_call_back_ = nullptr;
 
+    platform::RecordEvent record_event(method, p_ctx);
+
     auto call = s->stub_g_.PrepareUnaryCall(
         s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_);
     call->StartCall();
     call->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
+
+    if (UNLIKELY(platform::IsProfileEnabled())) {
+      h->Wait();
+    }
   });
   req_count_++;
 
@@ -122,10 +127,11 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep,
   const framework::Scope* p_scope = &scope;
   const auto ch = GetChannel(ep_val);
   GetProcessor* s = new GetProcessor(ch);
-  VarHandlePtr h(new VarHandle(ep, "Get", var_name_val, p_ctx, p_scope));
+  const std::string method = "GetRPC";
+  VarHandlePtr h(new VarHandle(ep, method, var_name_val, p_ctx, p_scope));
   s->Prepare(h, time_out);
 
-  framework::AsyncIO([var_name_val, s, this] {
+  framework::AsyncIO([var_name_val, s, method, p_ctx, h, this] {
     // prepare input
     sendrecv::VariableMessage req;
     req.set_varname(var_name_val);
@@ -137,10 +143,16 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep,
     // stub context
     s->response_call_back_ = ProcGetResponse;
 
+    platform::RecordEvent record_event(method, p_ctx);
+
     auto call = s->stub_g_.PrepareUnaryCall(
         s->context_.get(), "/sendrecv.SendRecvService/GetVariable", buf, &cq_);
     call->StartCall();
     call->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
+
+    if (UNLIKELY(platform::IsProfileEnabled())) {
+      h->Wait();
+    }
   });
 
   req_count_++;
@@ -161,12 +173,14 @@ VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep,
   const framework::Scope* p_scope = &scope;
   const auto ch = GetChannel(ep_val);
   GetProcessor* s = new GetProcessor(ch);
-  VarHandlePtr h(
-      new VarHandle(ep, "Prefetch", out_var_name_val, p_ctx, p_scope));
+
+  const std::string method = "PrefetchRPC";
+
+  VarHandlePtr h(new VarHandle(ep, method, out_var_name_val, p_ctx, p_scope));
   s->Prepare(h, time_out);
 
   framework::AsyncIO([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx,
-                      s, this] {
+                      s, method, h, this] {
     auto* var = p_scope->FindVar(in_var_name_val);
 
     ::grpc::ByteBuffer req;
@@ -177,11 +191,17 @@ VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep,
     // stub context
     s->response_call_back_ = ProcGetResponse;
 
+    platform::RecordEvent record_event(method, p_ctx);
+
     auto call = s->stub_g_.PrepareUnaryCall(
         s->context_.get(), "/sendrecv.SendRecvService/PrefetchVariable", req,
         &cq_);
     call->StartCall();
     call->Finish(&s->reply_, &s->status_, static_cast<void*>(s));
+
+    if (UNLIKELY(platform::IsProfileEnabled())) {
+      h->Wait();
+    }
   });
 
   req_count_++;
@@ -193,15 +213,24 @@ VarHandlePtr GRPCClient::AsyncSendBatchBarrier(const std::string& ep,
   const auto ch = GetChannel(ep);
 
   BatchBarrierProcessor* s = new BatchBarrierProcessor(ch);
-  VarHandlePtr h(new VarHandle(ep, "BatchBarrier", BATCH_BARRIER_MESSAGE,
-                               nullptr, nullptr));
+  const std::string method = "BatchBarrierRPC";
+  VarHandlePtr h(
+      new VarHandle(ep, method, BATCH_BARRIER_MESSAGE, nullptr, nullptr));
   s->Prepare(h, time_out);
 
   sendrecv::VariableMessage req;
   req.set_varname(BATCH_BARRIER_MESSAGE);
+
+  platform::RecordEvent record_event(method, nullptr);
+
   auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
   req_count_++;
+
+  if (UNLIKELY(platform::IsProfileEnabled())) {
+    h->Wait();
+  }
+
   return h;
 }
 
@@ -209,15 +238,24 @@ VarHandlePtr GRPCClient::AsyncSendFetchBarrier(const std::string& ep,
                                                int64_t time_out) {
   const auto ch = GetChannel(ep);
   FetchBarrierProcessor* s = new FetchBarrierProcessor(ch);
-  VarHandlePtr h(new VarHandle(ep, "FetchBarrier", FETCH_BARRIER_MESSAGE,
-                               nullptr, nullptr));
+  const std::string method = "FetchBarrierRPC";
+  VarHandlePtr h(
+      new VarHandle(ep, method, FETCH_BARRIER_MESSAGE, nullptr, nullptr));
   s->Prepare(h, time_out);
 
   sendrecv::VariableMessage req;
   req.set_varname(FETCH_BARRIER_MESSAGE);
+
+  platform::RecordEvent record_event(method, nullptr);
+
   auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
   req_count_++;
+
+  if (UNLIKELY(platform::IsProfileEnabled())) {
+    h->Wait();
+  }
+
   return h;
 }
 
@@ -226,15 +264,23 @@ VarHandlePtr GRPCClient::AsyncSendComplete(const std::string& ep,
   const auto ch = GetChannel(ep);
 
   BatchBarrierProcessor* s = new BatchBarrierProcessor(ch);
-  VarHandlePtr h(
-      new VarHandle(ep, "SendComplete", COMPLETE_MESSAGE, nullptr, nullptr));
+  const std::string method = "SendCompleteRPC";
+  VarHandlePtr h(new VarHandle(ep, method, COMPLETE_MESSAGE, nullptr, nullptr));
   s->Prepare(h, time_out);
 
   sendrecv::VariableMessage req;
   req.set_varname(COMPLETE_MESSAGE);
+
+  platform::RecordEvent record_event(method, nullptr);
+
   auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
   req_count_++;
+
+  if (UNLIKELY(platform::IsProfileEnabled())) {
+    h->Wait();
+  }
+
   return h;
 }
 
@@ -244,17 +290,27 @@ VarHandlePtr GRPCClient::AsyncCheckpointNotify(const std::string& ep,
   const auto ch = GetChannel(ep);
 
   CheckpointNotifyProcessor* s = new CheckpointNotifyProcessor(ch);
-  VarHandlePtr h(new VarHandle(ep, "CheckPointNotify", CHECKPOINT_SAVE_MESSAGE,
-                               nullptr, nullptr));
+
+  const std::string method = "CheckPointNotifyRPC";
+
+  VarHandlePtr h(
+      new VarHandle(ep, method, CHECKPOINT_SAVE_MESSAGE, nullptr, nullptr));
   s->Prepare(h, time_out);
 
   sendrecv::VariableMessage req;
   req.set_varname(CHECKPOINT_SAVE_MESSAGE);
   req.set_out_varname(dir);
 
+  platform::RecordEvent record_event(method, nullptr);
+
   auto rpc = s->stub_->AsyncCheckpointNotify(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
   req_count_++;
+
+  if (UNLIKELY(platform::IsProfileEnabled())) {
+    h->Wait();
+  }
+
   return h;
 }
 
@@ -273,12 +329,16 @@ void GRPCClient::Proceed() {
     BaseProcessor* c = static_cast<BaseProcessor*>(tag);
     GPR_ASSERT(ok);
     PADDLE_ENFORCE(c);
+
     if (c->status_.ok()) {
       VLOG(3) << c->GetVarHandlePtr()->String() << " process";
       c->Process();
     } else if (c->status_.error_code() == grpc::StatusCode::DEADLINE_EXCEEDED) {
+      // FIXME(gongwb): parse error_details?
       LOG(ERROR) << c->GetVarHandlePtr()->String()
-                 << " meets grpc error:" << c->status_.error_message();
+                 << " meets grpc error, error_code:" << c->status_.error_code()
+                 << " error_message:" << c->status_.error_message()
+                 << " error_details:" << c->status_.error_details();
       {
         std::lock_guard<std::mutex> lk(sync_mutex_);
         ok_ = false;
@@ -286,7 +346,10 @@ void GRPCClient::Proceed() {
       c->Finish(false);
     } else {
       LOG(FATAL) << c->GetVarHandlePtr()->String()
-                 << " meets grpc error:" << c->status_.error_message();
+                 << " meets grpc error, error_code:" << c->status_.error_code()
+                 << " error_message:" << c->status_.error_message()
+                 << " error_details:" << c->status_.error_details();
+
       c->Finish(false);
     }
 

paddle/fluid/operators/distributed/grpc_serde.cc

@@ -36,6 +36,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
                            const platform::DeviceContext& ctx,
                            ::grpc::ByteBuffer* msg,
                            const std::string& out_name) {
+  platform::RecordEvent record_event("serial", &ctx);
   // Default DestroyCallback does nothing, When using GPU
   // the CPU buffer need to be freed.
   DestroyCallback destroy_callback = [](void* backing) {};
@@ -147,6 +148,7 @@ void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg,
                                const platform::DeviceContext& ctx,
                                const framework::Scope* scope,
                                framework::Variable** var) {
+  platform::RecordEvent record_event("deserial", &ctx);
   operators::distributed::GRPCVariableResponse resp(scope, &ctx);
   PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!");
   *var = resp.GetVar();

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/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/fusion_seqexpand_concat_fc_op.cc

@@ -136,9 +136,9 @@ class FusionSeqExpandConcatFCOpKernel : public framework::OpKernel<T> {
     // since infershape can not get lod info
     PADDLE_ENFORCE_EQ(ref_lod.size(), 1UL, "Only support input lod size is 1.");
     PADDLE_ENFORCE_EQ(in1_lod.size(), 1UL, "Only support input lod size is 1.");
-    PADDLE_ENFORCE_EQ(in1_lod[0].size() - 1, N,
+    PADDLE_ENFORCE_EQ(static_cast<int>(in1_lod[0].size() - 1), N,
                       "Batch size of all inputs should be equal.");
-    PADDLE_ENFORCE_EQ(in1_lod[0][N], N,
+    PADDLE_ENFORCE_EQ(static_cast<int>(in1_lod[0][N]), N,
                       "Seq_length of other inputs should be 1.");
     PADDLE_ENFORCE_EQ(in1_dims[0], N, "input height should be batch size.");
     for (size_t i = 2; i < ins.size(); ++i) {

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/listen_and_serv_op.cc

@@ -66,7 +66,7 @@ static void ParallelExecuteBlocks(
                 << "pointer: " << prepared[run_block].get();
         executor->RunPreparedContext(prepared[run_block].get(), scope);
       } catch (const std::exception &e) {
-        LOG(ERROR) << "run sub program error " << e.what();
+        LOG(FATAL) << "run sub program:" << idx << " error " << e.what();
       }
     }));
   }

paddle/fluid/operators/math/CMakeLists.txt

@@ -45,8 +45,6 @@ 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)
@@ -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/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/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 <random>
+#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/momentum_op.cc

@@ -24,7 +24,7 @@ class MomentumOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContext *ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Param"),
                    "Input(param) of Momentum should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Grad"),
@@ -45,12 +45,15 @@ class MomentumOp : public framework::OperatorWithKernel {
                    "Output(VelocityOut) of Momentum should not be null.");
 
     auto param_dim = ctx->GetInputDim("Param");
-    PADDLE_ENFORCE_EQ(
-        param_dim, ctx->GetInputDim("Grad"),
-        "Param and Grad input of MomentumOp should have the same dimension.");
-    PADDLE_ENFORCE_EQ(
-        param_dim, ctx->GetInputDim("Velocity"),
-        "Param and Velocity of MomentumOp should have the same dimension.");
+    if (ctx->GetInputsVarType("Grad")[0] ==
+        framework::proto::VarType::LOD_TENSOR) {
+      PADDLE_ENFORCE_EQ(
+          param_dim, ctx->GetInputDim("Grad"),
+          "Param and Grad input of MomentumOp should have the same dimension.");
+      PADDLE_ENFORCE_EQ(
+          param_dim, ctx->GetInputDim("Velocity"),
+          "Param and Velocity of MomentumOp should have the same dimension.");
+    }
     PADDLE_ENFORCE_EQ(framework::product(ctx->GetInputDim("LearningRate")), 1,
                       "Learning_rate should be a scalar");
 
@@ -58,13 +61,34 @@ class MomentumOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("VelocityOut", param_dim);
   }
   framework::OpKernelType GetExpectedKernelType(
-      const framework::ExecutionContext &ctx) const override {
-    auto input_data_type =
-        framework::ToDataType(ctx.Input<Tensor>("Param")->type());
+      const framework::ExecutionContext& ctx) const override {
+    auto input_data_type = framework::GetDataTypeOfVar(ctx.InputVar("Param"));
     return framework::OpKernelType(input_data_type, ctx.GetPlace());
   }
 };
 
+class MomentumOpInferVarType : 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 if (block->FindRecursiveOrCreateVar(input_var).GetType() ==
+                 framework::proto::VarType::LOD_TENSOR) {
+        block->FindRecursiveOrCreateVar(out_var).SetType(
+            framework::proto::VarType::LOD_TENSOR);
+      } else {
+        PADDLE_THROW(
+            "Only support LodTensor and SelectedRows, Unexpected Input Type.");
+      }
+    }
+  }
+};
+
 class MomentumOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
@@ -115,6 +139,9 @@ $$
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(momentum, ops::MomentumOp, ops::MomentumOpMaker);
-REGISTER_OP_CPU_KERNEL(momentum, ops::MomentumOpKernel<float>,
-                       ops::MomentumOpKernel<double>);
+REGISTER_OPERATOR(momentum, ops::MomentumOp, ops::MomentumOpMaker,
+                  paddle::framework::EmptyGradOpMaker,
+                  ops::MomentumOpInferVarType);
+REGISTER_OP_CPU_KERNEL(
+    momentum, ops::MomentumOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::MomentumOpKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/momentum_op.cu

@@ -15,76 +15,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/momentum_op.h"
 
-namespace paddle {
-namespace operators {
-
-template <typename T>
-__global__ void MomentumKernel(const T* p, const T* g, const T* v,
-                               const T* learning_rate, const T mu,
-                               const int64_t num, bool use_nesterov, T* p_out,
-                               T* v_out) {
-  T lr = learning_rate[0];
-  if (use_nesterov) {
-    for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num;
-         i += blockDim.x * gridDim.x) {
-      T g_val = g[i];
-      T v_new = v[i] * mu + g_val;
-      v_out[i] = v_new;
-      p_out[i] = p[i] - (g_val + v_new * mu) * lr;
-    }
-  } else {
-    for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num;
-         i += blockDim.x * gridDim.x) {
-      T v_new = v[i] * mu + g[i];
-      v_out[i] = v_new;
-      p_out[i] = p[i] - lr * v_new;
-    }
-  }
-}
-
-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");
-    auto velocity = ctx.Input<framework::Tensor>("Velocity");
-    auto grad = ctx.Input<framework::Tensor>("Grad");
-    auto learning_rate = ctx.Input<framework::Tensor>("LearningRate");
-
-    T* p_out = param_out->mutable_data<T>(ctx.GetPlace());
-    T* v_out = velocity_out->mutable_data<T>(ctx.GetPlace());
-
-    T mu = static_cast<T>(ctx.Attr<float>("mu"));
-    bool use_nesterov = ctx.Attr<bool>("use_nesterov");
-
-    auto* p = param->data<T>();
-    auto* v = velocity->data<T>();
-    auto* g = grad->data<T>();
-    auto* lr = learning_rate->data<T>();
-
-    int block = 512;
-    int grid = (param->numel() + block - 1) / block;
-    MomentumKernel<T><<<grid, block, 0, ctx.cuda_device_context().stream()>>>(
-        p, g, v, lr, mu, param->numel(), use_nesterov, p_out, v_out);
-  }
-};
-
-}  // namespace operators
-}  // namespace paddle
-
 namespace ops = paddle::operators;
-REGISTER_OP_CUDA_KERNEL(momentum, ops::MomentumOpCUDAKernel<float>,
-                        ops::MomentumOpCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(
+    momentum, ops::MomentumOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::MomentumOpKernel<paddle::platform::CUDADeviceContext, double>);

paddle/fluid/operators/momentum_op.h

@@ -13,35 +13,48 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <string>
 #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 {
 
-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");
-    auto velocity = ctx.Input<framework::Tensor>("Velocity");
-    auto grad = ctx.Input<framework::Tensor>("Grad");
-    auto learning_rate = ctx.Input<framework::Tensor>("LearningRate");
+using framework::Tensor;
+using framework::SelectedRows;
+struct NoNesterov;
+struct UseNesterov;
 
-    param_out->mutable_data<T>(ctx.GetPlace());
-    velocity_out->mutable_data<T>(ctx.GetPlace());
+template <typename T>
+class CPUDenseMomentumFunctor {
+ private:
+  const Tensor* param;
+  const Tensor* grad;
+  const Tensor* velocity;
+  const Tensor* learning_rate;
+  const T mu;
+  const T use_nesterov;
+  Tensor* param_out;
+  Tensor* velocity_out;
 
-    T mu = static_cast<T>(ctx.Attr<float>("mu"));
-    bool use_nesterov = ctx.Attr<bool>("use_nesterov");
+ public:
+  CPUDenseMomentumFunctor(const Tensor* param, const Tensor* grad,
+                          const Tensor* velocity, const Tensor* learning_rate,
+                          const T mu, const bool use_nesterov,
+                          Tensor* param_out, Tensor* velocity_out)
+      : param(param),
+        grad(grad),
+        velocity(velocity),
+        learning_rate(learning_rate),
+        mu(mu),
+        use_nesterov(use_nesterov),
+        param_out(param_out),
+        velocity_out(velocity_out) {}
 
+  inline void operator()() {
     auto p_out = framework::EigenVector<T>::Flatten(*param_out);
     auto v_out = framework::EigenVector<T>::Flatten(*velocity_out);
 
@@ -59,5 +72,283 @@ class MomentumOpKernel : public framework::OpKernel<T> {
   }
 };
 
+template <typename T, typename UpdateMethod>
+class DenseMomentumFunctor;
+
+// NOTE(dzh) for performance.
+// avoid if/else in inside kernel, implement GPU UseNesterov/NoNesterov as two
+// functor.
+template <typename T>
+class DenseMomentumFunctor<T, UseNesterov> {
+ private:
+  const T* p_;
+  const T* g_;
+  const T* v_;
+  const T* lr_;
+  const T mu_;
+  const int64_t num_;
+  T* p_out_;
+  T* v_out_;
+
+ public:
+  DenseMomentumFunctor(const T* p, const T* g, const T* v,
+                       const T* learning_rate, const T mu, const int64_t num,
+                       T* p_out, T* v_out)
+      : p_(p),
+        g_(g),
+        v_(v),
+        lr_(learning_rate),
+        mu_(mu),
+        num_(num),
+        p_out_(p_out),
+        v_out_(v_out) {}
+  inline HOSTDEVICE void operator()(size_t i) const {
+    // put memory access in register
+    const T p = p_[i];
+    const T g = g_[i];
+    const T lr = lr_[0];
+    const T v = v_[i];
+    T v_out = v * mu_ + g;
+    T p_out = p - (g + v_out * mu_) * lr;
+    // write reigster to memory
+    v_out_[i] = v_out;
+    p_out_[i] = p_out;
+  }
+};
+
+template <typename T>
+class DenseMomentumFunctor<T, NoNesterov> {
+ private:
+  const T* p_;
+  const T* g_;
+  const T* v_;
+  const T* lr_;
+  const T mu_;
+  const int64_t num_;
+  T* p_out_;
+  T* v_out_;
+
+ public:
+  DenseMomentumFunctor(const T* p, const T* g, const T* v,
+                       const T* learning_rate, const T mu, const int64_t num,
+                       T* p_out, T* v_out)
+      : p_(p),
+        g_(g),
+        v_(v),
+        lr_(learning_rate),
+        mu_(mu),
+        num_(num),
+        p_out_(p_out),
+        v_out_(v_out) {}
+  inline HOSTDEVICE void operator()(size_t i) const {
+    // put memory access in register
+    const T p = p_[i];
+    const T g = g_[i];
+    const T lr = lr_[0];
+    const T v = v_[i];
+    T v_out = v * mu_ + g;
+    T p_out = p - lr * v_out;
+    // write reigster to memory
+    v_out_[i] = v_out;
+    p_out_[i] = p_out;
+  }
+};
+
+template <typename T, typename UpdateMethod>
+class SparseMomentumFunctor;
+
+template <typename T>
+class SparseMomentumFunctor<T, UseNesterov> {
+ private:
+  const T* p_;
+  const T* g_;
+  const T* v_;
+  const T* lr_;
+  const T mu_;
+  const int64_t* rows_;
+  const int64_t row_numel_;
+  const int64_t row_height_;
+  T* p_out_;
+  T* v_out_;
+
+ public:
+  SparseMomentumFunctor(const T* p, const T* g, const T* v, const T* lr,
+                        const T mu, const int64_t* rows, int64_t row_numel,
+                        int64_t row_height, T* p_out, T* v_out)
+      : p_(p),
+        g_(g),
+        v_(v),
+        lr_(lr),
+        mu_(mu),
+        rows_(rows),
+        row_numel_(row_numel),
+        row_height_(row_height),
+        p_out_(p_out),
+        v_out_(v_out) {}
+
+  inline HOSTDEVICE void operator()(size_t i) {
+    auto row_idx =
+        math::BinarySearch<int64_t>(rows_, row_height_, i / row_numel_);
+    T g = row_idx >= 0 ? g_[row_idx * row_numel_ + i % row_numel_] : 0;
+    // put memory access in register
+    const T p = p_[i];
+    const T lr = lr_[0];
+    const T v = v_[i];
+    T v_out = v * mu_ + g;
+    T p_out = p - (g + v_out * mu_) * lr;
+    // write reigster to memory
+    v_out_[i] = v_out;
+    p_out_[i] = p_out;
+  }
+};
+
+template <typename T>
+class SparseMomentumFunctor<T, NoNesterov> {
+ private:
+  const T* p_;
+  const T* g_;
+  const T* v_;
+  const T* lr_;
+  const T mu_;
+  const int64_t* rows_;
+  const int64_t row_numel_;
+  const int64_t row_height_;
+  T* p_out_;
+  T* v_out_;
+
+ public:
+  SparseMomentumFunctor(const T* p, const T* g, const T* v, const T* lr,
+                        const T mu, const int64_t* rows, int64_t row_numel,
+                        int64_t row_height, T* p_out, T* v_out)
+      : p_(p),
+        g_(g),
+        v_(v),
+        lr_(lr),
+        mu_(mu),
+        rows_(rows),
+        row_numel_(row_numel),
+        row_height_(row_height),
+        p_out_(p_out),
+        v_out_(v_out) {}
+
+  inline HOSTDEVICE void operator()(size_t i) {
+    auto row_idx =
+        math::BinarySearch<int64_t>(rows_, row_height_, i / row_numel_);
+    T g = row_idx >= 0 ? g_[row_idx * row_numel_ + i % row_numel_] : 0;
+    // put memory access in register
+    const T p = p_[i];
+    const T lr = lr_[0];
+    const T v = v_[i];
+    T v_out = v * mu_ + g;
+    T p_out = p - v_out * lr;
+    // write reigster to memory
+    v_out_[i] = v_out;
+    p_out_[i] = p_out;
+  }
+};
+
+template <typename DeviceContext, typename T>
+class MomentumOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    T mu = static_cast<T>(ctx.Attr<float>("mu"));
+    bool use_nesterov = ctx.Attr<bool>("use_nesterov");
+
+    auto learning_rate = ctx.Input<framework::Tensor>("LearningRate");
+    auto param = ctx.Input<framework::Tensor>("Param");
+    auto param_out = ctx.Output<framework::Tensor>("ParamOut");
+    auto* velocity = ctx.Input<framework::Tensor>("Velocity");
+    auto velocity_out = ctx.Output<framework::Tensor>("VelocityOut");
+    param_out->mutable_data<T>(ctx.GetPlace());
+    velocity_out->mutable_data<T>(ctx.GetPlace());
+
+    auto* grad_var = ctx.InputVar("Grad");
+    if (grad_var->IsType<framework::LoDTensor>()) {
+      auto grad = ctx.Input<framework::Tensor>("Grad");
+      if (platform::is_cpu_place(ctx.GetPlace())) {
+        CPUDenseMomentumFunctor<T> functor(param, grad, velocity, learning_rate,
+                                           mu, use_nesterov, param_out,
+                                           velocity_out);
+        functor();
+      } else if (platform::is_gpu_place(ctx.GetPlace())) {
+        platform::ForRange<DeviceContext> for_range(
+            static_cast<const DeviceContext&>(ctx.device_context()),
+            param->numel());
+        if (use_nesterov) {
+          DenseMomentumFunctor<T, UseNesterov> functor(
+              param->data<T>(), grad->data<T>(), velocity->data<T>(),
+              learning_rate->data<T>(), mu, param->numel(),
+              param_out->mutable_data<T>(ctx.GetPlace()),
+              velocity_out->mutable_data<T>(ctx.GetPlace()));
+          for_range(functor);
+
+        } else {
+          DenseMomentumFunctor<T, NoNesterov> functor(
+              param->data<T>(), grad->data<T>(), velocity->data<T>(),
+              learning_rate->data<T>(), mu, param->numel(),
+              param_out->mutable_data<T>(ctx.GetPlace()),
+              velocity_out->mutable_data<T>(ctx.GetPlace()));
+          for_range(functor);
+        }
+      }
+
+    } else if (grad_var->IsType<framework::SelectedRows>()) {
+      // sparse update embedding with selectedrows
+      auto grad = ctx.Input<framework::SelectedRows>("Grad");
+
+      // sparse update maybe empty.
+      if (grad->rows().size() == 0) {
+        VLOG(3) << "Grad SelectedRows contains no data!";
+        return;
+      }
+      auto* merged_grad = const_cast<framework::Scope&>(ctx.scope())
+                              .Var()
+                              ->GetMutable<framework::SelectedRows>();
+      math::scatter::MergeAdd<DeviceContext, T> merge_func;
+      merge_func(ctx.template device_context<DeviceContext>(), *grad,
+                 merged_grad);
+
+      const int64_t* rows = nullptr;
+#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
+      int64_t row_numel =
+          merged_grad->value().numel() / merged_grad->rows().size();
+      platform::ForRange<DeviceContext> for_range(
+          static_cast<const DeviceContext&>(ctx.device_context()),
+          param->numel());
+      if (use_nesterov) {
+        SparseMomentumFunctor<T, UseNesterov> functor(
+            param->data<T>(), merged_grad->value().data<T>(),
+            velocity->data<T>(), learning_rate->data<T>(), mu, rows, row_numel,
+            static_cast<int64_t>(merged_grad->rows().size()),
+            param_out->mutable_data<T>(ctx.GetPlace()),
+            velocity_out->mutable_data<T>(ctx.GetPlace()));
+        for_range(functor);
+
+      } else {
+        SparseMomentumFunctor<T, NoNesterov> functor(
+            param->data<T>(), merged_grad->value().data<T>(),
+            velocity->data<T>(), learning_rate->data<T>(), mu, rows, row_numel,
+            static_cast<int64_t>(merged_grad->rows().size()),
+            param_out->mutable_data<T>(ctx.GetPlace()),
+            velocity_out->mutable_data<T>(ctx.GetPlace()));
+        for_range(functor);
+      }
+    } else {
+      PADDLE_THROW(
+          string::Sprintf("MomentumOp only supports LoDTensor or SelectedRows "
+                          "gradient, but the received Variable Type is %s",
+                          grad_var->Type().name()));
+    }
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle

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/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/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_reverse_op.h

@@ -70,6 +70,8 @@ Y.data() = [
 ]
 
 This Operator is useful to build a reverse dynamic RNN network.
+
+This Operator only supports one-level lod currently.
     )DOC");
   }
 };

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/sum_op.h

@@ -43,17 +43,31 @@ class SumKernel : public framework::OpKernel<T> {
         out->mutable_data<T>(context.GetPlace());
       }
       auto result = EigenVector<T>::Flatten(*out);
+      auto &place =
+          *context.template device_context<DeviceContext>().eigen_device();
+      int start = in_place ? 1 : 0;
       if (!in_place) {
-        math::SetConstant<DeviceContext, T> constant_functor;
-        constant_functor(context.template device_context<DeviceContext>(), out,
-                         0.0);
+        if ((in_num >= 2) && in_vars[0]->IsType<framework::LoDTensor>() &&
+            in_vars[1]->IsType<framework::LoDTensor>()) {
+          auto &in_0 = in_vars[0]->Get<framework::LoDTensor>();
+          auto &in_1 = in_vars[1]->Get<framework::LoDTensor>();
+          if (in_0.numel() && in_1.numel()) {
+            auto in_0_e = EigenVector<T>::Flatten(in_0);
+            auto in_1_e = EigenVector<T>::Flatten(in_1);
+            result.device(place) = in_0_e + in_1_e;
+            start = 2;
+          }
+        }
+        if (start != 2) {
+          math::SetConstant<DeviceContext, T> constant_functor;
+          constant_functor(context.template device_context<DeviceContext>(),
+                           out, 0.0);
+        }
       }
 
       math::SelectedRowsAddToTensor<DeviceContext, T> functor;
-      auto &place =
-          *context.template device_context<DeviceContext>().eigen_device();
       // If in_place, just skip the first tensor
-      for (size_t i = in_place ? 1 : 0; i < in_num; i++) {
+      for (size_t i = start; i < in_num; i++) {
         if (in_vars[i]->IsType<framework::LoDTensor>()) {
           auto &in_t = in_vars[i]->Get<framework::LoDTensor>();
           if (in_t.numel() == 0) {

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/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

@@ -370,8 +370,8 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
   std::vector<std::vector<Event>> merged_events_list;
   if (merge_thread) {
     std::vector<Event> merged_events;
-    for (int i = 0; i < events.size(); ++i) {
-      for (int j = 0; j < events[i].size(); ++j) {
+    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]);
       }
     }

paddle/fluid/platform/profiler.h

@@ -71,6 +71,7 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx);
 
 #if !defined(_WIN32)
 struct RecordEvent {
+  // dev_ctx can be set to nullptr if device is cpu.
   RecordEvent(const std::string& name, const DeviceContext* dev_ctx);
 
   ~RecordEvent();

paddle/scripts/paddle_build.sh

@@ -390,7 +390,9 @@ function run_mac_test() {
     Running unit tests ...
     ========================================
 EOF
-
+        #remove proxy here to fix dist error on mac
+        export http_proxy=
+        export https_proxy=
         # TODO: jiabin need to refine this part when these tests fixed on mac
         ctest --output-on-failure -j $1     
         # make install should also be test when unittest 
@@ -659,6 +661,7 @@ function gen_fluid_lib() {
 EOF
         cmake .. -DWITH_DISTRIBUTE=OFF
         make -j `nproc` fluid_lib_dist
+        make -j `nproc` inference_lib_dist
       fi
 }
 
@@ -672,6 +675,8 @@ EOF
         cd ${PADDLE_ROOT}/build
         cp -r fluid_install_dir fluid
         tar -czf fluid.tgz fluid
+        cp -r fluid_inference_install_dir fluid_inference
+        tar -czf fluid_inference.tgz fluid_inference
       fi
 }
 
@@ -683,7 +688,9 @@ function test_fluid_lib() {
     ========================================
 EOF
         cd ${PADDLE_ROOT}/paddle/fluid/inference/api/demo_ci
-        ./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}
+        ./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
 }

python/paddle/fluid/framework.py

@@ -1522,13 +1522,17 @@ class Program(object):
             >>> with program.lr_schedule_guard():
             >>>     lr = lr * decay
         """
+
+        tmp_role = self._current_role
+        tmp_var = self._op_role_var
+
         OpRole = core.op_proto_and_checker_maker.OpRole
         self._current_role = OpRole.LRSched
         # TODO(typhoonzero): how to set target learning rate var
         self._op_role_var = []
         yield
-        self._op_role_var = []
-        self._current_role = OpRole.Forward
+        self._op_role_var = tmp_var
+        self._current_role = tmp_role
 
     def __str__(self):
         """

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',
@@ -152,6 +154,7 @@ __all__ = [
     'sigmoid_cross_entropy_with_logits',
     'maxout',
     'sequence_reverse',
+    'affine_channel',
 ]
 
 
@@ -1903,6 +1906,76 @@ 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,
@@ -2245,19 +2318,28 @@ def layer_norm(input,
     Args:
         input(Variable): The input tensor variable.
         scale(bool): Whether to learn the adaptive gain :math:`g` after
-            normalization.
+            normalization. Default True.
         shift(bool): Whether to learn the adaptive bias :math:`b` after
-            normalization.
-        begin_norm_axis(bool): The normalization will be performed along
+            normalization. Default True.
+        begin_norm_axis(int): The normalization will be performed along
             dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
+            Default 1.
         epsilon(float): The small value added to the variance to prevent
-            division by zero.
+            division by zero. Default 1e-05.
         param_attr(ParamAttr|None): The parameter attribute for the learnable
-            gain :math:`g`.
+            gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is
+            omitted. If :attr:`scale` is True and :attr:`param_attr` is None,
+            a default :code:`ParamAttr` would be added as scale. The 
+            :attr:`param_attr` is initialized as 1 if it is added. Default None. 
         bias_attr(ParamAttr|None): The parameter attribute for the learnable
-            bias :math:`b`.
+            bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is
+            omitted. If :attr:`shift` is True and :attr:`param_attr` is None,
+            a default :code:`ParamAttr` would be added as bias. The 
+            :attr:`bias_attr` is initialized as 0 if it is added. Default None.
         act(str): Activation to be applied to the output of layer normalizaiton.
-        name (str): The name of this layer. It is optional.
+                  Default None.
+        name(str): The name of this layer. It is optional. Default None, and a
+                   unique name would be generated automatically.
 
     Returns:
         ${y_comment}
@@ -2794,7 +2876,7 @@ 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}
 
@@ -2808,7 +2890,9 @@ 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
@@ -2845,6 +2929,66 @@ 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,
@@ -7150,7 +7294,6 @@ def sequence_reverse(x, name=None):
         out(${y_type}): ${y_comment}
     """
     helper = LayerHelper("sequence_reverse", **locals())
-
     if name is None:
         out = helper.create_tmp_variable(dtype=x.dtype)
     else:
@@ -7163,3 +7306,43 @@ def sequence_reverse(x, name=None):
         outputs={"Y": out},
         attrs=dict())
     return out
+
+
+def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None):
+    """
+    Applies a separate affine transformation to each channel of the input.
+    Useful for replacing spatial batch norm with its equivalent fixed
+    transformation. The input also can be 2D tensor and applies a affine
+    transformation in second dimension.
+    
+    Args:
+        x (Variable): Feature map input can be a 4D tensor with order NCHW
+            or NHWC. It also can be a 2D tensor and the affine transformation
+            is applied in the second dimension.
+        scale (Variable): 1D input of shape (C), the c-th element is the scale
+            factor of the affine transformation for the c-th channel of
+            the input.
+        bias (Variable): 1D input of shape (C), the c-th element is the bias
+            of the affine transformation for the c-th channel of the input.
+        data_layout (string, default NCHW): NCHW or NHWC. If input is 2D
+            tensor, you can ignore data_layout.
+        name (str, default None): The name of this layer.
+    Returns:
+        out (Variable): A tensor of the same shape and data layout with x.
+    """
+    helper = LayerHelper("affine_channel", **locals())
+
+    if name is None:
+        out = helper.create_tmp_variable(dtype=x.dtype)
+    else:
+        out = helper.create_variable(
+            name=name, dtype=x.dtype, persistable=False)
+
+    helper.append_op(
+        type="affine_channel",
+        inputs={"X": x,
+                'Scale': scale,
+                'Bias': bias},
+        attrs={"data_layout": data_layout},
+        outputs={"Out": out})
+    return out

python/paddle/fluid/optimizer.py

@@ -15,7 +15,7 @@
 from __future__ import print_function
 import re
 from collections import defaultdict
-from paddle.fluid.framework import Program, Variable, name_scope
+from paddle.fluid.framework import Program, Variable, name_scope, default_main_program
 from . import framework
 from . import layers
 from .backward import append_backward
@@ -111,7 +111,8 @@ class Optimizer(object):
             if param_lr == 1.0:
                 return self._global_learning_rate()
             else:
-                return self._global_learning_rate() * param_lr
+                with default_main_program()._lr_schedule_guard():
+                    return self._global_learning_rate() * param_lr
 
     def _create_accumulators(self, block, parameters):
         """Create all accumulators needed by the parameters

python/paddle/fluid/tests/CMakeLists.txt

-set(PYTHON_TESTS_DIR ${CMAKE_CURRENT_BINARY_DIR} CACHE PATH "python tests directory")
+if(NOT APPLE)
+  set(PYTHON_TESTS_DIR ${CMAKE_CURRENT_BINARY_DIR} CACHE PATH "python tests directory")
+else()
+  set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
+endif(NOT APPLE)
+
 file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
 string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
 

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_affine_channel_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
+import paddle.fluid.core as core
+
+
+def affine_channel(x, scale, bias, layout):
+    C = x.shape[1] if layout == 'NCHW' else x.shape[-1]
+    if len(x.shape) == 4:
+        new_shape = (1, C, 1, 1) if layout == 'NCHW' else (1, 1, 1, C)
+    else:
+        new_shape = (1, C)
+    scale = scale.reshape(new_shape)
+    bias = bias.reshape(new_shape)
+    return x * scale + bias
+
+
+class TestAffineChannelOp(OpTest):
+    def setUp(self):
+        self.op_type = "affine_channel"
+        self.init_test_case()
+
+        x = np.random.random(self.shape).astype("float32")
+        scale = np.random.random(self.C).astype("float32")
+        bias = np.random.random(self.C).astype("float32")
+
+        y = affine_channel(x, scale, bias, self.layout)
+
+        self.inputs = {'X': x, 'Scale': scale, 'Bias': bias}
+        self.attrs = {'data_layout': self.layout}
+        self.outputs = {'Out': y}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X', 'Scale', 'Bias'], 'Out')
+
+    def test_check_grad_stopgrad_dx(self):
+        self.check_grad(['Scale', 'Bias'], 'Out', no_grad_set=set('X'))
+
+    def test_check_grad_stopgrad_dscale_dbias(self):
+        self.check_grad(['X'], 'Out', no_grad_set=set(['Scale', 'Bias']))
+
+    def init_test_case(self):
+        self.shape = [2, 32, 14, 14]
+        self.C = 32
+        self.layout = 'NCHW'
+
+
+class TestAffineChannelNHWC(TestAffineChannelOp):
+    def init_test_case(self):
+        self.shape = [2, 14, 14, 32]
+        self.C = 32
+        self.layout = 'NHWC'
+
+
+class TestAffineChannel2D(TestAffineChannelOp):
+    def init_test_case(self):
+        self.shape = [16, 64]
+        self.C = 64
+        self.layout = 'NCHW'
+
+
+class TestAffineChannelNCHWLargeShape(TestAffineChannelOp):
+    def init_test_case(self):
+        self.shape = [64, 128, 112, 112]
+        self.C = 128
+        self.layout = 'NCHW'
+
+    # since the gradient check is very slow in large shape, so skip check_grad
+    def test_check_grad(self):
+        pass
+
+    def test_check_grad_stopgrad_dx(self):
+        pass
+
+    def test_check_grad_stopgrad_dscale_dbias(self):
+        pass
+
+
+class TestAffineChannelNCHWLargeShape(TestAffineChannelNCHWLargeShape):
+    def init_test_case(self):
+        self.shape = [64, 112, 112, 512]
+        self.C = 512
+        self.layout = 'NHWC'
+
+
+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 +91,61 @@ class TestDistSimnetBow2x2SparseAsync(TestDistBase):
             need_envs=need_envs)
 
 
+# FIXME(tangwei): Learningrate variable is not created on pserver.
+"""
+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=True,
+            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_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_momentum_op.py

@@ -16,6 +16,8 @@ from __future__ import print_function
 
 import unittest
 import numpy as np
+import paddle.fluid.core as core
+from paddle.fluid.op import Operator
 from op_test import OpTest
 
 
@@ -88,5 +90,97 @@ class TestMomentumOp2(OpTest):
         self.check_output()
 
 
+class TestSparseMomentumOp(unittest.TestCase):
+    def setUp(self):
+        self.use_nesterov = False
+
+    def check_with_place(self, place):
+        self.init_kernel()
+        scope = core.Scope()
+        # create and initialize Grad Variable
+        height = 10
+        rows = [0, 4, 7]
+        row_numel = 12
+        mu = 1.0
+        use_nesterov = self.use_nesterov
+
+        # create and initialize Param Variable
+        param = scope.var('Param').get_tensor()
+        param_array = np.full((height, row_numel), 5.0).astype("float32")
+        param.set(param_array, place)
+        param_out = scope.var("ParamOut").get_tensor()
+        param_out_array = np.full((height, row_numel), 0.0).astype("float32")
+        param_out.set(param_out_array, place)
+
+        grad_selected_rows = scope.var('Grad').get_selected_rows()
+        grad_selected_rows.set_height(height)
+        grad_selected_rows.set_rows(rows)
+        grad_np_array = np.ones((len(rows), row_numel)).astype("float32")
+        grad_np_array[0, 0] = 2.0
+        grad_np_array[2, 8] = 4.0
+        grad_tensor = grad_selected_rows.get_tensor()
+        grad_tensor.set(grad_np_array, place)
+
+        velocity = scope.var('Velocity').get_tensor()
+        velocity_np_array = np.ones((height, row_numel)).astype("float32")
+        velocity.set(velocity_np_array, place)
+        velocity_out = scope.var('VelocityOut').get_tensor()
+        velocity_out_np_array = np.full((height, row_numel),
+                                        0.0).astype("float32")
+        velocity_out.set(velocity_out_np_array, place)
+
+        # create and initialize LeraningRate Variable
+        lr = scope.var('LearningRate').get_tensor()
+        lr_array = np.full((1), 2.0).astype("float32")
+        lr.set(lr_array, place)
+
+        # create and run operator
+        op = Operator(
+            "momentum",
+            Param='Param',
+            Grad='Grad',
+            Velocity='Velocity',
+            ParamOut='ParamOut',
+            VelocityOut='VelocityOut',
+            LearningRate='LearningRate',
+            mu=mu,
+            use_nesterov=use_nesterov)
+        op.run(scope, place)
+
+        # get and compare result
+        param_out_np_array = np.array(param_out)
+        velocity_out_np_array = np.array(velocity_out)
+
+        # TODO(dzh): add a more suitable general numpy interface
+        # for sparse update.
+        _grad_np_array = np.full((height, row_numel), 0.0).astype("float32")
+        for i in range(len(rows)):
+            _grad_np_array[rows[i]] = grad_np_array[i]
+        _velocity_out = mu * velocity_np_array + _grad_np_array
+        _param = param_array
+        if use_nesterov:
+            _param_out = _param - (_grad_np_array + _velocity_out * mu
+                                   ) * lr_array
+        else:
+            _param_out = _param - lr_array * _velocity_out
+        self.assertTrue((_velocity_out == velocity_out_np_array).all())
+        self.assertTrue((_param_out == param_out_np_array).all())
+
+    def init_kernel(self):
+        pass
+
+    def test_sparse_momentum(self):
+        places = [core.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(core.CUDAPlace(0))
+        for place in places:
+            self.check_with_place(place)
+
+
+class TestSparseMomentumOp2(TestSparseMomentumOp):
+    def init_kernel(self):
+        self.use_nesterov = True
+
+
 if __name__ == "__main__":
     unittest.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

@@ -1119,6 +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)
@@ -1143,7 +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: [
@@ -1189,7 +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]
 
@@ -1205,19 +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
@@ -1249,8 +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],

python/paddle/utils/__init__.py

@@ -12,4 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__all__ = ['dump_config']
+from plot import Ploter
+__all__ = ['dump_config', 'Ploter']

python/paddle/utils/plot.py

+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+
+
+class PlotData(object):
+    def __init__(self):
+        self.step = []
+        self.value = []
+
+    def append(self, step, value):
+        self.step.append(step)
+        self.value.append(value)
+
+    def reset(self):
+        self.step = []
+        self.value = []
+
+
+class Ploter(object):
+    """
+        Plot input data in a 2D graph
+        
+        Args:
+            title: assign the title of input data.
+            step: x_axis of the data.
+            value: y_axis of the data.
+    """
+
+    def __init__(self, *args):
+        self.__args__ = args
+        self.__plot_data__ = {}
+        for title in args:
+            self.__plot_data__[title] = PlotData()
+        # demo in notebooks will use Ploter to plot figure, but when we convert
+        # the ipydb to py file for testing, the import of matplotlib will make the
+        # script crash. So we can use `export DISABLE_PLOT=True` to disable import
+        # these libs
+        self.__disable_plot__ = os.environ.get("DISABLE_PLOT")
+        if not self.__plot_is_disabled__():
+            import matplotlib.pyplot as plt
+            from IPython import display
+            self.plt = plt
+            self.display = display
+
+    def __plot_is_disabled__(self):
+        return self.__disable_plot__ == "True"
+
+    def append(self, title, step, value):
+        """
+	    Feed data
+	    
+            Args:
+                title: assign the group data to this subtitle.
+                step: the x_axis of data.
+                value: the y_axis of data.
+            
+            Examples:
+                .. code-block:: python
+                plot_curve = Ploter("Curve 1","Curve 2")
+                plot_curve.append(title="Curve 1",step=1,value=1)
+	"""
+        assert isinstance(title, basestring)
+        assert self.__plot_data__.has_key(title)
+        data = self.__plot_data__[title]
+        assert isinstance(data, PlotData)
+        data.append(step, value)
+
+    def plot(self, path=None):
+        """
+            Plot data in a 2D graph
+
+            Args:
+                path: store the figure to this file path. Defaul None. 
+              
+            Examples:
+                .. code-block:: python
+                plot_curve = Ploter()
+                plot_cure.plot()
+	"""
+        if self.__plot_is_disabled__():
+            return
+
+        titles = []
+        for title in self.__args__:
+            data = self.__plot_data__[title]
+            assert isinstance(data, PlotData)
+            if len(data.step) > 0:
+                titles.append(title)
+                self.plt.plot(data.step, data.value)
+        self.plt.legend(titles, loc='upper left')
+        if path is None:
+            self.display.clear_output(wait=True)
+            self.display.display(self.plt.gcf())
+        else:
+            self.plt.savefig(path)
+        self.plt.gcf().clear()
+
+    def reset(self):
+        for key in self.__plot_data__:
+            data = self.__plot_data__[key]
+            assert isinstance(data, PlotData)
+            data.reset()