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

test=develop

CMakeLists.txt

@@ -126,16 +126,12 @@ if(ANDROID OR IOS)
     add_definitions(-DPADDLE_MOBILE_INFERENCE)
 endif()
 
-if (APPLE OR WIN32)
+if (APPLE)
     set(WITH_MKL OFF CACHE STRING
-        "Disable MKL for building on mac and windows" FORCE)
+        "Disable MKL for building on mac" FORCE)
 endif()
 
 if (WIN32)
-    set(WITH_DSO OFF CACHE STRING
-            "Disable DSO when compiling for Windows" FORCE)
-    set(WITH_MKL OFF CACHE STRING
-            "Disable MKL when compiling for Windows" FORCE)
     set(WITH_DISTRIBUTE OFF CACHE STRING
             "Disable DISTRIBUTE when compiling for Windows" FORCE)
     set(WITH_C_API OFF CACHE STRING

cmake/external/mkldnn.cmake

@@ -23,15 +23,14 @@ SET(MKLDNN_SOURCES_DIR    ${THIRD_PARTY_PATH}/mkldnn)
 SET(MKLDNN_INSTALL_DIR    ${THIRD_PARTY_PATH}/install/mkldnn)
 SET(MKLDNN_INC_DIR        "${MKLDNN_INSTALL_DIR}/include" CACHE PATH "mkldnn include directory." FORCE)
 
-IF(WIN32 OR APPLE)
+IF(APPLE)
     MESSAGE(WARNING
-        "Windows or Mac is not supported with MKLDNN in Paddle yet."
+        "Mac is not supported with MKLDNN in Paddle yet."
         "Force WITH_MKLDNN=OFF")
-    SET(WITH_MKLDNN OFF CACHE STRING "Disable MKLDNN in Windows and MacOS" FORCE)
+    SET(WITH_MKLDNN OFF CACHE STRING "Disable MKLDNN in MacOS" FORCE)
     return()
 ENDIF()
 
-SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
 MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/lib to runtime path")
 SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
 SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
@@ -44,10 +43,14 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
 ELSE()
     MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN")
 ENDIF()
-SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-error=array-bounds")
-SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value")
-SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
-SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
+
+IF(NOT WIN32)
+    SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-error=array-bounds")
+    SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value")
+    SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
+    SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
+ENDIF(NOT WIN32)
+
 ExternalProject_Add(
     ${MKLDNN_PROJECT}
     ${EXTERNAL_PROJECT_LOG_ARGS}
@@ -58,8 +61,15 @@ ExternalProject_Add(
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
     CMAKE_ARGS          -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+    CMAKE_ARGS          -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+    CMAKE_ARGS          -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
+    CMAKE_ARGS          -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
+    CMAKE_ARGS          -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+    CMAKE_ARGS          -DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}
+    CMAKE_ARGS          -DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}
     CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
     CMAKE_ARGS          -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
+    CMAKE_ARGS          -DCMAKE_POSITION_INDEPENDENT_CODE=ON
     CMAKE_ARGS          -DMKLROOT=${MKLML_ROOT}
     CMAKE_ARGS          -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
     CMAKE_ARGS          -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
@@ -67,6 +77,11 @@ ExternalProject_Add(
     CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
                         -DMKLROOT:PATH=${MKLML_ROOT}
 )
+if(WIN32)
+    SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/mkldnn.lib" CACHE FILEPATH "mkldnn library." FORCE)
+else(WIN32)
+    SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
+endif(WIN32)
 
 ADD_LIBRARY(shared_mkldnn SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET shared_mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
@@ -85,10 +100,14 @@ ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
 
 # copy the real so.0 lib to install dir
 # it can be directly contained in wheel or capi
-SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
-ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}
-    COMMAND cp ${MKLDNN_LIB} ${MKLDNN_SHARED_LIB}
+if(WIN32)
+    SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/lib/mkldnn.dll)
+else(WIN32)
+    SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
+    ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}
+            COMMAND ${CMAKE_COMMAND} -E copy ${MKLDNN_LIB} ${MKLDNN_SHARED_LIB}
             DEPENDS mkldnn)
+endif(WIN32)
 ADD_CUSTOM_TARGET(mkldnn_shared_lib ALL DEPENDS ${MKLDNN_SHARED_LIB})
 
 IF(WITH_C_API)

cmake/external/mklml.cmake

@@ -16,56 +16,67 @@ IF(NOT ${WITH_MKLML})
   return()
 ENDIF(NOT ${WITH_MKLML})
 
-IF(WIN32 OR APPLE)
+IF(APPLE)
     MESSAGE(WARNING
-        "Windows or Mac is not supported with MKLML in Paddle yet."
+        "Mac is not supported with MKLML in Paddle yet."
         "Force WITH_MKLML=OFF")
     SET(WITH_MKLML OFF CACHE STRING "Disable MKLML package in Windows and MacOS" FORCE)
     return()
 ENDIF()
 
 INCLUDE(ExternalProject)
-
-SET(MKLML_PROJECT       "extern_mklml")
-IF((NOT DEFINED MKLML_VER) OR (NOT DEFINED MKLML_URL))
-  MESSAGE(STATUS "use pre defined download url")
-  SET(MKLML_VER "mklml_lnx_2019.0.20180710" CACHE STRING "" FORCE)
-  SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
-ENDIF()
-MESSAGE(STATUS "MKLML_VER: ${MKLML_VER}, MKLML_URL: ${MKLML_URL}")
-SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
-SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
 SET(MKLML_DST_DIR       "mklml")
 SET(MKLML_INSTALL_ROOT  "${THIRD_PARTY_PATH}/install")
 SET(MKLML_INSTALL_DIR   ${MKLML_INSTALL_ROOT}/${MKLML_DST_DIR})
 SET(MKLML_ROOT          ${MKLML_INSTALL_DIR})
 SET(MKLML_INC_DIR       ${MKLML_ROOT}/include)
 SET(MKLML_LIB_DIR       ${MKLML_ROOT}/lib)
-SET(MKLML_LIB           ${MKLML_LIB_DIR}/libmklml_intel.so)
-SET(MKLML_IOMP_LIB      ${MKLML_LIB_DIR}/libiomp5.so)
+if(WIN32)
+    SET(MKLML_LIB                 ${MKLML_LIB_DIR}/mklml.lib)
+    SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5md.lib)
+    SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/mklml.dll)
+    SET(MKLML_SHARED_IOMP_LIB     ${MKLML_LIB_DIR}/libiomp5md.dll)
+else()
+    SET(MKLML_LIB                 ${MKLML_LIB_DIR}/libmklml_intel.so)
+    SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5.so)
+    SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/libmklml_intel.so)
+    SET(MKLML_SHARED_IOMP_LIB     ${MKLML_LIB_DIR}/libiomp5.so)
+endif()
 SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLML_ROOT}/lib")
 
-INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
+IF((NOT DEFINED MKLML_VER) OR (NOT DEFINED MKLML_URL))
+    MESSAGE(STATUS "use pre defined download url")
+    if(WIN32)
+        SET(MKLML_VER "mklml_win_2019.0.20180710" CACHE STRING "" FORCE)
+        SET(MKLML_URL "https://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
+    else()
+        SET(MKLML_VER "mklml_lnx_2019.0.20180710" CACHE STRING "" FORCE)
+        SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
+    ENDIF()
+endif()
 
-FILE(WRITE ${MKLML_DOWNLOAD_DIR}/CMakeLists.txt
-  "PROJECT(MKLML)\n"
-  "cmake_minimum_required(VERSION 3.0)\n"
-  "install(DIRECTORY ${MKLML_VER}/include ${MKLML_VER}/lib \n"
-  "        DESTINATION ${MKLML_DST_DIR})\n")
+SET(MKLML_PROJECT       "extern_mklml")
+MESSAGE(STATUS "MKLML_VER: ${MKLML_VER}, MKLML_URL: ${MKLML_URL}")
+SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
+SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
 
 ExternalProject_Add(
     ${MKLML_PROJECT}
     ${EXTERNAL_PROJECT_LOG_ARGS}
     PREFIX                 ${MKLML_SOURCE_DIR}
+    URL                    ${MKLML_URL}
     DOWNLOAD_DIR          ${MKLML_DOWNLOAD_DIR}
-    DOWNLOAD_COMMAND      wget --no-check-certificate ${MKLML_URL} -c -q -O ${MKLML_VER}.tgz 
-                          && tar zxf ${MKLML_VER}.tgz
     DOWNLOAD_NO_PROGRESS  1
+    CONFIGURE_COMMAND     ""
+    BUILD_COMMAND         ""
     UPDATE_COMMAND ""
-    CMAKE_ARGS            -DCMAKE_INSTALL_PREFIX=${MKLML_INSTALL_ROOT}
-    CMAKE_CACHE_ARGS      -DCMAKE_INSTALL_PREFIX:PATH=${MKLML_INSTALL_ROOT}
+    INSTALL_COMMAND
+        ${CMAKE_COMMAND} -E copy_directory ${MKLML_DOWNLOAD_DIR}/include ${MKLML_INC_DIR} &&
+        ${CMAKE_COMMAND} -E copy_directory ${MKLML_DOWNLOAD_DIR}/lib ${MKLML_LIB_DIR}
 )
 
+INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
+
 ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB})
 ADD_DEPENDENCIES(mklml ${MKLML_PROJECT})

cmake/generic.cmake

@@ -267,7 +267,11 @@ function(cc_library TARGET_NAME)
           list(APPEND cc_library_DEPS dynload_mklml)
         endif()
         add_dependencies(${TARGET_NAME} mklml)
+        if(WIN32)
+          target_link_libraries(${TARGET_NAME} ${MKLML_IOMP_LIB})
+        else(WIN32)
           target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
+        endif(WIN32)
       endif()
       # remove link to python, see notes at:
       # https://github.com/pybind/pybind11/blob/master/docs/compiling.rst#building-manually

cmake/inference_lib.cmake

@@ -115,20 +115,20 @@ if (NOT PROTOBUF_FOUND OR WIN32)
             )
 endif ()
 
-if (NOT CBLAS_FOUND)
-    set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/openblas")
-    copy(openblas_lib
-            SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
-            DSTS ${dst_dir} ${dst_dir}
-            DEPS extern_openblas
-            )
-elseif (WITH_MKLML)
+if (WITH_MKLML)
     set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mklml")
     copy(mklml_lib
             SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR}
             DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir}
             DEPS mklml
             )
+elseif (NOT CBLAS_FOUND OR WIN32)
+    set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/openblas")
+    copy(openblas_lib
+            SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
+            DSTS ${dst_dir} ${dst_dir}
+            DEPS extern_openblas
+            )
 endif ()
 
 if (WITH_MKLDNN)

paddle/fluid/API.spec

@@ -208,6 +208,7 @@ paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act
 paddle.fluid.layers.merge_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.get_tensor_from_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'hidden_size', 'num_layers', 'dropout_prob', 'is_bidirec', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(0.0, False, False, None, None, -1))
+paddle.fluid.layers.py_func ArgSpec(args=['func', 'x', 'out', 'backward_func', 'skip_vars_in_backward_input'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.psroi_pool ArgSpec(args=['input', 'rois', 'output_channels', 'spatial_scale', 'pooled_height', 'pooled_width', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.huber_loss ArgSpec(args=['input', 'label', 'delta'], varargs=None, keywords=None, defaults=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/framework/details/multi_devices_graph_pass.cc

@@ -355,9 +355,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
           BuildStrategy::GradientScaleStrategy::kCustomized) {
         // TODO(paddle-dev): Why is there no input for this op_handle?
         auto loss_grad_name = node->Op()->OutputArgumentNames()[0];
-        auto out_dtype = all_vars_.at(loss_grad_name)->GetDataType();
-        CreateScaleLossGradOp(&result, loss_grad_name, node->outputs[0],
-                              out_dtype);
+        CreateScaleLossGradOp(&result, loss_grad_name, node->outputs[0]);
       }
       // This assumes the backward generating code will ensure IsScaleLossOp
       // is true only for the op that scale the final scalar loss.
@@ -660,13 +658,13 @@ int MultiDevSSAGraphBuilder::GetVarDeviceID(
 
 void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(
     ir::Graph *result, const std::string &loss_grad_name,
-    ir::Node *out_var_node, proto::VarType::Type dtype) const {
+    ir::Node *out_var_node) const {
   for (size_t i = 0; i < places_.size(); ++i) {
     // Insert ScaleCost OpHandle
     auto *dev_ctx = platform::DeviceContextPool::Instance().Get(places_[i]);
     auto *op_handle = new ScaleLossGradOpHandle(
         result->CreateEmptyNode("scale_loss_grad", ir::Node::Type::kOperation),
-        local_scopes_.size(), local_scopes_[i], places_[i], dev_ctx, dtype);
+        local_scopes_.size(), local_scopes_[i], places_[i], dev_ctx);
     result->Get<GraphOps>(kGraphOps).emplace_back(op_handle);
 
     // FIXME: Currently ScaleLossGradOp only use device_count as scale

paddle/fluid/framework/details/multi_devices_graph_pass.h

@@ -68,8 +68,7 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
 
   void CreateScaleLossGradOp(ir::Graph *result,
                              const std::string &loss_grad_name,
-                             ir::Node *out_var_node,
-                             proto::VarType::Type dtype) const;
+                             ir::Node *out_var_node) const;
 
   VarHandle *CreateReduceOp(ir::Graph *result, const std::string &og,
                             int dst_dev_id) const;

paddle/fluid/framework/details/scale_loss_grad_op_handle.cc

@@ -22,66 +22,39 @@ namespace details {
 ScaleLossGradOpHandle::ScaleLossGradOpHandle(ir::Node *node, size_t num_dev,
                                              Scope *scope,
                                              platform::Place place,
-                                             platform::DeviceContext *dev_ctx,
-                                             proto::VarType::Type dtype)
+                                             platform::DeviceContext *dev_ctx)
     : OpHandleBase(node),
       coeff_(static_cast<float>(1.0 / num_dev)),
       scope_(scope),
-      place_(place),
-      out_dtype_(dtype) {
+      place_(place) {
   this->SetDeviceContext(place_, dev_ctx);
 }
 
 ScaleLossGradOpHandle::~ScaleLossGradOpHandle() {}
 
-struct ScaleLossGradFunctor {
-  float coeff_;
-  Tensor *out_;
-  platform::Place place_;
-  OpHandleBase *op_handle_;
-  proto::VarType::Type out_dtype_;
-  platform::DeviceContext *ctx_;
-
-  ScaleLossGradFunctor(float coeff, Tensor *out, platform::Place place,
-                       OpHandleBase *op_handle, proto::VarType::Type dtype,
-                       platform::DeviceContext *ctx)
-      : coeff_(coeff), out_(out), place_(place), out_dtype_(dtype), ctx_(ctx) {}
-
-  template <typename OutT>
-  void apply() const {
-    auto *out_data = out_->mutable_data<OutT>(place_);
-    if (platform::is_cpu_place(place_)) {
-      *out_data = static_cast<OutT>(coeff_);
-    } else {
-#ifdef PADDLE_WITH_CUDA
-      OutT cast_coeff = static_cast<OutT>(coeff_);
-      auto stream = static_cast<platform::CUDADeviceContext *>(ctx_)->stream();
-      memory::Copy(boost::get<platform::CUDAPlace>(place_), out_data,
-                   platform::CPUPlace(), &cast_coeff, SizeOfType(out_dtype_),
-                   stream);
-      VLOG(10) << place_ << "RUN Scale loss grad op";
-
-#endif
-    }
-  }
-};
-
 void ScaleLossGradOpHandle::RunImpl() {
   // Doesn't wait any event
   std::string var_name = static_cast<VarHandle *>(this->outputs_[0])->name_;
   auto &local_scope = *scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
 
-  auto *tensor = local_scope.FindVar(var_name)->GetMutable<LoDTensor>();
-  tensor->Resize(make_ddim({1}));
+  float *tmp = local_scope.FindVar(var_name)
+                   ->GetMutable<LoDTensor>()
+                   ->mutable_data<float>(make_ddim({1}), place_);
 
+  if (platform::is_cpu_place(place_)) {
+    *tmp = coeff_;
+  } else {
 #ifdef PADDLE_WITH_CUDA
-  ScaleLossGradFunctor func(coeff_, tensor, place_, this, out_dtype_,
-                            this->dev_ctxes_.at(place_));
-  this->RunAndRecordEvent([&] { framework::VisitDataType(out_dtype_, func); });
-#else
-  ScaleLossGradFunctor func(coeff_, tensor, place_, this, out_dtype_, nullptr);
-  framework::VisitDataType(out_dtype_, func);
+    this->RunAndRecordEvent([&] {
+      auto stream = static_cast<platform::CUDADeviceContext *>(
+                        this->dev_ctxes_.at(place_))
+                        ->stream();
+      memory::Copy(boost::get<platform::CUDAPlace>(place_), tmp,
+                   platform::CPUPlace(), &coeff_, sizeof(float), stream);
+      VLOG(10) << place_ << "RUN Scale loss grad op";
+    });
 #endif
+  }
 }
 
 std::string ScaleLossGradOpHandle::Name() const { return "Scale LossGrad"; }

paddle/fluid/framework/details/scale_loss_grad_op_handle.h

@@ -26,8 +26,8 @@ namespace details {
 
 struct ScaleLossGradOpHandle : public OpHandleBase {
   ScaleLossGradOpHandle(ir::Node *node, size_t num_dev, Scope *scope,
-                        platform::Place place, platform::DeviceContext *context,
-                        proto::VarType::Type dtype);
+                        platform::Place place,
+                        platform::DeviceContext *context);
 
   ~ScaleLossGradOpHandle() final;
 
@@ -40,7 +40,6 @@ struct ScaleLossGradOpHandle : public OpHandleBase {
   float coeff_;
   Scope *scope_;
   platform::Place place_;
-  proto::VarType::Type out_dtype_;
 };
 
 }  // namespace details

paddle/fluid/framework/lod_tensor.cc

@@ -157,13 +157,8 @@ bool CheckLoD(const LoD &in, int tensor_height) {
     if (level.size() < 2) return false;
     // check: the first offset(the begin offset) of each level should be 0.
     if (level.front() != 0) return false;
-    // check: all the offsets in a level should be ascending(no same items
-    // allows).
-    if (!std::is_sorted(level.begin(), level.begin(), [](size_t a, size_t b) {
-          if (a < b) return true;
-          return false;
-        })) {
-      LOG(INFO) << "ascending error";
+    // check: all the offsets in a level should be ascending(allow same items)
+    if (!std::is_sorted(level.begin(), level.end())) {
       return false;
     }
   }

paddle/fluid/framework/lod_tensor_test.cc

@@ -217,6 +217,11 @@ TEST(LoD, CheckLoD) {
   // check with underlying tensor storage.
   ASSERT_TRUE(CheckLoD(relative_lod, 5));
   ASSERT_FALSE(CheckLoD(relative_lod, 9));
+
+  // check whether lod is ascending-sorted (allow same items)
+  ASSERT_TRUE(CheckLoD({{0, 1, 2, 3, 4, 5}}, 5));
+  ASSERT_TRUE(CheckLoD({{0, 1, 3, 3, 4, 5}}, 5));
+  ASSERT_FALSE(CheckLoD({{0, 1, 3, 2, 5}}, 5));
 }
 
 TEST(LoD, CheckAbsLoD) {

paddle/fluid/framework/op_desc.cc

@@ -110,22 +110,125 @@ class CompileTimeInferShapeContext : public InferShapeContext {
     }
   }
 
+  std::vector<InferShapeVarPtr> GetInputVarPtrs(
+      const std::string &name) override {
+    const std::vector<std::string> arg_names = Inputs(name);
+    std::vector<InferShapeVarPtr> res;
+    res.reserve(arg_names.size());
+    std::transform(arg_names.begin(), arg_names.end(), std::back_inserter(res),
+                   [this](const std::string &name) {
+                     return block_.FindVarRecursive(name);
+                   });
+    return res;
+  }
+
+  std::vector<InferShapeVarPtr> GetOutputVarPtrs(
+      const std::string &name) override {
+    const std::vector<std::string> arg_names = Outputs(name);
+    std::vector<InferShapeVarPtr> res;
+    res.reserve(arg_names.size());
+    std::transform(arg_names.begin(), arg_names.end(), std::back_inserter(res),
+                   [this](const std::string &name) {
+                     return block_.FindVarRecursive(name);
+                   });
+    return res;
+  }
+
+  DDim GetInputDim(const std::string &name) const override {
+    const std::vector<std::string> &arg_names = Inputs(name);
+    PADDLE_ENFORCE_EQ(arg_names.size(), 1UL,
+                      "Input(%s) should hold one element, but now it holds %d",
+                      name, arg_names.size());
+    return this->GetDim(arg_names[0]);
+  }
+
+  std::vector<DDim> GetInputsDim(const std::string &name) const override {
+    const std::vector<std::string> &arg_names = Inputs(name);
+    return GetDims(arg_names);
+  }
+
   bool IsRuntime() const override;
 
+  std::vector<proto::VarType::Type> GetInputsVarType(
+      const std::string &name) const override {
+    return GetVarTypes(Inputs(name));
+  }
+
+  std::vector<proto::VarType::Type> GetOutputsVarType(
+      const std::string &name) const override {
+    return GetVarTypes(Outputs(name));
+  }
+
+  void SetOutputDim(const std::string &name, const DDim &dim) override {
+    auto &arg_names = Outputs(name);
+    PADDLE_ENFORCE_EQ(arg_names.size(), 1UL,
+                      "Output(%s) should hold one element, but now it holds %d",
+                      name, arg_names.size());
+    SetDim(arg_names[0], dim);
+  }
+
+  void SetOutputsDim(const std::string &name,
+                     const std::vector<DDim> &dims) override {
+    auto &names = Outputs(name);
+    SetDims(names, dims);
+  }
+
  protected:
-  proto::VarType::Type GetVarType(const std::string &name) const override;
+  std::vector<proto::VarType::Type> GetVarTypes(
+      const std::vector<std::string> &names) const {
+    std::vector<proto::VarType::Type> retv;
+    retv.resize(names.size());
+    std::transform(
+        names.begin(), names.end(), retv.begin(),
+        std::bind(std::mem_fn(&CompileTimeInferShapeContext::GetVarType), this,
+                  std::placeholders::_1));
+    return retv;
+  }
 
-  DDim GetDim(const std::string &name) const override;
+  proto::VarType::Type GetVarType(const std::string &name) const;
 
-  void SetDim(const std::string &name, const DDim &dim) override;
+  DDim GetDim(const std::string &name) const {
+    auto var = block_.FindVarRecursive(name);
+    PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s", name);
+    DDim res;
+    try {
+      auto shape = var->GetShape();
+      res = shape.empty() ? make_ddim({0UL}) : make_ddim(shape);
+    } catch (...) {
+      VLOG(5) << "GetDim of variable " << name << " error";
+      std::rethrow_exception(std::current_exception());
+    }
+    return res;
+  }
+
+  std::vector<DDim> GetDims(const std::vector<std::string> &names) const {
+    std::vector<DDim> ret;
+    ret.reserve(names.size());
+    std::transform(
+        names.begin(), names.end(), std::back_inserter(ret),
+        [this](const std::string &name) { return this->GetDim(name); });
+    return ret;
+  }
+
+  void SetDim(const std::string &name, const DDim &dim);
+
+  void SetDims(const std::vector<std::string> &names,
+               const std::vector<DDim> &dims) {
+    size_t length = names.size();
+    PADDLE_ENFORCE_EQ(length, dims.size());
+    for (size_t i = 0; i < length; ++i) {
+      if (names[i] == framework::kEmptyVarName) {
+        continue;
+      }
+      SetDim(names[i], dims[i]);
+    }
+  }
 
   std::vector<DDim> GetRepeatedDims(const std::string &name) const override;
 
   void SetRepeatedDims(const std::string &name,
                        const std::vector<DDim> &dims) override;
 
-  InferShapeVarPtr GetVarPtr(const std::string &name) override;
-
   const OpDesc &op_;
   const BlockDesc &block_;
 };
@@ -644,20 +747,6 @@ const std::vector<std::string> &CompileTimeInferShapeContext::Outputs(
   return op_.Output(name);
 }
 
-DDim CompileTimeInferShapeContext::GetDim(const std::string &name) const {
-  auto var = block_.FindVarRecursive(name);
-  PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s", name);
-  DDim res;
-  try {
-    auto shape = var->GetShape();
-    res = shape.empty() ? make_ddim({0UL}) : make_ddim(shape);
-  } catch (...) {
-    VLOG(5) << "GetDim of variable " << name << " error";
-    std::rethrow_exception(std::current_exception());
-  }
-  return res;
-}
-
 std::vector<DDim> CompileTimeInferShapeContext::GetRepeatedDims(
     const std::string &name) const {
   auto var = block_.FindVarRecursive(name);
@@ -696,10 +785,5 @@ proto::VarType::Type CompileTimeInferShapeContext::GetVarType(
   return block_.FindVarRecursive(name)->GetType();
 }
 
-InferShapeVarPtr CompileTimeInferShapeContext::GetVarPtr(
-    const std::string &name) {
-  return block_.FindVarRecursive(name);
-}
-
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/op_desc.h

@@ -123,6 +123,8 @@ class OpDesc {
 
   BlockDesc *Block() { return this->block_; }
 
+  const BlockDesc *Block() const { return this->block_; }
+
  private:
   template <typename MapType>
   static std::vector<typename MapType::key_type> MapKeys(const MapType &map) {

paddle/fluid/framework/operator.cc

@@ -142,12 +142,14 @@ RuntimeContext::RuntimeContext(const VariableNameMap& innames,
                                const Scope& scope) {
   for (auto& var_name_item : innames) {
     std::vector<Variable*>& input_vars = inputs[var_name_item.first];
+    input_vars.reserve(var_name_item.second.size());
     for (auto& var_name : var_name_item.second) {
       input_vars.push_back(scope.FindVar(var_name));
     }
   }
   for (auto& var_name_item : outnames) {
     std::vector<Variable*>& output_vars = outputs[var_name_item.first];
+    output_vars.reserve(var_name_item.second.size());
     for (auto& var_name : var_name_item.second) {
       output_vars.push_back(scope.FindVar(var_name));
     }
@@ -474,6 +476,28 @@ const Tensor* ExecutionContext::LegacyInput<Tensor>(
 template <>
 const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
     const std::string& name) const {
+  auto it = ctx_.inputs.find(name);
+  if (it == ctx_.inputs.end()) {
+    return {};
+  }
+  const std::vector<Variable*>& vars = it->second;
+  std::vector<const Tensor*> res;
+  res.reserve(vars.size());
+  std::transform(vars.begin(), vars.end(), std::back_inserter(res),
+                 [&](Variable* var) -> const Tensor* {
+                   if (var == nullptr) return nullptr;
+                   PADDLE_ENFORCE(
+                       var->IsType<LoDTensor>(),
+                       "should be LoDTensor, but the received type is %s",
+                       var->Type().name());
+                   return &(var->Get<LoDTensor>());
+                 });
+  return res;
+}
+
+template <>
+const std::vector<const Tensor*> ExecutionContext::LegacyMultiInput<Tensor>(
+    const std::string& name) const {
   auto names = op().Inputs(name);
   std::vector<const Tensor*> res;
   res.reserve(names.size());
@@ -556,30 +580,28 @@ class RuntimeInferShapeContext : public InferShapeContext {
 
   bool HasOutput(const std::string& name) const override {
     // has only one output
-    const auto& outs = op_.Outputs();
+    const auto& outs = ctx_.outputs;
     auto it = outs.find(name);
     if (it == outs.end()) {
       return false;
     }
     const auto& out = it->second;
-    if (out.size() == 0 || out[0] == kEmptyVarName) {
+    if (out.size() == 0) {
       return false;
     }
     PADDLE_ENFORCE_EQ(out.size(), 1UL,
                       "Output %s should not have more than one outputs", name);
-    return scope_.FindVar(out[0]) != nullptr;
+    return out[0] != nullptr;
   }
 
   bool HasInputs(const std::string& name) const override {
-    if (!op_.HasInputs(name)) {
-      return false;
-    }
-    auto inputs = op_.Inputs(name);
-    if (inputs.empty()) {
+    const auto& ins = ctx_.inputs;
+    auto it = ins.find(name);
+    if (it == ins.end() || it->second.empty()) {
       return false;
     }
-    for (auto& input : inputs) {
-      if (scope_.FindVar(input) == nullptr) {
+    for (auto& input : it->second) {
+      if (input == nullptr) {
         return false;
       }
     }
@@ -587,15 +609,13 @@ class RuntimeInferShapeContext : public InferShapeContext {
   }
 
   bool HasOutputs(const std::string& name) const override {
-    if (!op_.HasOutputs(name)) {
-      return false;
-    }
-    auto outputs = op_.Outputs(name);
-    if (outputs.empty()) {
+    const auto& outs = ctx_.outputs;
+    auto it = outs.find(name);
+    if (it == outs.end() || it->second.empty()) {
       return false;
     }
-    for (auto& output : outputs) {
-      if (scope_.FindVar(output) == nullptr) {
+    for (auto& output : it->second) {
+      if (output == nullptr) {
         return false;
       }
     }
@@ -616,16 +636,18 @@ class RuntimeInferShapeContext : public InferShapeContext {
 
   void ShareDim(const std::string& in, const std::string& out, size_t i = 0,
                 size_t j = 0) override {
-    PADDLE_ENFORCE_LT(i, Inputs(in).size());
-    PADDLE_ENFORCE_LT(j, Outputs(out).size());
-    const std::string& input_n = Inputs(in)[i];
-    const std::string& output_n = Outputs(out)[j];
+    auto in_it = ctx_.inputs.find(in);
+    auto out_it = ctx_.outputs.find(out);
+    PADDLE_ENFORCE(in_it != ctx_.inputs.end() && in_it->second.size() > i,
+                   "Inputs %s should have %llu argument", in, i);
+    PADDLE_ENFORCE(out_it != ctx_.outputs.end() && out_it->second.size() > j,
+                   "Outputs %s should have %llu argument", out, j);
+
+    Variable* in_var = in_it->second[i];
+    Variable* out_var = out_it->second[j];
 
-    Variable* in_var = scope_.FindVar(input_n);
-    Variable* out_var = scope_.FindVar(output_n);
     PADDLE_ENFORCE(in_var->Type() == out_var->Type(),
-                   "The type of %s and %s is not the same.", output_n,
-                   GetDim(input_n));
+                   "The type of %s and %s is not the same.", in, out);
 
     if (in_var->IsType<framework::SelectedRows>()) {
       auto& in_sele_rows = in_var->Get<framework::SelectedRows>();
@@ -646,13 +668,16 @@ class RuntimeInferShapeContext : public InferShapeContext {
 
   void ShareLoD(const std::string& in, const std::string& out, size_t i = 0,
                 size_t j = 0) const override {
-    const std::vector<std::string>& inputs = Inputs(in);
-    const std::vector<std::string>& outputs = Outputs(out);
-    PADDLE_ENFORCE_LT(i, inputs.size());
-    PADDLE_ENFORCE_LT(j, outputs.size());
-    Variable* in_var = scope_.FindVar(inputs.at(i));
+    auto in_it = ctx_.inputs.find(in);
+    auto out_it = ctx_.outputs.find(out);
+    PADDLE_ENFORCE(in_it != ctx_.inputs.end() && in_it->second.size() > i,
+                   "Inputs %s should have %llu argument", in, i);
+    PADDLE_ENFORCE(out_it != ctx_.outputs.end() && out_it->second.size() > j,
+                   "Outputs %s should have %llu argument", out, j);
+
+    Variable* in_var = in_it->second.at(i);
     if (!in_var->IsType<LoDTensor>()) return;
-    Variable* out_var = scope_.FindVar(outputs.at(j));
+    Variable* out_var = out_it->second.at(j);
     PADDLE_ENFORCE(out_var->IsType<LoDTensor>(),
                    "The %d-th output of Output(%s) must be LoDTensor.", j, out);
     auto in_tensor = in_var->Get<LoDTensor>();
@@ -687,9 +712,64 @@ class RuntimeInferShapeContext : public InferShapeContext {
 
   bool IsRuntime() const override { return true; }
 
+  // TODO(paddle-dev): Can this be template?
+  std::vector<InferShapeVarPtr> GetInputVarPtrs(
+      const std::string& name) override {
+    const std::vector<Variable*>& vars = InputVars(name);
+    std::vector<InferShapeVarPtr> res;
+    res.reserve(vars.size());
+    res.insert(res.begin(), vars.begin(), vars.end());
+    return res;
+  }
+
+  std::vector<InferShapeVarPtr> GetOutputVarPtrs(
+      const std::string& name) override {
+    const std::vector<Variable*>& vars = OutputVars(name);
+    std::vector<InferShapeVarPtr> res;
+    res.reserve(vars.size());
+    res.insert(res.begin(), vars.begin(), vars.end());
+    return res;
+  }
+
+  DDim GetInputDim(const std::string& name) const override {
+    const std::vector<Variable*>& vars = InputVars(name);
+    PADDLE_ENFORCE_EQ(vars.size(), 1UL,
+                      "Input(%s) should hold one element, but now it holds %d",
+                      name, vars.size());
+    return this->GetDim(vars[0]);
+  }
+
+  std::vector<DDim> GetInputsDim(const std::string& name) const override {
+    const std::vector<Variable*>& vars = InputVars(name);
+    return GetDims(vars);
+  }
+
+  std::vector<proto::VarType::Type> GetInputsVarType(
+      const std::string& name) const override {
+    return GetVarTypes(InputVars(name));
+  }
+
+  std::vector<proto::VarType::Type> GetOutputsVarType(
+      const std::string& name) const override {
+    return GetVarTypes(OutputVars(name));
+  }
+
+  void SetOutputDim(const std::string& name, const DDim& dim) override {
+    auto& vars = OutputVars(name);
+    PADDLE_ENFORCE_EQ(vars.size(), 1UL,
+                      "Output(%s) should hold one element, but now it holds %d",
+                      name, vars.size());
+    SetDim(vars[0], dim);
+  }
+
+  void SetOutputsDim(const std::string& name,
+                     const std::vector<DDim>& dims) override {
+    auto& vars = OutputVars(name);
+    SetDims(vars, dims);
+  }
+
  protected:
-  DDim GetDim(const std::string& name) const override {
-    Variable* var = scope_.FindVar(name);
+  DDim GetDim(Variable* var) const {
     PADDLE_ENFORCE_NOT_NULL(var);
     if (var->IsType<LoDTensor>()) {
       return var->Get<LoDTensor>().dims();
@@ -697,25 +777,44 @@ class RuntimeInferShapeContext : public InferShapeContext {
       return var->Get<SelectedRows>().GetCompleteDims();
     } else {
       PADDLE_THROW(
-          "Only LoDTensor/SelectedRows support 'GetDim', but Variable %s's "
+          "Only LoDTensor/SelectedRows support 'GetDim', but Variables "
           "type_id is %s.",
-          name, var->Type().name());
+          var->Type().name());
+    }
   }
+
+  std::vector<DDim> GetDims(const std::vector<Variable*>& vars) const {
+    std::vector<DDim> ret;
+    ret.reserve(vars.size());
+    std::transform(vars.begin(), vars.end(), std::back_inserter(ret),
+                   [this](Variable* var) { return this->GetDim(var); });
+    return ret;
   }
 
   std::vector<DDim> GetRepeatedDims(const std::string& name) const override {
     PADDLE_THROW("Only compile time support this method");
   }
 
-  void SetDim(const std::string& name, const DDim& dim) override {
-    Variable* var = scope_.FindVar(name);
+  void SetDim(Variable* var, const DDim& dim) {
     if (var->IsType<LoDTensor>()) {
       var->GetMutable<LoDTensor>()->Resize(dim);
     } else if (var->IsType<SelectedRows>()) {
       var->GetMutable<SelectedRows>()->set_height(dim[0]);
     } else {
-      PADDLE_THROW("Variable %s type_id %s, expect LoDTensor/SelectedRows.",
-                   name, var->Type().name());
+      PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
+                   var->Type().name());
+    }
+  }
+
+  void SetDims(const std::vector<Variable*>& vars,
+               const std::vector<DDim>& dims) {
+    size_t length = vars.size();
+    PADDLE_ENFORCE_EQ(length, dims.size());
+    for (size_t i = 0; i < length; ++i) {
+      if (vars[i] == nullptr) {
+        continue;
+      }
+      SetDim(vars[i], dims[i]);
     }
   }
 
@@ -724,16 +823,36 @@ class RuntimeInferShapeContext : public InferShapeContext {
     PADDLE_THROW("Only compile time support this method");
   }
 
-  proto::VarType::Type GetVarType(const std::string& name) const override {
-    auto* var = scope_.FindVar(name);
-    return ToVarType(var->Type());
+  std::vector<proto::VarType::Type> GetVarTypes(
+      const std::vector<Variable*>& vars) const {
+    std::vector<proto::VarType::Type> retv;
+    retv.resize(vars.size());
+    std::transform(vars.begin(), vars.end(), retv.begin(),
+                   std::bind(std::mem_fn(&RuntimeInferShapeContext::GetVarType),
+                             this, std::placeholders::_1));
+    return retv;
   }
 
-  InferShapeVarPtr GetVarPtr(const std::string& name) override {
-    return scope_.FindVar(name);
+  proto::VarType::Type GetVarType(Variable* var) const {
+    return ToVarType(var->Type());
   }
 
  private:
+  const std::vector<Variable*>& InputVars(const std::string& name) const {
+    auto it = ctx_.inputs.find(name);
+    PADDLE_ENFORCE(it != ctx_.inputs.end(),
+                   "Operator %s does not have the input %s.", op_.Type(), name);
+    return it->second;
+  }
+
+  const std::vector<Variable*>& OutputVars(const std::string& name) const {
+    auto it = ctx_.outputs.find(name);
+    PADDLE_ENFORCE(it != ctx_.outputs.end(),
+                   "Operator %s does not have the outputs %s.", op_.Type(),
+                   name);
+    return it->second;
+  }
+
   const OperatorBase& op_;
   const Scope& scope_;
   const RuntimeContext& ctx_;
@@ -864,8 +983,7 @@ Scope* OperatorWithKernel::PrepareData(
 
     for (size_t i = 0; i < var_name_item.second.size(); ++i) {
       auto& var_name = var_name_item.second[i];
-      auto* var = scope.FindVar(var_name);
-      input_vars[i] = var;
+      auto* var = input_vars[i];
 
       // Only tensor can be tranfer to another device.
       if (var == nullptr || !VarIsTensor(*var)) {

paddle/fluid/framework/operator.h

@@ -197,8 +197,31 @@ class ExecutionContext {
 
   const std::vector<const Variable*> MultiInputVar(
       const std::string& name) const {
-    auto names = op_.Inputs(name);
+    auto it = ctx_.inputs.find(name);
+    if (it == ctx_.inputs.end()) {
+      return {};
+    }
     std::vector<const Variable*> res;
+    res.reserve(it->second.size());
+    std::transform(it->second.begin(), it->second.end(),
+                   std::back_inserter(res),
+                   [this](Variable* var) { return var; });
+    return res;
+  }
+
+  std::vector<Variable*> MultiOutputVar(const std::string& name) const {
+    auto names = op_.Outputs(name);
+    auto it = ctx_.outputs.find(name);
+    if (it == ctx_.outputs.end()) {
+      return {};
+    }
+    return it->second;
+  }
+
+  const std::vector<Variable*> LegacyMultiInputVar(
+      const std::string& name) const {
+    auto names = op_.Inputs(name);
+    std::vector<Variable*> res;
     res.reserve(names.size());
     std::transform(names.begin(), names.end(), std::back_inserter(res),
                    [this](const std::string& name) {
@@ -208,7 +231,7 @@ class ExecutionContext {
     return res;
   }
 
-  std::vector<Variable*> MultiOutputVar(const std::string& name) const {
+  std::vector<Variable*> LegacyMultiOutputVar(const std::string& name) const {
     auto names = op_.Outputs(name);
     std::vector<Variable*> res;
     res.reserve(names.size());
@@ -250,6 +273,38 @@ class ExecutionContext {
 
   template <typename T>
   const std::vector<const T*> MultiInput(const std::string& name) const {
+    auto it = ctx_.inputs.find(name);
+    if (it == ctx_.inputs.end()) {
+      return {};
+    }
+    const std::vector<Variable*>& vars = it->second;
+    std::vector<const T*> res;
+    res.reserve(vars.size());
+    std::transform(vars.begin(), vars.end(), std::back_inserter(res),
+                   [&](Variable* var) -> const T* {
+                     return var == nullptr ? nullptr : &var->Get<T>();
+                   });
+    return res;
+  }
+
+  template <typename T>
+  std::vector<T*> MultiOutput(const std::string& name) const {
+    auto it = ctx_.outputs.find(name);
+    if (it == ctx_.outputs.end()) {
+      return {};
+    }
+    const std::vector<Variable*>& vars = it->second;
+    std::vector<T*> res;
+    res.reserve(vars.size());
+    std::transform(vars.begin(), vars.end(), std::back_inserter(res),
+                   [&](Variable* var) -> T* {
+                     return var == nullptr ? nullptr : var->GetMutable<T>();
+                   });
+    return res;
+  }
+
+  template <typename T>
+  const std::vector<const T*> LegacyMultiInput(const std::string& name) const {
     auto names = op_.Inputs(name);
     std::vector<const T*> res;
     res.reserve(names.size());
@@ -262,7 +317,7 @@ class ExecutionContext {
   }
 
   template <typename T>
-  std::vector<T*> MultiOutput(const std::string& name) const {
+  std::vector<T*> LegacyMultiOutput(const std::string& name) const {
     auto names = op_.Outputs(name);
     std::vector<T*> res;
     res.reserve(names.size());
@@ -321,6 +376,10 @@ template <>
 const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
     const std::string& name) const;
 
+template <>
+const std::vector<const Tensor*> ExecutionContext::LegacyMultiInput<Tensor>(
+    const std::string& name) const;
+
 template <>
 Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;
 

paddle/fluid/framework/shape_inference.cc

@@ -22,20 +22,6 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-DDim InferShapeContext::GetInputDim(const std::string &name) const {
-  const std::vector<std::string> &arg_names = Inputs(name);
-  PADDLE_ENFORCE_EQ(arg_names.size(), 1UL,
-                    "Input(%s) should hold one element, but now it holds %d",
-                    name, arg_names.size());
-  return this->GetDim(arg_names[0]);
-}
-
-std::vector<DDim> InferShapeContext::GetInputsDim(
-    const std::string &name) const {
-  const std::vector<std::string> &arg_names = Inputs(name);
-  return GetDims(arg_names);
-}
-
 std::vector<DDim> InferShapeContext::GetReaderDims(
     const std::string &name) const {
   const std::vector<std::string> &arg_names = Inputs(name);
@@ -46,26 +32,6 @@ std::vector<DDim> InferShapeContext::GetReaderDims(
   return this->GetRepeatedDims(arg_names[0]);
 }
 
-DDim InferShapeContext::GetInputsElementDim(const std::string &name,
-                                            int idx) const {
-  const std::vector<std::string> &names = Inputs(name);
-  return this->GetDim(names[idx]);
-}
-
-void InferShapeContext::SetOutputDim(const std::string &name, const DDim &dim) {
-  auto &arg_names = Outputs(name);
-  PADDLE_ENFORCE_EQ(arg_names.size(), 1UL,
-                    "Output(%s) should hold one element, but now it holds %d",
-                    name, arg_names.size());
-  SetDim(arg_names[0], dim);
-}
-
-void InferShapeContext::SetOutputsDim(const std::string &name,
-                                      const std::vector<DDim> &dims) {
-  auto &names = Outputs(name);
-  SetDims(names, dims);
-}
-
 void InferShapeContext::SetReaderDims(const std::string &name,
                                       const std::vector<DDim> &dims) {
   const std::vector<std::string> &arg_names = Outputs(name);
@@ -76,69 +42,5 @@ void InferShapeContext::SetReaderDims(const std::string &name,
   return this->SetRepeatedDims(arg_names[0], dims);
 }
 
-std::vector<InferShapeVarPtr> InferShapeContext::GetInputVarPtrs(
-    const std::string &name) {
-  const std::vector<std::string> arg_names = Inputs(name);
-  std::vector<InferShapeVarPtr> res;
-  res.reserve(arg_names.size());
-  std::transform(
-      arg_names.begin(), arg_names.end(), std::back_inserter(res),
-      [this](const std::string &name) { return this->GetVarPtr(name); });
-  return res;
-}
-
-std::vector<InferShapeVarPtr> InferShapeContext::GetOutputVarPtrs(
-    const std::string &name) {
-  const std::vector<std::string> arg_names = Outputs(name);
-  std::vector<InferShapeVarPtr> res;
-  res.reserve(arg_names.size());
-  std::transform(
-      arg_names.begin(), arg_names.end(), std::back_inserter(res),
-      [this](const std::string &name) { return this->GetVarPtr(name); });
-  return res;
-}
-
-std::vector<DDim> InferShapeContext::GetDims(
-    const std::vector<std::string> &names) const {
-  std::vector<DDim> ret;
-  ret.reserve(names.size());
-  std::transform(
-      names.begin(), names.end(), std::back_inserter(ret),
-      [this](const std::string &name) { return this->GetDim(name); });
-  return ret;
-}
-
-void InferShapeContext::SetDims(const std::vector<std::string> &names,
-                                const std::vector<DDim> &dims) {
-  size_t length = names.size();
-  PADDLE_ENFORCE_EQ(length, dims.size());
-  for (size_t i = 0; i < length; ++i) {
-    if (names[i] == framework::kEmptyVarName) {
-      continue;
-    }
-    SetDim(names[i], dims[i]);
-  }
-}
-
-std::vector<proto::VarType::Type> InferShapeContext::GetInputsVarType(
-    const std::string &name) const {
-  return GetVarTypes(Inputs(name));
-}
-
-std::vector<proto::VarType::Type> InferShapeContext::GetOutputsVarType(
-    const std::string &name) const {
-  return GetVarTypes(Outputs(name));
-}
-
-std::vector<proto::VarType::Type> InferShapeContext::GetVarTypes(
-    const std::vector<std::string> &names) const {
-  std::vector<proto::VarType::Type> retv;
-  retv.resize(names.size());
-  std::transform(names.begin(), names.end(), retv.begin(),
-                 std::bind(std::mem_fn(&InferShapeContext::GetVarType), this,
-                           std::placeholders::_1));
-  return retv;
-}
-
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/shape_inference.h

@@ -25,6 +25,8 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+class OperatorBase;
+
 using InferShapeVarPtr = boost::variant<VarDesc *, Variable *>;
 
 class InferShapeContext {
@@ -33,22 +35,23 @@ class InferShapeContext {
   virtual bool HasInput(const std::string &name) const = 0;
   virtual bool HasOutput(const std::string &name) const = 0;
 
-  std::vector<proto::VarType::Type> GetInputsVarType(
-      const std::string &name) const;
-  std::vector<proto::VarType::Type> GetOutputsVarType(
-      const std::string &name) const;
+  virtual std::vector<proto::VarType::Type> GetInputsVarType(
+      const std::string &name) const = 0;
+  virtual std::vector<proto::VarType::Type> GetOutputsVarType(
+      const std::string &name) const = 0;
 
   virtual bool HasInputs(const std::string &name) const = 0;
   virtual bool HasOutputs(const std::string &name) const = 0;
 
-  DDim GetInputDim(const std::string &name) const;
-  std::vector<DDim> GetInputsDim(const std::string &name) const;
-  std::vector<DDim> GetReaderDims(const std::string &name) const;
-  DDim GetInputsElementDim(const std::string &name, int idx) const;
+  virtual DDim GetInputDim(const std::string &name) const = 0;
+  virtual std::vector<DDim> GetInputsDim(const std::string &name) const = 0;
+  virtual std::vector<DDim> GetReaderDims(const std::string &name) const;
 
-  void SetOutputDim(const std::string &name, const DDim &dim);
-  void SetOutputsDim(const std::string &name, const std::vector<DDim> &dims);
-  void SetReaderDims(const std::string &name, const std::vector<DDim> &dims);
+  virtual void SetOutputDim(const std::string &name, const DDim &dim) = 0;
+  virtual void SetOutputsDim(const std::string &name,
+                             const std::vector<DDim> &dims) = 0;
+  virtual void SetReaderDims(const std::string &name,
+                             const std::vector<DDim> &dims);
 
   virtual AttrReader Attrs() const = 0;
   virtual const std::vector<std::string> &Inputs(
@@ -67,27 +70,15 @@ class InferShapeContext {
 
   virtual bool IsRuntime() const = 0;
 
-  std::vector<InferShapeVarPtr> GetInputVarPtrs(const std::string &name);
-  std::vector<InferShapeVarPtr> GetOutputVarPtrs(const std::string &name);
-  virtual InferShapeVarPtr GetVarPtr(const std::string &name) = 0;
-
-  // Note: In while op, we need this to be public
-  void SetDims(const std::vector<std::string> &names,
-               const std::vector<DDim> &dims);
+  virtual std::vector<InferShapeVarPtr> GetInputVarPtrs(
+      const std::string &name) = 0;
+  virtual std::vector<InferShapeVarPtr> GetOutputVarPtrs(
+      const std::string &name) = 0;
 
  protected:
-  virtual DDim GetDim(const std::string &name) const = 0;
-  virtual void SetDim(const std::string &name, const DDim &dim) = 0;
   virtual std::vector<DDim> GetRepeatedDims(const std::string &name) const = 0;
   virtual void SetRepeatedDims(const std::string &name,
                                const std::vector<DDim> &dims) = 0;
-
-  std::vector<DDim> GetDims(const std::vector<std::string> &names) const;
-
-  std::vector<proto::VarType::Type> GetVarTypes(
-      const std::vector<std::string> &names) const;
-
-  virtual proto::VarType::Type GetVarType(const std::string &name) const = 0;
 };
 
 }  // namespace framework

paddle/fluid/framework/tensor.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/framework/var_type.h"
 
 namespace paddle {
 namespace framework {
@@ -27,6 +28,9 @@ void Tensor::check_memory_size() const {
       "or maybe the required data-type mismatches the data already stored.");
 }
 
+Tensor::Tensor(std::type_index type)
+    : type_(framework::ToDataType(type)), offset_(0) {}
+
 size_t Tensor::memory_size() const {
   return holder_ == nullptr ? 0UL : holder_->size() - offset_;
 }
@@ -101,5 +105,12 @@ const DDim& Tensor::dims() const { return dims_; }
 
 int64_t Tensor::numel() const { return product(dims_); }
 
+void Tensor::ResetHolder(std::shared_ptr<memory::Allocation> holder) {
+  if (holder_) {
+    PADDLE_ENFORCE_EQ(numel() * SizeOfType(type()), holder->size());
+  }
+  holder_ = holder;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/tensor.h

@@ -69,6 +69,8 @@ class Tensor {
  public:
   Tensor() : type_(proto::VarType::FP32), offset_(0) {}
 
+  explicit Tensor(std::type_index type);
+
   /*! Return a pointer to mutable memory block. */
   template <typename T>
   T* data();
@@ -162,6 +164,8 @@ class Tensor {
     return std::move(holder_);
   }
 
+  void ResetHolder(std::shared_ptr<memory::Allocation> holder);
+
  private:
   /*! holds the memory block if allocated. */
   std::shared_ptr<memory::Allocation> holder_;

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

@@ -89,12 +89,21 @@ endif()
 
 if(WITH_MKL)
   include_directories("${PADDLE_LIB}/third_party/install/mklml/include")
+  if(NOT WIN32)
     set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
                  ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX})
+  else(WIN32)
+    set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml${CMAKE_SHARED_LIBRARY_SUFFIX}
+            ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5md${CMAKE_SHARED_LIBRARY_SUFFIX})
+  endif(WIN32)
   set(MKLDNN_PATH "${PADDLE_LIB}/third_party/install/mkldnn")
   if(EXISTS ${MKLDNN_PATH})
     include_directories("${MKLDNN_PATH}/include")
+    if(WIN32)
+      set(MKLDNN_LIB ${MKLDNN_PATH}/lib/mkldnn.lib)
+    else(WIN32)
       set(MKLDNN_LIB ${MKLDNN_PATH}/lib/libmkldnn.so.0)
+    endif(WIN32)
   endif()
 else()
   set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas${CMAKE_STATIC_LIBRARY_SUFFIX})

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

@@ -254,5 +254,16 @@ TEST(Analyzer_dam, compare) { compare(); }
 TEST(Analyzer_dam, compare_mkldnn) { compare(true /* use_mkldnn */); }
 #endif
 
+// Compare Deterministic result
+TEST(Analyzer_dam, compare_determine) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                       input_slots_all);
+}
+
 }  // namespace inference
 }  // namespace paddle

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

@@ -180,6 +180,17 @@ TEST(Analyzer_LAC, compare) {
       reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
+// Compare Deterministic result
+TEST(Analyzer_LAC, compare_determine) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                       input_slots_all);
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

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

@@ -179,5 +179,16 @@ TEST(Analyzer_Chinese_ner, compare) {
       reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
+// Compare Deterministic result
+TEST(Analyzer_Chinese_ner, compare_determine) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                       input_slots_all);
+}
+
 }  // namespace inference
 }  // namespace paddle

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

@@ -85,6 +85,17 @@ TEST(Analyzer_resnet50, compare) { compare(); }
 TEST(Analyzer_resnet50, compare_mkldnn) { compare(true /* use_mkldnn */); }
 #endif
 
+// Compare Deterministic result
+TEST(Analyzer_resnet50, compare_determine) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                       input_slots_all);
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

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

@@ -265,6 +265,17 @@ TEST(Analyzer_rnn1, compare) {
       reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
+// Compare Deterministic result
+TEST(Analyzer_rnn1, compare_determine) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                       input_slots_all);
+}
+
 // Test Multi-Thread.
 TEST(Analyzer_rnn1, multi_thread) {
   contrib::AnalysisConfig cfg;

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

@@ -158,5 +158,16 @@ TEST(Analyzer_rnn2, compare) {
       reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
+// Compare Deterministic result
+TEST(Analyzer_rnn2, compare_determine) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                       input_slots_all);
+}
+
 }  // namespace inference
 }  // namespace paddle

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

@@ -204,5 +204,16 @@ TEST(Analyzer_seq_conv1, compare) {
       reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
+// Compare Deterministic result
+TEST(Analyzer_seq_conv1, compare_determine) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                       input_slots_all);
+}
+
 }  // namespace inference
 }  // namespace paddle

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

@@ -106,6 +106,17 @@ TEST(Analyzer_Text_Classification, compare) {
       reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
+// Compare Deterministic result
+TEST(Analyzer_Text_Classification, compare_determine) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                       input_slots_all);
+}
+
 TEST(Analyzer_Text_Classification, compare_against_embedding_fc_lstm_fused) {
   AnalysisConfig cfg;
   SetConfig(&cfg);

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

@@ -145,6 +145,17 @@ TEST(Analyzer_vis, compare) { compare(); }
 TEST(Analyzer_vis, compare_mkldnn) { compare(true /* use_mkldnn */); }
 #endif
 
+// Compare Deterministic result
+TEST(Analyzer_vis, compare_determine) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                       input_slots_all);
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

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

@@ -45,6 +45,7 @@ DEFINE_bool(use_analysis, true,
             "Running the inference program in analysis mode.");
 DEFINE_bool(record_benchmark, false,
             "Record benchmark after profiling the model");
+DEFINE_double(accuracy, 1e-3, "Result Accuracy.");
 
 DECLARE_bool(profile);
 DECLARE_int32(paddle_num_threads);
@@ -85,7 +86,7 @@ void CompareResult(const std::vector<PaddleTensor> &outputs,
         float *pdata = static_cast<float *>(out.data.data());
         float *pdata_ref = static_cast<float *>(ref_out.data.data());
         for (size_t j = 0; j < size; ++j) {
-          EXPECT_NEAR(pdata_ref[j], pdata[j], 1e-3);
+          EXPECT_NEAR(pdata_ref[j], pdata[j], FLAGS_accuracy);
         }
         break;
       }
@@ -283,6 +284,26 @@ void TestPrediction(const PaddlePredictor::Config *config,
   }
 }
 
+void CompareDeterministic(
+    const PaddlePredictor::Config *config,
+    const std::vector<std::vector<PaddleTensor>> &inputs) {
+  int batch_size = FLAGS_batch_size;
+  int num_times = FLAGS_repeat;
+  auto predictor = CreateTestPredictor(config, FLAGS_use_analysis);
+
+  // warmup run
+  std::vector<PaddleTensor> warmup_outputs, outputs;
+  predictor->Run(inputs[0], &warmup_outputs, batch_size);
+
+  // run num_times to Compare Deterministic Result.
+  for (int i = 0; i < num_times; i++) {
+    for (size_t j = 0; j < inputs.size(); j++) {
+      predictor->Run(inputs[j], &outputs, batch_size);
+      CompareResult(outputs, warmup_outputs);
+    }
+  }
+}
+
 void CompareNativeAndAnalysis(
     const PaddlePredictor::Config *config,
     const std::vector<std::vector<PaddleTensor>> &inputs) {

paddle/fluid/operators/CMakeLists.txt

@@ -16,6 +16,7 @@ add_subdirectory(metrics)
 add_subdirectory(optimizers)
 add_subdirectory(reduce_ops)
 add_subdirectory(sequence_ops)
+add_subdirectory(jit)
 
 if(WITH_DISTRIBUTE)
     add_subdirectory(distributed)
@@ -42,8 +43,7 @@ if (WITH_DISTRIBUTE)
     SET(OP_PREFETCH_DEPS ${OP_PREFETCH_DEPS} parameter_prefetch)
 endif()
 
-register_operators(EXCLUDES warpctc_op conv_fusion_op DEPS ${OP_HEADER_DEPS} ${OP_PREFETCH_DEPS})
-
+register_operators(EXCLUDES py_func_op warpctc_op conv_fusion_op DEPS ${OP_HEADER_DEPS} ${OP_PREFETCH_DEPS})
 
 # warpctc_op needs cudnn 7 above
 if (WITH_GPU AND NOT WIN32)
@@ -65,7 +65,7 @@ set(COMMON_OP_DEPS ${OP_HEADER_DEPS})
 
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} selected_rows_functor selected_rows lod_tensor maxouting unpooling pooling lod_rank_table context_project sequence_pooling executor)
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dynload_warpctc)
-set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel concat_and_split cross_entropy softmax vol2col im2col sampler)
+set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel_helper concat_and_split cross_entropy softmax vol2col im2col sampler)
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence2batch lstm_compute matrix_bit_code gru_compute activation_functions)
 if (WITH_GPU)
   set(COMMON_OP_DEPS ${COMMON_OP_DEPS} depthwise_conv prelu)
@@ -92,4 +92,8 @@ cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)
 cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op)
 nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor)
 
+if (WITH_PYTHON)
+  cc_library(py_func_op SRCS py_func_op.cc DEPS op_registry python pybind)
+endif()
+
 set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")

paddle/fluid/operators/controlflow/while_op.cc

@@ -399,26 +399,41 @@ class WhileGradOpShapeInference : public framework::InferShapeBase {
     ctx->HasInputs(kOutputs);
     ctx->HasInputs(framework::GradVarName(kOutputs));
 
-    auto p_names = ctx->Inputs(kX);
     auto pg_ig_names = ctx->Outputs(kXGRAD);
-    auto var_types = ctx->GetInputsVarType(kX);
-    std::vector<std::string> names_to_set;
-    std::vector<framework::DDim> dims_to_set;
-    for (size_t i = 0; i < p_names.size(); ++i) {
+    std::vector<framework::InferShapeVarPtr> in_var_ptrs =
+        ctx->GetInputVarPtrs(kX);
+    std::vector<framework::InferShapeVarPtr> out_var_ptrs =
+        ctx->GetOutputVarPtrs(kXGRAD);
+    PADDLE_ENFORCE(in_var_ptrs.size() == out_var_ptrs.size());
+
+    for (size_t i = 0; i < in_var_ptrs.size(); ++i) {
       if (pg_ig_names[i] == framework::kEmptyVarName) {
         continue;
       }
-      auto dims = ctx->GetInputsElementDim(kX, i);
-      if (var_types[i] == framework::proto::VarType::LOD_TENSOR) {
-        names_to_set.push_back(pg_ig_names[i]);
-        dims_to_set.push_back(dims);
-      } else if (var_types[i] == framework::proto::VarType::LOD_TENSOR_ARRAY) {
-        // not sure how to set the dim of LOD_TENSOR_ARRAY
-        names_to_set.push_back(pg_ig_names[i]);
-        dims_to_set.push_back(dims);
+      if (ctx->IsRuntime()) {
+        framework::Variable *in_var =
+            boost::get<framework::Variable *>(in_var_ptrs[i]);
+        framework::Variable *out_var =
+            boost::get<framework::Variable *>(out_var_ptrs[i]);
+
+        auto type = framework::ToVarType(in_var->Type());
+        if (type == framework::proto::VarType::LOD_TENSOR) {
+          out_var->GetMutable<LoDTensor>()->Resize(
+              in_var->Get<framework::LoDTensor>().dims());
+        } else if (type == framework::proto::VarType::SELECTED_ROWS) {
+          out_var->GetMutable<framework::SelectedRows>()->set_height(
+              in_var->Get<framework::SelectedRows>().GetCompleteDims()[0]);
+        } else if (type == framework::proto::VarType::LOD_TENSOR_ARRAY) {
+          PADDLE_THROW("WhileGradOp doesn't support type %d",
+                       static_cast<int>(type));
+        }
+      } else {
+        framework::VarDesc *in_var =
+            boost::get<framework::VarDesc *>(in_var_ptrs[i]);
+        boost::get<framework::VarDesc *>(out_var_ptrs[i])
+            ->SetShape(in_var->GetShape());
       }
     }
-    ctx->SetDims(names_to_set, dims_to_set);
   }
 };
 

paddle/fluid/operators/conv_mkldnn_op.cc

@@ -155,11 +155,14 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     auto chosen_memory_format =
         platform::data_format_to_memory_format(data_format);
 
+    weights_format = mkldnn::memory::format::any;
+    // Check the format for user's special output
+    if (chosen_memory_format != mkldnn::memory::format::any) {
       if (is_conv3d) {
         chosen_memory_format =
             platform::MKLDNNFormatForSize(src_tz.size(), chosen_memory_format);
       }
-    weights_format = GetWeightsFormat(chosen_memory_format, g, is_conv3d);
+    }
 
     auto src_md = platform::MKLDNNMemDesc(
         src_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
@@ -435,11 +438,14 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
     auto chosen_memory_format =
         platform::data_format_to_memory_format(data_format);
 
+    weights_format = mkldnn::memory::format::any;
+    // Check the format for user's special output
+    if (chosen_memory_format != mkldnn::memory::format::any) {
       if (is_conv3d) {
         chosen_memory_format =
             platform::MKLDNNFormatForSize(src_tz.size(), chosen_memory_format);
       }
-    weights_format = GetWeightsFormat(chosen_memory_format, g, is_conv3d);
+    }
 
     auto src_md = platform::MKLDNNMemDesc(
         src_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);

paddle/fluid/operators/conv_op.h

@@ -22,6 +22,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/math/depthwise_conv.h"
 #include "paddle/fluid/operators/math/im2col.h"
 #include "paddle/fluid/operators/math/vol2col.h"
+#include "paddle/fluid/platform/create_tensor_with_allocationptr.h"
 
 namespace paddle {
 namespace operators {
@@ -123,6 +124,8 @@ class GemmConvKernel : public framework::OpKernel<T> {
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
     std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
 
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+
     const int batch_size = static_cast<int>(input->dims()[0]);
 
     // filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w}
@@ -155,13 +158,19 @@ class GemmConvKernel : public framework::OpKernel<T> {
     // to call the matrix multiplication interface.
     Tensor col_matrix;
     if (is_expand) {
-      col.mutable_data<T>(col_shape, context.GetPlace());
+      auto tmp_allocation_ptr =
+          platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx).Allocate(
+              framework::product(col_shape) * sizeof(T));
+      Tensor tep_tensor =
+          platform::GetTensor<T>(std::move(tmp_allocation_ptr), col_shape);
+
+      col.ShareDataWith(tep_tensor);
       col_matrix.ShareDataWith(col);
       col_matrix.Resize(col_matrix_shape);
     }
 
-    framework::DDim input_shape = framework::slice_ddim(
-        input->dims(), 1, static_cast<int>(input->dims().size()));
+    framework::DDim input_shape =
+        framework::slice_ddim(input->dims(), 1, input->dims().size());
 
     framework::DDim filter_matrix_shape = {filter.dims()[0],
                                            filter.numel() / filter.dims()[0]};
@@ -178,7 +187,6 @@ class GemmConvKernel : public framework::OpKernel<T> {
     math::Vol2ColFunctor<DeviceContext, T> vol2col;
     math::Im2ColFunctor<math::ColFormat::kCFO, DeviceContext, T> im2col;
 
-    auto& dev_ctx = context.template device_context<DeviceContext>();
     auto blas = math::GetBlas<DeviceContext, T>(dev_ctx);
     for (int i = 0; i < batch_size; i++) {
       Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
@@ -237,6 +245,8 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
 
     const int batch_size = static_cast<int>(input->dims()[0]);
 
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+
     // filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w}
     std::vector<int64_t> filter_shape_vec(framework::vectorize(filter.dims()));
     // output_shape_vec: {o_n, o_c, o_h, o_w} or {o_n, o_c, o_d, o_h, o_w}
@@ -262,8 +272,8 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
     framework::DDim col_matrix_shape =
         framework::flatten_to_2d(col_shape, data_dim + 1);
 
-    framework::DDim input_shape = framework::slice_ddim(
-        input->dims(), 1, static_cast<int>(input->dims().size()));
+    framework::DDim input_shape =
+        framework::slice_ddim(input->dims(), 1, input->dims().size());
 
     framework::DDim filter_matrix_shape = {filter.dims()[0],
                                            filter.numel() / filter.dims()[0]};
@@ -286,13 +296,18 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
     // to call the matrix multiplication interface.
     Tensor col_matrix;
     if (is_expand) {
-      col.mutable_data<T>(col_shape, context.GetPlace());
+      auto tmp_allocation_ptr =
+          platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx).Allocate(
+              framework::product(col_shape) * sizeof(T));
+      Tensor tep_tensor =
+          platform::GetTensor<T>(std::move(tmp_allocation_ptr), col_shape);
+
+      col.ShareDataWith(tep_tensor);
       col_matrix.ShareDataWith(col);
       col_matrix.Resize(col_matrix_shape);
     }
 
     math::SetConstant<DeviceContext, T> set_zero;
-    auto& dev_ctx = context.template device_context<DeviceContext>();
     auto blas = math::GetBlas<DeviceContext, T>(dev_ctx);
 
     if (input_grad) {

paddle/fluid/operators/crf_decoding_op.h

@@ -16,7 +16,7 @@ limitations under the License. */
 #include <limits>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/operators/math/jit_kernel.h"
+#include "paddle/fluid/operators/jit/kernels.h"
 #include "paddle/fluid/operators/math/math_function.h"
 
 namespace paddle {
@@ -82,10 +82,9 @@ class CRFDecodingOpKernel : public framework::OpKernel<T> {
     Tensor track;
     int* track_value =
         track.mutable_data<int>(emission_dims, platform::CPUPlace());
-    const auto& ker = math::jitkernel::KernelPool::Instance()
-                          .template Get<math::jitkernel::CRFDecodeKernel<T>>(
-                              static_cast<int>(tag_num));
-    ker->Compute(static_cast<int>(seq_len), x, w, alpha_value, track_value);
+    auto ker = jit::Get<jit::kCRFDecoding, jit::CRFDecodingTuples<T>,
+                        platform::CPUPlace>(tag_num);
+    ker(static_cast<int>(seq_len), x, w, alpha_value, track_value, tag_num);
     T max_score = -std::numeric_limits<T>::max();
     int max_i = 0;
     for (size_t i = 0; i < tag_num; ++i) {

paddle/fluid/operators/elementwise/elementwise_div_op.cu

@@ -12,23 +12,18 @@ 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/elementwise/elementwise_div_op.h"
-#include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;
 
 REGISTER_OP_CUDA_KERNEL(
     elementwise_div,
     ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext,
-                              paddle::platform::float16>,
     ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext, double>,
     ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext, int>,
     ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext, int64_t>);
 REGISTER_OP_CUDA_KERNEL(
     elementwise_div_grad,
     ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext,
-                                  paddle::platform::float16>,
     ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext, double>,
     ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext, int>,
     ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext,

paddle/fluid/operators/elementwise/elementwise_mul_mkldnn_op.cc

@@ -16,11 +16,14 @@ limitations under the License. */
 #include "paddle/fluid/operators/elementwise/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 
+#include "paddle/fluid/operators/jit/kernels.h"
+#include "paddle/fluid/platform/cpu_info.h"
 #include "paddle/fluid/platform/mkldnn_helper.h"
 
-#include "paddle/fluid/operators/math/jit_kernel.h"
+#ifdef PADDLE_WITH_XBYAK
 #include "xbyak/xbyak.h"
 #include "xbyak/xbyak_util.h"
+#endif
 
 namespace paddle {
 namespace operators {
@@ -81,8 +84,7 @@ class ElementwiseMulMKLDNNKernel : public framework::OpKernel<T> {
     UpdateDataFormat(ctx, const_cast<Tensor*>(x), "x_data_format");
     UpdateDataFormat(ctx, const_cast<Tensor*>(y), "y_data_format");
 
-    Xbyak::util::Cpu cpu;
-    const bool is_avx512_enabled = cpu.has(Xbyak::util::Cpu::tAVX512F);
+    const bool is_avx512_enabled = platform::MayIUse(platform::avx512f);
     const bool are_dims_divisable = !(x_int_dims[1] % 16);
     const bool is_x_format_correct = x->format() == memory::format::nChw16c;
     const bool is_y_format_correct = y->format() == memory::format::nc;
@@ -108,10 +110,8 @@ class ElementwiseMulMKLDNNKernel : public framework::OpKernel<T> {
         constexpr int simd_width = 16;
         int C = c / simd_width;
 
-        const auto& multiply =
-            math::jitkernel::KernelPool::Instance()
-                .template Get<math::jitkernel::EltwiseMulnChw16cNCKernel<T>>(n);
-
+        auto multiply = jit::Get<jit::kNCHW16CMulNC, jit::NCHW16CMulNCTuples<T>,
+                                 platform::CPUPlace>(0);
 #pragma omp parallel for collapse(2)
         for (int ni = 0; ni < n; ni++) {
           for (int ci = 0; ci < C; ci++) {
@@ -122,7 +122,7 @@ class ElementwiseMulMKLDNNKernel : public framework::OpKernel<T> {
             auto ptr_z =
                 z_data + ni * C * h * w * simd_width + ci * h * w * simd_width;
 
-            multiply->Compute(ptr_x, ptr_y, ptr_z, h, w);
+            multiply(ptr_x, ptr_y, ptr_z, h, w);
           }
         }
       }

paddle/fluid/operators/elementwise/elementwise_mul_op.cu

@@ -12,21 +12,19 @@ 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/elementwise/elementwise_mul_op.h"
-#include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;
-namespace plat = paddle::platform;
 
 REGISTER_OP_CUDA_KERNEL(
-    elementwise_mul, ops::ElementwiseMulKernel<plat::CUDADeviceContext, float>,
-    ops::ElementwiseMulKernel<plat::CUDADeviceContext, double>,
-    ops::ElementwiseMulKernel<plat::CUDADeviceContext, int>,
-    ops::ElementwiseMulKernel<plat::CUDADeviceContext, int64_t>,
-    ops::ElementwiseMulKernel<plat::CUDADeviceContext, plat::float16>);
+    elementwise_mul,
+    ops::ElementwiseMulKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseMulKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseMulKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseMulKernel<paddle::platform::CUDADeviceContext, int64_t>);
 REGISTER_OP_CUDA_KERNEL(
     elementwise_mul_grad,
-    ops::ElementwiseMulGradKernel<plat::CUDADeviceContext, float>,
-    ops::ElementwiseMulGradKernel<plat::CUDADeviceContext, double>,
-    ops::ElementwiseMulGradKernel<plat::CUDADeviceContext, int>,
-    ops::ElementwiseMulGradKernel<plat::CUDADeviceContext, int64_t>,
-    ops::ElementwiseMulGradKernel<plat::CUDADeviceContext, plat::float16>);
+    ops::ElementwiseMulGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CUDADeviceContext,
+                                  int64_t>);

paddle/fluid/operators/fill_zeros_like_op.cu.cc

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/fill_zeros_like_op.h"
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;
 REGISTER_OP_CUDA_KERNEL(
@@ -23,6 +22,4 @@ REGISTER_OP_CUDA_KERNEL(
     ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext, int64_t>,
     ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext, float>,
     ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext, double>,
-    ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext,
-                             paddle::platform::float16>,
     ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext, bool>);

paddle/fluid/operators/fused/fusion_gru_op.cc

@@ -15,9 +15,9 @@ limitations under the License. */
 #include "paddle/fluid/operators/fused/fusion_gru_op.h"
 #include <cstring>  // for memcpy
 #include <string>
+#include "paddle/fluid/operators/jit/kernels.h"
 #include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/fc_compute.h"
-#include "paddle/fluid/operators/math/jit_kernel.h"
 #include "paddle/fluid/operators/math/sequence2batch.h"
 
 namespace paddle {
@@ -191,14 +191,16 @@ class FusionGRUKernel : public framework::OpKernel<T> {
   const int M = x_dims[1];                                                     \
   const int D = wh_dims[0];                                                    \
   const int D2 = D * 2;                                                        \
-  const math::jitkernel::gru_attr_t attr(                          \
-      D, ctx.Attr<std::string>("gate_activation"),                 \
-      ctx.Attr<std::string>("activation"));                        \
-  math::jitkernel::gru_t one_step;                                 \
-  const auto& ker =                                                \
-      math::jitkernel::KernelPool::Instance()                      \
-          .template Get<math::jitkernel::GRUKernel<T>,             \
-                        const math::jitkernel::gru_attr_t&>(attr); \
+  const jit::gru_attr_t attr(                                                  \
+      D, jit::to_kerneltype(ctx.Attr<std::string>("gate_activation")),         \
+      jit::to_kerneltype(ctx.Attr<std::string>("activation")));                \
+  jit::gru_t one_step;                                                         \
+  auto ComputeH1 =                                                             \
+      jit::Get<jit::kGRUH1, jit::GRUTuples<T>, platform::CPUPlace>(attr);      \
+  auto ComputeHtPart1 =                                                        \
+      jit::Get<jit::kGRUHtPart1, jit::GRUTuples<T>, platform::CPUPlace>(attr); \
+  auto ComputeHtPart2 =                                                        \
+      jit::Get<jit::kGRUHtPart2, jit::GRUTuples<T>, platform::CPUPlace>(attr); \
   const T* x_data = x->data<T>();                                              \
   const T* wx_data = wx->data<T>();                                            \
   const T* wh_data = wh->data<T>();                                            \
@@ -241,7 +243,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
       } else {
         one_step.gates = xx_data;
         one_step.ht = hidden_out_data;
-        ker->ComputeH1(&one_step, &attr);
+        ComputeH1(&one_step, &attr);
         prev_hidden_data = hidden_out_data;
         tstart = 1;
         move_step();
@@ -254,12 +256,12 @@ class FusionGRUKernel : public framework::OpKernel<T> {
         one_step.gates = xx_data;
         one_step.ht_1 = prev_hidden_data;
         one_step.ht = hidden_out_data;
-        ker->ComputeHtPart1(&one_step, &attr);
+        ComputeHtPart1(&one_step, &attr);
         // gemm rt * Ws
         blas.GEMM(CblasNoTrans, CblasNoTrans, 1, D, D, static_cast<T>(1),
                   hidden_out_data, D, wh_state_data, D, static_cast<T>(1),
                   xx_data + D2, D3);
-        ker->ComputeHtPart2(&one_step, &attr);
+        ComputeHtPart2(&one_step, &attr);
         // save prev
         prev_hidden_data = hidden_out_data;
         move_step();
@@ -323,7 +325,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
       for (int i = 0; i < max_bs; ++i) {
         one_step.gates = cur_in_data;
         one_step.ht = cur_out_data;
-        ker->ComputeH1(&one_step, &attr);
+        ComputeH1(&one_step, &attr);
         // add offset
         cur_in_data += D3;
         cur_out_data += D;
@@ -351,7 +353,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
         one_step.gates = cur_batched_data;
         one_step.ht_1 = cur_prev_hidden_data;
         one_step.ht = cur_out_data;
-        ker->ComputeHtPart1(&one_step, &attr);
+        ComputeHtPart1(&one_step, &attr);
 
         cur_batched_data += D3;
         cur_prev_hidden_data += D;
@@ -369,7 +371,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
         one_step.gates = cur_batched_data;
         one_step.ht_1 = cur_prev_hidden_data;
         one_step.ht = cur_out_data;
-        ker->ComputeHtPart2(&one_step, &attr);
+        ComputeHtPart2(&one_step, &attr);
         cur_batched_data += D3;
         cur_prev_hidden_data += D;
         cur_out_data += D;

paddle/fluid/operators/fused/fusion_lstm_op.cc

@@ -14,9 +14,9 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/fused/fusion_lstm_op.h"
 #include <string>
+#include "paddle/fluid/operators/jit/kernels.h"
 #include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/fc_compute.h"
-#include "paddle/fluid/operators/math/jit_kernel.h"
 #include "paddle/fluid/operators/math/sequence2batch.h"
 
 namespace paddle {
@@ -249,17 +249,18 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
     auto* checked_cell = ctx.Output<Tensor>("CheckedCell");                   \
     checked_cell_data = checked_cell->mutable_data<T>(place);                 \
   }                                                                           \
-  const math::jitkernel::lstm_attr_t attr(                      \
-      D, ctx.Attr<std::string>("gate_activation"),              \
-      ctx.Attr<std::string>("candidate_activation"),            \
-      ctx.Attr<std::string>("cell_activation"), use_peepholes); \
-  math::jitkernel::lstm_t one_step;                             \
+  const jit::lstm_attr_t attr(                                                \
+      D, jit::to_kerneltype(ctx.Attr<std::string>("gate_activation")),        \
+      jit::to_kerneltype(ctx.Attr<std::string>("candidate_activation")),      \
+      jit::to_kerneltype(ctx.Attr<std::string>("cell_activation")),           \
+      use_peepholes);                                                         \
+  jit::lstm_t one_step;                                                       \
   one_step.wp = wp_data;                                                      \
   one_step.checked = checked_cell_data;                                       \
-  const auto& ker =                                             \
-      math::jitkernel::KernelPool::Instance()                   \
-          .template Get<math::jitkernel::LSTMKernel<T>,         \
-                        const math::jitkernel::lstm_attr_t&>(attr)
+  auto ComputeC1H1 =                                                          \
+      jit::Get<jit::kLSTMC1H1, jit::LSTMTuples<T>, platform::CPUPlace>(attr); \
+  auto ComputeCtHt =                                                          \
+      jit::Get<jit::kLSTMCtHt, jit::LSTMTuples<T>, platform::CPUPlace>(attr)
 
 // Wh GEMM
 #define GEMM_WH_ADDON(bs, prev, out)                                           \
@@ -305,7 +306,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
         one_step.gates = xx_data;
         one_step.ct = c_out_data;
         one_step.ht = h_out_data;
-        ker->ComputeC1H1(&one_step, &attr);
+        ComputeC1H1(&one_step, &attr);
         tstart = 1;
         // move one step
         prev_h_data = h_out_data;
@@ -321,7 +322,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
         one_step.ct_1 = prev_c_data;
         one_step.ct = c_out_data;
         one_step.ht = h_out_data;
-        ker->ComputeCtHt(&one_step, &attr);
+        ComputeCtHt(&one_step, &attr);
         // move one step
         prev_h_data = h_out_data;
         prev_c_data = c_out_data;
@@ -401,7 +402,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
         one_step.gates = cur_in_data;
         one_step.ct = cur_c_out_data;
         one_step.ht = cur_h_out_data;
-        ker->ComputeC1H1(&one_step, &attr);
+        ComputeC1H1(&one_step, &attr);
 
         cur_in_data += D4;
         cur_c_out_data += D;
@@ -431,7 +432,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
         one_step.ct_1 = cur_prev_c_data;
         one_step.ct = cur_c_out_data;
         one_step.ht = cur_h_out_data;
-        ker->ComputeCtHt(&one_step, &attr);
+        ComputeCtHt(&one_step, &attr);
 
         // move one batch
         cur_in_data += D4;

paddle/fluid/operators/jit/CMakeLists.txt

+
+set(jit_file ${PADDLE_BINARY_DIR}/paddle/fluid/operators/jit/kernels.h)
+file(WRITE ${jit_file} "// Generated by the paddle/fluid/operators/jit/CMakeLists.txt.  DO NOT EDIT!\n\n")
+file(APPEND ${jit_file} "\#pragma once\n")
+file(APPEND ${jit_file} "\#include \"paddle/fluid/operators/jit/helper.h\"\n")
+file(APPEND ${jit_file} "\#include \"paddle/fluid/operators/jit/registry.h\"\n\n")
+
+set(JIT_KERNEL_DEPS cpu_info cblas gflags enforce place)
+
+file(GLOB jit_kernel_cc_srcs RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.cc")
+list(REMOVE_ITEM jit_kernel_cc_srcs test.cc benchmark.cc)
+cc_library(jit_kernel_base SRCS ${jit_kernel_cc_srcs} DEPS ${JIT_KERNEL_DEPS})
+
+# refer must go first
+add_subdirectory(refer)
+add_subdirectory(more)
+if(WITH_XBYAK)
+    add_subdirectory(gen)
+endif()
+
+cc_library(jit_kernel_helper SRCS ${jit_kernel_cc_srcs} DEPS ${JIT_KERNEL_DEPS})
+cc_test(jit_kernel_test SRCS test.cc DEPS jit_kernel_helper)
+if(NOT WIN32)
+    cc_binary(jit_kernel_benchmark SRCS benchmark.cc DEPS jit_kernel_helper device_tracer)
+endif()

paddle/fluid/operators/jit/README.en.md

+# JIT Kernel
+
+JIT(Just In Time) Kernel contains actually generated code and some other implemenations with the same logic.
+Each implementations has its own condition to use, defined in `UseMe`.
+They are combined together to get the best performance of one single independent function.
+They could be some very simple functions like vector multiply, or some complicated functions like LSTM.
+And they can be composed with some other exited jit kernels to build up a complex function. 
+Currently it's only supported on CPU yet.
+
+## Contents
+
+```txt
+PaddlePaddle/Paddle/paddle/fluid/
+├── ...
+└── operators/
+    ├── .../
+    └── jit/
+        ├── ...
+        ├── gen/
+        │   └── ...
+        |── more/
+        │   ├── ...
+        │   ├── mkl/
+        │   │   └── ...
+        │   ├── mkldnn/
+        │   │   └── ...
+        │   ├── mix/
+        │   │   └── ...
+        │   ├── intrinsic/
+        │   │   └── ...
+        │   └── openblas/
+        │       └── ...
+        └── refer/
+            └── ...
+```
+
+All basical definations of jit kernels are addressed in `paddle/fluid/operators/jit` including these three key folders `refer`, `gen`, `more`. There is only one unique name for each kernel while may have seraval implementations with same functionality.
+
+- `refer`: Each kernel must have one reference implementation on CPU, and it should only focus on the correctness and should not depends on any third-party libraries.
+- `gen`: The code generated should be kept here. They should be designed focusing on the best performance, which depends on Xbyak.
+- `more`: All other implementations should be kept in this folder with one directory corresponding to one library kind or method kind, such as mkl, mkldnn, openblas or intrinsic code. Each implementation should have it advantage. 
+
+## How to use
+
+One simple function `jit::Get`, which is very easy to use, is supported to get the kernel.
+It can automatically return the expected function with best performance under the given attributes. 
+All kernels are inlcuded in `paddle/fluid/operators/jit/kernels.h`, you can only include this one header to get all the registered kernels.
+
+## Solid Test
+
+- Unit Test
+    All functions should be compared with the corresponding reference functions, including data tyep `float` and `double`.
+- Benchmark
+    All functions should be tested, and make sure the `jit::Get` function obtain the best performance with all attributes.
+
+# How to add new kernel
+
+## Required
+
+1. Add `your_key` at `KernelType`.
+2. Add reference function of `your_key`. 
+Note:
+    - this should be run on CPU and do not depend on any third-party.
+    - Add `USE_JITKERNEL_REFER(your_key)` in `refer/CmakeLists.txt` to make sure this code can be used.
+3. Add unit test in `test.cc`, and verfiy at least `float` and `double`.
+Test more data type for some special functions if necessary, for example `int8`.
+4. Add functions in `benchmark.cc` to test all function of same `KernelType`. Make sure `jit::Get` always get the best one.
+
+## Optional
+
+Add more implementations of `your_kery` for performance enhancement.
+
+1. Add functions based on generated code in `gen`. It should be derived from `JitCode` and should have corepsonding creator from `JitCodeCreator` which will be registered on the `your_key`.
+Note: Add new `KernelTuples` if necessary，your can refer to `XYZNTuples`.
+Specialie method `JitCodeKey` when add new attribute type。
+2. Add more functions in `more`，you can use any third party you wish, like mkl, mkldnn or intrinsic code to reach the best performance.

paddle/fluid/operators/jit/README.md

+# JIT Kernel
+
+结合函数模板和JIT生成需要的kernel函数。
+这里的kernel是比Operator中kernel更小级别的算子单元，更侧重的是在不同硬件上的性能。可以有多重第三方库的实现，每种实现有自己的`UseMe`函数负责什么条件下可以被调用。
+这里实现的函数可以非常细粒度的函数方法，比如Vector MUL， 也可以是一个复杂的逻辑比如LSTM等。复杂的逻辑也可以由自己的底层函数拼接而成。
+目前仅支持CPU上的高性能计算。
+
+## 目录结构
+
+```txt
+PaddlePaddle/Paddle/paddle/fluid/
+├── ...
+└── operators/
+    ├── .../
+    └── jit/
+        ├── ...
+        ├── gen/
+        │   └── ...
+        |── more/
+        │   ├── ...
+        │   ├── mkl/
+        │   │   └── ...
+        │   ├── mkldnn/
+        │   │   └── ...
+        │   ├── mix/
+        │   │   └── ...
+        │   ├── intrinsic/
+        │   │   └── ...
+        │   └── openblas/
+        │       └── ...
+        └── refer/
+            └── ...
+```
+
+基本类的定义都放在根目录下，根目录下包括gen,more和refer三个目录。每个目录下都是一种或者多种实现，每种kernel算子都需要有reference的实现，用作单元测试的基准，其他的实现都是可选的。
+- gen: 代表使用jit生成的code，需要依赖xbyak库。该实现最关心的就是性能。
+- refer: 代表reference的实现，每种kernel算子都需要有在CPU上的reference的实现，他主要关心的算法逻辑的正确性。
+- more: 下面可以放入跟多实现，可以包括mkl，mkldnn，intrinsic，openblas等，也可以是自身已有的kernel组合。
+
+## 动态获取
+
+提供一个`jit::Get`方法，根据kernel类别获取，每种实现都有自己的使用范围，根据范围动态和当前条件选择需要的kernel函数。
+
+## 测试
+
+- 逻辑测试
+    所有实现都要与refer的code对比，需要满足精度要求， 包括float和double的数据类型
+- 性能测试
+    所有实现的性能对比，并且与最终的`jit::Get`方法对比，该方法拿到的性能需要在各种条件下都是最好的。
+
+# 如何添加新的算子
+
+- 在`KernelType` 中添加 `your_key` .
+- 实现Reference 的逻辑，这个是必须是在CPU上的实现，并且不能依赖任何第三方库。实现后在`refer/CmakeLists.txt`中添加`USE_JITKERNEL_REFER(your_key)`来使用该kernel.
+- (optional) 实现更多的算法在`more`目录下，可以依赖mkl，intrinsic或者mkldnn等第三方库。
+- (optional) 实现基于Xbyak的生成code，在`gen`目下。 jitcode需要实现自己的`JitCodeCreator`，并注册在与refer相同的`KernelType`上。
+- 必要时可以添加新的`KernelTuples`，可以参考`XYZNTuples`，新加的Attr类型需要特例化`JitCodeKey`方法。
+- 在`test.cc`中添加unit test，至少需要测试`float`和`double`两种数据类型，如有必要需要支持额外的数据类型，比如`int8`的相关函数。
+- 在`benchmark.cc`中添加相应的性能对比，同一种kernel需要对比所有实现，并且确保`jit::Get`得到的实现一直是速度最快的。
+
+# 优点
+- 统一的Get方法，接口简单。
+- 同一套逻辑可以有多套实现，可以依赖多套第三方库，互不影响。
+- 目录结构清晰，不会在某个文件中有多个宏定义，导致的可读性差问题。
+- 优化方便，可以直接针对某种属性针对性优化，并不影响其他属性下的性能。
+- 可以支持多种平台，包括Linux，Mac 和 Windows，至少可以保证每种平台都可以正常work。后期也可以针对不同平台有针对的优化。框架层面可以使用统一接口，不必关心底层实现。

paddle/fluid/operators/jit/benchmark.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 <iostream>
+#include <random>
+#include <string>
+#include <vector>
+#include "gflags/gflags.h"
+#include "glog/logging.h"
+#include "paddle/fluid/operators/jit/kernels.h"
+#include "paddle/fluid/platform/device_tracer.h"
+#include "paddle/fluid/platform/place.h"
+#include "paddle/fluid/platform/port.h"
+
+DEFINE_int32(burning, 10, "Burning times.");
+DEFINE_int32(repeat, 3000, "Repeat times.");
+DEFINE_int32(max_size, 1000, "The Max size would be tested.");
+
+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), 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);
+  }
+}
+
+std::vector<int> TestSizes() {
+  std::vector<int> s;
+  for (int i = 1; i <= FLAGS_max_size; ++i) {
+    s.push_back(i);
+  }
+  return s;
+}
+
+template <typename KernelTuples, typename... Args>
+struct BenchFunc {
+  // return this function avg time
+  double operator()(const typename KernelTuples::func_type tgt, Args... args) {
+    for (int i = 0; i < FLAGS_burning; ++i) {
+      tgt(args...);
+    }
+    auto start = paddle::platform::PosixInNsec() / 1e-3;
+    for (int i = 0; i < FLAGS_repeat; ++i) {
+      tgt(args...);
+    }
+    auto end = paddle::platform::PosixInNsec() / 1e-3;
+    return static_cast<double>(end - start) / FLAGS_repeat;
+  }
+};
+
+namespace jit = paddle::operators::jit;
+
+template <jit::KernelType KT, typename KernelTuples, typename PlaceType,
+          typename... Args>
+void BenchAllImpls(const typename KernelTuples::attr_type& attr, Args... args) {
+  BenchFunc<KernelTuples, Args...> benchmark;
+  std::vector<std::pair<std::string, double>> infos;
+  // test refer
+  auto refer = jit::GetRefer<KT, KernelTuples>();
+  if (!refer) {
+    LOG(FATAL) << "Refer can not be empty!";
+  }
+  infos.push_back(std::make_pair("Refer", benchmark(refer, args...)));
+
+  // test jitcode
+  auto jitcode = jit::GetJitCode<KT, KernelTuples, PlaceType>(attr);
+  if (jitcode) {
+    infos.push_back(std::make_pair("JitCode", benchmark(jitcode, args...)));
+  }
+  // test all impls in more
+  jit::KernelKey kkey(KT, PlaceType());
+  auto& pool = jit::KernelPool().Instance().AllKernels();
+  auto iter = pool.find(kkey);
+  if (iter != pool.end()) {
+    auto& impls = iter->second;
+    for (auto& impl : impls) {
+      auto i = dynamic_cast<const jit::KernelMore<KernelTuples>*>(impl.get());
+      if (i && i->UseMe(attr)) {
+        auto more = i->GetFunc();
+        infos.push_back(
+            std::make_pair(i->ImplType(), benchmark(more, args...)));
+      }
+    }
+  }
+  // Test result from Get function
+  auto tgt = jit::Get<KT, KernelTuples, PlaceType>(attr);
+  if (!tgt) {
+    LOG(FATAL) << "Target can not be empty!";
+  }
+  infos.push_back(std::make_pair("Target", benchmark(tgt, args...)));
+
+  // print
+  std::ostringstream loginfos;
+  loginfos << "Kernel Type " << jit::to_string(KT) << ": " << attr << ": ";
+  for (auto pair : infos) {
+    loginfos << pair.first << " takes " << pair.second << " us; ";
+  }
+  LOG(INFO) << loginfos.str();
+}
+
+template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+void BenchXYZNKernel() {
+  for (int d : TestSizes()) {
+    std::vector<T> x(d), y(d), z(d);
+    RandomVec<T>(d, x.data());
+    RandomVec<T>(d, y.data());
+    BenchAllImpls<KT, jit::XYZNTuples<T>, PlaceType>(d, x.data(), y.data(),
+                                                     z.data(), d);
+  }
+}
+
+template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+void BenchAXYNKernel() {
+  for (int d : TestSizes()) {
+    const T a = static_cast<T>(3);
+    std::vector<T> x(d), y(d);
+    RandomVec<T>(d, x.data());
+    BenchAllImpls<KT, jit::AXYNTuples<T>, PlaceType>(d, &a, x.data(), y.data(),
+                                                     d);
+  }
+}
+
+template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+void BenchXYNKernel() {
+  for (int d : TestSizes()) {
+    std::vector<T> x(d), y(d);
+    RandomVec<T>(d, x.data());
+    BenchAllImpls<KT, jit::XYNTuples<T>, PlaceType>(d, x.data(), y.data(), d);
+  }
+}
+
+template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+void BenchLSTMKernel() {
+  for (bool use_peephole : {true, false}) {
+    for (int d : TestSizes()) {
+      const jit::lstm_attr_t attr(d, jit::kVSigmoid, jit::kVTanh, jit::kVTanh,
+                                  use_peephole);
+      std::vector<T> x(4 * d), ct_1(d), ct(d), ht(d), wp(3 * d), checked(2 * d);
+      RandomVec<T>(4 * d, x.data(), -2.f, 2.f);
+      RandomVec<T>(3 * d, wp.data(), -2.f, 2.f);
+      RandomVec<T>(d, ct_1.data(), -2.f, 2.f);
+      const T* ct_1_data = ct_1.data();
+      const T* wp_data = wp.data();
+      T* x_data = x.data();
+      T* checked_data = checked.data();
+      T* ct_data = ct.data();
+      T* ht_data = ht.data();
+      jit::lstm_t step;
+      step.gates = x_data;
+      step.ct_1 = ct_1_data;
+      step.ct = ct_data;
+      step.ht = ht_data;
+      if (use_peephole) {
+        step.wp = wp_data;
+        step.checked = checked_data;
+      }
+      BenchAllImpls<KT, jit::LSTMTuples<T>, PlaceType>(attr, &step, &attr);
+    }
+  }
+}
+
+template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+void BenchGRUKernel() {
+  for (int d : TestSizes()) {
+    const jit::gru_attr_t attr(d, jit::kVSigmoid, jit::kVTanh);
+    std::vector<T> x(3 * d), ht_1(d), ht(d);
+    RandomVec<T>(3 * d, x.data(), -2.f, 2.f);
+    RandomVec<T>(d, ht_1.data(), -2.f, 2.f);
+    const T* ht_1_data = ht_1.data();
+    T* x_data = x.data();
+    T* ht_data = ht.data();
+    jit::gru_t step;
+    step.gates = x_data;
+    step.ht_1 = ht_1_data;
+    step.ht = ht_data;
+    BenchAllImpls<KT, jit::GRUTuples<T>, PlaceType>(attr, &step, &attr);
+  }
+}
+
+// Benchmark all jit kernels including jitcode, mkl and refer.
+// To use this tool, run command: ./benchmark [options...]
+// Options:
+//     --burning: the burning time before count
+//     --repeat: the repeat times
+//     --max_size: the max size would be tested
+int main(int argc, char* argv[]) {
+  gflags::ParseCommandLineFlags(&argc, &argv, true);
+  google::InitGoogleLogging(argv[0]);
+  LOG(INFO) << "Burning " << FLAGS_burning << " times, Repeat " << FLAGS_repeat
+            << " times.";
+  using T = float;
+  using PlaceType = paddle::platform::CPUPlace;
+  // xyzn
+  BenchXYZNKernel<jit::kVMul, T, PlaceType>();
+  BenchXYZNKernel<jit::kVAdd, T, PlaceType>();
+  BenchXYZNKernel<jit::kVAddRelu, T, PlaceType>();
+  BenchXYZNKernel<jit::kVSub, T, PlaceType>();
+
+  // axyn
+  BenchAXYNKernel<jit::kVScal, T, PlaceType>();
+  BenchAXYNKernel<jit::kVAddBias, T, PlaceType>();
+
+  // xyn
+  BenchXYNKernel<jit::kVRelu, T, PlaceType>();
+  BenchXYNKernel<jit::kVIdentity, T, PlaceType>();
+  BenchXYNKernel<jit::kVExp, T, PlaceType>();
+  BenchXYNKernel<jit::kVSigmoid, T, PlaceType>();
+  BenchXYNKernel<jit::kVTanh, T, PlaceType>();
+
+  // lstm and peephole
+  BenchLSTMKernel<jit::kLSTMCtHt, T, PlaceType>();
+  BenchLSTMKernel<jit::kLSTMC1H1, T, PlaceType>();
+
+  // gru functions
+  BenchGRUKernel<jit::kGRUH1, T, PlaceType>();
+  BenchGRUKernel<jit::kGRUHtPart1, T, PlaceType>();
+  BenchGRUKernel<jit::kGRUHtPart2, T, PlaceType>();
+}

paddle/fluid/operators/jit/gen/CMakeLists.txt

+
+file(GLOB jitcode_cc_srcs RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.cc")
+
+cc_library(jit_kernel_jitcode SRCS ${jitcode_cc_srcs} DEPS jit_kernel_base xbyak)
+set(JIT_KERNEL_DEPS ${JIT_KERNEL_DEPS} xbyak jit_kernel_jitcode PARENT_SCOPE)
+
+function(USE_JITKERNEL_GEN TARGET)
+    file(APPEND ${jit_file} "USE_JITKERNEL_GEN(${TARGET});\n")
+endfunction()
+
+# use gen jitcode kernel by name
+USE_JITKERNEL_GEN(kVMul)
+USE_JITKERNEL_GEN(kVAdd)
+#USE_JITKERNEL_GEN(kVSub) # TODO(TJ): enable me
+USE_JITKERNEL_GEN(kVAddRelu)
+USE_JITKERNEL_GEN(kVScal)
+USE_JITKERNEL_GEN(kVAddBias)
+USE_JITKERNEL_GEN(kVRelu)
+USE_JITKERNEL_GEN(kVIdentity)
+USE_JITKERNEL_GEN(kVExp)
+USE_JITKERNEL_GEN(kVSigmoid)
+USE_JITKERNEL_GEN(kVTanh)
+USE_JITKERNEL_GEN(kLSTMCtHt)
+USE_JITKERNEL_GEN(kLSTMC1H1)
+USE_JITKERNEL_GEN(kGRUH1)
+USE_JITKERNEL_GEN(kGRUHtPart1)
+USE_JITKERNEL_GEN(kGRUHtPart2)
+USE_JITKERNEL_GEN(kNCHW16CMulNC)

paddle/fluid/operators/jit/gen/act.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/jit/gen/act.h"
+#include "paddle/fluid/operators/jit/registry.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+const float ALIGN32_BEG exp_float_consts[] ALIGN32_END = {
+    REPEAT_8TIMES(1.f),
+    REPEAT_8TIMES(2.f),
+    REPEAT_8TIMES(0.5f),
+    REPEAT_8TIMES(EXP_HIG),
+    REPEAT_8TIMES(EXP_LOW),
+    REPEAT_8TIMES(CEPHES_LOG2EF),
+    REPEAT_8TIMES(CEPHES_EXP_C1),
+    REPEAT_8TIMES(CEPHES_EXP_C2),
+    REPEAT_8TIMES(CEPHES_EXP_P0),
+    REPEAT_8TIMES(CEPHES_EXP_P1),
+    REPEAT_8TIMES(CEPHES_EXP_P2),
+    REPEAT_8TIMES(CEPHES_EXP_P3),
+    REPEAT_8TIMES(CEPHES_EXP_P4),
+    REPEAT_8TIMES(CEPHES_EXP_P5),
+    REPEAT_8TIMES(EXP_MAX_INPUT),
+    REPEAT_8TIMES(SIGMOID_THRESHOLD_MAX),
+    REPEAT_8TIMES(SIGMOID_THRESHOLD_MIN)};
+
+const int ALIGN32_BEG exp_int_0x7f[] ALIGN32_END = {REPEAT_8TIMES(0x7f)};
+int ALIGN32_BEG g_tmp_mem[16] ALIGN32_END = {0};
+
+void VActJitCode::genCode() {
+  int offset = 0;
+  for (int i = 0; i < num_ / YMM_FLOAT_BLOCK; ++i) {
+    vmovups(ymm_src, ptr[param1 + offset]);
+    act<ymm_t>(ymm_dst, ymm_src, type_);
+    vmovups(ptr[param2 + offset], ymm_dst);
+    offset += sizeof(float) * YMM_FLOAT_BLOCK;
+  }
+  int rest = num_ % YMM_FLOAT_BLOCK;
+  while (rest > 0) {
+    int block = XMM_FLOAT_BLOCK;
+    if (rest >= 4) {
+      block = 4;
+      vmovups(xmm_src, ptr[param1 + offset]);
+    } else if (rest >= 2) {
+      block = 2;
+      vmovq(xmm_src, ptr[param1 + offset]);
+    } else {
+      block = 1;
+      vmovss(xmm_src, ptr[param1 + offset]);
+    }
+    act<xmm_t>(xmm_dst, xmm_src, type_);
+    if (rest >= 4) {
+      vmovups(ptr[param2 + offset], xmm_dst);
+    } else if (rest >= 2) {
+      vmovq(ptr[param2 + offset], xmm_dst);
+    } else {
+      vmovss(ptr[param2 + offset], xmm_dst);
+    }
+    offset += sizeof(float) * block;
+    rest -= block;
+  }
+  ret();
+}
+
+#define DECLARE_ACT_CREATOR(name)                                            \
+  class name##Creator : public JitCodeCreator<int> {                         \
+   public:                                                                   \
+    bool UseMe(const int& attr) const override {                             \
+      return platform::MayIUse(platform::avx);                               \
+    }                                                                        \
+    size_t CodeSize(const int& d) const override;                            \
+    std::unique_ptr<GenBase> CreateJitCode(const int& attr) const override { \
+      return make_unique<name##JitCode>(attr, CodeSize(attr));               \
+    }                                                                        \
+  }
+
+DECLARE_ACT_CREATOR(VRelu);
+DECLARE_ACT_CREATOR(VIdentity);
+DECLARE_ACT_CREATOR(VExp);
+DECLARE_ACT_CREATOR(VSigmoid);
+DECLARE_ACT_CREATOR(VTanh);
+
+// TODO(TJ): tuning use me
+size_t VReluCreator::CodeSize(const int& d) const {
+  return 96 /* init size */ +
+         (d / YMM_FLOAT_BLOCK + 3) * 4 /* instructions */ *
+             8 /* average bytes for each instruction */;
+}
+
+size_t VIdentityCreator::CodeSize(const int& d) const {
+  return 96 + (d / YMM_FLOAT_BLOCK + 3) * 4 * 8;
+}
+
+size_t VExpCreator::CodeSize(const int& d) const {
+  return 96 + (d / YMM_FLOAT_BLOCK + 3) * 70 * 8;
+}
+
+size_t VSigmoidCreator::CodeSize(const int& d) const {
+  return 96 + (d / YMM_FLOAT_BLOCK + 3) * 82 * 8;
+}
+
+size_t VTanhCreator::CodeSize(const int& d) const {
+  return 96 + (d / YMM_FLOAT_BLOCK + 3) * 84 * 8;
+}
+
+#undef DECLARE_ACT_CREATOR
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle
+
+namespace gen = paddle::operators::jit::gen;
+
+REGISTER_JITKERNEL_GEN(kVRelu, gen::VReluCreator);
+REGISTER_JITKERNEL_GEN(kVIdentity, gen::VIdentityCreator);
+REGISTER_JITKERNEL_GEN(kVExp, gen::VExpCreator);
+REGISTER_JITKERNEL_GEN(kVSigmoid, gen::VSigmoidCreator);
+REGISTER_JITKERNEL_GEN(kVTanh, gen::VTanhCreator);

paddle/fluid/operators/math/jit_code.h → paddle/fluid/operators/jit/gen/act.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. */
+ *
+ * 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/jit_gen.h"
-#include "paddle/fluid/operators/math/jit_kernel_impl.h"
-#include "paddle/fluid/platform/cpu_info.h"
+#include "glog/logging.h"
+#include "paddle/fluid/operators/jit/gen/jitcode.h"
 
 namespace paddle {
 namespace operators {
-namespace math {
-namespace jitkernel {
+namespace jit {
 namespace gen {
 
-using reg64_t = const Xbyak::Reg64;
-using reg32_t = const Xbyak::Reg32;
-using xmm_t = const Xbyak::Xmm;
-using ymm_t = const Xbyak::Ymm;
-using zmm_t = const Xbyak::Zmm;
-using Label = Xbyak::Label;
-
-typedef enum {
-  mul = 0,
-  add,
-  sub,
-  relu,
-  exp,
-  sigmoid,
-  tanh,
-  identity
-} operand_type;
-
 extern const float exp_float_consts[];
 extern const int exp_int_0x7f[];
 extern int g_tmp_mem[];
@@ -79,94 +59,15 @@ extern int g_tmp_mem[];
 #define OFFSET_SIGMOID_MAX 15 * YMM_FLOAT_BLOCK * sizeof(float)
 #define OFFSET_SIGMOID_MIN 16 * YMM_FLOAT_BLOCK * sizeof(float)
 
-// function: vec = Operand(vec(or scalar), vec(or scalar)) (maybe with relu)
-class VXXJitCode : public JitCode {
+class VActFunc : public JitCode {
  public:
-  const char* name() const override {
-    std::string base = "VXXJitCode";
-    if (scalar_index_ == 1) {
-      base += "_Scalar";
-    } else {
-      base += "_Vec";
-    }
-    if (type_ == operand_type::mul) {
-      base += "_Mul";
-    } else if (type_ == operand_type::add) {
-      base += "_Add";
-    }
-    if (scalar_index_ == 2) {
-      base += "_Scalar";
-    } else {
-      base += "_Vec";
-    }
-    base += (with_relu_ ? "_Relu" : "");
-    return base.c_str();
-  }
-  explicit VXXJitCode(int d, operand_type type, int scalar_index,
-                      bool with_relu, size_t code_size = 256 * 1024,
-                      void* code_ptr = nullptr)
-      : JitCode(code_size, code_ptr),
-        num_(d),
-        type_(type),
-        scalar_index_(scalar_index),
-        with_relu_(with_relu) {}
-  static bool init(int d, int scalar_index = 0);
-  void generate() override;
-
- private:
-  int num_;
-  operand_type type_;
-  int scalar_index_;
-  bool with_relu_;
-  reg64_t param1{abi_param1};
-  reg64_t param2{abi_param2};
-  reg64_t param3{abi_param3};
-
-  xmm_t xmm_src1 = xmm_t(0);
-  xmm_t xmm_src2 = xmm_t(1);
-  xmm_t xmm_dst = xmm_t(2);
-  xmm_t xmm_zero = xmm_t(3);
-
-  ymm_t ymm_src1 = ymm_t(0);
-  ymm_t ymm_src2 = ymm_t(1);
-  ymm_t ymm_dst = ymm_t(2);
-  ymm_t ymm_zero = ymm_t(3);
-};
-
-class VActJitCode : public JitCode {
- public:
-  const char* name() const override {
-    std::string base = "VActJitCode";
-    switch (type_) {
-      case operand_type::relu:
-        base += "_Relu";
-        break;
-      case operand_type::exp:
-        base += "_Exp";
-        break;
-      case operand_type::sigmoid:
-        base += "_Sigmoid";
-        break;
-      case operand_type::tanh:
-        base += "_Tanh";
-        break;
-      case operand_type::identity:
-        base += "_Identity";
-        break;
-      default:
-        break;
-    }
-    return base.c_str();
-  }
-
-  explicit VActJitCode(int d, operand_type type, size_t code_size = 256 * 1024,
-                       void* code_ptr = nullptr)
-      : JitCode(code_size, code_ptr), num_(d), type_(type) {}
-  static bool init(int d, operand_type type);
-  void generate() override;
+  explicit VActFunc(size_t code_size, void* code_ptr)
+      : JitCode(code_size, code_ptr) {}
+  virtual const char* name() const = 0;
+  virtual void genCode() = 0;
 
  protected:
-  // compute relu with ymm, xmm
+  // compute RELU with ymm, xmm
   template <typename JMM>
   void relu_jmm(JMM& dst, JMM& src, int zero_idx = 15) {  // NOLINT
     JMM zero = JMM(zero_idx);
@@ -174,7 +75,7 @@ class VActJitCode : public JitCode {
     vmaxps(dst, src, zero);
   }
 
-  // compute exp with ymm, xmm
+  // compute EXP with ymm, xmm
   template <typename JMM>
   void exp_jmm(JMM& dst, JMM& src, int src_idx = 11, int fx_idx = 12,  // NOLINT
                int fy_idx = 13, int mask_idx = 14, int tmp_idx = 15) {
@@ -258,7 +159,7 @@ class VActJitCode : public JitCode {
     pop(reg_ptr_global);
   }
 
-  // compute sigmoid with ymm, xmm
+  // compute SIGMOID with ymm, xmm
   template <typename JMM>
   void sigmoid_jmm(JMM& dst, JMM& src, int src_idx = 11,  // NOLINT
                    int fx_idx = 12, int fy_idx = 13, int mask_idx = 14,
@@ -283,7 +184,7 @@ class VActJitCode : public JitCode {
     pop(reg_ptr_global);
   }
 
-  // compute tanh with ymm, xmm
+  // compute TANH with ymm, xmm
   template <typename JMM>
   void tanh_jmm(JMM& dst, JMM& src, int src_idx = 11,  // NOLINT
                 int fx_idx = 12, int fy_idx = 13, int mask_idx = 14,
@@ -310,223 +211,109 @@ class VActJitCode : public JitCode {
     pop(reg_ptr_global);
   }
 
+  // compute IDENTITY with ymm, xmm
+  template <typename JMM>
+  void identity_jmm(JMM& dst, JMM& src, int zero_idx) {  // NOLINT
+    JMM zero = JMM(zero_idx);
+    vxorps(zero, zero, zero);
+    vaddps(dst, src, zero);
+    // TODO(TJ): use below
+    // dst.setIdx(src.getIdx());
+  }
+
   template <typename JMM>
   void act(JMM& dst, JMM& src, operand_type type) {  // NOLINT
     // use 11~15
     switch (type) {
-      case operand_type::relu:
+      case operand_type::RELU:
         relu_jmm<JMM>(dst, src, 15);
         break;
-      case operand_type::exp:
+      case operand_type::EXP:
         exp_jmm<JMM>(dst, src, 11, 12, 13, 14, 15);
         break;
-      case operand_type::sigmoid:
+      case operand_type::SIGMOID:
         sigmoid_jmm<JMM>(dst, src, 11, 12, 13, 14, 15);
         break;
-      case operand_type::tanh:
+      case operand_type::TANH:
         tanh_jmm<JMM>(dst, src, 11, 12, 13, 14, 15);
         break;
-      case operand_type::identity:
+      case operand_type::IDENTITY:
+        identity_jmm<JMM>(dst, src, 15);
         break;
       default:
-        // throw error
+        LOG(FATAL) << "Do not support this operand type: " << type;
         break;
     }
   }
-
- protected:
-  int num_;
-  operand_type type_;
-  reg64_t param1{abi_param1};
-  reg64_t param2{abi_param2};
-
-  xmm_t xmm_src = xmm_t(0);
-  ymm_t ymm_src = ymm_t(0);
-
-  xmm_t xmm_dst = xmm_t(1);
-  ymm_t ymm_dst = ymm_t(1);
 };
 
-class LSTMJitCode : public VActJitCode {
+class VActJitCode : public VActFunc {
  public:
-  const char* name() const override {
-    std::string base = "LSTMJitCode";
-    if (use_peephole_) {
-      base += "_Peephole";
-    }
-    if (compute_c1h1_) {
-      base += "_C1H1";
-    }
-    auto AddTypeStr = [&](operand_type type) {
-      switch (type) {
-        case operand_type::relu:
-          base += "_Relu";
-          break;
-        case operand_type::exp:
-          base += "_Exp";
-          break;
-        case operand_type::sigmoid:
-          base += "_Sigmoid";
-          break;
-        case operand_type::tanh:
-          base += "_Tanh";
-          break;
-        case operand_type::identity:
-          base += "_Identity";
-          break;
-        default:
-          break;
+  explicit VActJitCode(int d, operand_type type, size_t code_size,
+                       void* code_ptr = nullptr)
+      : VActFunc(code_size, code_ptr), num_(d), type_(type) {
+    if (!(type_ == operand_type::RELU || type_ == operand_type::EXP ||
+          type_ == operand_type::SIGMOID || type_ == operand_type::TANH ||
+          type_ == operand_type::IDENTITY)) {
+      LOG(FATAL) << "Do not support this operand type: " << type_;
     }
-    };
-    AddTypeStr(act_gate_);
-    AddTypeStr(act_cand_);
-    AddTypeStr(act_cell_);
-    return base.c_str();
+    this->genCode();
   }
 
-  explicit LSTMJitCode(bool compute_c1h1, const lstm_attr_t& attr,
-                       size_t code_size = 256 * 1024, void* code_ptr = nullptr)
-      : VActJitCode(attr.d, operand_type::sigmoid /* this is bugy*/, code_size,
-                    code_ptr),
-        compute_c1h1_(compute_c1h1) {
-    auto typeExchange = [](const std::string& type) -> gen::operand_type {
-      if (type == "sigmoid") {
-        return operand_type::sigmoid;
-      } else if (type == "relu") {
-        return operand_type::relu;
-      } else if (type == "tanh") {
-        return operand_type::tanh;
-      } else if (type == "identity" || type == "") {
-        return operand_type::identity;
-      }  // else throw error
-      return operand_type::identity;
-    };
-    num_ = attr.d;
-    use_peephole_ = attr.use_peephole;
-    act_gate_ = typeExchange(attr.act_gate);
-    act_cand_ = typeExchange(attr.act_cand);
-    act_cell_ = typeExchange(attr.act_cell);
-  }
-  static bool init(int d);
-  void generate() override;
-
- protected:
-  int num_;
-  bool compute_c1h1_;
-  bool use_peephole_;
-  operand_type act_gate_;
-  operand_type act_cand_;
-  operand_type act_cell_;
-  reg64_t param1{abi_param1};
-};
-
-class GRUJitCode : public VActJitCode {
- public:
   const char* name() const override {
-    std::string base = "GRUJitCode";
-    if (id_ == 0) {
-      base += "_H1";
-    } else if (id_ == 1) {
-      base += "_HtPart1";
-    } else if (id_ == 2) {
-      base += "_HtPart2";
-    }
-    auto AddTypeStr = [&](operand_type type) {
-      switch (type) {
-        case operand_type::relu:
+    std::string base = "VActJitCode";
+    switch (type_) {
+      case operand_type::RELU:
         base += "_Relu";
         break;
-        case operand_type::exp:
+      case operand_type::EXP:
         base += "_Exp";
         break;
-        case operand_type::sigmoid:
+      case operand_type::SIGMOID:
         base += "_Sigmoid";
         break;
-        case operand_type::tanh:
+      case operand_type::TANH:
         base += "_Tanh";
         break;
-        case operand_type::identity:
+      case operand_type::IDENTITY:
         base += "_Identity";
         break;
       default:
         break;
     }
-    };
-    AddTypeStr(act_gate_);
-    AddTypeStr(act_cand_);
     return base.c_str();
   }
-
-  explicit GRUJitCode(int id, const gru_attr_t& attr,
-                      size_t code_size = 256 * 1024, void* code_ptr = nullptr)
-      : VActJitCode(attr.d, operand_type::sigmoid /* this is bugy*/, code_size,
-                    code_ptr),
-        id_(id) {
-    auto typeExchange = [](const std::string& type) -> gen::operand_type {
-      if (type == "sigmoid") {
-        return operand_type::sigmoid;
-      } else if (type == "relu") {
-        return operand_type::relu;
-      } else if (type == "tanh") {
-        return operand_type::tanh;
-      } else if (type == "identity" || type == "") {
-        return operand_type::identity;
-      }  // else throw error
-      return operand_type::identity;
-    };
-    num_ = attr.d;
-    act_gate_ = typeExchange(attr.act_gate);
-    act_cand_ = typeExchange(attr.act_cand);
-  }
-  static bool init(int d);
-  void generate() override;
+  void genCode() override;
 
  protected:
-  int id_;
   int num_;
-  operand_type act_gate_;
-  operand_type act_cand_;
+  operand_type type_;
   reg64_t param1{abi_param1};
-};
+  reg64_t param2{abi_param2};
 
-#ifdef PADDLE_WITH_MKLDNN
-struct EltwiseMulnChw16cNC : public Xbyak::CodeGenerator {
-  explicit EltwiseMulnChw16cNC(size_t code_size = 256 * 1024)
-      : Xbyak::CodeGenerator(code_size) {
-    // RDI is ptr x_input
-    // RSI is ptr y_input
-    // RDX is ptr output
-    // RCX is height
-    // r8 is width
+  xmm_t xmm_src = xmm_t(0);
+  ymm_t ymm_src = ymm_t(0);
 
-    push(rbx);
+  xmm_t xmm_dst = xmm_t(1);
+  ymm_t ymm_dst = ymm_t(1);
+};
 
-    xor_(rax, rax);
-    xor_(r10, r10);
-    vmovups(zmm3, ptr[rsi]);
+#define DECLARE_ACT_JITCODE(name, op_type)                                    \
+  class name##JitCode : public VActJitCode {                                  \
+   public:                                                                    \
+    explicit name##JitCode(int d, size_t code_size, void* code_ptr = nullptr) \
+        : VActJitCode(d, op_type, code_size, code_ptr) {}                     \
+  };
 
-    L("h_loop");
-    xor_(rbx, rbx);
-    L("w_loop");
-    vmovups(zmm2, ptr[rdi + rax]);
-    vmulps(zmm1, zmm2, zmm3);
-    vmovups(ptr[rdx + rax], zmm1);
-    add(rax, 64);
-    inc(rbx);
-    cmp(r8, rbx);
-    jnz("w_loop");
-    inc(r10);
-    cmp(r10, rcx);
-    jnz("h_loop");
+DECLARE_ACT_JITCODE(VRelu, operand_type::RELU);
+DECLARE_ACT_JITCODE(VIdentity, operand_type::IDENTITY);
+DECLARE_ACT_JITCODE(VExp, operand_type::EXP);
+DECLARE_ACT_JITCODE(VSigmoid, operand_type::SIGMOID);
+DECLARE_ACT_JITCODE(VTanh, operand_type::TANH);
 
-    pop(rbx);
-    ret();
-  }
-};
-#endif
+#undef DECLARE_ACT_JITCODE
 
 }  // namespace gen
-}  // namespace jitkernel
-}  // namespace math
+}  // namespace jit
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/jit/gen/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/jit/gen/blas.h"
+#include "paddle/fluid/operators/jit/registry.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+void VXXJitCode::genCode() {
+  // do not need push stack, and do not need save avx512reg if do not use avx512
+  int offset = 0;
+  if (with_relu_) {
+    vxorps(ymm_zero, ymm_zero, ymm_zero);
+  }
+  if (scalar_index_ == 1) {
+    vbroadcastss(ymm_src1, ptr[param1]);
+  } else if (scalar_index_ == 2) {
+    vbroadcastss(ymm_src2, ptr[param2]);
+  }
+  for (int i = 0; i < num_ / YMM_FLOAT_BLOCK; ++i) {
+    if (scalar_index_ != 1) {
+      vmovups(ymm_src1, ptr[param1 + offset]);
+    }
+    if (scalar_index_ != 2) {
+      vmovups(ymm_src2, ptr[param2 + offset]);
+    }
+    if (type_ == operand_type::MUL) {
+      vmulps(ymm_dst, ymm_src1, ymm_src2);
+    } else if (type_ == operand_type::ADD) {
+      vaddps(ymm_dst, ymm_src1, ymm_src2);
+    }
+    if (with_relu_) {
+      vmaxps(ymm_dst, ymm_zero, ymm_dst);
+    }
+    vmovups(ptr[param3 + offset], ymm_dst);
+    offset += sizeof(float) * YMM_FLOAT_BLOCK;
+  }
+  int rest = num_ % YMM_FLOAT_BLOCK;
+  while (rest > 0) {
+    int block = XMM_FLOAT_BLOCK;
+    if (rest >= 4) {
+      block = 4;
+      if (scalar_index_ != 1) {
+        vmovups(xmm_src1, ptr[param1 + offset]);
+      }
+      if (scalar_index_ != 2) {
+        vmovups(xmm_src2, ptr[param2 + offset]);
+      }
+    } else if (rest >= 2) {
+      block = 2;
+      if (scalar_index_ != 1) {
+        vmovq(xmm_src1, ptr[param1 + offset]);
+      }
+      if (scalar_index_ != 2) {
+        vmovq(xmm_src2, ptr[param2 + offset]);
+      }
+    } else {
+      block = 1;
+      if (scalar_index_ != 1) {
+        vmovss(xmm_src1, ptr[param1 + offset]);
+      }
+      if (scalar_index_ != 2) {
+        vmovss(xmm_src2, ptr[param2 + offset]);
+      }
+    }
+    switch (type_) {
+      case operand_type::MUL:
+        vmulps(xmm_dst, xmm_src1, xmm_src2);
+        break;
+      case operand_type::ADD:
+        vaddps(xmm_dst, xmm_src1, xmm_src2);
+        break;
+      default:
+        break;
+    }
+    if (with_relu_) {
+      vmaxps(xmm_dst, xmm_zero, xmm_dst);
+    }
+    if (rest >= 4) {
+      vmovups(ptr[param3 + offset], xmm_dst);
+    } else if (rest >= 2) {
+      vmovq(ptr[param3 + offset], xmm_dst);
+    } else {
+      vmovss(ptr[param3 + offset], xmm_dst);
+    }
+    offset += sizeof(float) * block;
+    rest -= block;
+  }
+  ret();
+}
+
+void NCHW16CMulNCJitCode::genCode() {
+  // RDI is ptr x_input
+  // RSI is ptr y_input
+  // RDX is ptr output
+  // RCX is height
+  // r8 is width
+
+  push(rbx);
+
+  xor_(rax, rax);
+  xor_(r10, r10);
+  vmovups(zmm3, ptr[rsi]);
+
+  L("h_loop");
+  xor_(rbx, rbx);
+  L("w_loop");
+  vmovups(zmm2, ptr[rdi + rax]);
+  vmulps(zmm1, zmm2, zmm3);
+  vmovups(ptr[rdx + rax], zmm1);
+  add(rax, 64);
+  inc(rbx);
+  cmp(r8, rbx);
+  jnz("w_loop");
+  inc(r10);
+  cmp(r10, rcx);
+  jnz("h_loop");
+
+  pop(rbx);
+  ret();
+}
+
+class NCHW16CMulNCCreator : public JitCodeCreator<int> {
+ public:
+  bool UseMe(const int& attr) const override {
+    return platform::MayIUse(platform::avx512f);
+  }
+  size_t CodeSize(const int& d) const override { return 256 * 1024; }
+  std::unique_ptr<GenBase> CreateJitCode(const int& attr) const override {
+    return make_unique<NCHW16CMulNCJitCode>(attr, CodeSize(attr));
+  }
+};
+
+#define DECLARE_BLAS_CREATOR(name)                                           \
+  class name##Creator : public JitCodeCreator<int> {                         \
+   public:                                                                   \
+    bool UseMe(const int& attr) const override {                             \
+      return platform::MayIUse(platform::avx);                               \
+    }                                                                        \
+    size_t CodeSize(const int& d) const override {                           \
+      return 96 + d / YMM_FLOAT_BLOCK * 4 * 8;                               \
+    }                                                                        \
+    std::unique_ptr<GenBase> CreateJitCode(const int& attr) const override { \
+      return make_unique<name##JitCode>(attr, CodeSize(attr));               \
+    }                                                                        \
+  }
+
+DECLARE_BLAS_CREATOR(VMul);
+DECLARE_BLAS_CREATOR(VAdd);
+DECLARE_BLAS_CREATOR(VSub);
+DECLARE_BLAS_CREATOR(VAddRelu);
+DECLARE_BLAS_CREATOR(VScal);
+DECLARE_BLAS_CREATOR(VAddBias);
+
+#undef DECLARE_BLAS_CREATOR
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle
+
+namespace gen = paddle::operators::jit::gen;
+
+REGISTER_JITKERNEL_GEN(kVMul, gen::VMulCreator);
+REGISTER_JITKERNEL_GEN(kVAdd, gen::VAddCreator);
+// TODO(TJ): enable sub
+// REGISTER_JITKERNEL_GEN(kVSub, gen::VSubCreator);
+REGISTER_JITKERNEL_GEN(kVAddRelu, gen::VAddReluCreator);
+REGISTER_JITKERNEL_GEN(kVScal, gen::VScalCreator);
+REGISTER_JITKERNEL_GEN(kVAddBias, gen::VAddBiasCreator);
+REGISTER_JITKERNEL_GEN(kNCHW16CMulNC, gen::NCHW16CMulNCCreator);

paddle/fluid/operators/jit/gen/blas.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 "glog/logging.h"
+#include "paddle/fluid/operators/jit/gen/jitcode.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+// function: vec = Operand(vec(or scalar), vec(or scalar)) (maybe with relu)
+class VXXJitCode : public JitCode {
+ public:
+  explicit VXXJitCode(int d, operand_type type, int scalar_index,
+                      bool with_relu, size_t code_size = 256 * 1024,
+                      void* code_ptr = nullptr)
+      : JitCode(code_size, code_ptr),
+        num_(d),
+        type_(type),
+        scalar_index_(scalar_index),
+        with_relu_(with_relu) {
+    if (!(type_ == operand_type::MUL || type_ == operand_type::ADD)) {
+      LOG(FATAL) << "Do not support this operand type: " << type_;
+    }
+    this->genCode();
+  }
+
+  virtual const char* name() const {
+    std::string base = "VXXJitCode";
+    if (scalar_index_ == 1) {
+      base += "_Scalar";
+    } else {
+      base += "_Vec";
+    }
+    if (type_ == operand_type::MUL) {
+      base += "_Mul";
+    } else if (type_ == operand_type::ADD) {
+      base += "_Add";
+    }
+    if (scalar_index_ == 2) {
+      base += "_Scalar";
+    } else {
+      base += "_Vec";
+    }
+    base += (with_relu_ ? "_Relu" : "");
+    return base.c_str();
+  }
+  void genCode() override;
+
+ private:
+  int num_;
+  operand_type type_;
+  int scalar_index_;
+  bool with_relu_;
+  reg64_t param1{abi_param1};
+  reg64_t param2{abi_param2};
+  reg64_t param3{abi_param3};
+
+  xmm_t xmm_src1 = xmm_t(0);
+  xmm_t xmm_src2 = xmm_t(1);
+  xmm_t xmm_dst = xmm_t(2);
+  xmm_t xmm_zero = xmm_t(3);
+
+  ymm_t ymm_src1 = ymm_t(0);
+  ymm_t ymm_src2 = ymm_t(1);
+  ymm_t ymm_dst = ymm_t(2);
+  ymm_t ymm_zero = ymm_t(3);
+};
+
+#define DECLARE_BLAS_JITCODE(name, op_type, scalar_idx, with_relu)             \
+  class name##JitCode : public VXXJitCode {                                    \
+   public:                                                                     \
+    explicit name##JitCode(int d, size_t code_size, void* code_ptr = nullptr)  \
+        : VXXJitCode(d, op_type, scalar_idx, with_relu, code_size, code_ptr) { \
+    }                                                                          \
+  };
+
+DECLARE_BLAS_JITCODE(VMul, operand_type::MUL, 0, false);
+DECLARE_BLAS_JITCODE(VAdd, operand_type::ADD, 0, false);
+DECLARE_BLAS_JITCODE(VSub, operand_type::SUB, 0, false);
+DECLARE_BLAS_JITCODE(VAddRelu, operand_type::ADD, 0, true);
+DECLARE_BLAS_JITCODE(VScal, operand_type::MUL, 1, false);
+DECLARE_BLAS_JITCODE(VAddBias, operand_type::ADD, 1, false);
+
+#undef DECLARE_BLAS_JITCODE
+
+// nChw16c = nChw16c .* NC
+class NCHW16CMulNCJitCode : public JitCode {
+ public:
+  DECLARE_JIT_CODE(NCHW16CMulNCJitCode);
+  explicit NCHW16CMulNCJitCode(int d /*unused*/, size_t code_size,
+                               void* code_ptr = nullptr)
+      : JitCode(code_size, code_ptr) {
+    this->genCode();
+  }
+  void genCode() override;
+};
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/gen/gru.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/jit/gen/gru.h"
+#include <stddef.h>  // offsetof
+#include "paddle/fluid/operators/jit/registry.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+void GRUJitCode::genCode() {
+  reg64_t reg_ptr_gates = rax;
+  reg64_t reg_ptr_ht_1 = r9;
+  reg64_t reg_ptr_ht = r10;
+  mov(reg_ptr_gates, ptr[param1 + offsetof(gru_t, gates)]);
+  mov(reg_ptr_ht_1, ptr[param1 + offsetof(gru_t, ht_1)]);
+  mov(reg_ptr_ht, ptr[param1 + offsetof(gru_t, ht)]);
+  ymm_t ymm_one = ymm_t(0);
+
+  if (id_ == 2) {
+    reg64_t reg_ptr_tmp = r11;
+    mov(reg_ptr_tmp, reinterpret_cast<size_t>(exp_float_consts));
+    vmovaps(ymm_one, ptr[reg_ptr_tmp + OFFSET_EXP_ONE]);
+  }
+  int offset = 0;
+  int d = num_ * sizeof(float);
+  for (int i = 0; i < num_ / YMM_FLOAT_BLOCK; ++i) {
+    ymm_t ymm_u = ymm_t(1);
+    ymm_t ymm_r = ymm_t(2);
+    ymm_t ymm_s = ymm_t(3);
+    ymm_t ymm_ht_1 = ymm_t(4);
+    // W: {W_update, W_reset; W_state}
+    if (id_ == 0 || id_ == 2) {
+      vmovups(ymm_u, ptr[reg_ptr_gates + offset]);
+      vmovups(ymm_s, ptr[reg_ptr_gates + offset + 2 * d]);
+    }
+    if (id_ == 1) {
+      vmovups(ymm_r, ptr[reg_ptr_gates + offset + d]);
+    }
+    if (id_ == 1 || id_ == 2) {
+      vmovups(ymm_ht_1, ptr[reg_ptr_ht_1 + offset]);
+    }
+
+    if (id_ == 0) {
+      // ht = act_gate(u) * act_cand(s)
+      act<ymm_t>(ymm_u, ymm_u, act_gate_);
+      act<ymm_t>(ymm_s, ymm_s, act_cand_);
+      vmulps(ymm_s, ymm_s, ymm_u);
+      vmovups(ptr[reg_ptr_ht + offset], ymm_s);
+    } else if (id_ == 1) {
+      // ht = act_gate(r) * ht_1
+      act<ymm_t>(ymm_r, ymm_r, act_gate_);
+      vmulps(ymm_r, ymm_r, ymm_ht_1);
+      vmovups(ptr[reg_ptr_ht + offset], ymm_r);
+    } else if (id_ == 2) {
+      // ht = act_gate(u) * act_cand(s) + (1-act_gate(u)) * ht_1
+      ymm_t ymm_one_inner = ymm_t(ymm_one.getIdx());
+      act<ymm_t>(ymm_u, ymm_u, act_gate_);
+      act<ymm_t>(ymm_s, ymm_s, act_cand_);
+      vmulps(ymm_s, ymm_s, ymm_u);
+      vsubps(ymm_u, ymm_one_inner, ymm_u);
+      vmulps(ymm_u, ymm_ht_1, ymm_u);
+      vaddps(ymm_u, ymm_s, ymm_u);
+      vmovups(ptr[reg_ptr_ht + offset], ymm_u);
+    }
+    offset += sizeof(float) * YMM_FLOAT_BLOCK;
+  }
+  ret();
+}
+
+#define DECLARE_GRU_CREATOR(name)                                 \
+  class name##Creator : public JitCodeCreator<gru_attr_t> {       \
+   public:                                                        \
+    /* TODO(TJ): enable more */                                   \
+    bool UseMe(const gru_attr_t& attr) const override {           \
+      return platform::MayIUse(platform::avx) && attr.d % 8 == 0; \
+    }                                                             \
+    size_t CodeSize(const gru_attr_t& attr) const override {      \
+      return 96 + attr.d / YMM_FLOAT_BLOCK * 96 * 2 * 8;          \
+    }                                                             \
+    std::unique_ptr<GenBase> CreateJitCode(                       \
+        const gru_attr_t& attr) const override {                  \
+      return make_unique<name##JitCode>(attr, CodeSize(attr));    \
+    }                                                             \
+  }
+
+DECLARE_GRU_CREATOR(GRUH1);
+DECLARE_GRU_CREATOR(GRUHtPart1);
+DECLARE_GRU_CREATOR(GRUHtPart2);
+
+#undef DECLARE_GRU_CREATOR
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle
+
+namespace gen = paddle::operators::jit::gen;
+
+REGISTER_JITKERNEL_GEN(kGRUH1, gen::GRUH1Creator);
+REGISTER_JITKERNEL_GEN(kGRUHtPart1, gen::GRUHtPart1Creator);
+REGISTER_JITKERNEL_GEN(kGRUHtPart2, gen::GRUHtPart2Creator);

paddle/fluid/operators/jit/gen/gru.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 "glog/logging.h"
+#include "paddle/fluid/operators/jit/gen/act.h"
+#include "paddle/fluid/operators/jit/gen/jitcode.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+class GRUJitCode : public VActFunc {
+ public:
+  explicit GRUJitCode(int id, const gru_attr_t& attr, size_t code_size,
+                      void* code_ptr = nullptr)
+      : VActFunc(code_size, code_ptr), id_(id), num_(attr.d) {
+    auto typeExchange = [](KernelType type) -> gen::operand_type {
+      if (type == KernelType::kVSigmoid) {
+        return operand_type::SIGMOID;
+      } else if (type == KernelType::kVRelu) {
+        return operand_type::RELU;
+      } else if (type == KernelType::kVTanh) {
+        return operand_type::TANH;
+      } else if (type == KernelType::kVIdentity) {
+        return operand_type::IDENTITY;
+      } else {
+        LOG(FATAL) << "Do not support this jit::KernelType: " << type;
+      }
+      return operand_type::IDENTITY;
+    };
+    act_gate_ = typeExchange(attr.act_gate);
+    act_cand_ = typeExchange(attr.act_cand);
+
+    this->genCode();
+  }
+
+  const char* name() const override {
+    std::string base = "GRUJitCode";
+    if (id_ == 0) {
+      base += "_H1";
+    } else if (id_ == 1) {
+      base += "_HtPart1";
+    } else if (id_ == 2) {
+      base += "_HtPart2";
+    }
+    auto AddTypeStr = [&](operand_type type) {
+      switch (type) {
+        case operand_type::RELU:
+          base += "_Relu";
+          break;
+        case operand_type::EXP:
+          base += "_Exp";
+          break;
+        case operand_type::SIGMOID:
+          base += "_Sigmoid";
+          break;
+        case operand_type::TANH:
+          base += "_Tanh";
+          break;
+        case operand_type::IDENTITY:
+          base += "_Identity";
+          break;
+        default:
+          break;
+      }
+    };
+    AddTypeStr(act_gate_);
+    AddTypeStr(act_cand_);
+    return base.c_str();
+  }
+  void genCode() override;
+
+ protected:
+  int id_;
+  int num_;
+  operand_type act_gate_;
+  operand_type act_cand_;
+  reg64_t param1{abi_param1};
+};
+
+#define DECLARE_GRU_JITCODE(name, id)                                \
+  class name##JitCode : public GRUJitCode {                          \
+   public:                                                           \
+    explicit name##JitCode(const gru_attr_t& attr, size_t code_size, \
+                           void* code_ptr = nullptr)                 \
+        : GRUJitCode(id, attr, code_size, code_ptr) {}               \
+  };
+
+DECLARE_GRU_JITCODE(GRUH1, 0);
+DECLARE_GRU_JITCODE(GRUHtPart1, 1);
+DECLARE_GRU_JITCODE(GRUHtPart2, 2);
+
+#undef DECLARE_GRU_JITCODE
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/gen/jitcode.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 <type_traits>
+#include "paddle/fluid/operators/jit/gen_base.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+#define XBYAK_USE_MMAP_ALLOCATOR
+#include "xbyak/xbyak.h"
+#include "xbyak/xbyak_util.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+// Application Binary Interface
+constexpr Xbyak::Operand::Code abi_param1(Xbyak::Operand::RDI),
+    abi_param2(Xbyak::Operand::RSI), abi_param3(Xbyak::Operand::RDX),
+    abi_param4(Xbyak::Operand::RCX);
+
+constexpr Xbyak::Operand::Code g_abi_regs[] = {
+    Xbyak::Operand::RBX, Xbyak::Operand::RBP, Xbyak::Operand::R12,
+    Xbyak::Operand::R13, Xbyak::Operand::R14, Xbyak::Operand::R15};
+
+constexpr int num_g_abi_regs = sizeof(g_abi_regs) / sizeof(g_abi_regs[0]);
+
+using reg64_t = const Xbyak::Reg64;
+using reg32_t = const Xbyak::Reg32;
+using xmm_t = const Xbyak::Xmm;
+using ymm_t = const Xbyak::Ymm;
+using zmm_t = const Xbyak::Zmm;
+using Label = Xbyak::Label;
+
+typedef enum {
+  MUL = 0,
+  ADD,
+  SUB,
+  RELU,
+  EXP,
+  SIGMOID,
+  TANH,
+  IDENTITY
+} operand_type;
+
+#define DECLARE_JIT_CODE(codename) \
+  const char* name() const override { return #codename; }
+
+class JitCode : public GenBase, public Xbyak::CodeGenerator {
+ public:
+  explicit JitCode(size_t code_size, void* code_ptr = nullptr)
+      : Xbyak::CodeGenerator(
+            (code_size % 4096 != 0 ? (code_size / 4096 + 1) * 4096 : code_size),
+            code_ptr) {}
+
+  virtual const char* name() const = 0;
+  virtual void genCode() = 0;
+
+  size_t getSize() const override { return CodeGenerator::getSize(); }
+  const unsigned char* getCodeInternal() override {
+    const Xbyak::uint8* code = CodeGenerator::getCode();
+    return code;
+  }
+
+ protected:
+  Xbyak::Reg64 param1{abi_param1};
+  const int EVEX_max_8b_offt = 0x200;
+  const Xbyak::Reg64 reg_EVEX_max_8b_offt = rbp;
+
+  virtual void preCode() {
+    for (int i = 0; i < num_g_abi_regs; ++i) {
+      push(Xbyak::Reg64(g_abi_regs[i]));
+    }
+    if (platform::MayIUse(platform::avx512f)) {
+      mov(reg_EVEX_max_8b_offt, 2 * EVEX_max_8b_offt);
+    }
+  }
+  virtual void postCode() {
+    for (int i = 0; i < num_g_abi_regs; ++i) {
+      pop(Xbyak::Reg64(g_abi_regs[num_g_abi_regs - 1 - i]));
+    }
+    ret();
+  }
+  void L(const char* label) { Xbyak::CodeGenerator::L(label); }
+  void L(const Xbyak::Label& label) { Xbyak::CodeGenerator::L(label); }
+  // Enhanced vector extension
+  Xbyak::Address EVEX_compress_addr(Xbyak::Reg64 base, int offt,
+                                    bool bcast = false) {
+    int scale = 0;
+    // Learn from https://github.com/intel/mkl-dnn
+    if (EVEX_max_8b_offt <= offt && offt < 3 * EVEX_max_8b_offt) {
+      offt = offt - 2 * EVEX_max_8b_offt;
+      scale = 1;
+    } else if (3 * EVEX_max_8b_offt <= offt && offt < 5 * EVEX_max_8b_offt) {
+      offt = offt - 4 * EVEX_max_8b_offt;
+      scale = 2;
+    }
+    auto re = Xbyak::RegExp() + base + offt;
+    if (scale) {
+      re = re + reg_EVEX_max_8b_offt * scale;
+    }
+    if (bcast) {
+      return zword_b[re];
+    } else {
+      return zword[re];
+    }
+  }
+};
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/gen/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/jit/gen/lstm.h"
+#include <stddef.h>  // offsetof
+#include "paddle/fluid/operators/jit/registry.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+void LSTMJitCode::genCode() {
+  if (use_peephole_) {
+    preCode();
+  }
+  reg64_t reg_ptr_gates = rax;
+  reg64_t reg_ptr_ct_1 = r9;
+  reg64_t reg_ptr_ct = r10;
+  reg64_t reg_ptr_ht = r11;
+  reg64_t reg_ptr_wp = r12;
+  mov(reg_ptr_gates, ptr[param1 + offsetof(lstm_t, gates)]);
+  mov(reg_ptr_ct_1, ptr[param1 + offsetof(lstm_t, ct_1)]);
+  mov(reg_ptr_ct, ptr[param1 + offsetof(lstm_t, ct)]);
+  mov(reg_ptr_ht, ptr[param1 + offsetof(lstm_t, ht)]);
+  if (use_peephole_) {
+    mov(reg_ptr_wp, ptr[param1 + offsetof(lstm_t, wp)]);
+  }
+
+  int offset = 0;
+  int d = num_ * sizeof(float);
+  for (int i = 0; i < num_ / YMM_FLOAT_BLOCK; ++i) {
+    /* gates: W_ch, W_ih, W_fh, W_oh */
+    ymm_t ymm_c = ymm_t(0);
+    ymm_t ymm_i = ymm_t(1);
+    ymm_t ymm_f = ymm_t(2);
+    ymm_t ymm_o = ymm_t(3);
+    ymm_t ymm_ct_1 = ymm_t(4);
+    ymm_t ymm_wp0 = ymm_t(5);
+    ymm_t ymm_wp1 = ymm_t(6);
+    ymm_t ymm_wp2 = ymm_t(7);
+    vmovups(ymm_c, ptr[reg_ptr_gates + offset]);
+    vmovups(ymm_i, ptr[reg_ptr_gates + offset + d]);
+    vmovups(ymm_f, ptr[reg_ptr_gates + offset + 2 * d]);
+    vmovups(ymm_o, ptr[reg_ptr_gates + offset + 3 * d]);
+    if (!compute_c1h1_) {
+      vmovups(ymm_ct_1, ptr[reg_ptr_ct_1 + offset]);
+    }
+    if (use_peephole_) {
+      vmovups(ymm_wp0, ptr[reg_ptr_wp + offset]);
+      vmovups(ymm_wp1, ptr[reg_ptr_wp + offset + d]);
+      vmovups(ymm_wp2, ptr[reg_ptr_wp + offset + 2 * d]);
+    }
+    /* C_t = act_cand(c) * act_gate(i) + C_t-1 * act_gate(f) */
+    // act_cand(c)
+    act<ymm_t>(ymm_c, ymm_c, act_cand_);
+    // act_gate(i) or act_gate(ct_1 * wp0 + i)
+    if (!compute_c1h1_ && use_peephole_) {
+      vmulps(ymm_wp0, ymm_ct_1, ymm_wp0);
+      vaddps(ymm_i, ymm_i, ymm_wp0);
+    }
+    act<ymm_t>(ymm_i, ymm_i, act_gate_);
+    vmulps(ymm_c, ymm_c, ymm_i);
+    if (!compute_c1h1_) {
+      // act_gate(f) or act_gate(ct_1 * wp1 + f)
+      if (use_peephole_) {
+        vmulps(ymm_wp1, ymm_ct_1, ymm_wp1);
+        vaddps(ymm_f, ymm_f, ymm_wp1);
+      }
+      act<ymm_t>(ymm_f, ymm_f, act_gate_);
+      // ct
+      vmulps(ymm_f, ymm_f, ymm_ct_1);
+      vaddps(ymm_f, ymm_f, ymm_c);
+    }
+    /* H_t = act_cell(C_t) * act_gate(o) */
+    // act_cell(C_t)
+    ymm_t ymm_ct = compute_c1h1_ ? ymm_c : ymm_f;
+    ymm_t ymm_tmp = ymm_i;
+    act<ymm_t>(ymm_tmp, ymm_ct, act_cell_);
+    // act_gate(o) or act_gate(ct * wp2 + o)
+    if (use_peephole_) {
+      vmulps(ymm_wp2, ymm_ct, ymm_wp2);
+      vaddps(ymm_o, ymm_o, ymm_wp2);
+    }
+    act<ymm_t>(ymm_o, ymm_o, act_gate_);
+    // ht
+    vmulps(ymm_o, ymm_o, ymm_tmp);
+    // save ct and ht
+    vmovups(ptr[reg_ptr_ct + offset], ymm_ct);
+    vmovups(ptr[reg_ptr_ht + offset], ymm_o);
+    offset += sizeof(float) * YMM_FLOAT_BLOCK;
+  }
+
+  if (use_peephole_) {
+    postCode();
+  } else {
+    ret();
+  }
+}
+
+#define DECLARE_LSTM_CREATOR(name)                                \
+  class name##Creator : public JitCodeCreator<lstm_attr_t> {      \
+   public:                                                        \
+    /* TODO(TJ): enable more */                                   \
+    bool UseMe(const lstm_attr_t& attr) const override {          \
+      return platform::MayIUse(platform::avx) && attr.d % 8 == 0; \
+    }                                                             \
+    size_t CodeSize(const lstm_attr_t& attr) const override {     \
+      return 96 + attr.d / YMM_FLOAT_BLOCK * 90 * 4 * 8;          \
+    }                                                             \
+    std::unique_ptr<GenBase> CreateJitCode(                       \
+        const lstm_attr_t& attr) const override {                 \
+      return make_unique<name##JitCode>(attr, CodeSize(attr));    \
+    }                                                             \
+  }
+
+DECLARE_LSTM_CREATOR(LSTMCtHt);
+DECLARE_LSTM_CREATOR(LSTMC1H1);
+
+#undef DECLARE_LSTM_CREATOR
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle
+
+namespace gen = paddle::operators::jit::gen;
+
+REGISTER_JITKERNEL_GEN(kLSTMCtHt, gen::LSTMCtHtCreator);
+REGISTER_JITKERNEL_GEN(kLSTMC1H1, gen::LSTMC1H1Creator);

paddle/fluid/operators/jit/gen/lstm.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 "glog/logging.h"
+#include "paddle/fluid/operators/jit/gen/act.h"
+#include "paddle/fluid/operators/jit/gen/jitcode.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+class LSTMJitCode : public VActFunc {
+ public:
+  explicit LSTMJitCode(bool compute_c1h1, const lstm_attr_t& attr,
+                       size_t code_size, void* code_ptr = nullptr)
+      : VActFunc(code_size, code_ptr),
+        num_(attr.d),
+        compute_c1h1_(compute_c1h1),
+        use_peephole_(attr.use_peephole) {
+    auto typeExchange = [](KernelType type) -> gen::operand_type {
+      if (type == KernelType::kVSigmoid) {
+        return operand_type::SIGMOID;
+      } else if (type == KernelType::kVRelu) {
+        return operand_type::RELU;
+      } else if (type == KernelType::kVTanh) {
+        return operand_type::TANH;
+      } else if (type == KernelType::kVIdentity) {
+        return operand_type::IDENTITY;
+      } else {
+        LOG(FATAL) << "Do not support this jit::KernelType: " << type;
+      }
+      return operand_type::IDENTITY;
+    };
+    act_gate_ = typeExchange(attr.act_gate);
+    act_cand_ = typeExchange(attr.act_cand);
+    act_cell_ = typeExchange(attr.act_cell);
+
+    this->genCode();
+  }
+
+  const char* name() const override {
+    std::string base = "LSTMJitCode";
+    if (use_peephole_) {
+      base += "_Peephole";
+    }
+    if (compute_c1h1_) {
+      base += "_C1H1";
+    }
+    auto AddTypeStr = [&](operand_type type) {
+      switch (type) {
+        case operand_type::RELU:
+          base += "_Relu";
+          break;
+        case operand_type::EXP:
+          base += "_Exp";
+          break;
+        case operand_type::SIGMOID:
+          base += "_Sigmoid";
+          break;
+        case operand_type::TANH:
+          base += "_Tanh";
+          break;
+        case operand_type::IDENTITY:
+          base += "_Identity";
+          break;
+        default:
+          break;
+      }
+    };
+    AddTypeStr(act_gate_);
+    AddTypeStr(act_cand_);
+    AddTypeStr(act_cell_);
+    return base.c_str();
+  }
+  void genCode() override;
+
+ protected:
+  int num_;
+  bool compute_c1h1_;
+  bool use_peephole_;
+  operand_type act_gate_;
+  operand_type act_cand_;
+  operand_type act_cell_;
+  reg64_t param1{abi_param1};
+};
+
+#define DECLARE_LSTM_JITCODE(name, compute_c1h1)                      \
+  class name##JitCode : public LSTMJitCode {                          \
+   public:                                                            \
+    explicit name##JitCode(const lstm_attr_t& attr, size_t code_size, \
+                           void* code_ptr = nullptr)                  \
+        : LSTMJitCode(compute_c1h1, attr, code_size, code_ptr) {}     \
+  };
+
+DECLARE_LSTM_JITCODE(LSTMCtHt, false);
+DECLARE_LSTM_JITCODE(LSTMC1H1, true);
+
+#undef DECLARE_LSTM_JITCODE
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/gen_base.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License. */
+
+#include "paddle/fluid/operators/jit/gen_base.h"
+#include <fstream>
+#include <iostream>
+#include <sstream>
+
+DEFINE_bool(dump_jitcode, false, "Whether to dump the jitcode to file");
+
+namespace paddle {
+namespace operators {
+namespace jit {
+
+// refer do not need useme, it would be the last one.
+void GenBase::dumpCode(const unsigned char* code) const {
+  if (code) {
+    static int counter = 0;
+    std::ostringstream filename;
+    filename << "paddle_jitcode_" << name() << "." << counter << ".bin";
+    counter++;
+    std::ofstream fout(filename.str(), std::ios::out);
+    if (fout.is_open()) {
+      fout.write(reinterpret_cast<const char*>(code), this->getSize());
+      fout.close();
+    }
+  }
+}
+
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/gen_base.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 <gflags/gflags.h>
+#include <memory>  // for unique_ptr
+#include "paddle/fluid/operators/jit/kernel_base.h"
+
+DECLARE_bool(dump_jitcode);
+
+namespace paddle {
+namespace operators {
+namespace jit {
+
+class GenBase : public Kernel {
+ public:
+  virtual ~GenBase() = default;
+  virtual const char* name() const = 0;
+  virtual size_t getSize() const = 0;
+  virtual const unsigned char* getCodeInternal() = 0;
+  template <typename Func>
+  Func getCode() {
+    const unsigned char* code = this->getCodeInternal();
+    if (FLAGS_dump_jitcode) {
+      this->dumpCode(code);
+    }
+    // Note: failed to cast with reinterpret_cast<const Func> on Mac clang,
+    // then workaround with const_cast. Any better idea is appreciated.
+    return reinterpret_cast<Func>(const_cast<unsigned char*>(code));
+  }
+
+ protected:
+  void dumpCode(const unsigned char* code) const;
+};
+
+// Creator is used to creat the jitcode and save in pool.
+// Every JitCode should have one creator.
+class GenCreator {
+ public:
+  virtual ~GenCreator() = default;
+};
+
+template <typename Attr>
+class JitCodeCreator : public GenCreator {
+ public:
+  virtual ~JitCodeCreator() = default;
+
+  // condition when this jit code can be used.
+  virtual bool UseMe(const Attr& attr) const = 0;
+
+  // estimate this code size
+  virtual size_t CodeSize(const Attr& attr) const = 0;
+
+  // create this code
+  virtual std::unique_ptr<GenBase> CreateJitCode(const Attr& attr) const = 0;
+};
+
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/helper.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/jit/helper.h"
+#include <algorithm>  // tolower
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+
+#define ONE_CASE(key) \
+  case key:           \
+    return #key
+
+const char* to_string(KernelType kt) {
+  switch (kt) {
+    ONE_CASE(kVMul);
+    ONE_CASE(kVAdd);
+    ONE_CASE(kVAddRelu);
+    ONE_CASE(kVSub);
+    ONE_CASE(kVScal);
+    ONE_CASE(kVAddBias);
+    ONE_CASE(kVRelu);
+    ONE_CASE(kVIdentity);
+    ONE_CASE(kVExp);
+    ONE_CASE(kVSigmoid);
+    ONE_CASE(kVTanh);
+    ONE_CASE(kLSTMCtHt);
+    ONE_CASE(kLSTMC1H1);
+    ONE_CASE(kGRUH1);
+    ONE_CASE(kGRUHtPart1);
+    ONE_CASE(kGRUHtPart2);
+    ONE_CASE(kCRFDecoding);
+    ONE_CASE(kLayerNorm);
+    ONE_CASE(kNCHW16CMulNC);
+    default:
+      PADDLE_THROW("Not support type: %d, or forget to add it.", kt);
+      return "NOT JITKernel";
+  }
+  return nullptr;
+}
+#undef ONE_CASE
+
+KernelType to_kerneltype(const std::string& act) {
+  std::string lower = act;
+  std::transform(lower.begin(), lower.end(), lower.begin(), ::tolower);
+  if (lower == "relu" || lower == "vrelu") {
+    return kVRelu;
+  } else if (lower == "identity" || lower == "videntity" || lower == "") {
+    return kVIdentity;
+  } else if (lower == "exp" || lower == "vexp") {
+    return kVExp;
+  } else if (lower == "sigmoid" || lower == "vsigmoid") {
+    return kVSigmoid;
+  } else if (lower == "tanh" || lower == "vtanh") {
+    return kVTanh;
+  }
+  PADDLE_THROW("Not support type: %s, or forget to add this case", act);
+  return kNone;
+}
+
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/helper.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 <iostream>
+#include <string>
+#include <vector>
+#include "paddle/fluid/operators/jit/gen_base.h"
+#include "paddle/fluid/operators/jit/kernel_base.h"
+#include "paddle/fluid/operators/jit/kernel_key.h"
+#include "paddle/fluid/operators/jit/kernel_pool.h"
+#include "paddle/fluid/platform/place.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+
+template <KernelType KT, typename KernelTuples, typename PlaceType>
+inline typename std::enable_if<
+    std::is_same<typename KernelTuples::data_type, float>::value &&
+        std::is_same<PlaceType, platform::CPUPlace>::value,
+    typename KernelTuples::func_type>::type
+GetJitCode(const typename KernelTuples::attr_type& attr) {
+  using Func = typename KernelTuples::func_type;
+  using Attr = typename KernelTuples::attr_type;
+  size_t key = JitCodeKey<Attr>(attr);
+  auto& codes = JitCodePool<KT>().Instance();
+  if (codes.Has(key)) {
+    return codes.AllKernels().at(key)->template getCode<Func>();
+  }
+
+  // creator is not related with attr, so can use KernelKey as key
+  KernelKey kkey(KT, PlaceType());
+  // pool: (KernelKey(type, place), vector<GenCreatorPtr>)
+  auto& creator_map = JitCodeCreatorPool().Instance().AllCreators();
+  auto iter = creator_map.find(kkey);
+  if (iter != creator_map.end()) {
+    auto& creators = iter->second;
+    for (auto& cur : creators) {
+      auto i = dynamic_cast<const JitCodeCreator<Attr>*>(cur.get());
+      if (i && i->UseMe(attr)) {
+        auto p = i->CreateJitCode(attr);
+        if (p) {
+          auto f = p->template getCode<Func>();
+          codes.Insert(key, std::move(p));
+          return f;
+        }
+      }
+    }
+  }
+  return nullptr;
+}
+
+template <KernelType KT, typename KernelTuples, typename PlaceType>
+inline typename std::enable_if<
+    !std::is_same<typename KernelTuples::data_type, float>::value ||
+        !std::is_same<PlaceType, platform::CPUPlace>::value,
+    typename KernelTuples::func_type>::type
+GetJitCode(const typename KernelTuples::attr_type& attr) {
+  return nullptr;
+}
+
+// Refer code do not related with attr, which is just for cast
+// Refer is always on CPUPlace
+template <KernelType KT, typename KernelTuples>
+inline typename KernelTuples::func_type GetRefer() {
+  auto& ref_pool = ReferKernelPool().Instance().AllKernels();
+  KernelKey kkey(KT, platform::CPUPlace());
+  auto ref_iter = ref_pool.find(kkey);
+  PADDLE_ENFORCE(ref_iter != ref_pool.end(),
+                 "Every Kernel should have reference function.");
+  auto& ref_impls = ref_iter->second;
+  for (auto& impl : ref_impls) {
+    auto i = dynamic_cast<const ReferKernel<KernelTuples>*>(impl.get());
+    if (i) {
+      return i->GetFunc();
+    }
+  }
+  return nullptr;
+}
+
+template <KernelType KT, typename KernelTuples,
+          typename PlaceType = platform::CPUPlace>
+typename KernelTuples::func_type Get(
+    const typename KernelTuples::attr_type& attr) {
+  auto jitfunc = GetJitCode<KT, KernelTuples, PlaceType>(attr);
+  if (jitfunc) {
+    return jitfunc;
+  }
+
+  // pool: (KernelKey(type, place), vector<KernelPtr>)
+  KernelKey kkey(KT, PlaceType());
+  auto& pool = KernelPool().Instance().AllKernels();
+  auto iter = pool.find(kkey);
+  if (iter != pool.end()) {
+    auto& impls = iter->second;
+    for (auto& impl : impls) {
+      auto i = dynamic_cast<const KernelMore<KernelTuples>*>(impl.get());
+      if (i && i->UseMe(attr)) {
+        return i->GetFunc();
+      }
+    }
+  }
+
+  // The last implementation should be reference function on CPUPlace.
+  return GetRefer<KT, KernelTuples>();
+}
+
+const char* to_string(KernelType kt);
+
+KernelType to_kerneltype(const std::string& act);
+
+inline std::ostream& operator<<(std::ostream& os, const lstm_attr_t& attr) {
+  os << "dim_size[" << attr.d << "],act_gate[" << to_string(attr.act_gate)
+     << "],act_cand[" << to_string(attr.act_cand) << "],act_cell["
+     << to_string(attr.act_cell) << "],use_peephole["
+     << (attr.use_peephole ? "True" : "False") << "]";
+  return os;
+}
+inline std::ostream& operator<<(std::ostream& os, const gru_attr_t& attr) {
+  os << "dim_size[" << attr.d << "],act_gate[" << to_string(attr.act_gate)
+     << "],act_cand[" << to_string(attr.act_cand) << "]";
+  return os;
+}
+
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/kernel_base.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License. */
+
+#pragma once
+#include "paddle/fluid/operators/jit/macro.h"
+#include "paddle/fluid/platform/macros.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+
+typedef enum {
+  kNone = 0,
+  kVMul = 1,
+  kVAdd = 2,
+  kVAddRelu,
+  kVSub,
+  kVScal,
+  kVAddBias,
+  kVRelu,
+  kVIdentity,
+  kVExp,
+  kVSigmoid,
+  kVTanh,
+  kLSTMCtHt,
+  kLSTMC1H1,
+  kGRUH1,
+  kGRUHtPart1,
+  kGRUHtPart2,
+  kCRFDecoding,
+  kLayerNorm,
+  kNCHW16CMulNC,
+} KernelType;
+
+template <typename T>
+struct XYZNTuples {
+  typedef T data_type;
+  typedef int attr_type;
+  typedef void (*func_type)(const T*, const T*, T*, int);
+};
+
+template <typename T>
+struct AXYNTuples : public XYZNTuples<T> {};
+
+template <typename T>
+struct XYNTuples {
+  typedef T data_type;
+  typedef int attr_type;
+  typedef void (*func_type)(const T*, T*, int);
+};
+
+typedef struct {
+  void* gates;  // gates: x_ch, x_ih, x_fh, x_oh
+  const void* ct_1;
+  void* ct;
+  void* ht;
+  /* weight_peephole and checked data are only used in peephole*/
+  const void* wp{nullptr};  //  W_ic, W_fc, W_oc
+  void* checked{nullptr};   // size: 2 * d
+} lstm_t;
+
+typedef struct {
+  void* gates;  // gates: {x_update, x_reset; x_state}
+  const void* ht_1;
+  void* ht;
+} gru_t;
+
+struct rnn_attr_s {
+  int d;
+  KernelType act_gate, act_cand;
+  rnn_attr_s() = default;
+  explicit rnn_attr_s(int _d, KernelType _act_gate, KernelType _act_cand)
+      : d(_d), act_gate(_act_gate), act_cand(_act_cand) {}
+};
+
+struct lstm_attr_s : public rnn_attr_s {
+  bool use_peephole;
+  KernelType act_cell;
+  lstm_attr_s() = default;
+  explicit lstm_attr_s(int _d, KernelType _act_gate, KernelType _act_cand,
+                       KernelType _act_cell, bool _use_peephole = false)
+      : rnn_attr_s(_d, _act_gate, _act_cand),
+        use_peephole(_use_peephole),
+        act_cell(_act_cell) {}
+};
+
+typedef struct rnn_attr_s gru_attr_t;
+typedef struct lstm_attr_s lstm_attr_t;
+
+template <typename T>
+struct LSTMTuples {
+  typedef T data_type;
+  typedef lstm_attr_t attr_type;
+  typedef void (*func_type)(lstm_t*, const lstm_attr_t*);
+};
+
+template <typename T>
+struct GRUTuples {
+  typedef T data_type;
+  typedef gru_attr_t attr_type;
+  typedef void (*func_type)(gru_t*, const gru_attr_t*);
+};
+
+template <typename T>
+struct CRFDecodingTuples {
+  typedef T data_type;
+  typedef int attr_type;
+  typedef void (*func_type)(const int, const T*, const T*, T*, int*, int);
+};
+
+template <typename T>
+struct LayerNormTuples {
+  typedef T data_type;
+  typedef int attr_type;
+  typedef void (*func_type)(T*, T*, T*, T*, const T*, const T*, int,
+                            const float, int);
+};
+
+// nChw16c = nChw16c .* NC
+template <typename T>
+struct NCHW16CMulNCTuples {
+  typedef T data_type;
+  typedef int attr_type;
+  typedef void (*func_type)(const T*, const T*, T*, int, int);
+};
+
+// Just for adding to kernel pool without template
+class Kernel {
+ public:
+  Kernel() = default;
+  virtual ~Kernel() = default;
+  DISABLE_COPY_AND_ASSIGN(Kernel);
+};
+
+template <typename KernelTuples>
+class KernelMore : public Kernel {
+ public:
+  using T = typename KernelTuples::data_type;
+  using Func = typename KernelTuples::func_type;
+  using Attr = typename KernelTuples::attr_type;
+  virtual Func GetFunc() const { return func; }
+  virtual bool UseMe(const Attr& attr) const = 0;
+  virtual const char* ImplType() const = 0;
+
+ protected:
+  Func func{nullptr};
+};
+
+template <typename KernelTuples>
+class ReferKernel : public KernelMore<KernelTuples> {
+ public:
+  // Refer code can always be used
+  bool UseMe(const typename KernelTuples::attr_type& attr) const override {
+    return true;
+  }
+  const char* ImplType() const override { return "Refer"; }
+};
+
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/kernel_key.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/jit/kernel_key.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+
+template <>
+size_t JitCodeKey<int>(const int& d) {
+  return d;
+}
+
+constexpr int act_type_shift = 3;  // suppot 2^3 act types
+
+template <>
+size_t JitCodeKey<lstm_attr_t>(const lstm_attr_t& attr) {
+  size_t key = attr.d;
+  int gate_key = static_cast<int>(attr.act_gate) << 1;
+  int cand_key = static_cast<int>(attr.act_cand) << (1 + act_type_shift);
+  int cell_key = static_cast<int>(attr.act_cell) << (1 + act_type_shift * 2);
+  return (key << (1 + act_type_shift * 3)) + gate_key + cand_key + cell_key +
+         attr.use_peephole;
+}
+
+template <>
+size_t JitCodeKey<gru_attr_t>(const gru_attr_t& attr) {
+  size_t key = attr.d;
+  return (key << (act_type_shift * 2)) + static_cast<int>(attr.act_gate) +
+         (static_cast<int>(attr.act_cand) << act_type_shift);
+}
+
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/kernel_key.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License. */
+
+#pragma once
+#include "paddle/fluid/operators/jit/kernel_base.h"
+#include "paddle/fluid/platform/place.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+
+struct KernelKey {
+  struct Hash {
+    size_t operator()(const KernelKey& key) const {
+      int place = key.place_.which();               // less than 2^8
+      int type = static_cast<int>(key.type_) << 8;  // less than 2^(32-8)
+      std::hash<int> hasher;
+      return hasher(place + type);
+    }
+  };
+
+  KernelType type_;
+  platform::Place place_;
+
+  KernelKey(KernelType type, platform::Place place)
+      : type_(type), place_(place) {}
+  size_t hash_key() const { return Hash()(*this); }
+
+  bool operator==(const KernelKey& o) const {
+    return platform::places_are_same_class(place_, o.place_) &&
+           type_ == o.type_;
+  }
+  bool operator!=(const KernelKey& o) const { return !(*this == o); }
+};
+
+// Every JitCode should have a method to get the key from attribution
+template <typename Attr>
+size_t JitCodeKey(const Attr& attr);
+
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/kernel_pool.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/jit/kernel_pool.h"
+#include <memory>  // for shared_ptr
+#include <string>
+#include <unordered_map>
+
+namespace paddle {
+namespace operators {
+namespace jit {
+
+JitCodeCreatorPool& JitCodeCreatorPool::Instance() {
+  static JitCodeCreatorPool g_creator_pool;
+  return g_creator_pool;
+}
+
+KernelPool& KernelPool::Instance() {
+  static KernelPool g_kernel_pool;
+  return g_kernel_pool;
+}
+
+ReferKernelPool& ReferKernelPool::Instance() {
+  static ReferKernelPool g_refer_kernel_pool;
+  return g_refer_kernel_pool;
+}
+
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/kernel_pool.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 <memory>  // for unique_ptr
+#include <string>
+#include <unordered_map>
+#include <vector>
+#include "paddle/fluid/operators/jit/gen_base.h"
+#include "paddle/fluid/operators/jit/kernel_base.h"
+#include "paddle/fluid/operators/jit/kernel_key.h"
+#include "paddle/fluid/platform/place.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+
+template <KernelType KT>
+class JitCodePool {
+  typedef std::unique_ptr<GenBase> GenBasePtr;
+  typedef std::unordered_map<size_t, GenBasePtr> JitCodeMap;
+
+ public:
+  JitCodePool() = default;
+  static JitCodePool& Instance() {
+    static thread_local JitCodePool<KT> g_jit_codes;
+    return g_jit_codes;
+  }
+
+  const JitCodeMap& AllKernels() { return codes_; }
+
+  bool Has(size_t key) const { return codes_.find(key) != codes_.end(); }
+
+  void Insert(size_t key, GenBasePtr value) {
+    codes_.emplace(key, std::move(value));
+  }
+
+ private:
+  JitCodeMap codes_;
+  DISABLE_COPY_AND_ASSIGN(JitCodePool);
+};
+
+class JitCodeCreatorPool {
+  typedef std::unique_ptr<const GenCreator> GenCreatorPtr;
+  typedef std::unordered_map<KernelKey, std::vector<GenCreatorPtr>,
+                             KernelKey::Hash>
+      GenCreatorPtrMap;
+
+ public:
+  JitCodeCreatorPool() = default;
+  static JitCodeCreatorPool& Instance();
+  GenCreatorPtrMap& AllCreators() { return creators_; }
+  void Insert(const KernelKey& key, GenCreatorPtr value) {
+    if (creators_.find(key) == creators_.end()) {
+      creators_.emplace(key, std::vector<GenCreatorPtr>());
+    }
+    creators_.at(key).emplace_back(std::move(value));
+  }
+
+ private:
+  GenCreatorPtrMap creators_;
+  DISABLE_COPY_AND_ASSIGN(JitCodeCreatorPool);
+};
+
+typedef std::unique_ptr<const Kernel> KernelPtr;
+typedef std::unordered_map<KernelKey, std::vector<KernelPtr>, KernelKey::Hash>
+    KernelMap;
+
+class KernelPool {
+ public:
+  static KernelPool& Instance();
+  KernelPool() = default;
+  KernelMap& AllKernels() { return pool_; }
+  void Insert(const KernelKey& key, KernelPtr value) {
+    if (pool_.find(key) == pool_.end()) {
+      pool_.emplace(key, std::vector<KernelPtr>());
+    }
+    pool_.at(key).emplace_back(std::move(value));
+  }
+
+ private:
+  KernelMap pool_;
+  DISABLE_COPY_AND_ASSIGN(KernelPool);
+};
+
+// Every kernel should have refer code and it should be used in unit tests,
+// so refer kernels should have it's independent kernel pool
+class ReferKernelPool {
+ public:
+  static ReferKernelPool& Instance();
+  ReferKernelPool() = default;
+  KernelMap& AllKernels() { return pool_; }
+  void Insert(const KernelKey& key, KernelPtr value) {
+    if (pool_.find(key) == pool_.end()) {
+      pool_.emplace(key, std::vector<KernelPtr>());
+    }
+    pool_.at(key).emplace_back(std::move(value));
+  }
+
+ private:
+  KernelMap pool_;
+  DISABLE_COPY_AND_ASSIGN(ReferKernelPool);
+};
+
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/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 <type_traits>
+
+namespace paddle {
+namespace operators {
+namespace jit {
+
+#define SIGMOID_THRESHOLD_MIN -40.0
+#define SIGMOID_THRESHOLD_MAX 13.0
+#define EXP_MAX_INPUT 40.0
+
+#define XMM_FLOAT_BLOCK 4
+#define YMM_FLOAT_BLOCK 8
+#define ZMM_FLOAT_BLOCK 16
+
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/more/CMakeLists.txt

+
+function(USE_JITKERNEL_MORE TARGET TYPE)
+    file(APPEND ${jit_file} "USE_JITKERNEL_MORE(${TARGET} ${TYPE});\n")
+endfunction()
+
+if(WITH_MKLML)
+    add_subdirectory(mkl)
+endif()
+
+if(WITH_AVX)
+    add_subdirectory(intrinsic)
+endif()
+
+# mix should be last
+add_subdirectory(mix)
+
+set(JIT_KERNEL_DEPS ${JIT_KERNEL_DEPS} PARENT_SCOPE)

paddle/fluid/operators/jit/more/intrinsic/CMakeLists.txt

+
+file(GLOB jit_kernel_cc_intrinsic RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.cc")
+cc_library(jit_kernel_intrinsic SRCS ${jit_kernel_cc_intrinsic} DEPS jit_kernel_base)
+
+set(JIT_KERNEL_DEPS ${JIT_KERNEL_DEPS} jit_kernel_intrinsic PARENT_SCOPE)
+
+# use mkl kernels by name and type
+USE_JITKERNEL_MORE(kCRFDecoding, intrinsic)
+USE_JITKERNEL_MORE(kLayerNorm, intrinsic)

paddle/fluid/operators/jit/more/intrinsic/crf_decoding.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/jit/more/intrinsic/crf_decoding.h"
+#include <limits>
+#include "paddle/fluid/operators/jit/registry.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace more {
+namespace intrinsic {
+// Note: intrinsic code is not runtime build.
+// For example, if you build code on AVX, and run on AVX512 it can only use AVX
+
+void CRFDecoding(const int seq_len, const float* x, const float* w,
+                 float* alpha, int* track, int tag_num) {
+#ifdef __AVX512F__
+  const int step_size = ZMM_FLOAT_BLOCK;
+#else
+  const int step_size = YMM_FLOAT_BLOCK;
+#endif
+  const int end = tag_num / step_size;
+  const int rest = tag_num % step_size;
+  /* Setup the alpha initial value.*/
+  int i_offset = 0;
+  int last_offset = rest - step_size;
+  for (int i = 0; i <= end; ++i) {
+#ifdef __AVX512F__
+    // Declare the variable for the content of weights, input and alpha values.
+    __m512 w_content, x_content, alpha_content;
+    // Load the relevant data into the variables from un-aligned address.
+    w_content = _mm512_loadu_ps(w + i_offset);
+    x_content = _mm512_loadu_ps(x + i_offset);
+    alpha_content = _mm512_add_ps(w_content, x_content);
+    // Save the alpha value.
+    _mm512_storeu_ps(alpha_value + i_offset, alpha_content);
+#else
+    // AVX or AVX2
+    // weights, input and alpha values.
+    __m256 w_content, x_content, alpha_content;
+    // Load the relevant data into the variables from un-aligned address.
+    w_content = _mm256_loadu_ps(w + i_offset);
+    x_content = _mm256_loadu_ps(x + i_offset);
+    alpha_content = _mm256_add_ps(w_content, x_content);
+    _mm256_storeu_ps(alpha + i_offset, alpha_content);
+#endif
+    i_offset += step_size;
+    if (i == end - 1) {
+      if (rest > 0) {
+        i_offset += last_offset;
+      } else {
+        break;
+      }
+    }
+  }
+  // Use the column-major strategy to get the location of maximum score.
+  int seq_offset = 0;
+  constexpr int state_trans_base_idx = 2;
+  for (int k = 1; k < seq_len; ++k) {
+    int j_offset = 0;
+    for (int j = 0; j <= end; ++j) {
+/* Initialize the variables of maximum score and location.*/
+#ifdef __AVX512F__
+      __m512 max_score = _mm512_set1_ps(-std::numeric_limits<float>::max());
+      __m512i max_j = _mm512_setzero_si512();
+#else
+      __m256 max_score = _mm256_set1_ps(-std::numeric_limits<float>::max());
+      __m256i max_j = _mm256_set1_epi32(0);
+#endif
+      /* Calculate the offset of transition_weights.*/
+      int trans_offset = state_trans_base_idx * tag_num + j_offset;
+      for (int i = 0; i < tag_num; ++i) {
+/* Initalize the content of alpha variable with related offset.*/
+#ifdef __AVX512F__
+        __m512 alpha_content = _mm512_set1_ps(*(alpha + seq_offset + i));
+        /* Obtain the content of weights from un-aligned address.*/
+        __m512 w_content = _mm512_loadu_ps(w + trans_offset);
+        __m512 score_v = _mm512_add_ps(alpha_content, w_content);
+        __mmask16 mask = _mm512_cmp_ps_mask(score_v, max_score, _CMP_GT_OS);
+        /* AVX512 instructions.*/
+        max_j = _mm512_mask_set1_epi32(max_j, mask, i);
+        /* Update the max_score value.*/
+        max_score = _mm512_max_ps(max_score, score_v);
+
+#else
+        __m256 alpha_content = _mm256_broadcast_ss(alpha + seq_offset + i);
+        /* Obtain the content of weights from un-aligned address.*/
+        __m256 w_content = _mm256_loadu_ps(w + trans_offset);
+        __m256 score_v = _mm256_add_ps(alpha_content, w_content);
+        __m256 mask = _mm256_cmp_ps(score_v, max_score, _CMP_GT_OS);
+/* According to the mask value, update the index of the max_score.*/
+#ifdef __AVX2__
+        max_j = _mm256_or_si256(
+            _mm256_andnot_si256((__m256i)mask, max_j),
+            _mm256_and_si256((__m256i)mask, _mm256_set1_epi32(i)));
+#else
+        __m128i lo_max_j = _mm256_extractf128_si256(max_j, 0);
+        __m128i hi_max_j = _mm256_extractf128_si256(max_j, 1);
+        __m128i lo_mask =
+            _mm256_extractf128_si256(*(__m256i*)&mask, 0);  // NOLINT
+        __m128i hi_mask =
+            _mm256_extractf128_si256(*(__m256i*)&mask, 1);  // NOLINT
+        lo_max_j = _mm_andnot_si128(lo_mask, lo_max_j);
+        hi_max_j = _mm_andnot_si128(hi_mask, hi_max_j);
+        lo_mask = _mm_and_si128(lo_mask, _mm_set1_epi32(i));
+        hi_mask = _mm_and_si128(hi_mask, _mm_set1_epi32(i));
+        lo_max_j = _mm_or_si128(lo_mask, lo_max_j);
+        hi_max_j = _mm_or_si128(hi_mask, hi_max_j);
+        max_j = _mm256_insertf128_si256(max_j, lo_max_j, 0);
+        max_j = _mm256_insertf128_si256(max_j, hi_max_j, 1);
+#endif
+        /* Update the max_score value.*/
+        max_score = _mm256_max_ps(max_score, score_v);
+
+#endif
+
+        trans_offset += tag_num;
+      }
+/* Update the alpha and track values. */
+#ifdef __AVX512F__
+      __m512 x_content =
+          _mm512_loadu_ps(x + seq_offset + this->num_ + j_offset);
+      max_score = _mm512_add_ps(max_score, x_content);
+      _mm512_storeu_ps(alpha + seq_offset + this->num_ + j_offset, max_score);
+      _mm512_storeu_si512(reinterpret_cast<__m512i*>(track + seq_offset +
+                                                     this->num_ + j_offset),
+                          max_j);
+#else
+      __m256 x_content = _mm256_loadu_ps(x + seq_offset + tag_num + j_offset);
+      max_score = _mm256_add_ps(max_score, x_content);
+      _mm256_storeu_ps(alpha + seq_offset + tag_num + j_offset, max_score);
+      _mm256_storeu_si256(
+          reinterpret_cast<__m256i*>(track + seq_offset + tag_num + j_offset),
+          max_j);
+#endif
+
+      /* Calculate the offset of next step*/
+      j_offset += step_size;
+      if (j == end - 1) {
+        if (rest > 0) {
+          j_offset += last_offset;
+        } else {
+          break;
+        }
+      }
+    }
+    seq_offset += tag_num;
+  }
+}
+
+bool CRFDecodingKernel::UseMe(const int& d) const {
+#ifdef __AVX512F__
+  constexpr int block = ZMM_FLOAT_BLOCK;
+#else
+  constexpr int block = YMM_FLOAT_BLOCK;
+#endif
+  return platform::MayIUse(platform::avx) && d >= block;
+}
+
+}  // namespace intrinsic
+}  // namespace more
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle
+
+namespace intrinsic = paddle::operators::jit::more::intrinsic;
+
+REGISTER_JITKERNEL_MORE(kCRFDecoding, intrinsic, intrinsic::CRFDecodingKernel);

paddle/fluid/operators/jit/more/intrinsic/crf_decoding.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 <type_traits>
+#include "paddle/fluid/operators/jit/kernel_base.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace more {
+namespace intrinsic {
+
+void CRFDecoding(const int seq_len, const float* x, const float* w,
+                 float* alpha, int* track, int tag_num);
+
+class CRFDecodingKernel : public KernelMore<CRFDecodingTuples<float>> {
+ public:
+  CRFDecodingKernel() { this->func = CRFDecoding; }
+  bool UseMe(
+      const typename CRFDecodingTuples<float>::attr_type&) const override;
+  const char* ImplType() const override { return "Intrinsic"; }
+};
+
+}  // namespace intrinsic
+}  // namespace more
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/more/intrinsic/layer_norm.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/jit/more/intrinsic/layer_norm.h"
+#include <limits>
+#include "paddle/fluid/operators/jit/registry.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace more {
+namespace intrinsic {
+
+void LayerNorm(float* x, float* out, float* mean, float* var,
+               const float* scale, const float* bias, int height,
+               const float epsilon, int right) {
+  __m256 sum;
+  __m256 mean_vec, var_vec;
+  __m128 hi, lo;
+  __m256 tmp;
+  size_t offset;
+  size_t j;
+  int block = YMM_FLOAT_BLOCK;
+  const int rest = right % block;
+  const int end = right - rest;
+
+  __m256 reverse_num_vec =
+      _mm256_div_ps(_mm256_set1_ps(1.0), _mm256_set1_ps(right));
+  __m256 epsilon_vec = _mm256_set1_ps(epsilon);
+  int rest_mask =
+      ((-1) & (~((~0U) >> (sizeof(int) * 8 - (block - rest))))) & 0x0ff;
+  __m256i mask_vec = _mm256_set_epi32(
+      rest_mask & 0x80 ? 0xffffffff : 0, rest_mask & 0x40 ? 0xffffffff : 0,
+      rest_mask & 0x20 ? 0xffffffff : 0, rest_mask & 0x10 ? 0xffffffff : 0,
+      rest_mask & 0x8 ? 0xffffffff : 0, rest_mask & 0x4 ? 0xffffffff : 0,
+      rest_mask & 0x2 ? 0xffffffff : 0, rest_mask & 0x1 ? 0xffffffff : 0);
+
+  for (int i = 0; i < height; ++i) {
+    offset = i * right;
+
+    /* get mean */
+    sum = _mm256_setzero_ps();
+    for (j = offset; j < end + offset; j += block) {
+      sum = _mm256_add_ps(sum, _mm256_loadu_ps((const float*)x + j));
+    }
+    if (rest != 0) {
+      j = offset + right - block;
+      tmp = _mm256_loadu_ps((const float*)x + j);
+      tmp = _mm256_blendv_ps(_mm256_setzero_ps(), tmp,
+                             *(__m256*)&mask_vec);  // NOLINT
+      sum = _mm256_add_ps(sum, tmp);
+    }
+    hi = _mm256_extractf128_ps(sum, 1);
+    lo = _mm256_extractf128_ps(sum, 0);
+    sum = _mm256_add_ps(
+        sum, _mm256_insertf128_ps(
+                 _mm256_insertf128_ps(_mm256_setzero_ps(), hi, 0), lo, 1));
+    sum = _mm256_hadd_ps(sum, sum);
+    sum = _mm256_hadd_ps(sum, sum);
+    mean_vec = _mm256_mul_ps(sum, reverse_num_vec);
+    mean[i] = *reinterpret_cast<float*>(&mean_vec);
+
+    /* get variance */
+    sum = _mm256_setzero_ps();
+    for (j = offset; j < end + offset; j += block) {
+      tmp = _mm256_sub_ps(_mm256_loadu_ps((const float*)x + j), mean_vec);
+      tmp = _mm256_mul_ps(tmp, tmp);
+      sum = _mm256_add_ps(sum, tmp);
+    }
+    if (rest != 0) {
+      j = offset + right - block;
+      tmp = _mm256_sub_ps(_mm256_loadu_ps((const float*)x + j), mean_vec);
+      tmp = _mm256_mul_ps(tmp, tmp);
+      tmp = _mm256_blendv_ps(_mm256_setzero_ps(), tmp,
+                             *(__m256*)&mask_vec);  // NOLINT
+      sum = _mm256_add_ps(sum, tmp);
+    }
+    hi = _mm256_extractf128_ps(sum, 1);
+    lo = _mm256_extractf128_ps(sum, 0);
+    sum = _mm256_add_ps(
+        sum, _mm256_insertf128_ps(
+                 _mm256_insertf128_ps(_mm256_setzero_ps(), hi, 0), lo, 1));
+    sum = _mm256_hadd_ps(sum, sum);
+    sum = _mm256_hadd_ps(sum, sum);
+    var_vec = _mm256_mul_ps(sum, reverse_num_vec);
+    var[i] = *reinterpret_cast<float*>(&var_vec);
+
+    /* get x_norm and calculate output*/
+    for (j = offset; j < end + offset; j += block) {
+      tmp = _mm256_sub_ps(_mm256_loadu_ps((const float*)x + j), mean_vec);
+      tmp = _mm256_div_ps(tmp,
+                          _mm256_sqrt_ps(_mm256_add_ps(var_vec, epsilon_vec)));
+      _mm256_storeu_ps(reinterpret_cast<float*>(out) + j, tmp);
+    }
+    if (rest != 0) {
+      j = offset + right - block;
+      tmp = _mm256_sub_ps(_mm256_loadu_ps((const float*)x + j), mean_vec);
+      tmp = _mm256_div_ps(tmp,
+                          _mm256_sqrt_ps(_mm256_add_ps(var_vec, epsilon_vec)));
+      _mm256_storeu_ps(reinterpret_cast<float*>(out) + j, tmp);
+    }
+
+    if (scale) {
+      if (rest != 0) {
+        j = offset + right - block;
+        tmp = _mm256_loadu_ps((const float*)out + j);
+      }
+      for (j = offset; j < end + offset; j += block) {
+        _mm256_storeu_ps(
+            reinterpret_cast<float*>(out) + j,
+            _mm256_mul_ps(_mm256_loadu_ps((const float*)out + j),
+                          _mm256_loadu_ps((const float*)scale + j - offset)));
+      }
+      if (rest != 0) {
+        j = offset + right - block;
+        _mm256_storeu_ps(
+            reinterpret_cast<float*>(out) + j,
+            _mm256_mul_ps(tmp,
+                          _mm256_loadu_ps((const float*)scale + j - offset)));
+      }
+    }
+
+    if (bias) {
+      if (rest != 0) {
+        j = offset + right - block;
+        tmp = _mm256_loadu_ps((const float*)out + j);
+      }
+      for (j = offset; j < end + offset; j += block) {
+        _mm256_storeu_ps(
+            reinterpret_cast<float*>(out) + j,
+            _mm256_add_ps(_mm256_loadu_ps((const float*)out + j),
+                          _mm256_loadu_ps((const float*)bias + j - offset)));
+      }
+      if (rest != 0) {
+        j = offset + right - block;
+        _mm256_storeu_ps(reinterpret_cast<float*>(out) + j,
+                         _mm256_add_ps(tmp, _mm256_loadu_ps((const float*)bias +
+                                                            j - offset)));
+      }
+    }
+  }
+}
+
+bool LayerNormKernel::UseMe(const int& d) const {
+  return platform::MayIUse(platform::avx) && d >= YMM_FLOAT_BLOCK;
+}
+
+}  // namespace intrinsic
+}  // namespace more
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle
+
+namespace intrinsic = paddle::operators::jit::more::intrinsic;
+
+REGISTER_JITKERNEL_MORE(kLayerNorm, intrinsic, intrinsic::LayerNormKernel);

paddle/fluid/operators/jit/more/intrinsic/layer_norm.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 <type_traits>
+#include "paddle/fluid/operators/jit/kernel_base.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace more {
+namespace intrinsic {
+
+void LayerNorm(float* x, float* out, float* mean, float* var,
+               const float* scale, const float* bias, int height,
+               const float epsilon, int right);
+
+class LayerNormKernel : public KernelMore<LayerNormTuples<float>> {
+ public:
+  LayerNormKernel() { this->func = LayerNorm; }
+  bool UseMe(const typename LayerNormTuples<float>::attr_type&) const override;
+  const char* ImplType() const override { return "Intrinsic"; }
+};
+
+}  // namespace intrinsic
+}  // namespace more
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/more/mix/CMakeLists.txt

+
+
+file(GLOB jit_kernel_mix_cc RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.cc")
+cc_library(jit_kernel_mix SRCS ${jit_kernel_mix_cc} DEPS jit_kernel_base)
+
+set(JIT_KERNEL_DEPS ${JIT_KERNEL_DEPS} jit_kernel_mix PARENT_SCOPE)
+
+USE_JITKERNEL_MORE(kVSigmoid, mix)
+USE_JITKERNEL_MORE(kVTanh, mix)
+USE_JITKERNEL_MORE(kLSTMCtHt, mix)
+USE_JITKERNEL_MORE(kLSTMC1H1, mix)
+USE_JITKERNEL_MORE(kGRUH1, mix)
+USE_JITKERNEL_MORE(kGRUHtPart1, mix)
+USE_JITKERNEL_MORE(kGRUHtPart2, mix)

paddle/fluid/operators/jit/more/mix/mix.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/jit/more/mix/mix.h"
+#include "paddle/fluid/operators/jit/kernels.h"
+#include "paddle/fluid/operators/jit/registry.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace more {
+namespace mix {
+
+void VSigmoid(const T* x, T* y, int n) {
+  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] = static_cast<T>(0) - y[i];
+  }
+  auto compute = Get<KernelType::kVExp, XYNTuples<T>, platform::CPUPlace>(n);
+  compute(y, y, n);
+  for (int i = 0; i < n; ++i) {
+    y[i] = static_cast<T>(1) / (static_cast<T>(1) + y[i]);
+  }
+}
+
+void VTanh(const T* x, T* y, int n) {
+  const T a = 2, b = -1;
+  auto compute_scal = Get<kVScal, AXYNTuples<T>, platform::CPUPlace>(n);
+  auto compute_addbias = Get<kVAddBias, AXYNTuples<T>, platform::CPUPlace>(n);
+  auto compute_sigmoid = Get<kVSigmoid, XYNTuples<T>, platform::CPUPlace>(n);
+  compute_scal(&a, x, y, n);
+  compute_sigmoid(y, y, n);
+  compute_scal(&a, y, y, n);
+  compute_addbias(&b, y, y, n);
+}
+
+void (*getActFunc(KernelType type, int d))(const T*, T*, int) {  // NOLINT
+  if (type == kVSigmoid) {
+    return Get<kVSigmoid, XYNTuples<T>, platform::CPUPlace>(d);
+  } else if (type == kVRelu) {
+    return Get<kVRelu, XYNTuples<T>, platform::CPUPlace>(d);
+  } else if (type == kVTanh) {
+    return Get<kVTanh, XYNTuples<T>, platform::CPUPlace>(d);
+  } else if (type == kVIdentity) {
+    return Get<kVIdentity, XYNTuples<T>, platform::CPUPlace>(d);
+  }
+  PADDLE_THROW("Not support type: %s", type);
+  return nullptr;
+}
+
+void LSTMCtHt(lstm_t* step, const lstm_attr_t* attr) {
+  T* gates = reinterpret_cast<T*>(step->gates);
+  const T* ct_1 = reinterpret_cast<const T*>(step->ct_1);
+  T* ct = reinterpret_cast<T*>(step->ct);
+  T* ht = reinterpret_cast<T*>(step->ht);
+  const T* wp = reinterpret_cast<const T*>(step->wp);
+  T* checked = reinterpret_cast<T*>(step->checked);
+  const int d = attr->d;
+  const int d2 = d * 2;
+  const int d3 = d * 3;
+  auto vmul_d = Get<kVMul, XYZNTuples<T>, platform::CPUPlace>(d);
+  auto vadd_d = Get<kVAdd, XYZNTuples<T>, platform::CPUPlace>(d);
+  auto vadd_d2 = Get<kVAdd, XYZNTuples<T>, platform::CPUPlace>(d2);
+  auto act_gate_d = getActFunc(attr->act_gate, d);
+  auto act_gate_d2 = getActFunc(attr->act_gate, d2);
+  auto act_gate_d3 = getActFunc(attr->act_gate, d3);
+  auto act_cand_d = getActFunc(attr->act_cand, d);
+  auto act_cell_d = getActFunc(attr->act_cell, d);
+
+  if (attr->use_peephole) {
+    vmul_d(wp, ct_1, checked, d);
+    vmul_d(wp + d, ct_1, checked + d, d);
+    vadd_d2(checked, gates + d, gates + d, d2);
+    act_gate_d2(gates + d, gates + d, d2);
+  } else {
+    act_gate_d3(gates + d, gates + d, d3);
+  }
+
+  // C_t = C_t-1 * fgated + cand_gated * igated
+  act_cand_d(gates, gates, d);
+  vmul_d(gates, gates + d, gates + d, d);
+  vmul_d(ct_1, gates + d2, gates + d2, d);
+  vadd_d(gates + d, gates + d2, ct, d);
+
+  if (attr->use_peephole) {
+    // get ogated
+    vmul_d(wp + d2, ct, gates + d, d);
+    vadd_d(gates + d, gates + d3, gates + d3, d);
+    act_gate_d(gates + d3, gates + d3, d);
+  }
+  // H_t = act_cell(C_t) * ogated
+  act_cell_d(ct, gates + d2, d);
+  vmul_d(gates + d2, gates + d3, ht, d);
+}
+
+void LSTMC1H1(lstm_t* step, const lstm_attr_t* attr) {
+  T* gates = reinterpret_cast<T*>(step->gates);
+  T* ct = reinterpret_cast<T*>(step->ct);
+  T* ht = reinterpret_cast<T*>(step->ht);
+  int d = attr->d;
+  int d2 = d * 2;
+  int d3 = d * 3;
+  auto vmul_d = Get<kVMul, XYZNTuples<T>, platform::CPUPlace>(d);
+  auto vadd_d = Get<kVAdd, XYZNTuples<T>, platform::CPUPlace>(d);
+  auto act_gate_d = getActFunc(attr->act_gate, d);
+  auto act_cand_d = getActFunc(attr->act_cand, d);
+  auto act_cell_d = getActFunc(attr->act_cell, d);
+  /* C_t = igated * cgated*/
+  act_gate_d(gates + d, gates + d, d);
+  act_cand_d(gates, gates, d);
+  vmul_d(gates, gates + d, ct, d);
+  if (attr->use_peephole) {
+    // get outgated, put W_oc * C_t on igated
+    const T* wp = reinterpret_cast<const T*>(step->wp);
+    vmul_d(wp + d2, ct, gates + d, d);
+    vadd_d(gates + d, gates + d3, gates + d3, d);
+  }
+  /* H_t = act_cell(C_t) * ogated */
+  act_gate_d(gates + d3, gates + d3, d);
+  act_cell_d(ct, gates + d2, d);
+  vmul_d(gates + d2, gates + d3, ht, d);
+}
+
+// compute h1 without h0
+void GRUH1(gru_t* step, const gru_attr_t* attr) {
+  T* gates = reinterpret_cast<T*>(step->gates);
+  T* ht = reinterpret_cast<T*>(step->ht);
+  int d = attr->d;
+  int d2 = d * 2;
+  auto act_gate = getActFunc(attr->act_gate, d);
+  auto act_cand = getActFunc(attr->act_cand, d);
+  auto vmul_d = Get<kVMul, XYZNTuples<T>, platform::CPUPlace>(d);
+  act_gate(gates, gates, d);
+  act_cand(gates + d2, gates + d2, d);
+  vmul_d(gates, gates + d2, ht, d);
+}
+
+// compute the first part of GRU: ht = act_gate(r) * ht_1
+void GRUHtPart1(gru_t* step, const gru_attr_t* attr) {
+  // W: {W_update, W_reset; W_state}
+  T* gates = reinterpret_cast<T*>(step->gates);
+  T* ht = reinterpret_cast<T*>(step->ht);
+  const T* ht_1 = reinterpret_cast<const T*>(step->ht_1);
+  auto act_gate = getActFunc(attr->act_gate, attr->d);
+  auto vmul_d = Get<kVMul, XYZNTuples<T>, platform::CPUPlace>(attr->d);
+  act_gate(gates + attr->d, gates + attr->d, attr->d);
+  vmul_d(ht_1, gates + attr->d, ht, attr->d);
+}
+
+// compute the second part of GRU:
+// ht = act_gate(u) * act_cand(s) + (1-act_gate(u)) * ht_1
+void GRUHtPart2(gru_t* step, const gru_attr_t* attr) {
+  T* gates = reinterpret_cast<T*>(step->gates);
+  T* ht = reinterpret_cast<T*>(step->ht);
+  const T* ht_1 = reinterpret_cast<const T*>(step->ht_1);
+  int d = attr->d;
+  auto act_gate = getActFunc(attr->act_gate, d);
+  auto act_cand = getActFunc(attr->act_cand, d);
+  T* y = gates + d * 2;
+  act_gate(gates, gates, d);
+  act_cand(y, y, d);
+  // out = zt*ht~ + (1-zt)*ht_1
+  for (int i = 0; i < d; ++i) {
+    ht[i] = gates[i] * y[i] + (static_cast<T>(1) - gates[i]) * ht_1[i];
+  }
+}
+
+// TODO(TJ): tuning me
+bool VSigmoidKernel::UseMe(const int& d) const { return true; }
+
+bool VTanhKernel::UseMe(const int& d) const { return true; }
+
+bool LSTMCtHtKernel::UseMe(const lstm_attr_t& attr) const { return true; }
+
+bool LSTMC1H1Kernel::UseMe(const lstm_attr_t& attr) const { return true; }
+
+bool GRUH1Kernel::UseMe(const gru_attr_t& attr) const { return true; }
+
+bool GRUHtPart1Kernel::UseMe(const gru_attr_t& attr) const { return true; }
+
+bool GRUHtPart2Kernel::UseMe(const gru_attr_t& attr) const { return true; }
+
+}  // namespace mix
+}  // namespace more
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle
+
+namespace mix = paddle::operators::jit::more::mix;
+
+#define REGISTER_MORE_KERNEL(key, func) \
+  REGISTER_JITKERNEL_MORE(key, mix, mix::func##Kernel)
+
+REGISTER_MORE_KERNEL(kVSigmoid, VSigmoid);
+REGISTER_MORE_KERNEL(kVTanh, VTanh);
+REGISTER_MORE_KERNEL(kLSTMCtHt, LSTMCtHt);
+REGISTER_MORE_KERNEL(kLSTMC1H1, LSTMC1H1);
+REGISTER_MORE_KERNEL(kGRUH1, GRUH1);
+REGISTER_MORE_KERNEL(kGRUHtPart1, GRUHtPart1);
+REGISTER_MORE_KERNEL(kGRUHtPart2, GRUHtPart2);
+
+#undef REGISTER_MORE_KERNEL

paddle/fluid/operators/jit/more/mix/mix.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 <type_traits>
+#include "paddle/fluid/operators/jit/kernel_base.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace more {
+namespace mix {
+using T = float;
+
+void VSigmoid(const T* x, T* y, int n);
+void VTanh(const T* x, T* y, int n);
+
+void LSTMCtHt(lstm_t* step, const lstm_attr_t* attr);
+void LSTMC1H1(lstm_t* step, const lstm_attr_t* attr);
+void GRUH1(gru_t* step, const gru_attr_t* attr);
+void GRUHtPart1(gru_t* step, const gru_attr_t* attr);
+void GRUHtPart2(gru_t* step, const gru_attr_t* attr);
+
+#define DECLARE_MORE_KERNEL(name, tuples)                            \
+  class name##Kernel : public KernelMore<tuples<T>> {                \
+   public:                                                           \
+    name##Kernel() { this->func = name; }                            \
+    bool UseMe(const typename tuples<T>::attr_type&) const override; \
+    const char* ImplType() const override { return "Mixed"; }        \
+  }
+
+// XYN
+DECLARE_MORE_KERNEL(VSigmoid, XYNTuples);
+DECLARE_MORE_KERNEL(VTanh, XYNTuples);
+
+DECLARE_MORE_KERNEL(LSTMCtHt, LSTMTuples);
+DECLARE_MORE_KERNEL(LSTMC1H1, LSTMTuples);
+
+DECLARE_MORE_KERNEL(GRUH1, GRUTuples);
+DECLARE_MORE_KERNEL(GRUHtPart1, GRUTuples);
+DECLARE_MORE_KERNEL(GRUHtPart2, GRUTuples);
+
+#undef DECLARE_MORE_KERNEL
+
+}  // namespace mix
+}  // namespace more
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/more/mkl/CMakeLists.txt

+
+cc_library(jit_kernel_mkl SRCS mkl.cc DEPS jit_kernel_base dynload_mklml)
+set(JIT_KERNEL_DEPS ${JIT_KERNEL_DEPS} dynload_mklml jit_kernel_mkl PARENT_SCOPE)
+
+# use mkl kernels by name and type
+USE_JITKERNEL_MORE(kVMul, mkl)
+USE_JITKERNEL_MORE(kVAdd, mkl)
+USE_JITKERNEL_MORE(kVScal, mkl)
+USE_JITKERNEL_MORE(kVExp, mkl)
+USE_JITKERNEL_MORE(kVSigmoid, mkl)
+USE_JITKERNEL_MORE(kVTanh, mkl)

paddle/fluid/operators/jit/more/mkl/mkl.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/jit/more/mkl/mkl.h"
+#include "paddle/fluid/operators/jit/refer/refer.h"
+#include "paddle/fluid/operators/jit/registry.h"
+#include "paddle/fluid/platform/cpu_info.h"
+#include "paddle/fluid/platform/dynload/mklml.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace more {
+namespace mkl {
+
+template <>
+void VMul<float>(const float* x, const float* y, float* z, int n) {
+  platform::dynload::vsMul(n, x, y, z);
+}
+
+template <>
+void VMul<double>(const double* x, const double* y, double* z, int n) {
+  platform::dynload::vdMul(n, x, y, z);
+}
+
+template <>
+void VAdd<float>(const float* x, const float* y, float* z, int n) {
+  platform::dynload::vsAdd(n, x, y, z);
+}
+
+template <>
+void VAdd<double>(const double* x, const double* y, double* z, int n) {
+  platform::dynload::vdAdd(n, x, y, z);
+}
+
+template <>
+void VScal<float>(const float* a, const float* x, float* y, int n) {
+  if (x == y) {
+    platform::dynload::cblas_sscal(n, *a, y, 1);
+  } else {
+    refer::VScal<float>(a, x, y, n);
+  }
+}
+
+template <>
+void VScal<double>(const double* a, const double* x, double* y, int n) {
+  if (x == y) {
+    platform::dynload::cblas_dscal(n, *a, y, 1);
+  } else {
+    refer::VScal<double>(a, x, y, n);
+  }
+}
+
+template <>
+void VExp<float>(const float* x, float* y, int n) {
+  platform::dynload::vsExp(n, x, y);
+}
+
+template <>
+void VExp<double>(const double* x, double* y, int n) {
+  platform::dynload::vdExp(n, x, y);
+}
+
+// TODO(TJ): tuning me carefully on AVX, AVX2 and AVX512
+template <>
+bool VMulKernel<float>::UseMe(const int& d) const {
+  return platform::MayIUse(platform::avx512f) && d > 512;
+}
+
+template <>
+bool VAddKernel<float>::UseMe(const int& d) const {
+  return platform::MayIUse(platform::avx512f) && d > 512;
+}
+
+template <>
+bool VScalKernel<float>::UseMe(const int& d) const {
+  return platform::MayIUse(platform::avx512f) && d > 512;
+}
+
+template <>
+bool VExpKernel<float>::UseMe(const int& d) const {
+  return d > 7;
+}
+
+template <>
+bool VSigmoidKernel<float>::UseMe(const int& d) const {
+  return d > 7;
+}
+
+template <>
+bool VTanhKernel<float>::UseMe(const int& d) const {
+  return d > 7;
+}
+
+#define AWALYS_USE_ME_WITH_DOUBLE(func)                  \
+  template <>                                            \
+  bool func##Kernel<double>::UseMe(const int& d) const { \
+    return true;                                         \
+  }
+
+AWALYS_USE_ME_WITH_DOUBLE(VMul);
+AWALYS_USE_ME_WITH_DOUBLE(VAdd);
+AWALYS_USE_ME_WITH_DOUBLE(VScal);
+AWALYS_USE_ME_WITH_DOUBLE(VExp);
+AWALYS_USE_ME_WITH_DOUBLE(VSigmoid);
+AWALYS_USE_ME_WITH_DOUBLE(VTanh);
+
+#undef AWALYS_USE_ME_WITH_DOUBLE
+}  // namespace mkl
+}  // namespace more
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle
+
+namespace mkl = paddle::operators::jit::more::mkl;
+
+#define REGISTER_MKL_KERNEL(key, func)                        \
+  REGISTER_JITKERNEL_MORE(key, mkl, mkl::func##Kernel<float>, \
+                          mkl::func##Kernel<double>)
+
+REGISTER_MKL_KERNEL(kVMul, VMul);
+REGISTER_MKL_KERNEL(kVAdd, VAdd);
+REGISTER_MKL_KERNEL(kVScal, VScal);
+REGISTER_MKL_KERNEL(kVExp, VExp);
+REGISTER_MKL_KERNEL(kVSigmoid, VSigmoid);
+REGISTER_MKL_KERNEL(kVTanh, VTanh);
+
+#undef REGISTER_MKL_KERNEL

paddle/fluid/operators/jit/more/mkl/mkl.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 <type_traits>
+#include "paddle/fluid/operators/jit/kernel_base.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace more {
+namespace mkl {
+
+template <typename T>
+void VMul(const T* x, const T* y, T* z, int n);
+
+template <typename T>
+void VAdd(const T* x, const T* y, T* z, int n);
+
+template <typename T>
+void VScal(const T* a, const T* x, T* y, int n);
+
+template <typename T>
+void VExp(const T* x, T* y, int n);
+
+template <typename T>
+void VSigmoid(const T* x, T* y, int n) {
+  const T min = SIGMOID_THRESHOLD_MIN;
+  const T 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] = static_cast<T>(0) - y[i];
+  }
+  VExp(y, y, n);
+  for (int i = 0; i < n; ++i) {
+    y[i] = static_cast<T>(1) / (static_cast<T>(1) + y[i]);
+  }
+}
+
+template <typename T>
+void VTanh(const T* x, T* y, int n) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = static_cast<T>(2) * x[i];
+  }
+  VSigmoid(y, y, n);
+  for (int i = 0; i < n; ++i) {
+    y[i] = static_cast<T>(2) * y[i] - static_cast<T>(1);
+  }
+}
+
+#define DECLARE_MKL_KERNEL(name, tuples)                             \
+  template <typename T>                                              \
+  class name##Kernel : public KernelMore<tuples<T>> {                \
+   public:                                                           \
+    name##Kernel() { this->func = name<T>; }                         \
+    bool UseMe(const typename tuples<T>::attr_type&) const override; \
+    const char* ImplType() const override { return "MKL"; }          \
+  }
+
+// XYZN
+DECLARE_MKL_KERNEL(VMul, XYZNTuples);
+DECLARE_MKL_KERNEL(VAdd, XYZNTuples);
+
+// AXYN
+DECLARE_MKL_KERNEL(VScal, AXYNTuples);
+
+// XYN
+DECLARE_MKL_KERNEL(VExp, XYNTuples);
+DECLARE_MKL_KERNEL(VSigmoid, XYNTuples);
+DECLARE_MKL_KERNEL(VTanh, XYNTuples);
+
+#undef DECLARE_MKL_KERNEL
+
+}  // namespace mkl
+}  // namespace more
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/refer/CMakeLists.txt

+
+cc_library(jit_kernel_refer SRCS refer.cc DEPS jit_kernel_base)
+set(JIT_KERNEL_DEPS ${JIT_KERNEL_DEPS} jit_kernel_refer PARENT_SCOPE)
+
+function(USE_JITKERNEL_REFER TARGET)
+    file(APPEND ${jit_file} "USE_JITKERNEL_REFER(${TARGET});\n")
+endfunction()
+
+# use refer kernel by name
+USE_JITKERNEL_REFER(kVMul)
+USE_JITKERNEL_REFER(kVAdd)
+USE_JITKERNEL_REFER(kVAddRelu)
+USE_JITKERNEL_REFER(kVSub)
+USE_JITKERNEL_REFER(kVScal)
+USE_JITKERNEL_REFER(kVAddBias)
+USE_JITKERNEL_REFER(kVRelu)
+USE_JITKERNEL_REFER(kVIdentity)
+USE_JITKERNEL_REFER(kVExp)
+USE_JITKERNEL_REFER(kVSigmoid)
+USE_JITKERNEL_REFER(kVTanh)
+USE_JITKERNEL_REFER(kLSTMCtHt)
+USE_JITKERNEL_REFER(kLSTMC1H1)
+USE_JITKERNEL_REFER(kGRUH1)
+USE_JITKERNEL_REFER(kGRUHtPart1)
+USE_JITKERNEL_REFER(kGRUHtPart2)
+USE_JITKERNEL_REFER(kCRFDecoding)
+USE_JITKERNEL_REFER(kLayerNorm)
+USE_JITKERNEL_REFER(kNCHW16CMulNC)

paddle/fluid/operators/jit/refer/refer.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/jit/refer/refer.h"
+#include "paddle/fluid/operators/jit/registry.h"
+
+namespace refer = paddle::operators::jit::refer;
+
+#define REGISTER_REFER_KERNEL(key, func)                    \
+  REGISTER_JITKERNEL_REFER(key, refer::func##Kernel<float>, \
+                           refer::func##Kernel<double>)
+
+REGISTER_REFER_KERNEL(kVMul, VMul);
+REGISTER_REFER_KERNEL(kVAdd, VAdd);
+REGISTER_REFER_KERNEL(kVAddRelu, VAddRelu);
+REGISTER_REFER_KERNEL(kVSub, VSub);
+
+REGISTER_REFER_KERNEL(kVScal, VScal);
+REGISTER_REFER_KERNEL(kVAddBias, VAddBias);
+
+REGISTER_REFER_KERNEL(kVRelu, VRelu);
+REGISTER_REFER_KERNEL(kVIdentity, VIdentity);
+REGISTER_REFER_KERNEL(kVExp, VExp);
+REGISTER_REFER_KERNEL(kVSigmoid, VSigmoid);
+REGISTER_REFER_KERNEL(kVTanh, VTanh);
+
+REGISTER_REFER_KERNEL(kLSTMCtHt, LSTMCtHt);
+REGISTER_REFER_KERNEL(kLSTMC1H1, LSTMC1H1);
+
+REGISTER_REFER_KERNEL(kGRUH1, GRUH1);
+REGISTER_REFER_KERNEL(kGRUHtPart1, GRUHtPart1);
+REGISTER_REFER_KERNEL(kGRUHtPart2, GRUHtPart2);
+
+REGISTER_REFER_KERNEL(kCRFDecoding, CRFDecoding);
+REGISTER_REFER_KERNEL(kLayerNorm, LayerNorm);
+
+REGISTER_REFER_KERNEL(kNCHW16CMulNC, NCHW16CMulNC);
+
+#undef REGISTER_REFER_KERNEL

paddle/fluid/operators/math/jit_kernel_refer.h → paddle/fluid/operators/jit/refer/refer.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. */
+ *
+ * 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 <cmath>
-#include <string>
-#include "paddle/fluid/operators/math/jit_kernel_impl.h"
+#include <limits>
+#include "paddle/fluid/operators/jit/helper.h"
+#include "paddle/fluid/operators/jit/kernel_base.h"
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {
 namespace operators {
-namespace math {
-namespace jitkernel {
+namespace jit {
 namespace refer {
-/* Refer code only focus on correctness */
 
+// Refer code only focus on correctness
 template <typename T>
 void VMul(const T* x, const T* y, T* z, int n) {
   for (int i = 0; i < n; ++i) {
@@ -47,6 +48,13 @@ void VAddRelu(const T* x, const T* y, T* z, int n) {
   }
 }
 
+template <typename T>
+void VSub(const T* x, const T* y, T* z, int n) {
+  for (int i = 0; i < n; ++i) {
+    z[i] = x[i] - y[i];
+  }
+}
+
 template <typename T>
 void VScal(const T* a, const T* x, T* y, int n) {
   for (int i = 0; i < n; ++i) {
@@ -69,7 +77,11 @@ void VRelu(const T* x, T* y, int n) {
 }
 
 template <typename T>
-inline void VIdentity(const T* x, T* y, int n) {}
+inline void VIdentity(const T* x, T* y, int n) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i];
+  }
+}
 
 template <typename T>
 void VExp(const T* x, T* y, int n) {
@@ -102,20 +114,22 @@ void VTanh(const T* x, T* y, int n) {
 }
 
 template <typename T>
-void (*getActFunc(const std::string& type))(const T*, T*, int) {  // NOLINT
-  if (type == "sigmoid") {
+void (*getActFunc(KernelType type))(const T*, T*, int) {  // NOLINT
+  if (type == kVSigmoid) {
     return VSigmoid<T>;
-  } else if (type == "relu") {
+  } else if (type == kVRelu) {
     return VRelu<T>;
-  } else if (type == "tanh") {
+  } else if (type == kVTanh) {
     return VTanh<T>;
-  } else if (type == "identity" || type == "") {
+  } else if (type == kVIdentity) {
     return VIdentity<T>;
   }
   PADDLE_THROW("Not support type: %s", type);
   return nullptr;
 }
 
+// TODO(TJ): add refer gemm and make LSTM kernels combine as same GRU kernels
+
 // compute ct and ht
 template <typename T>
 void LSTMCtHt(lstm_t* step, const lstm_attr_t* attr) {
@@ -231,8 +245,134 @@ void GRUHtPart2(gru_t* step, const gru_attr_t* attr) {
   }
 }
 
+template <typename T>
+void CRFDecoding(const int seq_len, const T* x, const T* w, T* alpha,
+                 int* track, int right) {
+  constexpr int state_trans_base_idx = 2;
+  for (int i = 0; i < right; ++i) {
+    alpha[i] = w[i] + x[i];
+  }
+  for (int k = 1; k < seq_len; ++k) {
+    for (int i = 0; i < right; ++i) {
+      T max_score = -std::numeric_limits<T>::max();
+      int max_j = 0;
+      for (int j = 0; j < right; ++j) {
+        T score = alpha[(k - 1) * right + j] +
+                  w[(j + state_trans_base_idx) * right + i];
+        if (score > max_score) {
+          max_score = score;
+          max_j = j;
+        }
+      }
+      alpha[k * right + i] = max_score + x[k * right + i];
+      track[k * right + i] = max_j;
+    }
+  }
+}
+
+template <typename T>
+void LayerNorm(T* x, T* out, T* mean, T* var, const T* scale, const T* bias,
+               int height, const float epsilon, int right) {
+  // get mean
+  for (int i = 0; i < height; i++) {
+    T sum = 0.0;
+    int offset = i * right;
+    for (int j = 0; j < right; j++) {
+      sum += x[offset + j];
+    }
+    mean[i] = sum / right;
+  }
+
+  // get variance
+  for (int i = 0; i < height; i++) {
+    T sum = 0.0;
+    int offset = i * right;
+    for (int j = 0; j < right; j++) {
+      sum += (x[offset + j] - mean[i]) * (x[offset + j] - mean[i]);
+    }
+    var[i] = sum / right;
+  }
+
+  for (int i = 0; i < height; i++) {
+    int offset = i * right;
+    T sqrt_var = std::sqrt(var[i] + (T)epsilon);
+    for (int j = 0; j < right; j++) {
+      out[offset + j] = (x[offset + j] - mean[i]) / sqrt_var;
+    }
+  }
+  if (scale) {
+    for (int i = 0; i < height; i++) {
+      int offset = i * right;
+      for (int j = 0; j < right; j++) {
+        out[offset + j] *= scale[j];
+      }
+    }
+  }
+
+  if (bias) {
+    for (int i = 0; i < height; i++) {
+      int offset = i * right;
+      for (int j = 0; j < right; j++) {
+        out[offset + j] += bias[j];
+      }
+    }
+  }
+}
+
+template <typename T>
+void NCHW16CMulNC(const T* x, const T* y, T* z, int height, int width) {
+  int offset = 0;
+  for (int h = 0; h < height; ++h) {
+    for (int w = 0; w < width; ++w) {
+      for (int i = 0; i < 16; ++i) {
+        z[i + offset] = y[i] * x[i + offset];
+      }
+      offset += ZMM_FLOAT_BLOCK;
+    }
+  }
+}
+
+#define DECLARE_REFER_KERNEL(name, tuples)             \
+  template <typename T>                                \
+  class name##Kernel : public ReferKernel<tuples<T>> { \
+   public:                                             \
+    name##Kernel() { this->func = name<T>; }           \
+  }
+
+// const T* x, const T* y, T* z, int n
+DECLARE_REFER_KERNEL(VMul, XYZNTuples);
+DECLARE_REFER_KERNEL(VAdd, XYZNTuples);
+DECLARE_REFER_KERNEL(VAddRelu, XYZNTuples);
+DECLARE_REFER_KERNEL(VSub, XYZNTuples);
+
+// const T* a, const T* x, T* y, int n
+DECLARE_REFER_KERNEL(VScal, AXYNTuples);
+DECLARE_REFER_KERNEL(VAddBias, AXYNTuples);
+
+// const T* x, T* y, int n
+DECLARE_REFER_KERNEL(VRelu, XYNTuples);
+DECLARE_REFER_KERNEL(VIdentity, XYNTuples);
+DECLARE_REFER_KERNEL(VExp, XYNTuples);
+DECLARE_REFER_KERNEL(VSigmoid, XYNTuples);
+DECLARE_REFER_KERNEL(VTanh, XYNTuples);
+
+// lstm_t*, const lstm_attr_t*
+DECLARE_REFER_KERNEL(LSTMCtHt, LSTMTuples);
+DECLARE_REFER_KERNEL(LSTMC1H1, LSTMTuples);
+
+// gru_t*, const gru_attr_t*
+DECLARE_REFER_KERNEL(GRUH1, GRUTuples);
+DECLARE_REFER_KERNEL(GRUHtPart1, GRUTuples);
+DECLARE_REFER_KERNEL(GRUHtPart2, GRUTuples);
+
+DECLARE_REFER_KERNEL(CRFDecoding, CRFDecodingTuples);
+DECLARE_REFER_KERNEL(LayerNorm, LayerNormTuples);
+
+DECLARE_REFER_KERNEL(NCHW16CMulNC, NCHW16CMulNCTuples);
+
+#undef DECLARE_REFER_KERNEL
+
 }  // namespace refer
-}  // namespace jitkernel
-}  // namespace math
+}  // namespace jit
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/jit/registry.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 <memory>
+#include <tuple>
+#include <type_traits>
+#include "paddle/fluid/operators/jit/kernel_base.h"
+#include "paddle/fluid/operators/jit/kernel_pool.h"
+#include "paddle/fluid/platform/place.h"
+#include "paddle/fluid/platform/variant.h"  // for UNUSED
+
+namespace paddle {
+namespace operators {
+namespace jit {
+
+// make_unique is supported since c++14
+template <typename T, typename... Args>
+inline std::unique_ptr<T> make_unique(Args&&... args) {
+  static_assert(!std::is_array<T>::value, "T must not be array");
+  return std::unique_ptr<T>(new T(std::forward<Args>(args)...));
+}
+
+template <typename Pool, typename PlaceType, bool IsEnd, size_t I,
+          typename... KernelImpls>
+struct JitKernelRegistrarFunctor;
+
+template <typename Pool, typename PlaceType, size_t I, typename... KernelImpls>
+struct JitKernelRegistrarFunctor<Pool, PlaceType, true, I, KernelImpls...> {
+  void operator()(KernelType kt) const {}
+};
+
+template <typename Pool, typename PlaceType, size_t I, typename... KernelImpls>
+struct JitKernelRegistrarFunctor<Pool, PlaceType, false, I, KernelImpls...> {
+  using KERNEL_IMPL_TYPE =
+      typename std::tuple_element<I, std::tuple<KernelImpls...>>::type;
+
+  void operator()(KernelType kt) const {
+    KernelKey kkey(kt, PlaceType());
+    Pool().Instance().Insert(kkey,
+                             std::move(make_unique<const KERNEL_IMPL_TYPE>()));
+    constexpr auto size = std::tuple_size<std::tuple<KernelImpls...>>::value;
+    JitKernelRegistrarFunctor<Pool, PlaceType, I + 1 == size, I + 1,
+                              KernelImpls...>
+        func;
+    func(kt);
+  }
+};
+
+template <typename Pool, typename PlaceType, typename... KernelImpls>
+class JitKernelRegistrar {
+ public:
+  explicit JitKernelRegistrar(KernelType kt) {
+    JitKernelRegistrarFunctor<Pool, PlaceType, false, 0, KernelImpls...> func;
+    func(kt);
+  }
+  void Touch() {}
+};
+
+#define STATIC_ASSERT_JITKERNEL_GLOBAL_NAMESPACE(uniq_name, msg)              \
+  struct __test_global_namespace_##uniq_name##__ {};                          \
+  static_assert(std::is_same<::__test_global_namespace_##uniq_name##__,       \
+                             __test_global_namespace_##uniq_name##__>::value, \
+                msg)
+
+// Refer always on CPUPlace
+#define REGISTER_JITKERNEL_REFER(kernel_type, ...)                             \
+  STATIC_ASSERT_JITKERNEL_GLOBAL_NAMESPACE(                                    \
+      __reg_jitkernel_##kernel_type##_refer_CPUPlace,                          \
+      "REGISTER_KERNEL_REFER must be called in global namespace");             \
+  static ::paddle::operators::jit::JitKernelRegistrar<                         \
+      ::paddle::operators::jit::ReferKernelPool, ::paddle::platform::CPUPlace, \
+      __VA_ARGS__>                                                             \
+      __jit_kernel_registrar_##kernel_type##_refer_CPUPlace_(                  \
+          ::paddle::operators::jit::KernelType::kernel_type);                  \
+  int TouchJitKernelReg_##kernel_type##_refer_CPUPlace_() {                    \
+    __jit_kernel_registrar_##kernel_type##_refer_CPUPlace_.Touch();            \
+    return 0;                                                                  \
+  }
+
+// kernel_type: should be in paddle::operators::jit::KernelType
+// place_type: should be one of CPUPlace and GPUPlace in paddle::platform
+#define REGISTER_KERNEL_MORE(kernel_type, impl_type, place_type, ...)         \
+  STATIC_ASSERT_JITKERNEL_GLOBAL_NAMESPACE(                                   \
+      __reg_jitkernel_##kernel_type##_##impl_type##_##place_type,             \
+      "REGISTER_KERNEL_MORE must be called in global namespace");             \
+  extern int TouchJitKernelReg_##kernel_type##_refer_CPUPlace_();             \
+  static int __assert_##kernel_type##_##impl_type##_##place_type##_has_refer_ \
+      UNUSED = TouchJitKernelReg_##kernel_type##_refer_CPUPlace_();           \
+  static ::paddle::operators::jit::JitKernelRegistrar<                        \
+      ::paddle::operators::jit::KernelPool, ::paddle::platform::place_type,   \
+      __VA_ARGS__>                                                            \
+      __jit_kernel_registrar_##kernel_type##_##impl_type##_##place_type##_(   \
+          ::paddle::operators::jit::KernelType::kernel_type);                 \
+  int TouchJitKernelReg_##kernel_type##_##impl_type##_##place_type##_() {     \
+    __jit_kernel_registrar_##kernel_type##_##impl_type##_##place_type##_      \
+        .Touch();                                                             \
+    return 0;                                                                 \
+  }
+
+#define REGISTER_JITKERNEL_MORE(kernel_type, impl_type, ...) \
+  REGISTER_KERNEL_MORE(kernel_type, impl_type, CPUPlace, __VA_ARGS__)
+
+#define REGISTER_GPUKERNEL_MORE(kernel_type, impl_type, ...) \
+  REGISTER_KERNEL_MORE(kernel_type, impl_type, GPUPlace, __VA_ARGS__)
+
+#define REGISTER_JITKERNEL_GEN(kernel_type, ...)                    \
+  STATIC_ASSERT_JITKERNEL_GLOBAL_NAMESPACE(                         \
+      __reg_jitkernel_gen_##kernel_type##_CPUPlace_,                \
+      "REGISTER_JITKERNEL_GEN must be called in global namespace"); \
+  extern int TouchJitKernelReg_##kernel_type##_refer_CPUPlace_();   \
+  static int __assert_gen_##kernel_type##_has_refer_ UNUSED =       \
+      TouchJitKernelReg_##kernel_type##_refer_CPUPlace_();          \
+  static ::paddle::operators::jit::JitKernelRegistrar<              \
+      ::paddle::operators::jit::JitCodeCreatorPool,                 \
+      ::paddle::platform::CPUPlace, __VA_ARGS__>                    \
+      __jit_kernel_registrar_gen_##kernel_type##_CPUPlace_(         \
+          ::paddle::operators::jit::KernelType::kernel_type);       \
+  int TouchJitKernelReg_gen_##kernel_type##_CPUPlace_() {           \
+    __jit_kernel_registrar_gen_##kernel_type##_CPUPlace_.Touch();   \
+    return 0;                                                       \
+  }
+
+#define USE_JITKERNEL_GEN(kernel_type)                            \
+  STATIC_ASSERT_JITKERNEL_GLOBAL_NAMESPACE(                       \
+      __reg_jitkernel_gen_##kernel_type##_CPUPlace_,              \
+      "USE_JITKERNEL_GEN must be called in global namespace");    \
+  extern int TouchJitKernelReg_gen_##kernel_type##_CPUPlace_();   \
+  static int use_jitkernel_gen_##kernel_type##_CPUPlace_ UNUSED = \
+      TouchJitKernelReg_gen_##kernel_type##_CPUPlace_()
+
+#define USE_JITKERNEL_REFER(kernel_type)                            \
+  STATIC_ASSERT_JITKERNEL_GLOBAL_NAMESPACE(                         \
+      __reg_jitkernel_##kernel_type##_refer_CPUPlace_,              \
+      "USE_JITKERNEL_REFER must be called in global namespace");    \
+  extern int TouchJitKernelReg_##kernel_type##_refer_CPUPlace_();   \
+  static int use_jitkernel_##kernel_type##_refer_CPUPlace_ UNUSED = \
+      TouchJitKernelReg_##kernel_type##_refer_CPUPlace_()
+
+#define USE_KERNEL_MORE(kernel_type, impl_type, place_type)              \
+  STATIC_ASSERT_JITKERNEL_GLOBAL_NAMESPACE(                              \
+      __reg_jitkernel_##kernel_type##_##impl_type##_##place_type##_,     \
+      "USE_JITKERNEL_MORE must be called in global namespace");          \
+  extern int                                                             \
+      TouchJitKernelReg_##kernel_type##_##impl_type##_##place_type##_(); \
+  static int use_jitkernel_##kernel_type##_##impl_type##_##place_type##_ \
+      UNUSED =                                                           \
+          TouchJitKernelReg_##kernel_type##_##impl_type##_##place_type##_()
+
+#define USE_JITKERNEL_MORE(kernel_type, impl_type) \
+  USE_KERNEL_MORE(kernel_type, impl_type, CPUPlace)
+
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/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 <random>
+#include <string>
+#include <vector>
+#include "gflags/gflags.h"
+#include "glog/logging.h"
+#include "gtest/gtest.h"
+#include "paddle/fluid/operators/jit/kernels.h"
+#include "paddle/fluid/platform/cpu_info.h"
+#include "paddle/fluid/platform/place.h"
+
+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);
+  }
+}
+
+template <typename T>
+void ExpectEQ(const T* target, const T* refer, int n) {
+  if (std::is_floating_point<T>::value) {
+    for (int i = 0; i < n; ++i) {
+      EXPECT_NEAR(target[i], refer[i], 1e-5);
+    }
+  } else {
+    for (int i = 0; i < n; ++i) {
+      EXPECT_EQ(target[i], refer[i]);
+    }
+  }
+}
+
+std::vector<int> TestSizes() {
+  std::vector<int> s;
+  for (int i = 1; i < 32; ++i) {
+    s.push_back(i);
+  }
+  // test some large size
+  s.push_back(100);
+  s.push_back(1000);
+  s.push_back(2000);
+  return s;
+}
+
+namespace jit = paddle::operators::jit;
+
+template <typename KernelTuples, typename... Args>
+struct TestFuncWithRefer {
+  void operator()(const typename KernelTuples::func_type tgt, Args... args) {}
+};
+
+template <typename T>
+struct TestFuncWithRefer<jit::XYZNTuples<T>, std::vector<T>, std::vector<T>,
+                         std::vector<T>> {
+  void operator()(const typename jit::XYZNTuples<T>::func_type tgt,
+                  const std::vector<T>& x, const std::vector<T>& y,
+                  const std::vector<T>& zref) {
+    EXPECT_TRUE(tgt != nullptr);
+    EXPECT_EQ(zref.size(), x.size());
+    EXPECT_EQ(zref.size(), y.size());
+    const T* x_data = x.data();
+    const T* y_data = y.data();
+    const T* zref_data = zref.data();
+    const int d = zref.size();
+
+    std::vector<T> ztgt(d);
+    T* ztgt_data = ztgt.data();
+    // test normal
+    tgt(x_data, y_data, ztgt_data, d);
+    ExpectEQ<T>(ztgt_data, zref_data, d);
+    // test inplace x
+    std::copy(x.begin(), x.end(), ztgt.begin());
+    tgt(ztgt_data, y_data, ztgt_data, d);
+    ExpectEQ<T>(ztgt_data, zref_data, d);
+    // test inplace y
+    std::copy(y.begin(), y.end(), ztgt.begin());
+    tgt(x_data, ztgt_data, ztgt_data, d);
+    ExpectEQ<T>(ztgt_data, zref_data, d);
+  }
+};
+
+template <typename T>
+struct TestFuncWithRefer<jit::AXYNTuples<T>, T, std::vector<T>,
+                         std::vector<T>> {
+  void operator()(const typename jit::AXYNTuples<T>::func_type tgt, const T a,
+                  const std::vector<T>& x, const std::vector<T>& yref) {
+    EXPECT_TRUE(tgt != nullptr);
+    EXPECT_EQ(yref.size(), x.size());
+    const T* x_data = x.data();
+    const T* yref_data = yref.data();
+    const int d = yref.size();
+    std::vector<T> ytgt(d);
+    T* ytgt_data = ytgt.data();
+    // test normal
+    tgt(&a, x_data, ytgt_data, d);
+    ExpectEQ<T>(ytgt_data, yref_data, d);
+    // test inplace x
+    std::copy(x.begin(), x.end(), ytgt.begin());
+    tgt(&a, ytgt_data, ytgt_data, d);
+    ExpectEQ<T>(ytgt_data, yref_data, d);
+  }
+};
+
+template <typename T>
+struct TestFuncWithRefer<jit::XYNTuples<T>, std::vector<T>, std::vector<T>> {
+  void operator()(const typename jit::XYNTuples<T>::func_type tgt,
+                  const std::vector<T>& x, const std::vector<T>& yref) {
+    EXPECT_TRUE(tgt != nullptr);
+    EXPECT_EQ(yref.size(), x.size());
+    const T* x_data = x.data();
+    const T* yref_data = yref.data();
+    const int d = yref.size();
+    std::vector<T> ytgt(d);
+    T* ytgt_data = ytgt.data();
+    // test normal
+    tgt(x_data, ytgt_data, d);
+    ExpectEQ<T>(ytgt_data, yref_data, d);
+    // test inplace x
+    std::copy(x.begin(), x.end(), ytgt.begin());
+    tgt(ytgt_data, ytgt_data, d);
+    ExpectEQ<T>(ytgt_data, yref_data, d);
+  }
+};
+
+template <typename T>
+struct TestFuncWithRefer<jit::LSTMTuples<T>, std::vector<T>, std::vector<T>,
+                         std::vector<T>, std::vector<T>, std::vector<T>> {
+  void operator()(const typename jit::LSTMTuples<T>::func_type tgt,
+                  const std::vector<T>& xsrc, const std::vector<T>& wp,
+                  const std::vector<T>& ct_1, const std::vector<T>& ct_ref,
+                  const std::vector<T>& ht_ref,
+                  const typename jit::LSTMTuples<T>::attr_type& attr) {
+    EXPECT_TRUE(tgt != nullptr);
+    EXPECT_EQ(ct_ref.size(), ht_ref.size());
+    EXPECT_EQ(ct_1.size(), ht_ref.size());
+    EXPECT_EQ(xsrc.size(), 4 * ht_ref.size());
+    EXPECT_EQ(wp.size(), 3 * ht_ref.size());
+
+    // x could be changed after compute, so copy to save src
+    int d = ht_ref.size();
+    std::vector<T> x(xsrc.size()), ct(ct_ref.size()), ht(ht_ref.size());
+    std::vector<T> checked(2 * d);
+    std::copy(xsrc.begin(), xsrc.end(), x.begin());
+
+    const T* ct_1_data = ct_1.data();
+    const T* wp_data = wp.data();
+    const T* ct_ref_data = ct_ref.data();
+    const T* ht_ref_data = ht_ref.data();
+    T* x_data = x.data();
+    T* ct_data = ct.data();
+    T* ht_data = ht.data();
+    T* checked_data = checked.data();
+
+    paddle::operators::jit::lstm_t step;
+    step.gates = x_data;
+    step.ct_1 = ct_1_data;
+    step.ct = ct_data;
+    step.ht = ht_data;
+    if (attr.use_peephole) {
+      step.wp = wp_data;
+      step.checked = checked_data;
+    }
+
+    tgt(&step, &attr);
+    ExpectEQ<T>(ct_data, ct_ref_data, d);
+    ExpectEQ<T>(ht_data, ht_ref_data, d);
+  }
+};
+
+template <typename T>
+struct TestFuncWithRefer<jit::GRUTuples<T>, std::vector<T>, std::vector<T>,
+                         std::vector<T>> {
+  void operator()(const typename jit::GRUTuples<T>::func_type tgt,
+                  const std::vector<T>& xsrc, const std::vector<T>& ht_1,
+                  const std::vector<T>& ht_ref,
+                  const typename jit::GRUTuples<T>::attr_type& attr) {
+    EXPECT_TRUE(tgt != nullptr);
+    EXPECT_EQ(ht_1.size(), ht_ref.size());
+    EXPECT_EQ(xsrc.size(), 3 * ht_ref.size());
+
+    // x could be changed after compute, so copy to save src
+    int d = ht_ref.size();
+    std::vector<T> x(xsrc.size()), ht(ht_ref.size());
+    std::copy(xsrc.begin(), xsrc.end(), x.begin());
+    const T* ht_1_data = ht_1.data();
+    const T* ht_ref_data = ht_ref.data();
+    T* x_data = x.data();
+    T* ht_data = ht.data();
+    paddle::operators::jit::gru_t step;
+    step.gates = x_data;
+    step.ht_1 = ht_1_data;
+    step.ht = ht_data;
+    tgt(&step, &attr);
+    ExpectEQ<T>(ht_data, ht_ref_data, d);
+  }
+};
+
+template <paddle::operators::jit::KernelType KT, typename KernelTuples,
+          typename PlaceType, typename... Args>
+void TestAllImpls(const typename KernelTuples::attr_type& attr, Args... args) {
+  TestFuncWithRefer<KernelTuples, Args...> test;
+  // test jitcode
+  auto jitcode = jit::GetJitCode<KT, KernelTuples, PlaceType>(attr);
+  if (jitcode) {
+    VLOG(10) << "Test Jitcode Kernel ";
+    test(jitcode, args...);
+  }
+  // test all impls in more
+  jit::KernelKey kkey(KT, PlaceType());
+  auto& pool = jit::KernelPool().Instance().AllKernels();
+  auto iter = pool.find(kkey);
+  if (iter != pool.end()) {
+    auto& impls = iter->second;
+    for (auto& impl : impls) {
+      auto i = dynamic_cast<const jit::KernelMore<KernelTuples>*>(impl.get());
+      if (i && i->UseMe(attr)) {
+        auto more = i->GetFunc();
+        VLOG(10) << "Test More Kernel : " << i->ImplType();
+        test(more, args...);
+      }
+    }
+  }
+  // test result from Get function
+  // VLOG(10) << "Test Get function ";
+  auto tgt = jit::Get<KT, KernelTuples, PlaceType>(attr);
+  test(tgt, args...);
+}
+
+template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+void TestXYZNKernel() {
+  namespace jit = paddle::operators::jit;
+  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
+  for (int d : TestSizes()) {
+    auto ref = jit::GetRefer<KT, jit::XYZNTuples<T>>();
+    EXPECT_TRUE(ref != nullptr);
+
+    std::vector<T> x(d), y(d), zref(d);
+    RandomVec<T>(d, x.data());
+    RandomVec<T>(d, y.data());
+
+    std::vector<T> xinp(d), yinp(d);  // inplace test
+    std::copy(x.begin(), x.end(), xinp.begin());
+    std::copy(y.begin(), y.end(), yinp.begin());
+
+    const T* x_data = x.data();
+    const T* y_data = y.data();
+    T* zref_data = zref.data();
+    T* xinp_data = xinp.data();
+    T* yinp_data = yinp.data();
+
+    // test refer code inplace
+    ref(x_data, y_data, zref_data, d);
+    ref(x_data, yinp_data, yinp_data, d);
+    ref(xinp_data, y_data, xinp_data, d);
+    ExpectEQ<T>(xinp_data, zref_data, d);
+    ExpectEQ<T>(yinp_data, zref_data, d);
+
+    TestAllImpls<KT, jit::XYZNTuples<T>, PlaceType, std::vector<T>,
+                 std::vector<T>, std::vector<T>>(d, x, y, zref);
+  }
+}
+
+template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+void TestAXYNKernel() {
+  namespace jit = paddle::operators::jit;
+  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
+  for (int d : TestSizes()) {
+    auto ref = jit::GetRefer<KT, jit::AXYNTuples<T>>();
+    EXPECT_TRUE(ref != nullptr);
+
+    const T a = static_cast<T>(3);
+    std::vector<T> x(d), yref(d);
+    std::vector<T> xinp(d);  // inplace test
+    RandomVec<T>(d, x.data());
+    std::copy(x.begin(), x.end(), xinp.begin());
+
+    const T* x_data = x.data();
+    T* yref_data = yref.data();
+    T* xinp_data = xinp.data();
+    // test refer code inplace
+    ref(&a, x_data, yref_data, d);
+    ref(&a, xinp_data, xinp_data, d);
+    ExpectEQ<T>(xinp_data, yref_data, d);
+
+    TestAllImpls<KT, jit::AXYNTuples<T>, PlaceType, T, std::vector<T>,
+                 std::vector<T>>(d, a, x, yref);
+  }
+}
+
+template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+void TestXYNKernel() {
+  namespace jit = paddle::operators::jit;
+  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
+  for (int d : TestSizes()) {
+    auto ref = jit::GetRefer<KT, jit::XYNTuples<T>>();
+    EXPECT_TRUE(ref != nullptr);
+
+    std::vector<T> x(d), yref(d);
+    std::vector<T> xinp(d);  // inplace test
+    RandomVec<T>(d, x.data(), -2.f, 2.f);
+    std::copy(x.begin(), x.end(), xinp.begin());
+
+    const T* x_data = x.data();
+    T* yref_data = yref.data();
+    T* xinp_data = xinp.data();
+    // test refer code inplace
+    ref(x_data, yref_data, d);
+    ref(xinp_data, xinp_data, d);
+    ExpectEQ<T>(xinp_data, yref_data, d);
+
+    TestAllImpls<KT, jit::XYNTuples<T>, PlaceType, std::vector<T>,
+                 std::vector<T>>(d, x, yref);
+  }
+}
+
+template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+void TestLSTMKernel() {
+  namespace jit = paddle::operators::jit;
+  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
+  std::vector<std::string> all_acts = {"sigmoid", "tanh", "relu", "identity"};
+  for (int d : TestSizes()) {
+    for (bool use_peephole : {true, false}) {
+      for (auto& act_gate : all_acts) {
+        for (auto& act_cand : all_acts) {
+          for (auto& act_cell : all_acts) {
+            const jit::lstm_attr_t attr(
+                d, jit::to_kerneltype(act_gate), jit::to_kerneltype(act_cand),
+                jit::to_kerneltype(act_cell), use_peephole);
+            auto ref = jit::GetRefer<KT, jit::LSTMTuples<T>>();
+            EXPECT_TRUE(ref != nullptr);
+            std::vector<T> xsrc(4 * d), wp(3 * d), ct_1(d);
+            std::vector<T> ct_ref(d), ht_ref(d), checked(2 * d);
+            RandomVec<T>(4 * d, xsrc.data(), -2.f, 2.f);
+            RandomVec<T>(3 * d, wp.data(), -2.f, 2.f);
+            RandomVec<T>(d, ct_1.data(), -2.f, 2.f);
+            // x could be changed after compute, so copy to save src
+            std::vector<T> x(xsrc.size());
+            std::copy(xsrc.begin(), xsrc.end(), x.begin());
+            const T* ct_1_data = ct_1.data();
+            const T* wp_data = wp.data();
+            T* x_data = x.data();
+            T* checked_data = checked.data();
+            T* ct_ref_data = ct_ref.data();
+            T* ht_ref_data = ht_ref.data();
+            jit::lstm_t step;
+            step.gates = x_data;
+            step.ct_1 = ct_1_data;
+            step.ct = ct_ref_data;
+            step.ht = ht_ref_data;
+            if (use_peephole) {
+              step.wp = wp_data;
+              step.checked = checked_data;
+            }
+            ref(&step, &attr);
+            VLOG(10) << attr;
+            TestAllImpls<KT, jit::LSTMTuples<T>, PlaceType, std::vector<T>,
+                         std::vector<T>, std::vector<T>, std::vector<T>,
+                         std::vector<T>>(attr, xsrc, wp, ct_1, ct_ref, ht_ref,
+                                         attr);
+          }
+        }
+      }
+    }
+  }
+}
+
+template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+void TestGRUKernel() {
+  namespace jit = paddle::operators::jit;
+  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
+  std::vector<std::string> all_acts = {"sigmoid", "tanh", "relu", "identity"};
+  for (int d : TestSizes()) {
+    for (auto& act_gate : all_acts) {
+      for (auto& act_cand : all_acts) {
+        const jit::gru_attr_t attr(d, jit::to_kerneltype(act_gate),
+                                   jit::to_kerneltype(act_cand));
+        auto ref = jit::GetRefer<KT, jit::GRUTuples<T>>();
+        EXPECT_TRUE(ref != nullptr);
+        std::vector<T> xsrc(3 * d), ht_1(d), ht_ref(d);
+        RandomVec<T>(3 * d, xsrc.data(), -2.f, 2.f);
+        RandomVec<T>(d, ht_1.data(), -2.f, 2.f);
+        // x could be changed after compute, so copy to save src
+        std::vector<T> x(xsrc.size());
+        std::copy(xsrc.begin(), xsrc.end(), x.begin());
+        const T* ht_1_data = ht_1.data();
+        T* x_data = x.data();
+        T* ht_ref_data = ht_ref.data();
+        jit::gru_t step;
+        step.gates = x_data;
+        step.ht_1 = ht_1_data;
+        step.ht = ht_ref_data;
+        ref(&step, &attr);
+        VLOG(10) << attr;
+        TestAllImpls<KT, jit::GRUTuples<T>, PlaceType, std::vector<T>,
+                     std::vector<T>, std::vector<T>>(attr, xsrc, ht_1, ht_ref,
+                                                     attr);
+      }
+    }
+  }
+}
+
+template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+void TestNCHW16CMulNCKernel() {
+  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
+  const int n = 3, c = 16 * 4, h = 10, w = 10;
+  auto ref = jit::GetRefer<KT, jit::NCHW16CMulNCTuples<T>>();
+  EXPECT_TRUE(ref != nullptr);
+  int sz = n * c * h * w;
+  std::vector<T> x(sz), y(n * c), zref(sz);
+  std::vector<T> ztgt(sz), zjit(sz);
+  RandomVec<T>(sz, x.data(), -2.f, 2.f);
+  RandomVec<T>(n * c, y.data(), -2.f, 2.f);
+
+  const T* x_data = x.data();
+  const T* y_data = y.data();
+  T* zref_data = zref.data();
+  T* ztgt_data = ztgt.data();
+  T* zjit_data = zjit.data();
+  constexpr int simd_width = ZMM_FLOAT_BLOCK;
+  int C = c / simd_width;
+  auto tgt = jit::Get<KT, jit::NCHW16CMulNCTuples<T>, PlaceType>(0);
+  auto jitcode = jit::GetJitCode<KT, jit::NCHW16CMulNCTuples<T>, PlaceType>(0);
+  EXPECT_TRUE(tgt != nullptr);
+
+  if (std::is_same<T, float>::value &&
+      paddle::platform::MayIUse(paddle::platform::avx512f)) {
+    EXPECT_TRUE(jitcode != nullptr);
+  }
+  for (int ni = 0; ni < n; ni++) {
+    for (int ci = 0; ci < C; ci++) {
+      auto ptr_x =
+          x_data + ni * C * h * w * simd_width + ci * h * w * simd_width;
+      auto ptr_y = y_data + ni * C * simd_width + ci * simd_width;
+      auto ptr_zref =
+          zref_data + ni * C * h * w * simd_width + ci * h * w * simd_width;
+      auto ptr_ztgt =
+          ztgt_data + ni * C * h * w * simd_width + ci * h * w * simd_width;
+
+      ref(ptr_x, ptr_y, ptr_zref, h, w);
+      tgt(ptr_x, ptr_y, ptr_ztgt, h, w);
+
+      if (jitcode) {
+        auto ptr_zjit =
+            zjit_data + ni * C * h * w * simd_width + ci * h * w * simd_width;
+        jitcode(ptr_x, ptr_y, ptr_zjit, h, w);
+      }
+    }
+  }
+  ExpectEQ<T>(ztgt_data, zref_data, sz);
+  if (jitcode) {
+    ExpectEQ<T>(zjit_data, zref_data, sz);
+  }
+}
+
+// XYZNTuple
+TEST(JITKernel, kVMul) {
+  namespace jit = paddle::operators::jit;
+  TestXYZNKernel<jit::kVMul, float, paddle::platform::CPUPlace>();
+  TestXYZNKernel<jit::kVMul, double, paddle::platform::CPUPlace>();
+}
+
+TEST(JITKernel, kVAdd) {
+  namespace jit = paddle::operators::jit;
+  TestXYZNKernel<jit::kVAdd, float, paddle::platform::CPUPlace>();
+  TestXYZNKernel<jit::kVAdd, double, paddle::platform::CPUPlace>();
+}
+
+TEST(JITKernel, kVAddRelu) {
+  namespace jit = paddle::operators::jit;
+  TestXYZNKernel<jit::kVAddRelu, float, paddle::platform::CPUPlace>();
+  TestXYZNKernel<jit::kVAddRelu, double, paddle::platform::CPUPlace>();
+}
+
+TEST(JITKernel, kVSub) {
+  namespace jit = paddle::operators::jit;
+  TestXYZNKernel<jit::kVSub, float, paddle::platform::CPUPlace>();
+  TestXYZNKernel<jit::kVSub, double, paddle::platform::CPUPlace>();
+}
+
+// AXYNTuples
+TEST(JITKernel, kVScal) {
+  namespace jit = paddle::operators::jit;
+  TestAXYNKernel<jit::kVScal, float, paddle::platform::CPUPlace>();
+  TestAXYNKernel<jit::kVScal, double, paddle::platform::CPUPlace>();
+}
+
+TEST(JITKernel, kVAddBias) {
+  namespace jit = paddle::operators::jit;
+  TestAXYNKernel<jit::kVAddBias, float, paddle::platform::CPUPlace>();
+  TestAXYNKernel<jit::kVAddBias, double, paddle::platform::CPUPlace>();
+}
+
+// XYNTuples
+TEST(JITKernel, kVRelu) {
+  namespace jit = paddle::operators::jit;
+  TestXYNKernel<jit::kVRelu, float, paddle::platform::CPUPlace>();
+  TestXYNKernel<jit::kVRelu, double, paddle::platform::CPUPlace>();
+}
+
+TEST(JITKernel, kVIdentity) {
+  namespace jit = paddle::operators::jit;
+  TestXYNKernel<jit::kVIdentity, float, paddle::platform::CPUPlace>();
+  TestXYNKernel<jit::kVIdentity, double, paddle::platform::CPUPlace>();
+}
+
+TEST(JITKernel, kVExp) {
+  namespace jit = paddle::operators::jit;
+  TestXYNKernel<jit::kVExp, float, paddle::platform::CPUPlace>();
+  TestXYNKernel<jit::kVExp, double, paddle::platform::CPUPlace>();
+}
+
+TEST(JITKernel, kVSigmoid) {
+  namespace jit = paddle::operators::jit;
+  TestXYNKernel<jit::kVSigmoid, float, paddle::platform::CPUPlace>();
+  TestXYNKernel<jit::kVSigmoid, double, paddle::platform::CPUPlace>();
+}
+
+TEST(JITKernel, kVTanh) {
+  namespace jit = paddle::operators::jit;
+  TestXYNKernel<jit::kVTanh, float, paddle::platform::CPUPlace>();
+  TestXYNKernel<jit::kVTanh, double, paddle::platform::CPUPlace>();
+}
+
+// LSTM
+TEST(JITKernel, kLSTMCtHt) {
+  namespace jit = paddle::operators::jit;
+  TestLSTMKernel<jit::kLSTMCtHt, float, paddle::platform::CPUPlace>();
+  TestLSTMKernel<jit::kLSTMCtHt, double, paddle::platform::CPUPlace>();
+}
+
+TEST(JITKernel, kLSTMC1H1) {
+  namespace jit = paddle::operators::jit;
+  TestLSTMKernel<jit::kLSTMC1H1, float, paddle::platform::CPUPlace>();
+  TestLSTMKernel<jit::kLSTMC1H1, double, paddle::platform::CPUPlace>();
+}
+
+// GRU
+TEST(JITKernel, kGRUH1) {
+  namespace jit = paddle::operators::jit;
+  TestGRUKernel<jit::kGRUH1, float, paddle::platform::CPUPlace>();
+  TestGRUKernel<jit::kGRUH1, double, paddle::platform::CPUPlace>();
+}
+
+TEST(JITKernel, kGRUHtPart1) {
+  namespace jit = paddle::operators::jit;
+  TestGRUKernel<jit::kGRUHtPart1, float, paddle::platform::CPUPlace>();
+  TestGRUKernel<jit::kGRUHtPart1, double, paddle::platform::CPUPlace>();
+}
+
+TEST(JITKernel, kGRUHtPart2) {
+  namespace jit = paddle::operators::jit;
+  TestGRUKernel<jit::kGRUHtPart2, float, paddle::platform::CPUPlace>();
+  TestGRUKernel<jit::kGRUHtPart2, double, paddle::platform::CPUPlace>();
+}
+
+TEST(JITKernel, kNCHW16CMulNC) {
+  namespace jit = paddle::operators::jit;
+  TestNCHW16CMulNCKernel<jit::kNCHW16CMulNC, float,
+                         paddle::platform::CPUPlace>();
+  TestNCHW16CMulNCKernel<jit::kNCHW16CMulNC, double,
+                         paddle::platform::CPUPlace>();
+}
+
+// TODO(yihua/TJ): add crf decoding and layer norm unit tests
+
+TEST(JITKernel, pool) {
+  // TODO(TJ): add some test
+}

paddle/fluid/operators/layer_norm_op.h

@@ -19,7 +19,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/math/blas.h"
 #if !defined(PADDLE_WITH_CUDA) && !defined(_WIN32) && !defined(__APPLE__) && \
     !defined(__OSX__)
-#include "paddle/fluid/operators/math/jit_kernel.h"
+#include "paddle/fluid/operators/jit/kernels.h"
 #endif
 #include "paddle/fluid/operators/math/math_function.h"
 
@@ -229,12 +229,12 @@ class LayerNormKernel : public framework::OpKernel<T> {
     PADDLE_ENFORCE_EQ(scale->numel(), right);
     PADDLE_ENFORCE_EQ(bias->numel(), right);
 
-    const auto& ker = math::jitkernel::KernelPool::Instance()
-                          .template Get<math::jitkernel::LayerNormKernel<T>>(
-                              static_cast<int>(right));
-    ker->Compute(x.data<T>(), out.data<T>(), mean->data<T>(), var->data<T>(),
+    auto ker =
+        jit::Get<jit::kLayerNorm, jit::LayerNormTuples<T>, platform::CPUPlace>(
+            right);
+    ker(x.data<T>(), out.data<T>(), mean->data<T>(), var->data<T>(),
         scale->data<T>(), bias->data<T>(), static_cast<int>(left),
-                 static_cast<const float>(epsilon));
+        static_cast<const float>(epsilon), right);
 #endif
   }
 };

paddle/fluid/operators/math/CMakeLists.txt

@@ -73,12 +73,3 @@ if(WITH_GPU)
 endif()
 cc_test(concat_test SRCS concat_test.cc DEPS concat_and_split)
 cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info)
-
-set(JIT_KERNEL_SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_rnn.cc jit_kernel_crf_decode.cc jit_kernel_layer_norm.cc)
-set(JIT_KERNEL_DEPS cpu_info cblas gflags enforce)
-if(WITH_XBYAK)
-    list(APPEND JIT_KERNEL_SRCS jit_gen.cc jit_code.cc)
-    list(APPEND JIT_KERNEL_DEPS xbyak)
-endif()
-cc_library(jit_kernel SRCS ${JIT_KERNEL_SRCS} DEPS ${JIT_KERNEL_DEPS})
-cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel)

paddle/fluid/operators/math/concat_and_split.cu

@@ -131,9 +131,8 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
     int in_col = input[0].numel() / in_row;
     int out_row = in_row, out_col = 0;
 
-    framework::Vector<int16_t> inputs_data(in_num * sizeof(T*) / 2);
-    framework::Vector<int> inputs_col(in_num + 1);
-    T** inputs_ptr = reinterpret_cast<T**>(inputs_data.data());
+    std::vector<T*> inputs_data(in_num);
+    std::vector<int> inputs_col(in_num + 1);
 
     inputs_col[0] = 0;
     bool sameShape = true;
@@ -144,12 +143,9 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
       }
       out_col += t_cols;
       inputs_col[i + 1] = out_col;
-      inputs_ptr[i] = const_cast<T*>(input[i].data<T>());
+      inputs_data[i] = const_cast<T*>(input[i].data<T>());
     }
 
-    T** dev_ins_data =
-        reinterpret_cast<T**>(inputs_data.CUDAMutableData(context.GetPlace()));
-
     // computation
     // set the thread block and grid according to CurrentDeviceId
     const int kThreadsPerBlock = 1024;
@@ -169,18 +165,32 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
         std::min(max_blocks / grid_cols, std::max(out_row / block_rows, 1));
     dim3 grid_size = dim3(grid_cols, grid_rows, 1);
 
+    auto tmp_dev_ins_data =
+        platform::DeviceTemporaryAllocator::Instance().Get(context).Allocate(
+            inputs_data.size() * sizeof(T*));
+    memory::Copy(boost::get<platform::CUDAPlace>(context.GetPlace()),
+                 tmp_dev_ins_data->ptr(), platform::CPUPlace(),
+                 static_cast<void*>(inputs_data.data()),
+                 inputs_data.size() * sizeof(T*), context.stream());
+    T** dev_ins_data = reinterpret_cast<T**>(tmp_dev_ins_data->ptr());
+
     if (sameShape) {
       ConcatKernel<<<grid_size, block_size, 0, context.stream()>>>(
           dev_ins_data, in_col, out_row, out_col, output->data<T>());
     } else {
-      const int* dev_ins_col_data = inputs_col.CUDAData(context.GetPlace());
+      auto tmp_dev_ins_col_data =
+          platform::DeviceTemporaryAllocator::Instance().Get(context).Allocate(
+              inputs_col.size() * sizeof(int));
+      memory::Copy(boost::get<platform::CUDAPlace>(context.GetPlace()),
+                   tmp_dev_ins_col_data->ptr(), platform::CPUPlace(),
+                   static_cast<void*>(inputs_col.data()),
+                   inputs_col.size() * sizeof(int), context.stream());
+      int* dev_ins_col_data = static_cast<int*>(tmp_dev_ins_col_data->ptr());
+
       ConcatKernel<<<grid_size, block_size, 0, context.stream()>>>(
           dev_ins_data, dev_ins_col_data, static_cast<int>(inputs_col.size()),
           out_row, out_col, output->data<T>());
     }
-    // Wait() must be called because `inputs_data` may be destructed before
-    // kernel ends
-    context.Wait();
   }
 };
 
@@ -207,9 +217,8 @@ class SplitFunctor<platform::CUDADeviceContext, T> {
     int in_col = 0, in_row = out_row;
     bool sameShape = true;
 
-    framework::Vector<int16_t> outputs_data(o_num * sizeof(T*) / 2);
-    framework::Vector<int> outputs_cols(o_num + 1);
-    T** outputs_ptr = reinterpret_cast<T**>(outputs_data.data());
+    std::vector<T*> outputs_data(o_num);
+    std::vector<int> outputs_cols(o_num + 1);
 
     outputs_cols[0] = 0;
     for (int i = 0; i < o_num; ++i) {
@@ -220,15 +229,12 @@ class SplitFunctor<platform::CUDADeviceContext, T> {
       in_col += t_col;
       outputs_cols[i + 1] = in_col;
       if (outputs->at(i) != nullptr) {
-        outputs_ptr[i] = outputs->at(i)->data<T>();
+        outputs_data[i] = outputs->at(i)->data<T>();
       } else {
-        outputs_ptr[i] = nullptr;
+        outputs_data[i] = nullptr;
       }
     }
 
-    T** dev_out_gpu_data =
-        reinterpret_cast<T**>(outputs_data.CUDAMutableData(context.GetPlace()));
-
     // computation
     const int kThreadsPerBlock = 1024;
     int block_cols = kThreadsPerBlock;
@@ -247,18 +253,33 @@ class SplitFunctor<platform::CUDADeviceContext, T> {
         std::min(max_blocks / grid_cols, std::max(out_row / block_rows, 1));
     dim3 grid_size = dim3(grid_cols, grid_rows, 1);
 
+    auto tmp_dev_outs_data =
+        platform::DeviceTemporaryAllocator::Instance().Get(context).Allocate(
+            outputs_data.size() * sizeof(T*));
+    memory::Copy(boost::get<platform::CUDAPlace>(context.GetPlace()),
+                 tmp_dev_outs_data->ptr(), platform::CPUPlace(),
+                 reinterpret_cast<void*>(outputs_data.data()),
+                 outputs_data.size() * sizeof(T*), context.stream());
+    T** dev_out_gpu_data = reinterpret_cast<T**>(tmp_dev_outs_data->ptr());
+
     if (sameShape) {
       SplitKernel<<<grid_size, block_size, 0, context.stream()>>>(
           input.data<T>(), in_row, in_col, out0_col, dev_out_gpu_data);
     } else {
-      const int* dev_outs_col_data = outputs_cols.CUDAData(context.GetPlace());
+      auto tmp_dev_ins_col_data =
+          platform::DeviceTemporaryAllocator::Instance().Get(context).Allocate(
+              outputs_cols.size() * sizeof(int));
+      memory::Copy(boost::get<platform::CUDAPlace>(context.GetPlace()),
+                   tmp_dev_ins_col_data->ptr(), platform::CPUPlace(),
+                   reinterpret_cast<void*>(outputs_cols.data()),
+                   outputs_cols.size() * sizeof(int), context.stream());
+      int* dev_outs_col_data =
+          reinterpret_cast<int*>(tmp_dev_ins_col_data->ptr());
+
       SplitKernel<<<grid_size, block_size, 0, context.stream()>>>(
           input.data<T>(), in_row, in_col, dev_outs_col_data,
           static_cast<int>(outputs_cols.size()), dev_out_gpu_data);
     }
-    // Wait() must be called because `outputs_data` may be destructed before
-    // kernel ends
-    context.Wait();
   }
 };
 

paddle/fluid/operators/math/detail/lstm_cpu_kernel.h

@@ -17,6 +17,12 @@ limitations under the License. */
 #include "paddle/fluid/operators/math/detail/activation_functions.h"
 #include "paddle/fluid/operators/math/lstm_compute.h"
 
+#if defined(_WIN32)
+#if defined(__AVX2__) || defined(__AVX__)
+inline __m256 operator+=(__m256 a, __m256 b) { return _mm256_add_ps(a, b); }
+#endif
+#endif
+
 namespace paddle {
 namespace operators {
 namespace math {

paddle/fluid/operators/math/fc_compute.h

@@ -14,8 +14,8 @@ limitations under the License. */
 
 #pragma once
 
+#include "paddle/fluid/operators/jit/kernels.h"
 #include "paddle/fluid/operators/math/blas.h"
-#include "paddle/fluid/operators/math/jit_kernel.h"
 
 namespace paddle {
 namespace operators {
@@ -30,22 +30,21 @@ inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
     return;
   }
   if (relu) {
-    const auto& vaddrelu = jitkernel::KernelPool::Instance()
-                               .template Get<jitkernel::VAddReluKernel<T>>(N);
+    auto compute =
+        jit::Get<jit::kVAddRelu, jit::XYZNTuples<T>, platform::CPUPlace>(N);
     for (int i = 0; i < M; i++) {
       T* dst = Y + i * N;
-      vaddrelu->Compute(B, dst, dst, N);
+      compute(B, dst, dst, N);
     }
   } else {
-    const auto& vadd = jitkernel::KernelPool::Instance()
-                           .template Get<jitkernel::VAddKernel<T>>(N);
-
+    auto compute =
+        jit::Get<jit::kVAdd, jit::XYZNTuples<T>, platform::CPUPlace>(N);
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for
 #endif
     for (int i = 0; i < M; i++) {
       T* dst = Y + i * N;
-      vadd->Compute(B, dst, dst, N);
+      compute(B, dst, dst, N);
     }
   }
 }

paddle/fluid/operators/math/jit_code.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_code.h"
-#include <stddef.h>                                  // offsetof
-#include "paddle/fluid/operators/math/jit_kernel.h"  // TODO(TJ): remove me
-
-namespace paddle {
-namespace operators {
-namespace math {
-namespace jitkernel {
-namespace gen {
-
-using namespace platform;  // NOLINT
-
-bool VXXJitCode::init(int d, int scalar_index) {
-  // It's not necessary to use avx512 since it would slow down the frequency
-  // and this kernel is not compute bound.
-  return MayIUse(avx) && scalar_index >= 0 && scalar_index <= 2;
-}
-
-void VXXJitCode::generate() {
-  // do not need push stack, and do not need save avx512reg if do not use avx512
-  int offset = 0;
-  if (with_relu_) {
-    vxorps(ymm_zero, ymm_zero, ymm_zero);
-  }
-  if (scalar_index_ == 1) {
-    vbroadcastss(ymm_src1, ptr[param1]);
-  } else if (scalar_index_ == 2) {
-    vbroadcastss(ymm_src2, ptr[param2]);
-  }
-  for (int i = 0; i < num_ / YMM_FLOAT_BLOCK; ++i) {
-    if (scalar_index_ != 1) {
-      vmovups(ymm_src1, ptr[param1 + offset]);
-    }
-    if (scalar_index_ != 2) {
-      vmovups(ymm_src2, ptr[param2 + offset]);
-    }
-    if (type_ == operand_type::mul) {
-      vmulps(ymm_dst, ymm_src1, ymm_src2);
-    } else if (type_ == operand_type::add) {
-      vaddps(ymm_dst, ymm_src1, ymm_src2);
-    }
-    if (with_relu_) {
-      vmaxps(ymm_dst, ymm_zero, ymm_dst);
-    }
-    vmovups(ptr[param3 + offset], ymm_dst);
-    offset += sizeof(float) * YMM_FLOAT_BLOCK;
-  }
-  int rest = num_ % YMM_FLOAT_BLOCK;
-  while (rest > 0) {
-    int block = XMM_FLOAT_BLOCK;
-    if (rest >= 4) {
-      block = 4;
-      if (scalar_index_ != 1) {
-        vmovups(xmm_src1, ptr[param1 + offset]);
-      }
-      if (scalar_index_ != 2) {
-        vmovups(xmm_src2, ptr[param2 + offset]);
-      }
-    } else if (rest >= 2) {
-      block = 2;
-      if (scalar_index_ != 1) {
-        vmovq(xmm_src1, ptr[param1 + offset]);
-      }
-      if (scalar_index_ != 2) {
-        vmovq(xmm_src2, ptr[param2 + offset]);
-      }
-    } else {
-      block = 1;
-      if (scalar_index_ != 1) {
-        vmovss(xmm_src1, ptr[param1 + offset]);
-      }
-      if (scalar_index_ != 2) {
-        vmovss(xmm_src2, ptr[param2 + offset]);
-      }
-    }
-    switch (type_) {
-      case operand_type::mul:
-        vmulps(xmm_dst, xmm_src1, xmm_src2);
-        break;
-      case operand_type::add:
-        vaddps(xmm_dst, xmm_src1, xmm_src2);
-        break;
-      default:
-        break;
-    }
-    if (with_relu_) {
-      vmaxps(xmm_dst, xmm_zero, xmm_dst);
-    }
-    if (rest >= 4) {
-      vmovups(ptr[param3 + offset], xmm_dst);
-    } else if (rest >= 2) {
-      vmovq(ptr[param3 + offset], xmm_dst);
-    } else {
-      vmovss(ptr[param3 + offset], xmm_dst);
-    }
-    offset += sizeof(float) * block;
-    rest -= block;
-  }
-  ret();
-}
-
-const float ALIGN32_BEG exp_float_consts[] ALIGN32_END = {
-    REPEAT_8TIMES(1.f),
-    REPEAT_8TIMES(2.f),
-    REPEAT_8TIMES(0.5f),
-    REPEAT_8TIMES(EXP_HIG),
-    REPEAT_8TIMES(EXP_LOW),
-    REPEAT_8TIMES(CEPHES_LOG2EF),
-    REPEAT_8TIMES(CEPHES_EXP_C1),
-    REPEAT_8TIMES(CEPHES_EXP_C2),
-    REPEAT_8TIMES(CEPHES_EXP_P0),
-    REPEAT_8TIMES(CEPHES_EXP_P1),
-    REPEAT_8TIMES(CEPHES_EXP_P2),
-    REPEAT_8TIMES(CEPHES_EXP_P3),
-    REPEAT_8TIMES(CEPHES_EXP_P4),
-    REPEAT_8TIMES(CEPHES_EXP_P5),
-    REPEAT_8TIMES(EXP_MAX_INPUT),
-    REPEAT_8TIMES(SIGMOID_THRESHOLD_MAX),
-    REPEAT_8TIMES(SIGMOID_THRESHOLD_MIN)};
-
-const int ALIGN32_BEG exp_int_0x7f[] ALIGN32_END = {REPEAT_8TIMES(0x7f)};
-int ALIGN32_BEG g_tmp_mem[16] ALIGN32_END = {0};
-
-bool VActJitCode::init(int d, operand_type type) {
-  // TODO(TJ): implement avx512, avx_exp is slower than mkl when d >= 256
-  return MayIUse(avx);
-}
-
-void VActJitCode::generate() {
-  int offset = 0;
-  for (int i = 0; i < num_ / YMM_FLOAT_BLOCK; ++i) {
-    vmovups(ymm_src, ptr[param1 + offset]);
-    act<ymm_t>(ymm_dst, ymm_src, type_);
-    vmovups(ptr[param2 + offset], ymm_dst);
-    offset += sizeof(float) * YMM_FLOAT_BLOCK;
-  }
-  int rest = num_ % YMM_FLOAT_BLOCK;
-  while (rest > 0) {
-    int block = XMM_FLOAT_BLOCK;
-    if (rest >= 4) {
-      block = 4;
-      vmovups(xmm_src, ptr[param1 + offset]);
-    } else if (rest >= 2) {
-      block = 2;
-      vmovq(xmm_src, ptr[param1 + offset]);
-    } else {
-      block = 1;
-      vmovss(xmm_src, ptr[param1 + offset]);
-    }
-    act<xmm_t>(xmm_dst, xmm_src, type_);
-    if (rest >= 4) {
-      vmovups(ptr[param2 + offset], xmm_dst);
-    } else if (rest >= 2) {
-      vmovq(ptr[param2 + offset], xmm_dst);
-    } else {
-      vmovss(ptr[param2 + offset], xmm_dst);
-    }
-    offset += sizeof(float) * block;
-    rest -= block;
-  }
-  ret();
-}
-
-bool LSTMJitCode::init(int d) { return MayIUse(avx) && d % 8 == 0; }
-
-void LSTMJitCode::generate() {
-  if (use_peephole_) {
-    preCode();
-  }
-  reg64_t reg_ptr_gates = rax;
-  reg64_t reg_ptr_ct_1 = r9;
-  reg64_t reg_ptr_ct = r10;
-  reg64_t reg_ptr_ht = r11;
-  reg64_t reg_ptr_wp = r12;
-  mov(reg_ptr_gates, ptr[param1 + offsetof(lstm_t, gates)]);
-  mov(reg_ptr_ct_1, ptr[param1 + offsetof(lstm_t, ct_1)]);
-  mov(reg_ptr_ct, ptr[param1 + offsetof(lstm_t, ct)]);
-  mov(reg_ptr_ht, ptr[param1 + offsetof(lstm_t, ht)]);
-  if (use_peephole_) {
-    mov(reg_ptr_wp, ptr[param1 + offsetof(lstm_t, wp)]);
-  }
-
-  int offset = 0;
-  int d = num_ * sizeof(float);
-  for (int i = 0; i < num_ / YMM_FLOAT_BLOCK; ++i) {
-    /* gates: W_ch, W_ih, W_fh, W_oh */
-    ymm_t ymm_c = ymm_t(0);
-    ymm_t ymm_i = ymm_t(1);
-    ymm_t ymm_f = ymm_t(2);
-    ymm_t ymm_o = ymm_t(3);
-    ymm_t ymm_ct_1 = ymm_t(4);
-    ymm_t ymm_wp0 = ymm_t(5);
-    ymm_t ymm_wp1 = ymm_t(6);
-    ymm_t ymm_wp2 = ymm_t(7);
-    vmovups(ymm_c, ptr[reg_ptr_gates + offset]);
-    vmovups(ymm_i, ptr[reg_ptr_gates + offset + d]);
-    vmovups(ymm_f, ptr[reg_ptr_gates + offset + 2 * d]);
-    vmovups(ymm_o, ptr[reg_ptr_gates + offset + 3 * d]);
-    if (!compute_c1h1_) {
-      vmovups(ymm_ct_1, ptr[reg_ptr_ct_1 + offset]);
-    }
-    if (use_peephole_) {
-      vmovups(ymm_wp0, ptr[reg_ptr_wp + offset]);
-      vmovups(ymm_wp1, ptr[reg_ptr_wp + offset + d]);
-      vmovups(ymm_wp2, ptr[reg_ptr_wp + offset + 2 * d]);
-    }
-    /* C_t = act_cand(c) * act_gate(i) + C_t-1 * act_gate(f) */
-    // act_cand(c)
-    act<ymm_t>(ymm_c, ymm_c, act_cand_);
-    // act_gate(i) or act_gate(ct_1 * wp0 + i)
-    if (!compute_c1h1_ && use_peephole_) {
-      vmulps(ymm_wp0, ymm_ct_1, ymm_wp0);
-      vaddps(ymm_i, ymm_i, ymm_wp0);
-    }
-    act<ymm_t>(ymm_i, ymm_i, act_gate_);
-    vmulps(ymm_c, ymm_c, ymm_i);
-    if (!compute_c1h1_) {
-      // act_gate(f) or act_gate(ct_1 * wp1 + f)
-      if (use_peephole_) {
-        vmulps(ymm_wp1, ymm_ct_1, ymm_wp1);
-        vaddps(ymm_f, ymm_f, ymm_wp1);
-      }
-      act<ymm_t>(ymm_f, ymm_f, act_gate_);
-      // ct
-      vmulps(ymm_f, ymm_f, ymm_ct_1);
-      vaddps(ymm_f, ymm_f, ymm_c);
-    }
-    /* H_t = act_cell(C_t) * act_gate(o) */
-    // act_cell(C_t)
-    ymm_t ymm_ct = compute_c1h1_ ? ymm_c : ymm_f;
-    ymm_t ymm_tmp = ymm_i;
-    act<ymm_t>(ymm_tmp, ymm_ct, act_cell_);
-    // act_gate(o) or act_gate(ct * wp2 + o)
-    if (use_peephole_) {
-      vmulps(ymm_wp2, ymm_ct, ymm_wp2);
-      vaddps(ymm_o, ymm_o, ymm_wp2);
-    }
-    act<ymm_t>(ymm_o, ymm_o, act_gate_);
-    // ht
-    vmulps(ymm_o, ymm_o, ymm_tmp);
-    // save ct and ht
-    vmovups(ptr[reg_ptr_ct + offset], ymm_ct);
-    vmovups(ptr[reg_ptr_ht + offset], ymm_o);
-    offset += sizeof(float) * YMM_FLOAT_BLOCK;
-  }
-
-  if (use_peephole_) {
-    postCode();
-  } else {
-    ret();
-  }
-}
-
-bool GRUJitCode::init(int d) { return MayIUse(avx) && d % 8 == 0; }
-
-void GRUJitCode::generate() {
-  reg64_t reg_ptr_gates = rax;
-  reg64_t reg_ptr_ht_1 = r9;
-  reg64_t reg_ptr_ht = r10;
-  mov(reg_ptr_gates, ptr[param1 + offsetof(gru_t, gates)]);
-  mov(reg_ptr_ht_1, ptr[param1 + offsetof(gru_t, ht_1)]);
-  mov(reg_ptr_ht, ptr[param1 + offsetof(gru_t, ht)]);
-  ymm_t ymm_one = ymm_t(0);
-
-  if (id_ == 2) {
-    reg64_t reg_ptr_tmp = r11;
-    mov(reg_ptr_tmp, reinterpret_cast<size_t>(exp_float_consts));
-    vmovaps(ymm_one, ptr[reg_ptr_tmp + OFFSET_EXP_ONE]);
-  }
-  int offset = 0;
-  int d = num_ * sizeof(float);
-  for (int i = 0; i < num_ / YMM_FLOAT_BLOCK; ++i) {
-    ymm_t ymm_u = ymm_t(1);
-    ymm_t ymm_r = ymm_t(2);
-    ymm_t ymm_s = ymm_t(3);
-    ymm_t ymm_ht_1 = ymm_t(4);
-    // W: {W_update, W_reset; W_state}
-    if (id_ == 0 || id_ == 2) {
-      vmovups(ymm_u, ptr[reg_ptr_gates + offset]);
-      vmovups(ymm_s, ptr[reg_ptr_gates + offset + 2 * d]);
-    }
-    if (id_ == 1) {
-      vmovups(ymm_r, ptr[reg_ptr_gates + offset + d]);
-    }
-    if (id_ == 1 || id_ == 2) {
-      vmovups(ymm_ht_1, ptr[reg_ptr_ht_1 + offset]);
-    }
-
-    if (id_ == 0) {
-      // ht = act_gate(u) * act_cand(s)
-      act<ymm_t>(ymm_u, ymm_u, act_gate_);
-      act<ymm_t>(ymm_s, ymm_s, act_cand_);
-      vmulps(ymm_s, ymm_s, ymm_u);
-      vmovups(ptr[reg_ptr_ht + offset], ymm_s);
-    } else if (id_ == 1) {
-      // ht = act_gate(r) * ht_1
-      act<ymm_t>(ymm_r, ymm_r, act_gate_);
-      vmulps(ymm_r, ymm_r, ymm_ht_1);
-      vmovups(ptr[reg_ptr_ht + offset], ymm_r);
-    } else if (id_ == 2) {
-      // ht = act_gate(u) * act_cand(s) + (1-act_gate(u)) * ht_1
-      ymm_t ymm_one_inner = ymm_t(ymm_one.getIdx());
-      act<ymm_t>(ymm_u, ymm_u, act_gate_);
-      act<ymm_t>(ymm_s, ymm_s, act_cand_);
-      vmulps(ymm_s, ymm_s, ymm_u);
-      vsubps(ymm_u, ymm_one_inner, ymm_u);
-      vmulps(ymm_u, ymm_ht_1, ymm_u);
-      vaddps(ymm_u, ymm_s, ymm_u);
-      vmovups(ptr[reg_ptr_ht + offset], ymm_u);
-    }
-    offset += sizeof(float) * YMM_FLOAT_BLOCK;
-  }
-
-  ret();
-}
-}  // namespace gen
-}  // namespace jitkernel
-}  // namespace math
-}  // namespace operators
-}  // namespace paddle

paddle/fluid/operators/math/jit_gen.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_gen.h"
-#include <fstream>
-#include <iostream>
-#include <sstream>
-#include "paddle/fluid/platform/cpu_info.h"
-
-DEFINE_bool(dump_jitcode, false, "Whether to dump the jitcode to file");
-
-namespace paddle {
-namespace operators {
-namespace math {
-namespace jitkernel {
-namespace gen {
-
-constexpr Xbyak::Operand::Code g_abi_regs[] = {
-    Xbyak::Operand::RBX, Xbyak::Operand::RBP, Xbyak::Operand::R12,
-    Xbyak::Operand::R13, Xbyak::Operand::R14, Xbyak::Operand::R15};
-
-constexpr int num_g_abi_regs = sizeof(g_abi_regs) / sizeof(g_abi_regs[0]);
-
-void JitCode::preCode() {
-  for (int i = 0; i < num_g_abi_regs; ++i) {
-    push(Xbyak::Reg64(g_abi_regs[i]));
-  }
-  if (platform::MayIUse(platform::avx512f)) {
-    mov(reg_EVEX_max_8b_offt, 2 * EVEX_max_8b_offt);
-  }
-}
-
-void JitCode::postCode() {
-  for (int i = 0; i < num_g_abi_regs; ++i) {
-    pop(Xbyak::Reg64(g_abi_regs[num_g_abi_regs - 1 - i]));
-  }
-  ret();
-}
-
-void JitCode::dumpCode(const Xbyak::uint8 *code) const {
-  if (code) {
-    static int counter = 0;
-    std::ostringstream filename;
-    filename << "paddle_jitcode_" << name() << "." << counter << ".bin";
-    counter++;
-    std::ofstream fout(filename.str(), std::ios::out);
-    if (fout.is_open()) {
-      fout.write(reinterpret_cast<const char *>(code), getSize());
-      fout.close();
-    }
-  }
-}
-
-Xbyak::Address JitCode::EVEX_compress_addr(Xbyak::Reg64 base, int offt,
-                                           bool bcast) {
-  int scale = 0;
-  if (EVEX_max_8b_offt <= offt && offt < 3 * EVEX_max_8b_offt) {
-    offt = offt - 2 * EVEX_max_8b_offt;
-    scale = 1;
-  } else if (3 * EVEX_max_8b_offt <= offt && offt < 5 * EVEX_max_8b_offt) {
-    offt = offt - 4 * EVEX_max_8b_offt;
-    scale = 2;
-  }
-  auto re = Xbyak::RegExp() + base + offt;
-  if (scale) {
-    re = re + reg_EVEX_max_8b_offt * scale;
-  }
-  if (bcast) {
-    return zword_b[re];
-  } else {
-    return zword[re];
-  }
-}
-
-}  // namespace gen
-}  // namespace jitkernel
-}  // namespace math
-}  // namespace operators
-}  // namespace paddle

paddle/fluid/operators/math/jit_gen.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 <gflags/gflags.h>
-#include <type_traits>
-#include "paddle/fluid/platform/macros.h"
-
-#define XBYAK_USE_MMAP_ALLOCATOR
-#include "xbyak/xbyak.h"
-#include "xbyak/xbyak_util.h"
-
-DECLARE_bool(dump_jitcode);
-
-namespace paddle {
-namespace operators {
-namespace math {
-namespace jitkernel {
-namespace gen {
-
-#define DECLARE_JIT_CODE(codename) \
-  const char *name() const override { return #codename; }
-
-// Application Binary Interface
-constexpr Xbyak::Operand::Code abi_param1(Xbyak::Operand::RDI),
-    abi_param2(Xbyak::Operand::RSI), abi_param3(Xbyak::Operand::RDX),
-    abi_param4(Xbyak::Operand::RCX), abi_not_param1(Xbyak::Operand::RCX);
-
-class JitCode : public Xbyak::CodeGenerator {
- public:
-  explicit JitCode(size_t code_size = 256 * 1024, void *code_ptr = nullptr)
-      : Xbyak::CodeGenerator(code_size, code_ptr) {}
-
-  virtual ~JitCode() {}
-  virtual const char *name() const = 0;
-  virtual void generate() = 0;
-
-  template <typename FUNC>
-  const FUNC getCode() {
-    this->generate();
-    const Xbyak::uint8 *code = CodeGenerator::getCode();
-    if (FLAGS_dump_jitcode) {
-      this->dumpCode(code);
-    }
-    return reinterpret_cast<const FUNC>(code);
-  }
-  DISABLE_COPY_AND_ASSIGN(JitCode);
-
- protected:
-  Xbyak::Reg64 param1{abi_param1};
-  const int EVEX_max_8b_offt = 0x200;
-  const Xbyak::Reg64 reg_EVEX_max_8b_offt = rbp;
-
-  void preCode();
-  void postCode();
-  void dumpCode(const Xbyak::uint8 *code) const;
-  void L(const char *label) { Xbyak::CodeGenerator::L(label); }
-  void L(const Xbyak::Label &label) { Xbyak::CodeGenerator::L(label); }
-  // Enhanced vector extension
-  Xbyak::Address EVEX_compress_addr(Xbyak::Reg64 base, int offt,
-                                    bool bcast = false);
-};
-
-}  // namespace gen
-}  // namespace jitkernel
-}  // namespace math
-}  // namespace operators
-}  // namespace paddle

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/jit_kernel.h"
-#include <iostream>
-#include <string>
-
-namespace paddle {
-namespace operators {
-namespace math {
-namespace jitkernel {
-
-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);
-}
-
-}  // 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/operators/math/jit_kernel_impl.h"
-#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 {
-
-// TODO(TJ): remove me
-typedef enum { kLT8, kEQ8, kGT8LT16, kEQ16, kGT16 } jit_block;
-
-class Kernel {
- public:
-  Kernel() = default;
-  virtual ~Kernel() = default;
-  // TODO(TJ): below members should be deprecated.
-  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:
-  void (*Compute)(const T *, const T *, T *, int);
-};
-
-template <typename T>
-class VAddKernel : public Kernel {
- public:
-  void (*Compute)(const T *, const T *, T *, int);
-};
-
-template <typename T>
-class VAddReluKernel : public Kernel {
- public:
-  void (*Compute)(const T *, const T *, T *, int);
-};
-
-template <typename T>
-class VScalKernel : public Kernel {
- public:
-  // y = a.*x
-  void (*Compute)(const T *, const T *, T *, int);
-};
-
-template <typename T>
-class VAddBiasKernel : public Kernel {
- public:
-  // y = a.+x
-  void (*Compute)(const T *, const T *, T *, int);
-};
-
-#ifdef PADDLE_WITH_MKLDNN
-template <typename T>
-class EltwiseMulnChw16cNCKernel : public Kernel {
- public:
-  // nChw16c = nChw16c .* NC
-  void (*Compute)(const float *, const float *, float *, int, int);
-};
-#endif
-
-template <typename T>
-class VActKernel : public Kernel {
- public:
-  void (*Compute)(const T *, T *, int);
-};
-
-template <typename T>
-class VReluKernel : public VActKernel<T> {};
-
-template <typename T>
-class VIdentityKernel : public VActKernel<T> {};
-
-template <typename T>
-class VExpKernel : public VActKernel<T> {};
-
-template <typename T>
-class VSigmoidKernel : public VActKernel<T> {};
-
-template <typename T>
-class VTanhKernel : public VActKernel<T> {};
-
-template <typename T>
-class LSTMKernel : public Kernel {
- public:
-  // compute c1 and h1 without c0 or h0
-  void (*ComputeC1H1)(lstm_t *, const lstm_attr_t *);
-  void (*ComputeCtHt)(lstm_t *, const lstm_attr_t *);
-};
-
-template <typename T>
-class GRUKernel : public Kernel {
- public:
-  // compute h1 without h0
-  void (*ComputeH1)(gru_t *, const gru_attr_t *);
-  void (*ComputeHtPart1)(gru_t *, const gru_attr_t *);
-  void (*ComputeHtPart2)(gru_t *, const gru_attr_t *);
-};
-
-template <typename T>
-class CRFDecodeKernel : public Kernel {
- public:
-  virtual void Compute(const int seq_len, const T *x, const T *w, T *alpha,
-                       int *track) const = 0;
-};
-
-template <typename T>
-class LayerNormKernel : public Kernel {
- public:
-  virtual void Compute(T *x, T *out, T *mean, T *var, const T *scale,
-                       const T *bias, int height,
-                       const float epsilon) 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"
-#include "paddle/fluid/operators/math/jit_kernel_refer.h"
-#include "paddle/fluid/platform/enforce.h"
-
-#ifdef PADDLE_WITH_XBYAK
-#include "paddle/fluid/operators/math/jit_code.h"
-#endif
-
-#ifdef PADDLE_WITH_MKLML
-#include "paddle/fluid/platform/dynload/mklml.h"
-#endif
-
-namespace paddle {
-namespace operators {
-namespace math {
-namespace jitkernel {
-
-#ifdef PADDLE_WITH_MKLML
-template <typename T>
-void VMulMKL(const T* x, const T* y, T* z, int n);
-
-template <>
-void VMulMKL<float>(const float* x, const float* y, float* z, int n) {
-  platform::dynload::vsMul(n, x, y, z);
-}
-
-template <>
-void VMulMKL<double>(const double* x, const double* y, double* z, int n) {
-  platform::dynload::vdMul(n, x, y, z);
-}
-
-template <typename T>
-void VAddMKL(const T* x, const T* y, T* z, int n);
-
-template <>
-void VAddMKL<float>(const float* x, const float* y, float* z, int n) {
-  platform::dynload::vsAdd(n, x, y, z);
-}
-
-template <>
-void VAddMKL<double>(const double* x, const double* y, double* z, int n) {
-  platform::dynload::vdAdd(n, x, y, z);
-}
-
-template <typename T>
-void VScalMKL(const T* a, const T* x, T* y, int n);
-
-template <>
-void VScalMKL<float>(const float* a, const float* x, float* y, int n) {
-  if (x == y) {
-    platform::dynload::cblas_sscal(n, *a, y, 1);
-  } else {
-    refer::VScal<float>(a, x, y, n);
-  }
-}
-
-template <>
-void VScalMKL<double>(const double* a, const double* x, double* y, int n) {
-  if (x == y) {
-    platform::dynload::cblas_dscal(n, *a, y, 1);
-  } else {
-    refer::VScal<double>(a, x, y, n);
-  }
-}
-
-#endif
-
-/* VMUL JitKernel */
-template <typename T>
-class VMulKernelImpl : public VMulKernel<T> {
- public:
-  JITKERNEL_DECLARE_STATIC_FUNC;
-  explicit VMulKernelImpl(int d) : VMulKernel<T>() {
-#ifdef PADDLE_WITH_XBYAK
-    if (useJIT(d)) {
-      // roughly estimate the size of code
-      size_t sz = 96 + d / YMM_FLOAT_BLOCK * 4 * 8;
-      jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::mul, 0, false,
-                                         sz > 4096 ? sz : 4096));
-      this->Compute =
-          jitcode_->getCode<void (*)(const T*, const T*, T*, int)>();
-      return;
-    }
-#endif
-#ifdef PADDLE_WITH_MKLML
-    if (useMKL(d)) {
-      this->Compute = VMulMKL<T>;
-      return;
-    }
-#endif
-    this->Compute = refer::VMul<T>;
-  }
-
-#ifdef PADDLE_WITH_XBYAK
-
- private:
-  std::unique_ptr<gen::VXXJitCode> jitcode_{nullptr};
-#endif
-};
-
-#ifdef PADDLE_WITH_XBYAK
-template <>
-bool VMulKernelImpl<float>::useJIT(int d) {
-  return gen::VXXJitCode::init(d);
-}
-#endif
-
-#ifdef PADDLE_WITH_MKLML
-template <>
-bool VMulKernelImpl<float>::useMKL(int d) {
-  return platform::MayIUse(platform::avx512f) && d > 512;
-}
-
-template <>
-bool VMulKernelImpl<double>::useMKL(int d) {
-  return true;
-}
-#endif
-
-/* VAdd JitKernel */
-template <typename T>
-class VAddKernelImpl : public VAddKernel<T> {
- public:
-  JITKERNEL_DECLARE_STATIC_FUNC;
-  explicit VAddKernelImpl(int d) : VAddKernel<T>() {
-#ifdef PADDLE_WITH_XBYAK
-    if (useJIT(d)) {
-      size_t sz = 96 + d / YMM_FLOAT_BLOCK * 4 * 8;
-      jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::add, 0, false,
-                                         sz > 4096 ? sz : 4096));
-      this->Compute =
-          jitcode_->getCode<void (*)(const T*, const T*, T*, int)>();
-      return;
-    }
-#endif
-#ifdef PADDLE_WITH_MKLML
-    if (useMKL(d)) {
-      this->Compute = VAddMKL<T>;
-      return;
-    }
-#endif
-    this->Compute = refer::VAdd<T>;
-  }
-#ifdef PADDLE_WITH_XBYAK
-
- private:
-  std::unique_ptr<gen::VXXJitCode> jitcode_{nullptr};
-#endif
-};
-
-#ifdef PADDLE_WITH_XBYAK
-template <>
-bool VAddKernelImpl<float>::useJIT(int d) {
-  return gen::VXXJitCode::init(d);
-}
-#endif
-
-#ifdef PADDLE_WITH_MKLML
-template <>
-bool VAddKernelImpl<float>::useMKL(int d) {
-  return d > 512;
-}
-
-template <>
-bool VAddKernelImpl<double>::useMKL(int d) {
-  return true;
-}
-#endif
-
-#ifdef PADDLE_WITH_MKLDNN
-/* EltwiseMul for nChw16c & NC inputs JitKernel */
-template <typename T>
-class EltwiseMulnChw16cNCKernelImpl
-    : public math::jitkernel::EltwiseMulnChw16cNCKernel<T> {
- public:
-  JITKERNEL_DECLARE_STATIC_FUNC;
-  explicit EltwiseMulnChw16cNCKernelImpl(int d)
-      : EltwiseMulnChw16cNCKernel<T>() {
-    using mul_func_t = void (*)(const float*, const float*, float*, int, int);
-#ifdef PADDLE_WITH_XBYAK
-    if (useJIT(d)) {
-      // roughly estimate the size of code
-      size_t sz = 96 + d / YMM_FLOAT_BLOCK * 4 * 8;
-      sz = sz > 4096 ? sz : 4096;
-      jitcode_.reset(new gen::EltwiseMulnChw16cNC(sz));
-      this->Compute = (mul_func_t)jitcode_->getCode();
-      return;
-    }
-#endif
-    PADDLE_THROW(
-        "This kernel shouldn't be used in Non-Xbyak, Non-MKL-DNN "
-        "environemnt");
-  }
-
-#ifdef PADDLE_WITH_XBYAK
-
- private:
-  std::unique_ptr<gen::EltwiseMulnChw16cNC> jitcode_{nullptr};
-};
-
-template <>
-bool EltwiseMulnChw16cNCKernelImpl<float>::useJIT(int d) {
-  return true;
-}
-#endif
-#endif
-
-/* VAddRelu JitKernel */
-template <typename T>
-class VAddReluKernelImpl : public VAddReluKernel<T> {
- public:
-  JITKERNEL_DECLARE_STATIC_FUNC;
-  explicit VAddReluKernelImpl(int d) : VAddReluKernel<T>() {
-#ifdef PADDLE_WITH_XBYAK
-    if (useJIT(d)) {
-      size_t sz = 96 + d / YMM_FLOAT_BLOCK * 4 * 8;
-      jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::add, 0, true,
-                                         sz > 4096 ? sz : 4096));
-      this->Compute =
-          jitcode_->getCode<void (*)(const T*, const T*, T*, int)>();
-      return;
-    }
-#endif
-    this->Compute = refer::VAddRelu<T>;
-  }
-#ifdef PADDLE_WITH_XBYAK
-
- private:
-  std::unique_ptr<gen::VXXJitCode> jitcode_{nullptr};
-#endif
-};
-
-#ifdef PADDLE_WITH_XBYAK
-template <>
-bool VAddReluKernelImpl<float>::useJIT(int d) {
-  return gen::VXXJitCode::init(d);
-}
-#endif
-
-/* VScal JitKernel */
-template <typename T>
-class VScalKernelImpl : public VScalKernel<T> {
- public:
-  JITKERNEL_DECLARE_STATIC_FUNC;
-  explicit VScalKernelImpl(int d) : VScalKernel<T>() {
-#ifdef PADDLE_WITH_XBYAK
-    if (useJIT(d)) {
-      size_t sz = 96 + d / YMM_FLOAT_BLOCK * 4 * 8;
-      jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::mul, 1, false,
-                                         sz > 4096 ? sz : 4096));
-      this->Compute =
-          jitcode_->getCode<void (*)(const T*, const T*, T*, int)>();
-      return;
-    }
-#endif
-#ifdef PADDLE_WITH_MKLML
-    if (useMKL(d)) {
-      this->Compute = VScalMKL<T>;
-      return;
-    }
-#endif
-    this->Compute = refer::VScal<T>;
-  }
-#ifdef PADDLE_WITH_XBYAK
-
- private:
-  std::unique_ptr<gen::VXXJitCode> jitcode_{nullptr};
-#endif
-};
-
-#ifdef PADDLE_WITH_XBYAK
-template <>
-bool VScalKernelImpl<float>::useJIT(int d) {
-  return gen::VXXJitCode::init(d, 1);
-}
-#endif
-
-#ifdef PADDLE_WITH_MKLML
-template <>
-bool VScalKernelImpl<float>::useMKL(int d) {
-  return d > 512;
-}
-template <>
-bool VScalKernelImpl<double>::useMKL(int d) {
-  return true;
-}
-#endif
-
-/* VAddBias JitKernel */
-template <typename T>
-class VAddBiasKernelImpl : public VAddBiasKernel<T> {
- public:
-  JITKERNEL_DECLARE_STATIC_FUNC;
-  explicit VAddBiasKernelImpl(int d) : VAddBiasKernel<T>() {
-#ifdef PADDLE_WITH_XBYAK
-    if (useJIT(d)) {
-      size_t sz = 96 + d / YMM_FLOAT_BLOCK * 4 * 8;
-      jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::add, 1, false,
-                                         sz > 4096 ? sz : 4096));
-      this->Compute =
-          jitcode_->getCode<void (*)(const T*, const T*, T*, int)>();
-      return;
-    }
-#endif
-
-    this->Compute = refer::VAddBias<T>;
-  }
-#ifdef PADDLE_WITH_XBYAK
-
- private:
-  std::unique_ptr<gen::VXXJitCode> jitcode_{nullptr};
-#endif
-};
-
-#ifdef PADDLE_WITH_XBYAK
-template <>
-bool VAddBiasKernelImpl<float>::useJIT(int d) {
-  return gen::VXXJitCode::init(d, 1);
-}
-#endif
-
-/* VRelu JitKernel */
-template <typename T>
-class VReluKernelImpl : public VReluKernel<T> {
- public:
-  JITKERNEL_DECLARE_STATIC_FUNC;
-  explicit VReluKernelImpl(int d) : VReluKernel<T>() {
-#ifdef PADDLE_WITH_XBYAK
-    if (useJIT(d)) {
-      size_t sz = 96 /* init size */ +
-                  d / YMM_FLOAT_BLOCK * 4 /* instructions */ *
-                      8 /* average bytes for each instruction */;
-      jitcode_.reset(new gen::VActJitCode(d, gen::operand_type::relu,
-                                          sz > 4096 ? sz : 4096));
-      this->Compute = jitcode_->getCode<void (*)(const T*, T*, int)>();
-      return;
-    }
-#endif
-
-    this->Compute = refer::VRelu<T>;
-  }
-#ifdef PADDLE_WITH_XBYAK
-
- private:
-  std::unique_ptr<gen::VActJitCode> jitcode_{nullptr};
-#endif
-};
-
-#ifdef PADDLE_WITH_XBYAK
-template <>
-bool VReluKernelImpl<float>::useJIT(int d) {
-  return gen::VActJitCode::init(d, gen::operand_type::relu);
-}
-#endif
-
-/* An empty JitKernel */
-template <typename T>
-class VIdentityKernelImpl : public VIdentityKernel<T> {
- public:
-  JITKERNEL_DECLARE_STATIC_FUNC;
-  explicit VIdentityKernelImpl(int d) : VIdentityKernel<T>() {
-    this->Compute = refer::VIdentity<T>;
-  }
-};
-
-REGISTER_JITKERNEL(vmul, VMulKernel);
-REGISTER_JITKERNEL(vadd, VAddKernel);
-REGISTER_JITKERNEL(vaddrelu, VAddReluKernel);
-REGISTER_JITKERNEL(vscal, VScalKernel);
-REGISTER_JITKERNEL(vaddbias, VAddBiasKernel);
-REGISTER_JITKERNEL(vrelu, VReluKernel);
-REGISTER_JITKERNEL(videntity, VIdentityKernel);
-#ifdef PADDLE_WITH_MKLDNN
-REGISTER_JITKERNEL(eltwise_mul_nchw16c, EltwiseMulnChw16cNCKernel);
-#endif
-
-}  // namespace jitkernel
-}  // namespace math
-}  // namespace operators
-}  // namespace paddle

paddle/fluid/operators/math/jit_kernel_crf_decode.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 <limits>
-#include <string>
-#include "paddle/fluid/operators/math/jit_kernel_macro.h"
-
-namespace paddle {
-namespace operators {
-namespace math {
-namespace jitkernel {
-
-/* CRF Decode JitKernel */
-template <typename T, platform::cpu_isa_t isa, jit_block>
-class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
- public:
-  explicit CRFDecodeKernelImpl(int tag_num) : CRFDecodeKernel<T>() {
-    this->num_ = tag_num;
-  }
-  void Compute(const int seq_len, const T* x, const T* w, T* alpha,
-               int* track) const override {
-    constexpr int state_trans_base_idx = 2;
-    for (int i = 0; i < this->num_; ++i) {
-      alpha[i] = w[i] + x[i];
-    }
-    for (int k = 1; k < seq_len; ++k) {
-      for (int i = 0; i < this->num_; ++i) {
-        T max_score = -std::numeric_limits<T>::max();
-        int max_j = 0;
-        for (int j = 0; j < this->num_; ++j) {
-          T score = alpha[(k - 1) * this->num_ + j] +
-                    w[(j + state_trans_base_idx) * this->num_ + i];
-          if (score > max_score) {
-            max_score = score;
-            max_j = j;
-          }
-        }
-        alpha[k * this->num_ + i] = max_score + x[k * this->num_ + i];
-        track[k * this->num_ + i] = max_j;
-      }
-    }
-  }
-};
-
-#define INIT_ALPHA(step_size)                                               \
-  /* Setup the alpha initial value.*/                                       \
-  int i_offset = 0;                                                         \
-  int last_offset = this->rest_ - step_size;                                \
-  for (int i = 0; i <= this->end_; ++i) {                                   \
-    /* weights, input and alpha values. */                                  \
-    __m256 w_content, x_content, alpha_content;                             \
-    /* Load the relevant data into the variables from un-aligned address.*/ \
-    w_content = _mm256_loadu_ps(w + i_offset);                              \
-    x_content = _mm256_loadu_ps(x + i_offset);                              \
-    alpha_content = _mm256_add_ps(w_content, x_content);                    \
-    _mm256_storeu_ps(alpha + i_offset, alpha_content);                      \
-    i_offset += step_size;                                                  \
-    if (i == this->end_ - 1) {                                              \
-      if (this->rest_ > 0) {                                                \
-        i_offset += last_offset;                                            \
-      } else {                                                              \
-        break;                                                              \
-      }                                                                     \
-    }                                                                       \
-  }
-
-#define UPDATE_ALPHA(step_size)                                               \
-  /* Update the alpha and track values. */                                    \
-  __m256 x_content = _mm256_loadu_ps(x + seq_offset + this->num_ + j_offset); \
-  max_score = _mm256_add_ps(max_score, x_content);                            \
-  _mm256_storeu_ps(alpha + seq_offset + this->num_ + j_offset, max_score);    \
-  _mm256_storeu_si256(                                                        \
-      reinterpret_cast<__m256i*>(track + seq_offset + this->num_ + j_offset), \
-      max_j);                                                                 \
-  /* Calculate the offset of next step*/                                      \
-  j_offset += step_size;                                                      \
-  if (j == this->end_ - 1) {                                                  \
-    if (this->rest_ > 0) {                                                    \
-      j_offset += last_offset;                                                \
-    } else {                                                                  \
-      break;                                                                  \
-    }                                                                         \
-  }
-
-#define INTRIAVX_FLOAT(block)                                                  \
-  template <>                                                                  \
-  CRFDecodeKernelImpl<float, platform::avx, block>::CRFDecodeKernelImpl(       \
-      int tag_num)                                                             \
-      : CRFDecodeKernel<float>() {                                             \
-    this->num_ = tag_num;                                                      \
-    this->end_ = this->num_ / YMM_FLOAT_BLOCK;                                 \
-    this->rest_ = this->num_ % YMM_FLOAT_BLOCK;                                \
-  }                                                                            \
-  template <>                                                                  \
-  void CRFDecodeKernelImpl<float, platform::avx, block>::Compute(              \
-      const int seq_len, const float* x, const float* w, float* alpha,         \
-      int* track) const {                                                      \
-    INIT_ALPHA(YMM_FLOAT_BLOCK)                                                \
-    /* Use the column-major strategy to get the location of maximum score.*/   \
-    int seq_offset = 0;                                                        \
-    constexpr int state_trans_base_idx = 2;                                    \
-    for (int k = 1; k < seq_len; ++k) {                                        \
-      int j_offset = 0;                                                        \
-      for (int j = 0; j <= this->end_; ++j) {                                  \
-        /* Initialize the variables of maximum score and location.*/           \
-        __m256 max_score = _mm256_set1_ps(-std::numeric_limits<float>::max()); \
-        __m256i max_j = _mm256_set1_epi32(0);                                  \
-        /* Calculate the offset of transition_weights.*/                       \
-        int trans_offset = state_trans_base_idx * this->num_ + j_offset;       \
-        for (int i = 0; i < this->num_; ++i) {                                 \
-          /* Initalize the content of alpha variable with related offset.*/    \
-          __m256 alpha_content = _mm256_broadcast_ss(alpha + seq_offset + i);  \
-          /* Obtain the content of weights from un-aligned address.*/          \
-          __m256 w_content = _mm256_loadu_ps(w + trans_offset);                \
-          __m256 score_v = _mm256_add_ps(alpha_content, w_content);            \
-          __m256 mask = _mm256_cmp_ps(score_v, max_score, _CMP_GT_OS);         \
-          /* According to the mask value, update the index of the max_score.*/ \
-          /* AVX instructions.*/                                               \
-          __m128i lo_max_j = _mm256_extractf128_si256(max_j, 0);               \
-          __m128i hi_max_j = _mm256_extractf128_si256(max_j, 1);               \
-          __m128i lo_mask = _mm256_extractf128_si256(*(__m256i*)&mask, 0);     \
-          __m128i hi_mask = _mm256_extractf128_si256(*(__m256i*)&mask, 1);     \
-          lo_max_j = _mm_andnot_si128(lo_mask, lo_max_j);                      \
-          hi_max_j = _mm_andnot_si128(hi_mask, hi_max_j);                      \
-          lo_mask = _mm_and_si128(lo_mask, _mm_set1_epi32(i));                 \
-          hi_mask = _mm_and_si128(hi_mask, _mm_set1_epi32(i));                 \
-          lo_max_j = _mm_or_si128(lo_mask, lo_max_j);                          \
-          hi_max_j = _mm_or_si128(hi_mask, hi_max_j);                          \
-          max_j = _mm256_insertf128_si256(max_j, lo_max_j, 0);                 \
-          max_j = _mm256_insertf128_si256(max_j, hi_max_j, 1);                 \
-          /* AVX done*/                                                        \
-          /* Update the max_score value.*/                                     \
-          max_score = _mm256_max_ps(max_score, score_v);                       \
-          trans_offset += this->num_;                                          \
-        }                                                                      \
-        UPDATE_ALPHA(YMM_FLOAT_BLOCK)                                          \
-      }                                                                        \
-      seq_offset += this->num_;                                                \
-    }                                                                          \
-  }
-
-#define INTRIAVX2_FLOAT(isa, block)                                            \
-  template <>                                                                  \
-  CRFDecodeKernelImpl<float, isa, block>::CRFDecodeKernelImpl(int tag_num)     \
-      : CRFDecodeKernel<float>() {                                             \
-    this->num_ = tag_num;                                                      \
-    this->end_ = this->num_ / YMM_FLOAT_BLOCK;                                 \
-    this->rest_ = this->num_ % YMM_FLOAT_BLOCK;                                \
-  }                                                                            \
-  template <>                                                                  \
-  void CRFDecodeKernelImpl<float, isa, block>::Compute(                        \
-      const int seq_len, const float* x, const float* w, float* alpha,         \
-      int* track) const {                                                      \
-    INIT_ALPHA(YMM_FLOAT_BLOCK)                                                \
-    /* Use the column-major strategy to get the location of maximum score.*/   \
-    int seq_offset = 0;                                                        \
-    constexpr int state_trans_base_idx = 2;                                    \
-    for (int k = 1; k < seq_len; ++k) {                                        \
-      int j_offset = 0;                                                        \
-      for (int j = 0; j <= this->end_; ++j) {                                  \
-        /* Initialize the variables of maximum score and location.*/           \
-        __m256 max_score = _mm256_set1_ps(-std::numeric_limits<float>::max()); \
-        __m256i max_j = _mm256_set1_epi32(0);                                  \
-        /* Calculate the offset of transition_weights.*/                       \
-        int trans_offset = state_trans_base_idx * this->num_ + j_offset;       \
-        for (int i = 0; i < this->num_; ++i) {                                 \
-          /* Initalize the content of alpha variable with related offset.*/    \
-          __m256 alpha_content = _mm256_broadcast_ss(alpha + seq_offset + i);  \
-          /* Obtain the content of weights from un-aligned address.*/          \
-          __m256 w_content = _mm256_loadu_ps(w + trans_offset);                \
-          __m256 score_v = _mm256_add_ps(alpha_content, w_content);            \
-          __m256 mask = _mm256_cmp_ps(score_v, max_score, _CMP_GT_OS);         \
-          /* According to the mask value, update the index of the max_score.*/ \
-          /* AVX2 instructions.*/                                              \
-          max_j = _mm256_or_si256(                                             \
-              _mm256_andnot_si256((__m256i)mask, max_j),                       \
-              _mm256_and_si256((__m256i)mask, _mm256_set1_epi32(i)));          \
-          /* Update the max_score value.*/                                     \
-          max_score = _mm256_max_ps(max_score, score_v);                       \
-          trans_offset += this->num_;                                          \
-        }                                                                      \
-        UPDATE_ALPHA(YMM_FLOAT_BLOCK)                                          \
-      }                                                                        \
-      seq_offset += this->num_;                                                \
-    }                                                                          \
-  }
-
-#define INTRIAVX512_FLOAT(block)                                               \
-  template <>                                                                  \
-  CRFDecodeKernelImpl<float, platform::avx512f, block>::CRFDecodeKernelImpl(   \
-      int tag_num)                                                             \
-      : CRFDecodeKernel<float>() {                                             \
-    this->num_ = tag_num;                                                      \
-    this->end_ = this->num_ / ZMM_FLOAT_BLOCK;                                 \
-    this->rest_ = this->num_ % ZMM_FLOAT_BLOCK;                                \
-  }                                                                            \
-  template <>                                                                  \
-  void CRFDecodeKernelImpl<float, platform::avx512f, block>::Compute(          \
-      const int seq_len, const float* x, const float* w, float* alpha,         \
-      int* track) const {                                                      \
-    INIT_ALPHA(ZMM_FLOAT_BLOCK)                                                \
-    /* Use the column-major strategy to get the location of maximum score.*/   \
-    int seq_offset = 0;                                                        \
-    constexpr int state_trans_base_idx = 2;                                    \
-    for (int k = 1; k < seq_len; ++k) {                                        \
-      int j_offset = 0;                                                        \
-      for (int j = 0; j <= this->end_; ++j) {                                  \
-        /* Initialize the variables of maximum score and location.*/           \
-        __m512 max_score = _mm512_set1_ps(-std::numeric_limits<float>::max()); \
-        __m512i max_j = _mm512_setzero_si512();                                \
-        /* Calculate the offset of transition_weights.*/                       \
-        int trans_offset = state_trans_base_idx * this->num_ + j_offset;       \
-        for (int i = 0; i < this->num_; ++i) {                                 \
-          /* Initalize the content of alpha variable with related offset.*/    \
-          __m512 alpha_content = _mm512_set1_ps(*(alpha + seq_offset + i));    \
-          /* Obtain the content of weights from un-aligned address.*/          \
-          __m512 w_content = _mm512_loadu_ps(w + trans_offset);                \
-          __m512 score_v = _mm512_add_ps(alpha_content, w_content);            \
-          __mmask16 mask = _mm512_cmp_ps_mask(score_v, max_score, _CMP_GT_OS); \
-          /* AVX512 instructions.*/                                            \
-          max_j = _mm512_mask_set1_epi32(max_j, mask, i);                      \
-          /* Update the max_score value.*/                                     \
-          max_score = _mm512_max_ps(max_score, score_v);                       \
-          trans_offset += this->num_;                                          \
-        }                                                                      \
-        /* Update the alpha and track values.*/                                \
-        __m512 x_content =                                                     \
-            _mm512_loadu_ps(x + seq_offset + this->num_ + j_offset);           \
-        max_score = _mm512_add_ps(max_score, x_content);                       \
-        _mm512_storeu_ps(alpha + seq_offset + this->num_ + j_offset,           \
-                         max_score);                                           \
-        _mm512_storeu_si512(reinterpret_cast<__m512i*>(track + seq_offset +    \
-                                                       this->num_ + j_offset), \
-                            max_j);                                            \
-        /* Calculate the offset of next step*/                                 \
-        j_offset += ZMM_FLOAT_BLOCK;                                           \
-        if (j == this->end_ - 1) {                                             \
-          if (this->rest_ > 0) {                                               \
-            j_offset += last_offset;                                           \
-          } else {                                                             \
-            break;                                                             \
-          }                                                                    \
-        }                                                                      \
-      }                                                                        \
-      seq_offset += this->num_;                                                \
-    }                                                                          \
-  }
-
-#ifdef __AVX__
-INTRIAVX_FLOAT(kEQ8);
-INTRIAVX_FLOAT(kGT8LT16);
-INTRIAVX_FLOAT(kEQ16);
-INTRIAVX_FLOAT(kGT16);
-#endif
-#ifdef __AVX2__
-INTRIAVX2_FLOAT(platform::avx2, kEQ8);
-INTRIAVX2_FLOAT(platform::avx2, kGT8LT16);
-INTRIAVX2_FLOAT(platform::avx2, kEQ16);
-INTRIAVX2_FLOAT(platform::avx2, kGT16);
-#endif
-#ifdef __AVX512F__
-INTRIAVX2_FLOAT(platform::avx512f, kEQ8);
-INTRIAVX2_FLOAT(platform::avx512f, kGT8LT16);
-INTRIAVX512_FLOAT(kEQ16);
-INTRIAVX512_FLOAT(kGT16);
-#endif
-
-#undef INTRIAVX512_FLOAT
-#undef INTRIAVX2_FLOAT
-#undef INTRIAVX_FLOAT
-#undef INIT_ALPHA
-#undef UPDATE_ALPHA
-
-REGISTER_JITKERNEL_DEPRECATED(crf_decode, CRFDecodeKernel);
-
-}  // 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 <string>
-#include "paddle/fluid/operators/math/jit_kernel_macro.h"
-#include "paddle/fluid/operators/math/jit_kernel_refer.h"
-
-#ifdef PADDLE_WITH_XBYAK
-#include "paddle/fluid/operators/math/jit_code.h"
-#endif
-
-#ifdef PADDLE_WITH_MKLML
-#include "paddle/fluid/platform/dynload/mklml.h"
-#endif
-
-namespace paddle {
-namespace operators {
-namespace math {
-namespace jitkernel {
-
-#ifdef PADDLE_WITH_MKLML
-// try to use MKL to speedup
-template <typename T>
-void VExpMKL(const T* x, T* y, int n);
-
-template <>
-void VExpMKL<float>(const float* x, float* y, int n) {
-  platform::dynload::vsExp(n, x, y);
-}
-
-template <>
-void VExpMKL<double>(const double* x, double* y, int n) {
-  platform::dynload::vdExp(n, x, y);
-}
-
-template <typename T>
-void VSigmoidMKL(const T* x, T* y, int n) {
-  const T min = SIGMOID_THRESHOLD_MIN;
-  const T 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] = static_cast<T>(0) - y[i];
-  }
-  VExpMKL(y, y, n);
-  for (int i = 0; i < n; ++i) {
-    y[i] = static_cast<T>(1) / (static_cast<T>(1) + y[i]);
-  }
-}
-
-template <typename T>
-void VTanhMKL(const T* x, T* y, int n) {
-  for (int i = 0; i < n; ++i) {
-    y[i] = static_cast<T>(2) * x[i];
-  }
-  VSigmoidMKL(y, y, n);
-  for (int i = 0; i < n; ++i) {
-    y[i] = static_cast<T>(2) * y[i] - static_cast<T>(1);
-  }
-}
-#endif
-
-/* VExp JitKernel */
-template <typename T>
-class VExpKernelImpl : public VExpKernel<T> {
- public:
-  JITKERNEL_DECLARE_STATIC_FUNC;
-  explicit VExpKernelImpl(int d) : VExpKernel<T>() {
-#ifdef PADDLE_WITH_XBYAK
-    if (useJIT(d)) {
-      size_t sz = 96 + d / YMM_FLOAT_BLOCK * 70 * 8;
-      jitcode_.reset(new gen::VActJitCode(d, gen::operand_type::exp,
-                                          sz > 4096 ? sz : 4096));
-      this->Compute = jitcode_->getCode<void (*)(const T*, T*, int)>();
-      return;
-    }
-#endif
-#ifdef PADDLE_WITH_MKLML
-    if (useMKL(d)) {
-      this->Compute = VExpMKL<T>;
-      return;
-    }
-#endif
-    this->Compute = refer::VExp<T>;
-  }
-
-#ifdef PADDLE_WITH_XBYAK
-
- private:
-  std::unique_ptr<gen::VActJitCode> jitcode_{nullptr};
-#endif
-};
-
-#ifdef PADDLE_WITH_XBYAK
-template <>
-bool VExpKernelImpl<float>::useJIT(int d) {
-  return gen::VActJitCode::init(d, gen::operand_type::exp);
-}
-#endif
-
-#ifdef PADDLE_WITH_MKLML
-template <>
-bool VExpKernelImpl<float>::useMKL(int d) {
-  return d > 512;
-}
-
-template <>
-bool VExpKernelImpl<double>::useMKL(int d) {
-  return true;
-}
-
-#endif
-
-/* VSigmoid JitKernel */
-template <typename T>
-class VSigmoidKernelImpl : public VSigmoidKernel<T> {
- public:
-  JITKERNEL_DECLARE_STATIC_FUNC;
-  explicit VSigmoidKernelImpl(int d) : VSigmoidKernel<T>() {
-#ifdef PADDLE_WITH_XBYAK
-    if (useJIT(d)) {
-      size_t sz = 96 + d / YMM_FLOAT_BLOCK * 82 * 8;
-      jitcode_.reset(new gen::VActJitCode(d, gen::operand_type::sigmoid,
-                                          sz > 4096 ? sz : 4096));
-      this->Compute = jitcode_->getCode<void (*)(const T*, T*, int)>();
-      return;
-    }
-#endif
-
-#ifdef PADDLE_WITH_MKLML
-    // strictly it's a better impl with MKL, then is refer
-    if (useMKL(d)) {
-      this->Compute = VSigmoidMKL<T>;
-      return;
-    }
-#endif
-    this->Compute = refer::VSigmoid<T>;
-  }
-
-#ifdef PADDLE_WITH_XBYAK
-
- private:
-  std::unique_ptr<gen::VActJitCode> jitcode_{nullptr};
-#endif
-};
-
-#ifdef PADDLE_WITH_XBYAK
-template <>
-bool VSigmoidKernelImpl<float>::useJIT(int d) {
-  return gen::VActJitCode::init(d, gen::operand_type::sigmoid);
-}
-#endif
-
-#ifdef PADDLE_WITH_MKLML
-template <>
-bool VSigmoidKernelImpl<float>::useMKL(int d) {
-  return d > 512;
-}
-
-template <>
-bool VSigmoidKernelImpl<double>::useMKL(int d) {
-  return true;
-}
-#endif
-
-/* VTanh JitKernel */
-template <typename T>
-class VTanhKernelImpl : public VTanhKernel<T> {
- public:
-  JITKERNEL_DECLARE_STATIC_FUNC;
-  explicit VTanhKernelImpl(int d) : VTanhKernel<T>() {
-#ifdef PADDLE_WITH_XBYAK
-    if (useJIT(d)) {
-      size_t sz = 96 + d / YMM_FLOAT_BLOCK * 84 * 8;
-      jitcode_.reset(new gen::VActJitCode(d, gen::operand_type::tanh,
-                                          sz > 4096 ? sz : 4096));
-      this->Compute = jitcode_->getCode<void (*)(const T*, T*, int)>();
-      return;
-    }
-#endif
-
-#ifdef PADDLE_WITH_MKLML
-    // strictly it's a better impl with MKL, then is refer
-    if (useMKL(d)) {
-      this->Compute = VTanhMKL<T>;
-      return;
-    }
-#endif
-    this->Compute = refer::VTanh<T>;
-  }
-
-#ifdef PADDLE_WITH_XBYAK
-
- private:
-  std::unique_ptr<gen::VActJitCode> jitcode_{nullptr};
-#endif
-};
-
-#ifdef PADDLE_WITH_XBYAK
-template <>
-bool VTanhKernelImpl<float>::useJIT(int d) {
-  return gen::VActJitCode::init(d, gen::operand_type::tanh);
-}
-#endif
-
-#ifdef PADDLE_WITH_MKLML
-template <>
-bool VTanhKernelImpl<float>::useMKL(int d) {
-  return d > 512;
-}
-
-template <>
-bool VTanhKernelImpl<double>::useMKL(int d) {
-  return true;
-}
-#endif
-
-REGISTER_JITKERNEL(vexp, VExpKernel);
-REGISTER_JITKERNEL(vsigmoid, VSigmoidKernel);
-REGISTER_JITKERNEL(vtanh, VTanhKernel);
-
-}  // namespace jitkernel
-}  // namespace math
-}  // namespace operators
-}  // namespace paddle

paddle/fluid/operators/math/jit_kernel_impl.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 <type_traits>
-
-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 XMM_FLOAT_BLOCK 4
-#define YMM_FLOAT_BLOCK 8
-#define ZMM_FLOAT_BLOCK 16
-
-typedef struct {
-  void* gates;  // gates: W_ch, W_ih, W_fh, W_oh
-  const void* ct_1;
-  void* ct;
-  void* ht;
-  /* weight_peephole and checked data are only used in peephole*/
-  const void* wp{nullptr};
-  void* checked{nullptr};
-} lstm_t;
-
-typedef struct {
-  void* gates;  // gates: {W_update, W_reset; W_state}
-  const void* ht_1;
-  void* ht;
-} gru_t;
-
-struct rnn_attr_s {
-  int d;
-  std::string act_gate, act_cand;
-  rnn_attr_s() = default;
-  rnn_attr_s(int _d, const std::string& _act_gate, const std::string& _act_cand)
-      : d(_d), act_gate(_act_gate), act_cand(_act_cand) {}
-};
-
-struct lstm_attr_s : public rnn_attr_s {
-  bool use_peephole;
-  std::string act_cell;
-  lstm_attr_s() = default;
-  lstm_attr_s(int _d, const std::string& _act_gate,
-              const std::string& _act_cand, const std::string& _act_cell,
-              bool _use_peephole = false)
-      : rnn_attr_s(_d, _act_gate, _act_cand),
-        use_peephole(_use_peephole),
-        act_cell(_act_cell) {}
-};
-
-typedef struct rnn_attr_s gru_attr_t;
-typedef struct lstm_attr_s lstm_attr_t;
-
-}  // namespace jitkernel
-}  // namespace math
-}  // namespace operators
-}  // namespace paddle

paddle/fluid/operators/math/jit_kernel_layer_norm.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 <math.h>
-#include <limits>
-#include <string>
-#include "paddle/fluid/operators/math/jit_kernel_macro.h"
-
-namespace paddle {
-namespace operators {
-namespace math {
-namespace jitkernel {
-
-/* Layer Norm JitKernel */
-template <typename T, platform::cpu_isa_t isa, jit_block>
-class LayerNormKernelImpl : public LayerNormKernel<T> {
- public:
-  explicit LayerNormKernelImpl(int right) : LayerNormKernel<T>() {
-    this->num_ = right;
-  }
-
-  void Compute(T* x, T* out, T* mean, T* var, const T* scale, const T* bias,
-               int height, const float epsilon) const override {
-    // get mean
-    for (int i = 0; i < height; i++) {
-      T sum = 0.0;
-      int offset = i * this->num_;
-      for (int j = 0; j < this->num_; j++) {
-        sum += x[offset + j];
-      }
-      mean[i] = sum / this->num_;
-    }
-
-    // get variance
-    for (int i = 0; i < height; i++) {
-      T sum = 0.0;
-      int offset = i * this->num_;
-      for (int j = 0; j < this->num_; j++) {
-        sum += (x[offset + j] - mean[i]) * (x[offset + j] - mean[i]);
-      }
-      var[i] = sum / this->num_;
-    }
-
-    for (int i = 0; i < height; i++) {
-      int offset = i * this->num_;
-      T sqrt_var = sqrt(var[i] + (T)epsilon);
-      for (int j = 0; j < this->num_; j++) {
-        out[offset + j] = (x[offset + j] - mean[i]) / sqrt_var;
-      }
-    }
-    if (scale) {
-      for (int i = 0; i < height; i++) {
-        int offset = i * this->num_;
-        for (int j = 0; j < this->num_; j++) {
-          out[offset + j] *= scale[j];
-        }
-      }
-    }
-
-    if (bias) {
-      for (int i = 0; i < height; i++) {
-        int offset = i * this->num_;
-        for (int j = 0; j < this->num_; j++) {
-          out[offset + j] += bias[j];
-        }
-      }
-    }
-  }
-};
-
-#define INTRIAVX_FLOAT(isa, jit_block)                                         \
-  template <>                                                                  \
-  LayerNormKernelImpl<float, isa, jit_block>::LayerNormKernelImpl(int right)   \
-      : LayerNormKernel<float>() {                                             \
-    this->num_ = right;                                                        \
-    this->rest_ = this->num_ % YMM_FLOAT_BLOCK;                                \
-    this->end_ = this->num_ - this->rest_;                                     \
-  }                                                                            \
-  template <>                                                                  \
-  void LayerNormKernelImpl<float, isa, jit_block>::Compute(                    \
-      float* x, float* out, float* mean, float* var, const float* scale,       \
-      const float* bias, int height, const float epsilon) const {              \
-    __m256 sum;                                                                \
-    __m256 mean_vec, var_vec;                                                  \
-    __m128 hi, lo;                                                             \
-    __m256 tmp;                                                                \
-    size_t offset;                                                             \
-    size_t j;                                                                  \
-    size_t block = YMM_FLOAT_BLOCK;                                            \
-    __m256 reverse_num_vec =                                                   \
-        _mm256_div_ps(_mm256_set1_ps(1.0), _mm256_set1_ps(this->num_));        \
-    __m256 epsilon_vec = _mm256_set1_ps(epsilon);                              \
-    int rest_mask =                                                            \
-        ((-1) & (~((~0U) >> (sizeof(int) * 8 - (YMM_FLOAT_BLOCK - rest_))))) & \
-        0x0ff;                                                                 \
-    __m256i mask_vec = _mm256_set_epi32(                                       \
-        rest_mask & 0x80 ? 0xffffffff : 0, rest_mask & 0x40 ? 0xffffffff : 0,  \
-        rest_mask & 0x20 ? 0xffffffff : 0, rest_mask & 0x10 ? 0xffffffff : 0,  \
-        rest_mask & 0x8 ? 0xffffffff : 0, rest_mask & 0x4 ? 0xffffffff : 0,    \
-        rest_mask & 0x2 ? 0xffffffff : 0, rest_mask & 0x1 ? 0xffffffff : 0);   \
-                                                                               \
-    for (int i = 0; i < height; ++i) {                                         \
-      offset = i * this->num_;                                                 \
-                                                                               \
-      /* get mean */                                                           \
-      sum = _mm256_setzero_ps();                                               \
-      for (j = offset; j < end_ + offset; j += block) {                        \
-        sum = _mm256_add_ps(sum, _mm256_loadu_ps((const float*)x + j));        \
-      }                                                                        \
-      if (rest_ != 0) {                                                        \
-        j = offset + this->num_ - block;                                       \
-        tmp = _mm256_loadu_ps((const float*)x + j);                            \
-        tmp = _mm256_blendv_ps(_mm256_setzero_ps(), tmp, *(__m256*)&mask_vec); \
-        sum = _mm256_add_ps(sum, tmp);                                         \
-      }                                                                        \
-      hi = _mm256_extractf128_ps(sum, 1);                                      \
-      lo = _mm256_extractf128_ps(sum, 0);                                      \
-      sum = _mm256_add_ps(                                                     \
-          sum, _mm256_insertf128_ps(                                           \
-                   _mm256_insertf128_ps(_mm256_setzero_ps(), hi, 0), lo, 1));  \
-      sum = _mm256_hadd_ps(sum, sum);                                          \
-      sum = _mm256_hadd_ps(sum, sum);                                          \
-      mean_vec = _mm256_mul_ps(sum, reverse_num_vec);                          \
-      mean[i] = *reinterpret_cast<float*>(&mean_vec);                          \
-                                                                               \
-      /* get variance */                                                       \
-      sum = _mm256_setzero_ps();                                               \
-      for (j = offset; j < end_ + offset; j += block) {                        \
-        tmp = _mm256_sub_ps(_mm256_loadu_ps((const float*)x + j), mean_vec);   \
-        tmp = _mm256_mul_ps(tmp, tmp);                                         \
-        sum = _mm256_add_ps(sum, tmp);                                         \
-      }                                                                        \
-      if (rest_ != 0) {                                                        \
-        j = offset + this->num_ - block;                                       \
-        tmp = _mm256_sub_ps(_mm256_loadu_ps((const float*)x + j), mean_vec);   \
-        tmp = _mm256_mul_ps(tmp, tmp);                                         \
-        tmp = _mm256_blendv_ps(_mm256_setzero_ps(), tmp, *(__m256*)&mask_vec); \
-        sum = _mm256_add_ps(sum, tmp);                                         \
-      }                                                                        \
-      hi = _mm256_extractf128_ps(sum, 1);                                      \
-      lo = _mm256_extractf128_ps(sum, 0);                                      \
-      sum = _mm256_add_ps(                                                     \
-          sum, _mm256_insertf128_ps(                                           \
-                   _mm256_insertf128_ps(_mm256_setzero_ps(), hi, 0), lo, 1));  \
-      sum = _mm256_hadd_ps(sum, sum);                                          \
-      sum = _mm256_hadd_ps(sum, sum);                                          \
-      var_vec = _mm256_mul_ps(sum, reverse_num_vec);                           \
-      var[i] = *reinterpret_cast<float*>(&var_vec);                            \
-                                                                               \
-      /* get x_norm and calculate output*/                                     \
-      for (j = offset; j < end_ + offset; j += block) {                        \
-        tmp = _mm256_sub_ps(_mm256_loadu_ps((const float*)x + j), mean_vec);   \
-        tmp = _mm256_div_ps(                                                   \
-            tmp, _mm256_sqrt_ps(_mm256_add_ps(var_vec, epsilon_vec)));         \
-        _mm256_storeu_ps(reinterpret_cast<float*>(out) + j, tmp);              \
-      }                                                                        \
-      if (rest_ != 0) {                                                        \
-        j = offset + num_ - block;                                             \
-        tmp = _mm256_sub_ps(_mm256_loadu_ps((const float*)x + j), mean_vec);   \
-        tmp = _mm256_div_ps(                                                   \
-            tmp, _mm256_sqrt_ps(_mm256_add_ps(var_vec, epsilon_vec)));         \
-        _mm256_storeu_ps(reinterpret_cast<float*>(out) + j, tmp);              \
-      }                                                                        \
-                                                                               \
-      if (scale) {                                                             \
-        if (rest_ != 0) {                                                      \
-          j = offset + this->num_ - block;                                     \
-          tmp = _mm256_loadu_ps((const float*)out + j);                        \
-        }                                                                      \
-        for (j = offset; j < end_ + offset; j += block) {                      \
-          _mm256_storeu_ps(                                                    \
-              reinterpret_cast<float*>(out) + j,                               \
-              _mm256_mul_ps(                                                   \
-                  _mm256_loadu_ps((const float*)out + j),                      \
-                  _mm256_loadu_ps((const float*)scale + j - offset)));         \
-        }                                                                      \
-        if (rest_ != 0) {                                                      \
-          j = offset + this->num_ - block;                                     \
-          _mm256_storeu_ps(                                                    \
-              reinterpret_cast<float*>(out) + j,                               \
-              _mm256_mul_ps(                                                   \
-                  tmp, _mm256_loadu_ps((const float*)scale + j - offset)));    \
-        }                                                                      \
-      }                                                                        \
-                                                                               \
-      if (bias) {                                                              \
-        if (rest_ != 0) {                                                      \
-          j = offset + this->num_ - block;                                     \
-          tmp = _mm256_loadu_ps((const float*)out + j);                        \
-        }                                                                      \
-        for (j = offset; j < end_ + offset; j += block) {                      \
-          _mm256_storeu_ps(                                                    \
-              reinterpret_cast<float*>(out) + j,                               \
-              _mm256_add_ps(                                                   \
-                  _mm256_loadu_ps((const float*)out + j),                      \
-                  _mm256_loadu_ps((const float*)bias + j - offset)));          \
-        }                                                                      \
-        if (rest_ != 0) {                                                      \
-          j = offset + this->num_ - block;                                     \
-          _mm256_storeu_ps(                                                    \
-              reinterpret_cast<float*>(out) + j,                               \
-              _mm256_add_ps(                                                   \
-                  tmp, _mm256_loadu_ps((const float*)bias + j - offset)));     \
-        }                                                                      \
-      }                                                                        \
-    }                                                                          \
-  }
-
-#ifdef __AVX__
-INTRIAVX_FLOAT(platform::avx, kEQ8);
-INTRIAVX_FLOAT(platform::avx, kGT8LT16);
-INTRIAVX_FLOAT(platform::avx, kEQ16);
-INTRIAVX_FLOAT(platform::avx, kGT16);
-INTRIAVX_FLOAT(platform::avx2, kEQ8);
-INTRIAVX_FLOAT(platform::avx2, kGT8LT16);
-INTRIAVX_FLOAT(platform::avx2, kEQ16);
-INTRIAVX_FLOAT(platform::avx2, kGT16);
-INTRIAVX_FLOAT(platform::avx512f, kEQ8);
-INTRIAVX_FLOAT(platform::avx512f, kGT8LT16);
-INTRIAVX_FLOAT(platform::avx512f, kEQ16);
-INTRIAVX_FLOAT(platform::avx512f, kGT16);
-#endif
-
-#undef INTRIAVX_FLOAT
-
-REGISTER_JITKERNEL_DEPRECATED(layer_norm, LayerNormKernel);
-
-}  // 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"
-#include "paddle/fluid/platform/enforce.h"
-
-namespace paddle {
-namespace operators {
-namespace math {
-namespace jitkernel {
-
-#define JITKERNEL_DECLARE_STATIC_FUNC                       \
-  static inline std::string name(int d) {                   \
-    PADDLE_THROW("DType should be either float or double"); \
-  }                                                         \
-  static inline bool useJIT(int d) { return false; }        \
-  static inline bool useMKL(int d) { return false; }
-
-#define JITKERNEL_DEFINE_NAME(ker_key, ker_class)    \
-  template <>                                        \
-  std::string ker_class##Impl<float>::name(int d) {  \
-    std::string key(#ker_key "f");                   \
-    if (useJIT(d)) {                                 \
-      /* only jit code need record d*/               \
-      return key + "jit" + std::to_string(d);        \
-    } else if (useMKL(d)) {                          \
-      return key + "mkl";                            \
-    } else {                                         \
-      return key + "any";                            \
-    }                                                \
-  }                                                  \
-  template <>                                        \
-  std::string ker_class##Impl<double>::name(int d) { \
-    std::string key(#ker_key "d");                   \
-    /* jit code do not support double yet*/          \
-    if (useMKL(d)) {                                 \
-      return key + "mkl";                            \
-    } else {                                         \
-      return key + "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_FIND_KEY(ker_class, ker_dtype) \
-  std::string key = ker_class##Impl<ker_dtype>::name(d)
-
-#define JITKERNEL_IMPL(ker_class, ker_dtype)           \
-  p = std::dynamic_pointer_cast<ker_class<ker_dtype>>( \
-      std::make_shared<ker_class##Impl<ker_dtype>>(d))
-
-#define REGISTER_JITKERNEL_WITH_DTYPE(ker_class, ker_dtype, marco_declare, \
-                                      macro_find_key, macro_impl)          \
-  marco_declare(ker_class, ker_dtype) {                                    \
-    macro_find_key(ker_class, ker_dtype);                                  \
-    if (kers_.find(key) == kers_.end()) {                                  \
-      std::shared_ptr<ker_class<ker_dtype>> p;                             \
-      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_ARGS(ker_key, ker_class, marco_define_name,     \
-                                marco_declare, macro_find_key, macro_impl) \
-  marco_define_name(ker_key, ker_class);                                   \
-  REGISTER_JITKERNEL_WITH_DTYPE(ker_class, float, marco_declare,           \
-                                macro_find_key, macro_impl);               \
-  REGISTER_JITKERNEL_WITH_DTYPE(ker_class, double, marco_declare,          \
-                                macro_find_key, macro_impl)
-
-#define REGISTER_JITKERNEL(ker_key, ker_class)                       \
-  REGISTER_JITKERNEL_ARGS(ker_key, ker_class, JITKERNEL_DEFINE_NAME, \
-                          JITKERNEL_DECLARE, JITKERNEL_FIND_KEY,     \
-                          JITKERNEL_IMPL)
-
-// TODO(TJ): below defines are deprecated, would be remove recently
-#define SEARCH_BLOCK(macro_, ker, dtype, isa)              \
-  if (d < YMM_FLOAT_BLOCK) {                               \
-    macro_(ker, dtype, isa, kLT8);                         \
-  } else if (d == YMM_FLOAT_BLOCK) {                       \
-    macro_(ker, dtype, isa, kEQ8);                         \
-  } else if (d > YMM_FLOAT_BLOCK && d < ZMM_FLOAT_BLOCK) { \
-    macro_(ker, dtype, isa, kGT8LT16);                     \
-  } else if (d == ZMM_FLOAT_BLOCK) {                       \
-    macro_(ker, dtype, isa, kEQ16);                        \
-  } else {                                                 \
-    macro_(ker, dtype, isa, kGT16);                        \
-  }
-
-#define SEARCH_ISA_BLOCK(macro_, ker, dtype)             \
-  if (platform::MayIUse(platform::avx512f)) {            \
-    SEARCH_BLOCK(macro_, ker, dtype, platform::avx512f); \
-  } else if (platform::MayIUse(platform::avx2)) {        \
-    SEARCH_BLOCK(macro_, ker, dtype, platform::avx2);    \
-  } else if (platform::MayIUse(platform::avx)) {         \
-    SEARCH_BLOCK(macro_, ker, dtype, platform::avx);     \
-  } else {                                               \
-    SEARCH_BLOCK(macro_, ker, dtype, platform::isa_any); \
-  }
-
-#define JITKERNEL_KEY(ker_key, dtype_key) \
-  #ker_key #dtype_key + std::to_string(d)
-
-#define JITKERNEL_NEW_IMPL_DEPRECATED(ker, dtype, isa, k) \
-  p = std::dynamic_pointer_cast<ker<dtype>>(              \
-      std::make_shared<ker##Impl<dtype, isa, k>>(d))
-
-#define JITKERNEL_WITH_DTYPE_DEPRECATED(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_DEPRECATED(ker_key, ker_class)           \
-  JITKERNEL_WITH_DTYPE_DEPRECATED(ker_key, ker_class, float, f,     \
-                                  JITKERNEL_DECLARE, JITKERNEL_KEY, \
-                                  JITKERNEL_NEW_IMPL_DEPRECATED);   \
-  JITKERNEL_WITH_DTYPE_DEPRECATED(ker_key, ker_class, double, d,    \
-                                  JITKERNEL_DECLARE, JITKERNEL_KEY, \
-                                  JITKERNEL_NEW_IMPL_DEPRECATED)
-
-#define REGISTER_JITKERNEL_ARGS_DEPRECATED(ker_key, ker_class, marco_declare,  \
-                                           macro_key, macro_impl)              \
-  JITKERNEL_WITH_DTYPE_DEPRECATED(ker_key, ker_class, float, f, marco_declare, \
-                                  macro_key, macro_impl);                      \
-  JITKERNEL_WITH_DTYPE_DEPRECATED(ker_key, ker_class, double, d,               \
-                                  marco_declare, macro_key, macro_impl)
-
-#define FOR_EACH_ISA(macro_, block) \
-  macro_(platform::avx512f, block); \
-  macro_(platform::avx2, block);    \
-  macro_(platform::avx, block);     \
-  macro_(platform::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_, platform::avx512f); \
-  FOR_EACH_BLOCK(macro_, platform::avx2);    \
-  FOR_EACH_BLOCK(macro_, platform::avx);     \
-  FOR_EACH_BLOCK(macro_, platform::isa_any)
-
-}  // namespace jitkernel
-}  // namespace math
-}  // namespace operators
-}  // namespace paddle

paddle/fluid/operators/math/jit_kernel_rnn.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/operators/math/jit_kernel_refer.h"
-#include "paddle/fluid/platform/enforce.h"
-#include "paddle/fluid/platform/macros.h"
-
-#ifdef PADDLE_WITH_XBYAK
-#include "paddle/fluid/operators/math/jit_code.h"
-#endif
-
-namespace paddle {
-namespace operators {
-namespace math {
-namespace jitkernel {
-
-/* LSTM JitKernel */
-template <typename T>
-class LSTMKernelImpl : public LSTMKernel<T> {
- public:
-  static inline std::string name(const lstm_attr_t& attr) {
-    PADDLE_THROW("DType should be either float or double");
-  }
-  static inline bool useJIT(int d) { return false; }
-  static inline bool useMKL(int d) { return false; }
-  explicit LSTMKernelImpl(const lstm_attr_t& attr) : LSTMKernel<T>() {
-#ifdef PADDLE_WITH_XBYAK
-    if (useJIT(attr.d)) {
-      size_t sz = 96 + attr.d / YMM_FLOAT_BLOCK * 90 * 4 * 8;
-      jitcode0_.reset(new gen::LSTMJitCode(false, attr, sz > 4096 ? sz : 4096));
-      this->ComputeCtHt =
-          jitcode0_->getCode<void (*)(lstm_t*, const lstm_attr_t*)>();
-
-      jitcode1_.reset(new gen::LSTMJitCode(true, attr, sz > 4096 ? sz : 4096));
-      this->ComputeC1H1 =
-          jitcode1_->getCode<void (*)(lstm_t*, const lstm_attr_t*)>();
-      return;
-    }
-#endif
-
-    this->ComputeCtHt = refer::LSTMCtHt<T>;
-    this->ComputeC1H1 = refer::LSTMC1H1<T>;
-  }
-
-#ifdef PADDLE_WITH_XBYAK
-
- private:
-  std::unique_ptr<gen::LSTMJitCode> jitcode0_{nullptr}, jitcode1_{nullptr};
-#endif
-};
-
-#ifdef PADDLE_WITH_XBYAK
-template <>
-bool LSTMKernelImpl<float>::useJIT(int d) {
-  return gen::LSTMJitCode::init(d);
-}
-#endif
-
-/* Peephole JitKernel */
-template <typename T>
-class PeepholeKernelImpl : public LSTMKernel<T> {
- public:
-  static inline std::string name(const lstm_attr_t& attr) {
-    PADDLE_THROW("DType should be either float or double");
-  }
-  static inline bool useJIT(int d) { return false; }
-  static inline bool useMKL(int d) { return false; }
-  explicit PeepholeKernelImpl(const lstm_attr_t& attr) : LSTMKernel<T>() {
-#ifdef PADDLE_WITH_XBYAK
-    if (useJIT(attr.d)) {
-      size_t sz = 96 + attr.d / YMM_FLOAT_BLOCK * 96 * 4 * 8;
-      jitcode0_.reset(new gen::LSTMJitCode(false, attr, sz > 4096 ? sz : 4096));
-      this->ComputeCtHt =
-          jitcode0_->getCode<void (*)(lstm_t*, const lstm_attr_t*)>();
-
-      jitcode1_.reset(new gen::LSTMJitCode(true, attr, sz > 4096 ? sz : 4096));
-      this->ComputeC1H1 =
-          jitcode1_->getCode<void (*)(lstm_t*, const lstm_attr_t*)>();
-      return;
-    }
-#endif
-
-    this->ComputeCtHt = refer::LSTMCtHt<T>;
-    this->ComputeC1H1 = refer::LSTMC1H1<T>;
-  }
-
-#ifdef PADDLE_WITH_XBYAK
-
- private:
-  std::unique_ptr<gen::LSTMJitCode> jitcode0_{nullptr}, jitcode1_{nullptr};
-#endif
-};
-
-#ifdef PADDLE_WITH_XBYAK
-template <>
-bool PeepholeKernelImpl<float>::useJIT(int d) {
-  return gen::LSTMJitCode::init(d);
-}
-#endif
-
-#define JITKERNEL_DEFINE_NAME_LSTM(ker_key, ker_class)                 \
-  template <>                                                          \
-  std::string ker_class##Impl<float>::name(const lstm_attr_t& attr) {  \
-    std::string key(#ker_key "f");                                     \
-    key += (attr.act_gate + attr.act_cand + attr.act_cell +            \
-            (attr.use_peephole ? "p" : "n"));                          \
-    if (useJIT(attr.d)) {                                              \
-      /* only jit code need record d*/                                 \
-      return key + "jit" + std::to_string(attr.d);                     \
-    } else if (useMKL(attr.d)) {                                       \
-      return key + "mkl";                                              \
-    } else {                                                           \
-      return key + "any";                                              \
-    }                                                                  \
-  }                                                                    \
-  template <>                                                          \
-  std::string ker_class##Impl<double>::name(const lstm_attr_t& attr) { \
-    std::string key(#ker_key "d");                                     \
-    /* jit code do not support double yet*/                            \
-    if (useMKL(attr.d)) {                                              \
-      return key + "mkl";                                              \
-    } else {                                                           \
-      return key + "any";                                              \
-    }                                                                  \
-  }
-
-#define JITKERNEL_DECLARE_LSTM(ker_class, ker_dtype)          \
-  template <>                                                 \
-  std::shared_ptr<const LSTMKernel<ker_dtype>>                \
-  KernelPool::Get<LSTMKernel<ker_dtype>, const lstm_attr_t&>( \
-      const lstm_attr_t& attr)
-
-#define JITKERNEL_FIND_KEY_LSTM(ker_class, ker_dtype) \
-  std::string key = ker_class##Impl<ker_dtype>::name(attr)
-
-#define JITKERNEL_LSTM_IMPL(ker, dtype)                     \
-  if (attr.use_peephole) {                                  \
-    p = std::dynamic_pointer_cast<ker<dtype>>(              \
-        std::make_shared<PeepholeKernelImpl<dtype>>(attr)); \
-  } else {                                                  \
-    p = std::dynamic_pointer_cast<ker<dtype>>(              \
-        std::make_shared<ker##Impl<dtype>>(attr));          \
-  }
-
-REGISTER_JITKERNEL_ARGS(lstm, LSTMKernel, JITKERNEL_DEFINE_NAME_LSTM,
-                        JITKERNEL_DECLARE_LSTM, JITKERNEL_FIND_KEY_LSTM,
-                        JITKERNEL_LSTM_IMPL);
-
-#undef JITKERNEL_LSTM_IMPL
-#undef JITKERNEL_FIND_KEY_LSTM
-#undef JITKERNEL_DECLARE_LSTM
-#undef JITKERNEL_DEFINE_NAME_LSTM
-
-/* GRU JitKernel */
-template <typename T>
-class GRUKernelImpl : public GRUKernel<T> {
- public:
-  static inline std::string name(const gru_attr_t& attr) {
-    PADDLE_THROW("DType should be either float or double");
-  }
-  static inline bool useJIT(int d) { return false; }
-  static inline bool useMKL(int d) { return false; }
-  explicit GRUKernelImpl(const gru_attr_t& attr) : GRUKernel<T>() {
-#ifdef PADDLE_WITH_XBYAK
-    if (useJIT(attr.d)) {
-      size_t sz = 96 + attr.d / YMM_FLOAT_BLOCK * 96 * 2 * 8;
-      jitcode0_.reset(new gen::GRUJitCode(0, attr, sz > 4096 ? sz : 4096));
-      this->ComputeH1 =
-          jitcode0_->getCode<void (*)(gru_t*, const gru_attr_t*)>();
-
-      jitcode1_.reset(new gen::GRUJitCode(1, attr, sz > 4096 ? sz : 4096));
-      this->ComputeHtPart1 =
-          jitcode1_->getCode<void (*)(gru_t*, const gru_attr_t*)>();
-
-      jitcode2_.reset(new gen::GRUJitCode(2, attr, sz > 4096 ? sz : 4096));
-      this->ComputeHtPart2 =
-          jitcode2_->getCode<void (*)(gru_t*, const gru_attr_t*)>();
-      return;
-    }
-#endif
-    this->ComputeH1 = refer::GRUH1<T>;
-    this->ComputeHtPart1 = refer::GRUHtPart1<T>;
-    this->ComputeHtPart2 = refer::GRUHtPart2<T>;
-  }
-#ifdef PADDLE_WITH_XBYAK
-
- private:
-  std::unique_ptr<gen::GRUJitCode> jitcode0_{nullptr}, jitcode1_{nullptr},
-      jitcode2_{nullptr};
-#endif
-};
-
-#ifdef PADDLE_WITH_XBYAK
-template <>
-bool GRUKernelImpl<float>::useJIT(int d) {
-  return gen::GRUJitCode::init(d);
-}
-#endif
-
-#define JITKERNEL_DEFINE_NAME_GRU(ker_key, ker_class)                 \
-  template <>                                                         \
-  std::string ker_class##Impl<float>::name(const gru_attr_t& attr) {  \
-    std::string key(#ker_key "f");                                    \
-    key += (attr.act_gate + attr.act_cand);                           \
-    if (useJIT(attr.d)) {                                             \
-      /* only jit code need record d*/                                \
-      return key + "jit" + std::to_string(attr.d);                    \
-    } else if (useMKL(attr.d)) {                                      \
-      return key + "mkl";                                             \
-    } else {                                                          \
-      return key + "any";                                             \
-    }                                                                 \
-  }                                                                   \
-  template <>                                                         \
-  std::string ker_class##Impl<double>::name(const gru_attr_t& attr) { \
-    std::string key(#ker_key "d");                                    \
-    /* jit code do not support double yet*/                           \
-    if (useMKL(attr.d)) {                                             \
-      return key + "mkl";                                             \
-    } else {                                                          \
-      return key + "any";                                             \
-    }                                                                 \
-  }
-
-#define JITKERNEL_DECLARE_GRU(ker_class, ker_dtype)         \
-  template <>                                               \
-  std::shared_ptr<const ker_class<ker_dtype>>               \
-  KernelPool::Get<ker_class<ker_dtype>, const gru_attr_t&>( \
-      const gru_attr_t& attr)
-
-#define JITKERNEL_FIND_KEY_GRU(ker_class, ker_dtype) \
-  std::string key = ker_class##Impl<ker_dtype>::name(attr)
-
-#define JITKERNEL_GRU_IMPL(ker, dtype)       \
-  p = std::dynamic_pointer_cast<ker<dtype>>( \
-      std::make_shared<ker##Impl<dtype>>(attr));
-
-REGISTER_JITKERNEL_ARGS(gru, GRUKernel, JITKERNEL_DEFINE_NAME_GRU,
-                        JITKERNEL_DECLARE_GRU, JITKERNEL_FIND_KEY_GRU,
-                        JITKERNEL_GRU_IMPL);
-
-#undef JITKERNEL_GRU_IMPL
-#undef JITKERNEL_FIND_KEY_GRU
-#undef JITKERNEL_DECLARE_GRU
-#undef JITKERNEL_DEFINE_NAME_GRU
-}  // 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 <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"
-#include "paddle/fluid/operators/math/jit_kernel_refer.h"
-#include "paddle/fluid/platform/port.h"
-
-#ifdef PADDLE_WITH_MKLML
-#include "paddle/fluid/platform/dynload/mklml.h"
-#endif
-
-#ifdef __AVX__
-#include <immintrin.h>
-#endif
-
-constexpr int repeat = 20000;
-
-// TODO(TJ): benchmark and test should be seperated,
-// benchmark should verify more sizes
-
-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);
-  }
-}
-
-#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;
-  namespace refer = paddle::operators::math::jitkernel::refer;
-  for (int d : {3, 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) {
-      refer::VRelu<float>(x_data, zref_data, d);
-    }
-    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 << " us";
-    }
-#endif
-    auto ttgts = GetCurrentUS();
-    for (int i = 0; i < repeat; ++i) {
-      ker->Compute(x_data, ztgt_data, d);
-    }
-    auto ttgte = GetCurrentUS();
-    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
-            << " us, tgt takes: " << (ttgte - ttgts) / repeat << " us";
-    for (int i = 0; i < d; ++i) {
-      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
-    }
-  }
-}
-
-TEST(JitKernel, vaddbias) {
-  namespace jit = paddle::operators::math::jitkernel;
-  namespace refer = paddle::operators::math::jitkernel::refer;
-  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) {
-      refer::VAddBias<float>(&a, x_data, zref_data, d);
-    }
-    auto trefe = GetCurrentUS();
-    auto ttgts = GetCurrentUS();
-    for (int i = 0; i < repeat; ++i) {
-      ker->Compute(&a, x_data, ztgt_data, d);
-    }
-    auto ttgte = GetCurrentUS();
-
-    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
-            << " us, tgt takes: " << (ttgte - ttgts) / repeat << " us";
-    for (int i = 0; i < d; ++i) {
-      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
-    }
-  }
-}
-
-#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;
-  namespace refer = paddle::operators::math::jitkernel::refer;
-  for (int d : {1, 3, 4, 6, 7, 8, 12, 15, 16, 20, 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) {
-      refer::VExp<float>(x_data, zref_data, d);
-    }
-    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);
-      ker->Compute(x_data, ztgt_data, d);
-    }
-    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 << " us";
-    for (int i = 0; i < d; ++i) {
-      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
-    }
-  }
-}
-
-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, n);
-  for (int i = 0; i < n; ++i) {
-    y[i] = 1.f / (1.f + y[i]);
-  }
-}
-
-TEST(JitKernel, vsigmoid) {
-  namespace jit = paddle::operators::math::jitkernel;
-  namespace refer = paddle::operators::math::jitkernel::refer;
-  for (int d : {1, 3, 4, 6, 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) {
-      refer::VSigmoid<float>(x_data, zref_data, d);
-    }
-    auto trefe = GetCurrentUS();
-    auto ttgts = GetCurrentUS();
-    for (int i = 0; i < repeat; ++i) {
-      ker->Compute(x_data, ztgt_data, d);
-    }
-    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 << " us";
-    for (int i = 0; i < d; ++i) {
-      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
-    }
-  }
-}
-
-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) {
-  const float a = 2.f, b = -1.f;
-  vscal->Compute(&a, x, y, n);
-  vsigmoid->Compute(y, y, n);
-  vscal->Compute(&a, y, y, n);
-  vaddbias->Compute(&b, y, y, n);
-}
-
-TEST(JitKernel, vtanh) {
-  namespace jit = paddle::operators::math::jitkernel;
-  namespace refer = paddle::operators::math::jitkernel::refer;
-  for (int d : {1, 2, 3, 4, 5, 6, 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) {
-      refer::VTanh<float>(x_data, zref_data, d);
-    }
-    auto trefe = GetCurrentUS();
-    auto ttgts = GetCurrentUS();
-    for (int i = 0; i < repeat; ++i) {
-      ker->Compute(x_data, ztgt_data, d);
-    }
-    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 << " us";
-    for (int i = 0; i < d; ++i) {
-      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
-    }
-  }
-}
-
-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, 3 * d);
-  vtanh_d->Compute(gates, gates, d);
-  vmul_d->Compute(gates, gates + d, gates + d, d);
-  vmul_d->Compute(ct_1, gates + d2, gates + d2, d);
-  vadd_d->Compute(gates + d, gates + d2, ct, d);
-  /* H_t = act_cell(C_t) * ogated */
-  vtanh_d->Compute(ct, gates + d2, d);
-  vmul_d->Compute(gates + d2, gates + d * 3, ht, d);
-}
-
-TEST(JitKernel, lstm) {
-  namespace jit = paddle::operators::math::jitkernel;
-  namespace refer = paddle::operators::math::jitkernel::refer;
-  for (int d : {1, 2, 3, 4, 5, 6, 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 jit::lstm_attr_t attr(d, act_gate, act_cand, act_cell, false);
-    const auto& ker =
-        jit::KernelPool::Instance()
-            .template Get<jit::LSTMKernel<float>, const jit::lstm_attr_t&>(
-                attr);
-    // 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& 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
-    jit::lstm_t step;
-    step.gates = xref_data;
-    step.ct_1 = ct_1_data;
-    step.ct = ct_ref_data;
-    step.ht = ht_ref_data;
-    refer::LSTMCtHt<float>(&step, &attr);
-
-    step.gates = x_data;
-    step.ct = ct_tgt_data;
-    step.ht = ht_tgt_data;
-    ker->ComputeCtHt(&step, &attr);
-    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) {
-      refer::LSTMCtHt<float>(&step, &attr);
-    }
-    auto trefe = GetCurrentUS();
-    auto ttgts = GetCurrentUS();
-    for (int i = 0; i < repeat; ++i) {
-      ker->ComputeCtHt(&step, &attr);
-    }
-    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 << " us";
-  }
-}
-
-#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;
-  namespace refer = paddle::operators::math::jitkernel::refer;
-  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) {
-      refer::VScal<float>(&a, x_data, zref_data, d);
-    }
-    auto trefe = GetCurrentUS();
-    auto trefs1 = GetCurrentUS();
-    for (int i = 0; i < repeat; ++i) {
-      refer::VScal<float>(&a, y_data, y_data, d);
-    }
-    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 << " us";
-    }
-#endif
-
-    auto ttgts = GetCurrentUS();
-    for (int i = 0; i < repeat; ++i) {
-      ker->Compute(&a, x_data, ztgt_data, d);
-    }
-    auto ttgte = GetCurrentUS();
-    auto ttgts1 = GetCurrentUS();
-    for (int i = 0; i < repeat; ++i) {
-      ker->Compute(&a, y_data, y_data, d);
-    }
-    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 << " us";
-    for (int i = 0; i < d; ++i) {
-      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
-    }
-  }
-}
-
-#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;
-  namespace refer = paddle::operators::math::jitkernel::refer;
-  for (int d : {7, 8, 15, 16, 20, 30, 256, 512, 1000, 1024}) {
-    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) {
-      refer::VMul<float>(x_data, y_data, zref_data, d);
-    }
-    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, d);
-    }
-    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 << " us";
-    for (int i = 0; i < d; ++i) {
-      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
-    }
-  }
-}
-
-#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;
-  namespace refer = paddle::operators::math::jitkernel::refer;
-  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) {
-      refer::VAdd<float>(x_data, y_data, zref_data, d);
-    }
-    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, d);
-    }
-    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 << " us";
-    for (int i = 0; i < d; ++i) {
-      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
-    }
-  }
-}
-
-void vaddrelu_better(
-    const std::shared_ptr<
-        const paddle::operators::math::jitkernel::VAddKernel<float>>& vadd,
-    const std::shared_ptr<
-        const paddle::operators::math::jitkernel::VReluKernel<float>>& vrelu,
-    const float* x, const float* y, float* z, int d) {
-  vadd->Compute(x, y, z, d);
-  vrelu->Compute(z, z, d);
-}
-
-TEST(JitKernel, vaddrelu) {
-  namespace jit = paddle::operators::math::jitkernel;
-  namespace refer = paddle::operators::math::jitkernel::refer;
-  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
-    std::vector<float> x(d), y(d);
-    std::vector<float> zref(d), ztgt(d);
-    RandomVec<float>(d, x.data());
-    RandomVec<float>(d, y.data());
-    const auto& ker =
-        jit::KernelPool::Instance().template Get<jit::VAddReluKernel<float>>(d);
-    const auto& vadd =
-        jit::KernelPool::Instance().template Get<jit::VAddKernel<float>>(d);
-    const auto& vrelu =
-        jit::KernelPool::Instance().template Get<jit::VReluKernel<float>>(d);
-    const float* x_data = x.data();
-    const float* y_data = y.data();
-    float* ztgt_data = ztgt.data();
-    float* zref_data = zref.data();
-    auto trefs = GetCurrentUS();
-    for (int i = 0; i < repeat; ++i) {
-      refer::VAddRelu<float>(x_data, y_data, zref_data, d);
-    }
-    auto trefe = GetCurrentUS();
-    auto tmkls = GetCurrentUS();
-    for (int i = 0; i < repeat; ++i) {
-      vaddrelu_better(vadd, vrelu, x_data, y_data, zref_data, d);
-    }
-    auto tmkle = GetCurrentUS();
-    auto ttgts = GetCurrentUS();
-    for (int i = 0; i < repeat; ++i) {
-      ker->Compute(x_data, y_data, ztgt_data, d);
-    }
-    auto ttgte = GetCurrentUS();
-    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
-            << " us, better takes: " << (tmkle - tmkls) / repeat << " us, "
-            << "tgt takes: " << (ttgte - ttgts) / repeat << " us";
-    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";
-  jit::lstm_attr_t attr(frame_size, act_gate, act_cand, act_cell, false);
-
-  // empty call it to avoid unknown flag 'use_pinned_memory' on Mac
-  paddle::platform::MayIUse(paddle::platform::avx);
-  const auto& plstm1 =
-      jit::KernelPool::Instance()
-          .template Get<jit::LSTMKernel<float>, const jit::lstm_attr_t&>(attr);
-
-  const auto& plstm2 =
-      jit::KernelPool::Instance()
-          .template Get<jit::LSTMKernel<float>, const jit::lstm_attr_t&>(attr);
-  EXPECT_EQ(plstm1, plstm2);
-
-  const auto& peephole =
-      jit::KernelPool::Instance()
-          .template Get<jit::LSTMKernel<float>, const jit::lstm_attr_t&>(
-              jit::lstm_attr_t(frame_size, act_gate, act_cand, act_cell, 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("vmulfjit4");
-#if defined(__APPLE__) || defined(__OSX__) || defined(_WIN32)
-  EXPECT_EQ(pvmul_from_key, nullptr);
-#else
-  EXPECT_EQ(pvmul_from_key, pvmul_f);
-#endif
-  const auto& pvmul_from_key2 = jit::KernelPool::Instance().Get("vmulfjit");
-  EXPECT_TRUE(pvmul_from_key2 == nullptr);
-}

paddle/fluid/operators/mean_iou_op.cu

@@ -92,8 +92,8 @@ template <typename T>
 class MeanIoUCUDAOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto& place = *ctx.template device_context<platform::CUDADeviceContext>()
-                       .eigen_device();
+    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+    auto& place = *dev_ctx.eigen_device();
     // get input and output tensor
     auto* predictions = ctx.Input<Tensor>("Predictions");
     auto* labels = ctx.Input<Tensor>("Labels");
@@ -115,11 +115,11 @@ class MeanIoUCUDAOpKernel : public framework::OpKernel<T> {
     auto out_wrong_t = EigenTensor<int, 1>::From(*out_wrong);
     auto out_correct_t = EigenTensor<int, 1>::From(*out_correct);
 
-    // Temporary tensor
-    Tensor ious;
-    float* ious_data = ious.mutable_data<float>(
-        {static_cast<int64_t>(num_classes)}, ctx.GetPlace());
-    auto ious_t = EigenTensor<float, 1>::From(ious);
+    // Temporary memory
+    auto& allocator =
+        platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx);
+    auto tmp_ious_data = allocator.Allocate(num_classes * sizeof(float));
+    float* ious_data = static_cast<float*>(tmp_ious_data->ptr());
 
     // Init out_wrong, out_correct and out_mean_iou
     out_wrong_t.device(place) = out_wrong_t.constant(0);
@@ -148,7 +148,7 @@ class MeanIoUCUDAOpKernel : public framework::OpKernel<T> {
     CountCUDAKernel<T><<<grid, block, cache_size, stream>>>(
         num_classes, predictions->numel(), predictions_data, labels_data,
         out_wrong_data, out_correct_data);
-    ctx.device_context().Wait();
+
     ComputeIoUCUDAKernel<<<1, block, 0, stream>>>(num_classes, out_wrong_data,
                                                   out_correct_data, ious_data,
                                                   out_mean_iou_data);

paddle/fluid/operators/merge_selected_rows_op.cc

@@ -26,6 +26,13 @@ class MergeSelectedRowsOp : public framework::OperatorWithKernel {
                    "Input(X) of MergeSelectedRowsOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of MergeSelectedRowsOp should not be null.");
+    PADDLE_ENFORCE_EQ(ctx->GetInputsVarType("X").front(),
+                      framework::proto::VarType::SELECTED_ROWS,
+                      "Input X only should be SelectedRows.");
+    PADDLE_ENFORCE_EQ(ctx->GetOutputsVarType("Out").front(),
+                      framework::proto::VarType::SELECTED_ROWS,
+                      "Output Y only should be SelectedRows.");
+
     ctx->ShareDim("X", /*->*/ "Out");
   }
 };
@@ -43,7 +50,28 @@ class MergeSelectedRowsOpMaker : public framework::OpProtoAndCheckerMaker {
         R"DOC(
 MergeSelectedRows Operator.
 
-MergeSelectedRows is used to merge the duplicated rows of the input.
+MergeSelectedRows is used to merge the duplicated rows of the input. The
+output's row has no duplicated, and it's order is incremental.
+
+Example:
+  Input:
+    X.rows is [0, 5, 5, 4, 19]
+    X.height is 20
+    X.value is:
+        [[1, 1]
+         [2, 2]
+         [3, 3]
+         [4, 4]
+         [6, 6]]
+
+   Output:
+    Out.row is [0, 4, 5, 19]
+    Out.height is 20
+    Out.value is:
+        [[1, 1]
+         [4, 4]
+         [5, 5]
+         [6, 6]]
 )DOC");
   }
 };

paddle/fluid/operators/metrics/accuracy_op.cu

@@ -16,7 +16,6 @@ limitations under the License. */
 #include <thrust/reduce.h>
 #include "paddle/fluid/operators/metrics/accuracy_op.h"
 #include "paddle/fluid/platform/cuda_primitives.h"
-#include "paddle/fluid/platform/float16.h"
 #include "paddle/fluid/platform/gpu_info.h"
 
 namespace paddle {
@@ -95,7 +94,6 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
 
 // FIXME(typhoonzero): types of T is for inference data.
 // label data is always int64
-REGISTER_OP_CUDA_KERNEL(
-    accuracy, paddle::operators::AccuracyOpCUDAKernel<float>,
-    paddle::operators::AccuracyOpCUDAKernel<double>,
-    paddle::operators::AccuracyOpCUDAKernel<paddle::platform::float16>);
+REGISTER_OP_CUDA_KERNEL(accuracy,
+                        paddle::operators::AccuracyOpCUDAKernel<float>,
+                        paddle::operators::AccuracyOpCUDAKernel<double>);

paddle/fluid/operators/optimizers/momentum_op.cu

@@ -14,11 +14,8 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/optimizers/momentum_op.h"
-#include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;
 REGISTER_OP_CUDA_KERNEL(
     momentum, ops::MomentumOpKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::MomentumOpKernel<paddle::platform::CUDADeviceContext, double>,
-    ops::MomentumOpKernel<paddle::platform::CUDADeviceContext,
-                          paddle::platform::float16>);
+    ops::MomentumOpKernel<paddle::platform::CUDADeviceContext, double>);

paddle/fluid/operators/optimizers/momentum_op.h

@@ -237,8 +237,7 @@ class SparseMomentumFunctor<T, UseNesterov> {
   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_]
-                       : static_cast<T>(0);
+    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];
@@ -283,8 +282,7 @@ class SparseMomentumFunctor<T, NoNesterov> {
   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_]
-                       : static_cast<T>(0);
+    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];

paddle/fluid/operators/py_func_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/py_func_op.h"
+#include <set>
+#include <string>
+#include <vector>
+#include "Python.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+namespace py = ::pybind11;
+
+static std::vector<py::object> g_py_callables;
+
+const char kForwardPythonCallableId[] = "forward_callable_id";
+const char kBackwardPythonCallableId[] = "backward_callable_id";
+const char kPyFuncBackwardSkipVars[] = "backward_skip_vars";
+
+size_t AppendPythonCallableObjectAndReturnId(const py::object &py_obj) {
+  g_py_callables.emplace_back(py_obj);
+  return g_py_callables.size() - 1;
+}
+
+// Return py::object* instead of py::object
+// Returning py::object would cause reference count increasing
+// but without GIL, reference count in Python may not be safe
+static py::object *GetPythonCallableObject(size_t i) {
+  PADDLE_ENFORCE_LT(i, g_py_callables.size(), "Invalid python callable id");
+  return &g_py_callables[i];
+}
+
+static std::string PythonFuncDebugString(const py::object &py_callable) {
+  py::gil_scoped_acquire guard;
+  std::string wrapper_func_str = py::str(py_callable);
+  auto inner_func = py_callable.attr("_func");
+  std::string inner_func_str = py::str(inner_func);
+  return inner_func_str + " wrapped by " + wrapper_func_str;
+}
+
+static void CallPythonFunc(py::object *callable,
+                           const std::vector<framework::LoDTensor> &ins,
+                           std::vector<framework::LoDTensor *> *outs) {
+  py::gil_scoped_acquire guard;
+  py::tuple in_args(ins.size());
+  for (size_t i = 0; i < ins.size(); ++i) {
+    in_args[i] = ins[i].IsInitialized() ? py::cast(ins[i]) : py::cast(nullptr);
+  }
+
+  auto ret = (*callable)(*in_args);
+  auto ret_tuple = py::cast<py::tuple>(ret);
+  size_t ret_num = py::len(ret_tuple);
+  size_t out_num = outs->size();
+  if (UNLIKELY(ret_num != out_num)) {
+    // Python function has no return values or returns None
+    // In this case, ret_num = 1 && ret[0] == None && out_num should be 0
+    // Otherwise, ret_num must be equal to out_num
+    PADDLE_ENFORCE(
+        ret_num == 1 && out_num == 0 &&
+            py::cast<framework::LoDTensor *>(ret_tuple[0]) == nullptr,
+        "Output number not match. Expected %d, actual %d", out_num, ret_num);
+  }
+
+  for (size_t i = 0; i < out_num; ++i) {
+    auto *out = (*outs)[i];
+    if (out == nullptr) {
+      continue;
+    }
+    try {
+      auto *py_out_tensor = py::cast<framework::LoDTensor *>(ret_tuple[i]);
+      PADDLE_ENFORCE_NOT_NULL(py_out_tensor,
+                              "Output tensor %d should not be nullptr", i);
+      out->set_lod(py_out_tensor->lod());
+      out->ShareDataWith(*py_out_tensor);
+    } catch (py::cast_error &) {
+      PADDLE_THROW("The %d-th output must be LoDTensor", i);
+    }
+  }
+}
+
+class PyFuncOpVarTypInference : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc &op,
+                  framework::BlockDesc *block) const override {
+    auto &outs = op.Outputs();
+    bool has_out = (outs.count("Out") > 0 && !outs.at("Out").empty());
+
+    auto &ins = op.Inputs();
+    bool has_in = (ins.count("X") > 0 && !ins.at("X").empty());
+
+    /**
+     * X or Out can be empty, so that py_func can be more flexible
+     * to support Python functions with no input or no output
+     */
+    PADDLE_ENFORCE(has_in || has_out, "Input(X) or Output(Out) must exist");
+
+    PADDLE_ENFORCE_GE(boost::get<int>(op.GetAttr(kForwardPythonCallableId)), 0,
+                      "Function id cannot be less than 0");
+
+    if (!has_out) return;
+
+    /**
+     * Traverse all outputs, check if name of any output ends with @GRAD.
+     * If found, set its shape, dtype, lod_level, type to be the same as
+     * the corresponding forward variable
+     */
+    const std::string kGradVarSuffix = framework::kGradVarSuffix;
+    auto &out_var_names = outs.at("Out");
+    for (auto &out_var_name : out_var_names) {
+      if (out_var_name == framework::kEmptyVarName ||
+          out_var_name.size() < kGradVarSuffix.size()) {
+        continue;
+      }
+
+      size_t len = out_var_name.size() - kGradVarSuffix.size();
+      if (out_var_name.substr(len) == kGradVarSuffix) {
+        auto fwd_var_name = out_var_name.substr(0, len);
+        auto *out_var_desc = block->FindVarRecursive(out_var_name);
+        auto *fwd_var_desc = block->FindVarRecursive(fwd_var_name);
+        PADDLE_ENFORCE_NOT_NULL(out_var_desc, "Backward variable %s not found",
+                                out_var_name);
+        PADDLE_ENFORCE_NOT_NULL(fwd_var_desc, "Forward variable %s not found",
+                                fwd_var_name);
+        VLOG(10) << "Infer var_desc of Output(" << out_var_name << ") as Input("
+                 << fwd_var_name << ")";
+        out_var_desc->SetShape(fwd_var_desc->GetShape());
+        out_var_desc->SetDataType(fwd_var_desc->GetDataType());
+        out_var_desc->SetLoDLevel(fwd_var_desc->GetLoDLevel());
+        out_var_desc->SetType(fwd_var_desc->GetType());
+      }
+    }
+  }
+};
+
+class PyFuncOpShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(!ctx->IsRuntime(),
+                   "Infer shape cannot be called in runtime.");
+  }
+};
+
+class PyFuncOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X", "Inputs of py_func op.").AsDuplicable();
+    AddOutput("Out", "Outputs of py_func op").AsDuplicable();
+    AddAttr<int>(kForwardPythonCallableId,
+                 "Index of registered forward Python function.")
+        .SetDefault(0);
+    AddAttr<int>(kBackwardPythonCallableId,
+                 "Index of registered backward Python function.")
+        .SetDefault(-1);
+    AddAttr<std::vector<std::string>>(kPyFuncBackwardSkipVars,
+                                      "Unused forward in/out in backward op")
+        .SetDefault(std::vector<std::string>());
+    AddComment(R"DOC("PyFunc Op")DOC");
+  }
+};
+
+/**
+ * There are several benefits when backward op of py_func op is
+ * still py_func op.
+ *
+ *  - Less codes are needed, since codes of backward is almost
+ *    the same as forward.
+ *
+ *  - To support high order derivative, so that py_func is
+ *    infinite-order differentiable
+ */
+class PyFuncOpGradDescMaker : public framework::GradOpDescMakerBase {
+ private:
+  static std::string DebugString(const std::vector<std::string> &strs) {
+    if (strs.empty()) return "";
+    std::string ret = strs[0];
+    for (size_t i = 1; i < strs.size(); ++i) {
+      ret += " ";
+      ret += strs[i];
+    }
+    return ret;
+  }
+
+ public:
+  using framework::GradOpDescMakerBase::GradOpDescMakerBase;
+
+  std::vector<std::unique_ptr<framework::OpDesc>> operator()() const override {
+    auto &fwd_attrs = Attrs();
+    // no backward op when backward_id is less than 0
+    if (boost::get<int>(fwd_attrs.at(kBackwardPythonCallableId)) < 0) {
+      return {};
+    }
+
+    std::unique_ptr<framework::OpDesc> grad_op(new framework::OpDesc());
+    grad_op->SetType("py_func");
+
+    framework::AttributeMap bwd_attrs;
+    bwd_attrs[kForwardPythonCallableId] =
+        fwd_attrs.at(kBackwardPythonCallableId);
+    bwd_attrs[kBackwardPythonCallableId] = -1;
+    grad_op->SetAttrMap(bwd_attrs);
+
+    // All forward inputs
+    auto fwd_ins = Input("X");
+    // All forward outputs
+    auto fwd_outs = Output("Out");
+
+    // For memory reused, some inputs/output in forward part may be not needed
+    // in backward part. Skipping these vars helps to save memory
+    auto &backward_skip_var_list = boost::get<std::vector<std::string>>(
+        fwd_attrs.at(kPyFuncBackwardSkipVars));
+    std::unordered_set<std::string> backward_skip_var_set(
+        backward_skip_var_list.begin(), backward_skip_var_list.end());
+    std::vector<std::string> bwd_ins;
+    bwd_ins.reserve(fwd_ins.size() + fwd_outs.size());
+    for (auto &fwd_in : fwd_ins) {
+      if (backward_skip_var_set.count(fwd_in) == 0) {
+        bwd_ins.emplace_back(fwd_in);
+      }
+    }
+
+    for (auto &fwd_out : fwd_outs) {
+      if (backward_skip_var_set.count(fwd_out) == 0) {
+        bwd_ins.emplace_back(fwd_out);
+      }
+    }
+
+    // Backward OG cannot be skipped
+    // But in Python side, if OG is kEmptyVarName, input tensor would be None
+    auto fwd_out_grads = OutputGrad("Out");
+    bwd_ins.reserve(bwd_ins.size() + fwd_out_grads.size());
+    bwd_ins.insert(bwd_ins.end(), fwd_out_grads.begin(), fwd_out_grads.end());
+
+    // Backward IG cannot be skipped
+    // But in Python side, if IG is not needed, users can just return None
+    auto bwd_outs = InputGrad("X", false);
+
+    VLOG(10) << "PyFunc Grad Input: " << DebugString(bwd_ins);
+    VLOG(10) << "PyFunc Grad Output: " << DebugString(bwd_outs);
+
+    grad_op->SetInput("X", bwd_ins);
+    grad_op->SetOutput("Out", bwd_outs);
+
+    std::vector<std::unique_ptr<framework::OpDesc>> ret(1);
+    ret[0] = std::move(grad_op);
+    return ret;
+  }
+};
+
+class PyFuncOp : public framework::OperatorBase {
+ public:
+  using framework::OperatorBase::OperatorBase;
+
+ protected:
+  void RunImpl(const framework::Scope &scope,
+               const platform::Place &place) const override {
+    auto &in_arg_names = Inputs("X");
+    auto &out_arg_names = Outputs("Out");
+
+    std::vector<framework::LoDTensor> inputs(in_arg_names.size());
+    for (size_t i = 0; i < in_arg_names.size(); ++i) {
+      auto in_var = scope.FindVar(in_arg_names[i]);
+      // When py_func op is called in backward, in_var may be null
+      if (in_var == nullptr) {
+        continue;
+      }
+      auto &in_tensor = in_var->Get<framework::LoDTensor>();
+      if (!in_tensor.IsInitialized()) {
+        continue;
+      }
+      if (platform::is_gpu_place(in_tensor.place())) {
+        framework::TensorCopySync(in_tensor, platform::CPUPlace(), &inputs[i]);
+      } else {
+        inputs[i].ShareDataWith(in_tensor);
+      }
+      inputs[i].set_lod(in_tensor.lod());
+    }
+
+    std::vector<framework::LoDTensor *> outputs(out_arg_names.size());
+    for (size_t i = 0; i < out_arg_names.size(); ++i) {
+      auto *out_var = scope.FindVar(out_arg_names[i]);
+      outputs[i] =
+          out_var ? out_var->GetMutable<framework::LoDTensor>() : nullptr;
+    }
+
+    auto callable_id = static_cast<size_t>(Attr<int>(kForwardPythonCallableId));
+    auto *py_callable = GetPythonCallableObject(callable_id);
+    VLOG(10) << "Call Python function with id " << callable_id << ": "
+             << PythonFuncDebugString(*py_callable);
+    CallPythonFunc(py_callable, inputs, &outputs);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(py_func, ops::PyFuncOp, ops::PyFuncOpMaker,
+                  ops::PyFuncOpVarTypInference, ops::PyFuncOpShapeInference,
+                  ops::PyFuncOpGradDescMaker);

paddle/fluid/operators/py_func_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 "pybind11/pybind11.h"
+
+namespace paddle {
+namespace operators {
+
+size_t AppendPythonCallableObjectAndReturnId(const ::pybind11::object &py_obj);
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/top_k_op.cu

@@ -16,7 +16,6 @@ limitations under the License. */
 #include "paddle/fluid/operators/top_k_op.h"
 #include "paddle/fluid/platform/assert.h"
 #include "paddle/fluid/platform/cuda_device_function.h"
-#include "paddle/fluid/platform/float16.h"
 
 namespace paddle {
 namespace operators {
@@ -151,7 +150,7 @@ __device__ __forceinline__ void ThreadGetTopK(Pair<T> topk[], int* beam,
         if (k < MaxLength - (*beam)) {
           topk[k] = topk[k + *beam];
         } else {
-          topk[k].set(-static_cast<T>(INFINITY), -1);
+          topk[k].set(-INFINITY, -1);
         }
       }
       if (!(*is_empty)) {
@@ -161,7 +160,7 @@ __device__ __forceinline__ void ThreadGetTopK(Pair<T> topk[], int* beam,
     }
 
     *max = topk[MaxLength - 1];
-    if ((*max).v == -static_cast<T>(1)) *is_empty = true;
+    if ((*max).v == -1) *is_empty = true;
     *beam = 0;
   }
 }
@@ -182,7 +181,7 @@ __device__ __forceinline__ void ThreadGetTopK(Pair<T> topk[], int* beam,
         if (k < MaxLength - *beam) {
           topk[k] = topk[k + *beam];
         } else {
-          topk[k].set(-static_cast<T>(INFINITY), -1);
+          topk[k].set(-INFINITY, -1);
         }
       }
       if (!(*is_empty)) {
@@ -279,7 +278,7 @@ __global__ void KeMatrixTopK(T* output, int output_stride, int64_t* indices,
     bool firststep = true;
 
     for (int j = 0; j < MaxLength; j++) {
-      topk[j].set(-static_cast<T>(INFINITY), -1);
+      topk[j].set(-INFINITY, -1);
     }
     while (top_num) {
       ThreadGetTopK<T, MaxLength, BlockSize>(
@@ -363,7 +362,5 @@ class TopkOpCUDAKernel : public framework::OpKernel<T> {
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OP_CUDA_KERNEL(
-    top_k, paddle::operators::TopkOpCUDAKernel<float>,
-    paddle::operators::TopkOpCUDAKernel<double>,
-    paddle::operators::TopkOpCUDAKernel<paddle::platform::float16>);
+REGISTER_OP_CUDA_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel<float>,
+                        paddle::operators::TopkOpCUDAKernel<double>);

paddle/fluid/operators/transpose_mkldnn_op.cc

@@ -29,10 +29,6 @@ class TransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
   void Compute(const paddle::framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
                    "It must use CPUPlace.");
-    const bool is_test = ctx.Attr<bool>("is_test");
-    PADDLE_ENFORCE(
-        is_test == true,
-        "ConvTransposeMKLDNN works only for inference!. Set is_test = True");
     auto& dev_ctx =
         ctx.template device_context<paddle::platform::MKLDNNDeviceContext>();
     const auto& mkldnn_engine = dev_ctx.GetEngine();
@@ -47,69 +43,75 @@ class TransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
       return;
     }
 
-    std::vector<int> nchw_axis(ndims, 0);
-    for (size_t i = 0; i < nchw_axis.size(); ++i) {
-      nchw_axis[i] = i;
-    }
-
     std::vector<int> nchw_tz = paddle::framework::vectorize2int(input->dims());
-    std::string data_format = ctx.Attr<std::string>("data_format");
 
-    auto src_md =
-        input->format() != mkldnn::memory::format::nchw
-            ? platform::MKLDNNMemDesc(nchw_tz, platform::MKLDNNGetDataType<T>(),
-                                      input->format())
-            : Axis2MemoryDesc(nchw_tz, nchw_axis);
+    const std::string key = platform::TransposeMKLDNNHandler::GetHash(
+        nchw_tz, axis, ctx.op().Output("Out"));
 
-    this->TransposeKernel(ctx.GetPlace(), Axis2MemoryDesc(nchw_tz, axis),
-                          src_md, output, input_data, nchw_tz, mkldnn_engine);
-  }
+    platform::TransposeMKLDNNHandler handler(nchw_tz, axis, dev_ctx,
+                                             mkldnn_engine, key);
 
- protected:
-  mkldnn::memory::desc Axis2MemoryDesc(std::vector<int>& nchw_tz,
-                                       std::vector<int>& axis) const {
-    mkldnn_memory_desc_t mem_fmt;
+    auto transpose_src_memory_p = handler.AcquireSrcMemory(
+        input->format(), platform::to_void_cast<T>(input_data));
+    auto transpose_dst_memory_p =
+        handler.AcquireDstMemory(output, ctx.GetPlace());
+    auto transpose_p = handler.AcquireTranspose(transpose_dst_memory_p,
+                                                transpose_src_memory_p);
 
-    mem_fmt.primitive_kind = mkldnn_memory;
-    mem_fmt.ndims = axis.size();
-    for (unsigned int i = 0; i < nchw_tz.size(); ++i) {
-      mem_fmt.dims[i] = nchw_tz[i];  // logical dimensions (nchw format,
-                                     // regardless physical layout)
+    std::vector<mkldnn::primitive> pipeline;
+    pipeline.push_back(*transpose_p);
+    mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
   }
-    mem_fmt.data_type = mkldnn_f32;
-    mem_fmt.format = mkldnn_blocked;
-
-    unsigned int total_stride = 1;
-    for (int i = nchw_tz.size() - 1; i >= 0; --i) {
-      mem_fmt.layout_desc.blocking.padding_dims[i] =
-          nchw_tz[i];  // logical dimensions (nchw format, regardless physical
-                       // layout)
-      mem_fmt.layout_desc.blocking.block_dims[i] = 1;
-      mem_fmt.layout_desc.blocking.offset_padding_to_data[i] = 0;  // no offset
-      mem_fmt.layout_desc.blocking.strides[0][axis[i]] = total_stride;
-      mem_fmt.layout_desc.blocking.strides[1][axis[i]] = 1;
-      total_stride *= nchw_tz[axis[i]];
+};
+
+template <typename T>
+class TransposeMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
+                   "It must use CPUPlace.");
+    auto* out_grad =
+        ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto* x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    if (!x_grad) return;
+
+    auto& dev_ctx =
+        ctx.template device_context<paddle::platform::MKLDNNDeviceContext>();
+    const auto& mkldnn_engine = dev_ctx.GetEngine();
+    std::vector<int> axis = ctx.Attr<std::vector<int>>("axis");
+    std::vector<int> reversed_axis(axis);
+    int ndims = axis.size();
+    if (ndims == 1) {
+      x_grad->ShareDataWith(*out_grad);
+      return;
     }
-    mem_fmt.layout_desc.blocking.offset_padding = 0;  // no initial offset
-    return mem_fmt;
+
+    for (size_t i = 0; i < axis.size(); i++) {
+      reversed_axis[axis[i]] = i;
     }
 
-  void TransposeKernel(platform::Place place, mkldnn::memory::desc md_o,
-                       mkldnn::memory::desc md_i, Tensor* output,
-                       const T* data_i, std::vector<int>& nchw_dims,
-                       const mkldnn::engine& eng) const {
-    // Make Memory primitive descriptors
-    auto mpd_o = mkldnn::memory::primitive_desc(md_o, eng);
-    auto mpd_i = mkldnn::memory::primitive_desc(md_i, eng);
+    const T* out_grad_data = out_grad->data<T>();
+    x_grad->mutable_data<T>(ctx.GetPlace());
+
+    std::vector<int> nchw_tz =
+        paddle::framework::vectorize2int(out_grad->dims());
 
-    auto data_o = output->mutable_data<T>(
-        place, paddle::memory::Allocator::kDefault, mpd_o.get_size());
+    const std::string key = platform::TransposeMKLDNNHandler::GetHash(
+        nchw_tz, axis, ctx.op().Output(framework::GradVarName("X")));
 
-    auto src = mkldnn::memory(mpd_i, (T*)(data_i));
-    auto dst = mkldnn::memory(mpd_o, data_o);
+    platform::TransposeMKLDNNHandler handler(nchw_tz, reversed_axis, dev_ctx,
+                                             mkldnn_engine, key);
 
-    auto r = mkldnn::reorder(src, dst);
-    mkldnn::stream(mkldnn::stream::kind::eager).submit({r}).wait();
+    auto transpose_src_memory_p = handler.AcquireSrcMemory(
+        out_grad->format(), platform::to_void_cast<T>(out_grad_data));
+    auto transpose_dst_memory_p =
+        handler.AcquireDstMemory(x_grad, ctx.GetPlace());
+    auto transpose_p = handler.AcquireTranspose(transpose_dst_memory_p,
+                                                transpose_src_memory_p);
+
+    std::vector<mkldnn::primitive> pipeline;
+    pipeline.push_back(*transpose_p);
+    mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
   }
 };
 
@@ -122,3 +124,8 @@ REGISTER_OP_KERNEL(transpose2, MKLDNN, ::paddle::platform::CPUPlace,
                    ops::TransposeMKLDNNOpKernel<float>);
 REGISTER_OP_KERNEL(transpose, MKLDNN, ::paddle::platform::CPUPlace,
                    ops::TransposeMKLDNNOpKernel<float>);
+
+REGISTER_OP_KERNEL(transpose_grad, MKLDNN, ::paddle::platform::CPUPlace,
+                   ops::TransposeMKLDNNGradOpKernel<float>);
+REGISTER_OP_KERNEL(transpose2_grad, MKLDNN, ::paddle::platform::CPUPlace,
+                   ops::TransposeMKLDNNGradOpKernel<float>);

paddle/fluid/operators/transpose_op.cc

@@ -79,10 +79,6 @@ class TransposeOp : public framework::OperatorWithKernel {
 class TransposeOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
-    AddAttr<bool>("is_test",
-                  "(bool, default false) Set to true for inference only, false "
-                  "for training. Some layers may run faster when this is true.")
-        .SetDefault(false);
     AddInput(
         "X",
         "(Tensor) The input tensor, tensors with rank up to 6 are supported.");
@@ -147,6 +143,24 @@ class TransposeOpGrad : public framework::OperatorWithKernel {
       ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
     }
   }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    framework::LibraryType library_{framework::LibraryType::kPlain};
+    std::string data_format = ctx.Attr<std::string>("data_format");
+    framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
+#ifdef PADDLE_WITH_MKLDNN
+    if (library_ == framework::LibraryType::kPlain &&
+        platform::CanMKLDNNBeUsed(ctx)) {
+      library_ = framework::LibraryType::kMKLDNN;
+      layout_ = framework::DataLayout::kMKLDNN;
+    }
+#endif
+    return framework::OpKernelType(
+        ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"))->type(),
+        ctx.GetPlace(), layout_, library_);
+  }
 };
 
 // FIXME(zcd): transpose2 adds an intermediate output(XShape) based on
@@ -237,9 +251,19 @@ class Transpose2OpGrad : public framework::OperatorWithKernel {
  protected:
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
+    framework::LibraryType library_{framework::LibraryType::kPlain};
+    std::string data_format = ctx.Attr<std::string>("data_format");
+    framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
+#ifdef PADDLE_WITH_MKLDNN
+    if (library_ == framework::LibraryType::kPlain &&
+        platform::CanMKLDNNBeUsed(ctx)) {
+      library_ = framework::LibraryType::kMKLDNN;
+      layout_ = framework::DataLayout::kMKLDNN;
+    }
+#endif
     return framework::OpKernelType(
         ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"))->type(),
-        ctx.device_context());
+        ctx.GetPlace(), layout_, library_);
   }
 };
 

paddle/fluid/platform/CMakeLists.txt

@@ -56,6 +56,8 @@ ELSE()
     set(MKLDNN_CTX_DEPS)
 ENDIF()
 
+cc_library(temp_allocator SRCS temporary_allocator.cc DEPS  allocator_facade)
+
 nv_library(stream_callback_manager SRCS stream_callback_manager.cc DEPS simple_threadpool enforce) 
 IF(WITH_GPU)
   set(STREAM_CALLBACK_DEPS stream_callback_manager)
@@ -66,7 +68,8 @@ ENDIF()
 # memcpy depends on device_context, here add deps individually for
 # avoiding cycle dependencies
 cc_library(device_context SRCS device_context.cc init.cc DEPS simple_threadpool malloc ${STREAM_CALLBACK_DEPS}
-    place eigen3 stringpiece cpu_helper cpu_info framework_proto ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS})
+    place eigen3 stringpiece cpu_helper cpu_info framework_proto ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS}  temp_allocator)
+
 if(WIN32)
     if(WITH_GPU AND NOT WITH_DSO)
         get_property(cuda_modules GLOBAL PROPERTY CUDA_MODULES)
@@ -92,3 +95,9 @@ IF(WITH_GPU)
   nv_test(cuda_helper_test SRCS cuda_helper_test.cu)
 ENDIF()
 nv_library(cuda_device_guard SRCS cuda_device_guard.cc DEPS gpu_info)
+
+if(WITH_GPU)
+    nv_test(temporal_allocator_test SRCS temporary_allocator_test.cc DEPS temp_allocator tensor)
+else()
+    cc_test(temporal_allocator_test SRCS temporary_allocator_test.cc DEPS temp_allocator tensor)
+endif()

paddle/fluid/platform/cpu_info.cc

@@ -22,7 +22,6 @@ limitations under the License. */
 #ifdef __APPLE__
 #include <sys/sysctl.h>
 #include <sys/types.h>
-
 #elif defined(_WIN32)
 #define NOMINMAX  // msvc max/min macro conflict with std::min/max
 #include <windows.h>

paddle/fluid/platform/create_tensor_with_allocationptr.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/platform/temporary_allocator.h"
+namespace paddle {
+namespace platform {
+
+template <typename T>
+paddle::framework::Tensor GetTensor(
+    memory::allocation::AllocationPtr temp_allocation_ptr,
+    const framework::DDim &dim) {
+  auto &deleter = temp_allocation_ptr.get_deleter();
+  auto *allocation_ptr = temp_allocation_ptr.release();
+  auto shared_allocation =
+      std::shared_ptr<memory::allocation::Allocation>(allocation_ptr, deleter);
+
+  PADDLE_ENFORCE(dynamic_cast<TemporaryAllocation *>(allocation_ptr) != nullptr,
+                 "The AllocationPtr must be TemporaryAllocation.");
+  PADDLE_ENFORCE_EQ(allocation_ptr->size(),
+                    framework::product(dim) * sizeof(T));
+
+  paddle::framework::Tensor temp_tensor(std::type_index(typeid(T)));
+  temp_tensor.Resize(dim);
+  temp_tensor.ResetHolder(std::move(shared_allocation));
+  return temp_tensor;
+}
+
+}  // namespace platform
+}  // namespace paddle

paddle/fluid/platform/device_context.cc

@@ -85,6 +85,49 @@ DeviceContextPool::DeviceContextPool(
   }
 }
 
+DeviceTemporaryAllocator* DeviceTemporaryAllocator::allocators = nullptr;
+
+#ifdef PADDLE_WITH_CUDA
+platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
+    const platform::Place& place, const cudaStream_t& stream) {
+  PADDLE_ENFORCE(platform::is_gpu_place(place));
+  auto place_stream = std::make_pair(place, stream);
+  {
+    std::unique_lock<std::mutex> lock(mtx_);
+    if (!device_allocator_.count(place_stream)) {
+      device_allocator_[place_stream].reset(new TemporaryAllocator(place));
+      device_allocator_[place_stream]->SetCallback([stream]() {
+        PADDLE_ENFORCE(cudaStreamSynchronize(stream));
+        PADDLE_ENFORCE(cudaGetLastError());
+      });
+    }
+  }
+  return *device_allocator_.at(place_stream);
+}
+
+template <>
+platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
+    const platform::CUDADeviceContext& dev_ctx) {
+  auto place_stream = std::make_pair(dev_ctx.GetPlace(), dev_ctx.stream());
+  if (device_allocator_.count(place_stream)) {
+    return *device_allocator_.at(place_stream);
+  }
+  return Get(dev_ctx.GetPlace(), dev_ctx.stream());
+}
+#endif
+
+template <>
+platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
+    const platform::CPUDeviceContext& dev_ctx) {
+  return cpu_allocator_;
+}
+
+platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
+    const platform::Place& place) {
+  PADDLE_ENFORCE(platform::is_cpu_place(place), "You should pass CPUPlace");
+  return cpu_allocator_;
+}
+
 CPUDeviceContext::CPUDeviceContext() {
   eigen_device_.reset(new Eigen::DefaultDevice());
 }
@@ -271,8 +314,12 @@ CUDADeviceContext::~CUDADeviceContext() {
 Place CUDADeviceContext::GetPlace() const { return place_; }
 
 void CUDADeviceContext::Wait() const {
+  auto& allocator =
+      DeviceTemporaryAllocator::Instance().Get<CUDADeviceContext>(*this);
+  allocator.Release([=]() {
     PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
     PADDLE_ENFORCE(cudaGetLastError());
+  });
 }
 
 int CUDADeviceContext::GetComputeCapability() const {

paddle/fluid/platform/device_context.h

@@ -15,8 +15,10 @@ limitations under the License. */
 #include <mutex>  // NOLINT
 #include <string>
 #include <unordered_map>
+#include <utility>
 #include <vector>
 #include "paddle/fluid/memory/malloc.h"
+#include "paddle/fluid/platform/temporary_allocator.h"
 #ifdef PADDLE_WITH_CUDA
 #include "paddle/fluid/platform/dynload/cublas.h"
 #include "paddle/fluid/platform/dynload/cudnn.h"
@@ -39,6 +41,50 @@ limitations under the License. */
 namespace paddle {
 namespace platform {
 
+/*! \brief device temporary allocator singleton */
+class DeviceTemporaryAllocator {
+ public:
+  static DeviceTemporaryAllocator& Instance() {
+    PADDLE_ENFORCE_NOT_NULL(allocators,
+                            "Need to Create DeviceTemporaryAllocator first!");
+    return *allocators;
+  }
+
+  static DeviceTemporaryAllocator& Init() {
+    if (allocators == nullptr) {
+      allocators = new DeviceTemporaryAllocator();
+    }
+    return *allocators;
+  }
+
+/*! \brief  Return handle of single temporary allocator. */
+#ifdef PADDLE_WITH_CUDA
+  platform::TemporaryAllocator& Get(const platform::Place& place,
+                                    const cudaStream_t& stream);
+#endif
+  template <typename DeviceContext>
+  platform::TemporaryAllocator& Get(const DeviceContext& dev_ctx);
+
+  platform::TemporaryAllocator& Get(const platform::Place& place);
+
+ private:
+  DeviceTemporaryAllocator() : cpu_allocator_(platform::CPUPlace()) {}
+
+  static DeviceTemporaryAllocator* allocators;
+
+  platform::TemporaryAllocator cpu_allocator_;
+
+#ifdef PADDLE_WITH_CUDA
+  std::map<std::pair<platform::Place, cudaStream_t>,
+           std::unique_ptr<platform::TemporaryAllocator>>
+      device_allocator_;
+#endif
+
+  std::mutex mtx_;
+
+  DISABLE_COPY_AND_ASSIGN(DeviceTemporaryAllocator);
+};
+
 class DeviceContext {
  public:
   virtual ~DeviceContext() {}

paddle/fluid/platform/dynload/dynamic_loader.cc

@@ -227,6 +227,8 @@ void* GetTensorRtDsoHandle() {
 void* GetMKLMLDsoHandle() {
 #if defined(__APPLE__) || defined(__OSX__)
   return GetDsoHandleFromSearchPath(FLAGS_mklml_dir, "libmklml_intel.dylib");
+#elif defined(_WIN32)
+  return GetDsoHandleFromSearchPath(FLAGS_mklml_dir, "mklml.dll");
 #else
   return GetDsoHandleFromSearchPath(FLAGS_mklml_dir, "libmklml_intel.so");
 #endif

paddle/fluid/platform/init.cc

@@ -110,7 +110,7 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
   }
   places.emplace_back(platform::CPUPlace());
   platform::DeviceContextPool::Init(places);
-
+  platform::DeviceTemporaryAllocator::Init();
 #ifndef PADDLE_WITH_MKLDNN
   platform::SetNumThreads(FLAGS_paddle_num_threads);
 #endif

paddle/fluid/platform/mkldnn_reuse.h

@@ -197,6 +197,130 @@ class MKLDNNHandler {
   bool is_reusing_;
 };
 
+class TransposeMKLDNNHandler : public MKLDNNHandler {
+ public:
+  TransposeMKLDNNHandler(std::vector<int>& dims, std::vector<int>& axis,
+                         const platform::MKLDNNDeviceContext& dev_ctx,
+                         mkldnn::engine engine, const std::string& base_key)
+      : platform::MKLDNNHandler(dev_ctx, engine, base_key),
+        dims_(dims),
+        axis_(axis),
+        logical_axis_(dims.size(), 0) {}
+
+  std::shared_ptr<mkldnn::memory> AcquireSrcMemory(
+      const mkldnn::memory::format& fmt, void* ptr) {
+    auto local_key = key_ + "@user_src_mem_p";
+    auto mem_p =
+        std::static_pointer_cast<mkldnn::memory>(dev_ctx_.GetBlob(local_key));
+    PADDLE_ENFORCE((mem_p != nullptr) || (is_reusing_ == false),
+                   " find mem primitive in device context");
+    if (mem_p == nullptr) {
+      // Make memory descriptor using input format, unless it
+      // cannot be trusted (nchw) then make up memory fmt manually
+      for (size_t i = 0; i < logical_axis_.size(); ++i) {
+        logical_axis_[i] = i;
+      }
+      auto src_md = fmt != mkldnn::memory::format::nchw
+                        ? platform::MKLDNNMemDesc(
+                              dims_, platform::MKLDNNGetDataType<float>(), fmt)
+                        : Axis2MemoryDesc(dims_, logical_axis_);
+      mem_p = std::make_shared<mkldnn::memory>(
+          mkldnn::memory::primitive_desc{src_md, engine_}, ptr);
+      dev_ctx_.SetBlob(local_key, mem_p);
+    } else {
+      mem_p->set_data_handle(ptr);
+      // Mark that reusing happenned. All primitives from operator instance
+      // should be reused or none of them. So we check consistency
+      is_reusing_ = true;
+    }
+    return mem_p;
+  }
+
+  std::shared_ptr<mkldnn::memory> AcquireDstMemory(framework::Tensor* output,
+                                                   platform::Place place) {
+    auto local_key = key_ + "@user_dst_mem_p";
+    auto mem_p =
+        std::static_pointer_cast<mkldnn::memory>(dev_ctx_.GetBlob(local_key));
+    PADDLE_ENFORCE((mem_p != nullptr) || (is_reusing_ == false),
+                   " find mem primitive in device context");
+    if (mem_p == nullptr) {
+      auto dst_mdp = mkldnn::memory::primitive_desc{
+          Axis2MemoryDesc(dims_, axis_), engine_};
+
+      auto dst_data = output->mutable_data<float>(
+          place, paddle::memory::Allocator::kDefault, dst_mdp.get_size());
+
+      mem_p = std::make_shared<mkldnn::memory>(dst_mdp, dst_data);
+      dev_ctx_.SetBlob(local_key, mem_p);
+    } else {
+      auto dst_data = output->mutable_data<float>(place);
+      mem_p->set_data_handle(dst_data);
+      // Mark that reusing happenned. All primitives from operator instance
+      // should be reused or none of them. So we check consistency
+      is_reusing_ = true;
+    }
+    return mem_p;
+  }
+
+  std::shared_ptr<mkldnn::reorder> AcquireTranspose(
+      std::shared_ptr<mkldnn::memory> dst_memory_p,
+      std::shared_ptr<mkldnn::memory> src_memory_p) {
+    auto prim_key = key_ + "@transpose_p";
+    auto transpose_p =
+        std::static_pointer_cast<mkldnn::reorder>(dev_ctx_.GetBlob(prim_key));
+    PADDLE_ENFORCE((transpose_p != nullptr) || (is_reusing_ == false),
+                   "Fail to find convolution primitive in device context");
+    if (transpose_p == nullptr) {
+      transpose_p =
+          std::make_shared<mkldnn::reorder>(*(src_memory_p), *(dst_memory_p));
+      dev_ctx_.SetBlob(prim_key, transpose_p);
+    } else {
+      is_reusing_ = true;
+    }
+    return transpose_p;
+  }
+
+  static std::string GetHash(std::vector<int>& shape,  // NOLINT
+                             std::vector<int>& axis,   // NOLINT
+                             const std::string& suffix) {
+    return dims2str(shape) + dims2str(axis) + suffix;
+  }
+
+ protected:
+  mkldnn_memory_desc_t Axis2MemoryDesc(std::vector<int>& nchw_tz,
+                                       std::vector<int>& axis) {
+    mkldnn_memory_desc_t mem_fmt;
+
+    mem_fmt.primitive_kind = mkldnn_memory;
+    mem_fmt.ndims = axis.size();
+    for (unsigned int i = 0; i < nchw_tz.size(); ++i) {
+      mem_fmt.dims[i] = nchw_tz[i];  // logical dimensions (nchw format,
+                                     // regardless physical layout)
+    }
+    mem_fmt.data_type = mkldnn_f32;
+    mem_fmt.format = mkldnn_blocked;
+
+    unsigned int total_stride = 1;
+    for (int i = nchw_tz.size() - 1; i >= 0; --i) {
+      mem_fmt.layout_desc.blocking.padding_dims[i] =
+          nchw_tz[i];  // logical dimensions (nchw format, regardless physical
+                       // layout)
+      mem_fmt.layout_desc.blocking.block_dims[i] = 1;
+      mem_fmt.layout_desc.blocking.offset_padding_to_data[i] = 0;  // no offset
+      mem_fmt.layout_desc.blocking.strides[0][axis[i]] = total_stride;
+      mem_fmt.layout_desc.blocking.strides[1][axis[i]] = 1;
+      total_stride *= nchw_tz[axis[i]];
+    }
+    mem_fmt.layout_desc.blocking.offset_padding = 0;  // no initial offset
+    return mem_fmt;
+  }
+
+ private:
+  std::vector<int> dims_;
+  std::vector<int> axis_;
+  std::vector<int> logical_axis_;
+};
+
 template <class forward_t, class backward_data_t, class backward_weights_t>
 class ConvMKLDNNTemplateHandler : public MKLDNNHandler {
  public:

paddle/fluid/platform/nccl_helper.h

@@ -23,7 +23,6 @@
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/platform/dynload/nccl.h"
 #include "paddle/fluid/platform/enforce.h"
-#include "paddle/fluid/platform/float16.h"
 
 #define NCCL_ID_VARNAME "NCCLID"
 
@@ -39,8 +38,6 @@ inline ncclDataType_t ToNCCLDataType(framework::proto::VarType::Type type) {
     return ncclInt;
   } else if (type == framework::proto::VarType::INT64) {
     return ncclInt64;
-  } else if (type == framework::proto::VarType::FP16) {
-    return ncclFloat16;
   } else {
     PADDLE_THROW("Not supported");
   }

paddle/fluid/platform/temporary_allocator.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/platform/temporary_allocator.h"
+#include "paddle/fluid/memory/allocation/allocator_facade.h"
+
+DEFINE_double(limit_of_temporary_allocation, -1,
+              "The up limit of temporary_allocation size.");
+
+namespace paddle {
+namespace platform {
+namespace alloc = memory::allocation;
+
+TemporaryAllocation::TemporaryAllocation(
+    alloc::AllocationPtr &&underlying_allocation)
+    : Allocation(underlying_allocation->ptr(), underlying_allocation->size(),
+                 underlying_allocation->place()),
+      underlying_allocation_(std::move(underlying_allocation)) {}
+
+TemporaryAllocator::TemporaryAllocator(platform::Place place) : place_(place) {
+  temp_mem_queue_.reset(new std::deque<TemporaryAllocation *>());
+}
+
+bool TemporaryAllocator::IsAllocThreadSafe() const { return true; }
+
+void TemporaryAllocator::Release(const std::function<void()> &callback) {
+  std::shared_ptr<std::deque<TemporaryAllocation *>> t_allocations;
+  {
+    std::unique_lock<std::mutex> lock(mtx_);
+    callback();
+    t_allocations = temp_mem_queue_;
+    temp_mem_queue_.reset(new std::deque<TemporaryAllocation *>());
+    wait_delete_mem_ = 0;
+  }
+  for (auto tmp : *t_allocations) {
+    VLOG(10) << "Delete temporary allocation " << tmp->ptr()
+             << " size: " << tmp->size();
+    delete tmp;
+  }
+}
+
+void TemporaryAllocator::Free(alloc::Allocation *allocation) {
+  auto *temp_allocation = dynamic_cast<TemporaryAllocation *>(allocation);
+  PADDLE_ENFORCE_NOT_NULL(temp_allocation);
+  if (platform::is_gpu_place(temp_allocation->place())) {
+    size_t wait_delete_mem = 0;
+    {
+      std::unique_lock<std::mutex> lock(mtx_);
+      temp_mem_queue_->emplace_back(temp_allocation);
+      wait_delete_mem_ += temp_allocation->size();
+      wait_delete_mem = wait_delete_mem_;
+      VLOG(10) << "Move temporary allocation: " << temp_allocation->ptr()
+               << " to delete queue: " << temp_allocation->size() << "; "
+               << "wait_delete_mem: " << wait_delete_mem_;
+    }
+    if (FLAGS_limit_of_temporary_allocation > 0 &&
+        wait_delete_mem > FLAGS_limit_of_temporary_allocation) {
+      Release(callback_);
+    }
+    return;
+  }
+  delete temp_allocation;
+}
+
+size_t TemporaryAllocator::TemporaryAllocationQueueSize() {
+  std::unique_lock<std::mutex> lock(mtx_);
+  return temp_mem_queue_ ? temp_mem_queue_->size() : 0;
+}
+
+void TemporaryAllocator::SetCallback(const std::function<void()> &callback) {
+  callback_ = callback;
+}
+
+alloc::Allocation *TemporaryAllocator::AllocateImpl(
+    size_t size, alloc::Allocator::Attr attr) {
+  auto raw_allocation =
+      alloc::AllocatorFacade::Instance().Alloc(place_, size, attr);
+  auto temp_mem = new TemporaryAllocation(std::move(raw_allocation));
+  VLOG(10) << "Alloc temporary allocation: " << temp_mem->ptr() << ": " << size;
+  return temp_mem;
+}
+
+}  // namespace platform
+}  // namespace paddle

paddle/fluid/platform/temporary_allocator.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 <condition_variable>  // NOLINT
+#include <deque>
+#include <mutex>  // NOLINT
+#include "paddle/fluid/memory/allocation/allocator.h"
+#include "paddle/fluid/platform/lock_guard_ptr.h"
+namespace paddle {
+namespace platform {
+
+class TemporaryAllocation : public memory::allocation::Allocation {
+ public:
+  explicit TemporaryAllocation(
+      memory::allocation::AllocationPtr &&underlying_allocation);
+
+  memory::allocation::AllocationPtr underlying_allocation_;
+};
+
+class TemporaryAllocator : public memory::allocation::Allocator {
+ public:
+  explicit TemporaryAllocator(platform::Place place);
+
+  void Release(const std::function<void()> &callback);
+
+  size_t TemporaryAllocationQueueSize();
+
+  bool IsAllocThreadSafe() const override;
+
+  void SetCallback(const std::function<void()> &callback);
+
+ protected:
+  void Free(memory::allocation::Allocation *allocation) override;
+
+  memory::allocation::Allocation *AllocateImpl(
+      size_t size, memory::allocation::Allocator::Attr attr) override;
+
+ private:
+  platform::Place place_;
+
+  // When the allocation is not held by any variable, it should be placed
+  // to temp_mem_queue immediately.
+  std::shared_ptr<std::deque<TemporaryAllocation *>> temp_mem_queue_{nullptr};
+
+  std::mutex mtx_;
+  size_t wait_delete_mem_{0};
+  std::function<void()> callback_;
+};
+
+}  // namespace platform
+}  // namespace paddle

paddle/fluid/platform/temporary_allocator_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/platform/temporary_allocator.h"
+#include <gtest/gtest.h>
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/platform/create_tensor_with_allocationptr.h"
+DECLARE_double(limit_of_temporary_allocation);
+
+namespace paddle {
+namespace platform {
+
+TEST(temporary_allocator, temporary_allocator) {
+  platform::CPUPlace cpu_place;
+  TemporaryAllocator alloc(cpu_place);
+  alloc.Allocate(100);
+
+#ifdef PADDLE_WITH_CUDA
+  platform::CUDAPlace gpu_place(0);
+  TemporaryAllocator gpu_alloc(gpu_place);
+
+  auto allocation = gpu_alloc.Allocate(101);
+  PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
+  gpu_alloc.Release([]() {});
+  PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
+
+  {
+    auto allocation = gpu_alloc.Allocate(102);
+    PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
+  }
+  PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 1);
+  gpu_alloc.Release([]() {});
+  PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
+#endif
+}
+
+TEST(temporary_allocator, add_callback) {
+#ifdef PADDLE_WITH_CUDA
+  FLAGS_limit_of_temporary_allocation = 10;
+  platform::CUDAPlace gpu_place(0);
+  TemporaryAllocator gpu_alloc(gpu_place);
+
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+  auto* dev_ctx =
+      static_cast<platform::CUDADeviceContext*>(pool.Get(gpu_place));
+  auto stream = dev_ctx->stream();
+  bool deleted = false;
+  gpu_alloc.SetCallback([stream, &deleted]() {
+    PADDLE_ENFORCE(cudaStreamSynchronize(stream));
+    PADDLE_ENFORCE(cudaGetLastError());
+    deleted = true;
+  });
+  { gpu_alloc.Allocate(100); }
+  PADDLE_ENFORCE(deleted);
+  FLAGS_limit_of_temporary_allocation = -1;
+#endif
+}
+
+TEST(temporary_allocator, create_tensor_with_allocationptr) {
+  platform::CPUPlace cpu_place;
+  TemporaryAllocator cpu_alloc(cpu_place);
+  {
+    size_t memory_size = 200;
+    auto allocation = cpu_alloc.Allocate(memory_size);
+    void* address = allocation->ptr();
+    int numel = memory_size / sizeof(float);
+    framework::Tensor tensor =
+        GetTensor<float>(std::move(allocation), framework::make_ddim({numel}));
+    PADDLE_ENFORCE_EQ(address, tensor.data<float>());
+    PADDLE_ENFORCE_EQ(tensor.numel(), numel);
+  }
+
+#ifdef PADDLE_WITH_CUDA
+  platform::CUDAPlace gpu_place(0);
+  TemporaryAllocator gpu_alloc(gpu_place);
+
+  {
+    size_t memory_size = 300;
+    auto allocation = gpu_alloc.Allocate(memory_size);
+    void* address = allocation->ptr();
+    int numel = memory_size / sizeof(float);
+    framework::Tensor tensor =
+        GetTensor<float>(std::move(allocation), framework::make_ddim({numel}));
+    PADDLE_ENFORCE_EQ(address, tensor.data<float>());
+    PADDLE_ENFORCE_EQ(tensor.numel(), numel);
+  }
+
+  // The allocation is not holded now, it should be placed to
+  // TemporaryAllocationQueue.
+  PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 1);
+  gpu_alloc.Release([]() {});
+  PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
+#endif
+}
+
+TEST(temporary_allocator, create_tensor_with_allocationptr2) {
+  platform::CPUPlace cpu_place;
+  TemporaryAllocator cpu_alloc(cpu_place);
+  {
+    size_t memory_size = 400;
+    int numel = memory_size / sizeof(float);
+
+    framework::Tensor out_side_tensor;
+    void* address;
+    {
+      auto allocation = cpu_alloc.Allocate(memory_size);
+      address = allocation->ptr();
+      framework::Tensor tensor = GetTensor<float>(
+          std::move(allocation), framework::make_ddim({numel}));
+      PADDLE_ENFORCE_EQ(address, tensor.data<float>());
+      PADDLE_ENFORCE_EQ(tensor.numel(), numel);
+
+      out_side_tensor.ShareDataWith(tensor);
+    }
+    PADDLE_ENFORCE_EQ(address, out_side_tensor.data<float>());
+    PADDLE_ENFORCE_EQ(out_side_tensor.numel(), numel);
+  }
+
+#ifdef PADDLE_WITH_CUDA
+  platform::CUDAPlace gpu_place(0);
+  TemporaryAllocator gpu_alloc(gpu_place);
+  {
+    void* address;
+    size_t memory_size = 500;
+    int numel = memory_size / sizeof(float);
+    framework::Tensor out_side_tensor;
+    {
+      auto allocation = gpu_alloc.Allocate(memory_size);
+      address = allocation->ptr();
+      framework::Tensor tensor = GetTensor<float>(
+          std::move(allocation), framework::make_ddim({numel}));
+      PADDLE_ENFORCE_EQ(address, tensor.data<float>());
+      PADDLE_ENFORCE_EQ(tensor.numel(), numel);
+
+      out_side_tensor.ShareDataWith(tensor);
+    }
+    PADDLE_ENFORCE_EQ(address, out_side_tensor.data<float>());
+    PADDLE_ENFORCE_EQ(out_side_tensor.numel(), numel);
+    // The allocation is holded by out_side_tensor.
+    PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
+    gpu_alloc.Release([]() {});
+    PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
+  }
+
+  // The allocation is not holded now, it should be placed to
+  // TemporaryAllocationQueue.
+  PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 1);
+  gpu_alloc.Release([]() {});
+  PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
+#endif
+}
+
+}  //  namespace platform
+}  //  namespace paddle

paddle/fluid/pybind/CMakeLists.txt

 
 set(PYBIND_DEPS pybind python proto_desc memory executor async_executor prune feed_fetch_method pass_builder parallel_executor profiler layer)
+if(WITH_PYTHON)
+  list(APPEND PYBIND_DEPS py_func_op)
+endif()
 set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc async_executor_py.cc imperative.cc)
 
 if(WITH_PYTHON)

paddle/fluid/pybind/protobuf.cc

@@ -328,7 +328,7 @@ void BindOpDesc(pybind11::module *m) {
       .def("infer_var_type", &pd::OpDesc::InferVarType)
       .def("set_is_target", &pd::OpDesc::SetIsTarget)
       .def("serialize_to_string", SerializeMessage<pd::OpDesc>)
-      .def("block", &pd::OpDesc::Block,
+      .def("block", [](pd::OpDesc &self) { return self.Block(); },
            pybind11::return_value_policy::reference);
 }
 

paddle/fluid/pybind/pybind.cc

@@ -37,6 +37,7 @@ limitations under the License. */
 #include "paddle/fluid/imperative/layer.h"
 #include "paddle/fluid/memory/allocation/allocator_strategy.h"
 #include "paddle/fluid/operators/activation_op.h"
+#include "paddle/fluid/operators/py_func_op.h"
 #include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h"
 #include "paddle/fluid/platform/cpu_info.h"
 #include "paddle/fluid/platform/enforce.h"
@@ -110,6 +111,12 @@ PYBIND11_MODULE(core, m) {
 
   BindException(&m);
 
+  m.def(
+      "_append_python_callable_object_and_return_id",
+      [](py::object py_obj) -> size_t {
+        return paddle::operators::AppendPythonCallableObjectAndReturnId(py_obj);
+      });
+
   py::class_<imperative::VarBase, PyVarBase>(m, "VarBase", R"DOC()DOC")
       .def(py::init<>())
       .def("_run_backward",

paddle/fluid/train/demo/CMakeLists.txt

@@ -35,16 +35,26 @@ add_executable(demo_trainer demo_trainer.cc)
 
 if(WITH_MKLDNN)
   include_directories("${PADDLE_LIB}/third_party/install/mkldnn/include")
+  if(WIN32)
+    set(MKLDNN_LIB ${PADDLE_LIB}/third_party/install/mkldnn/lib/mkldnn.lib)
+  else(WIN32)
     set(MKLDNN_LIB ${PADDLE_LIB}/third_party/install/mkldnn/lib/libmkldnn.so.0)
-endif()
+  endif(WIN32)
+endif(WITH_MKLDNN)
 
 if(WITH_MKL)
   include_directories("${PADDLE_LIB}/third_party/install/mklml/include")
+  if(WIN32)
+    set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/mklml.lib)
+  else(WIN32)
     set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel.so)
+  endif(WIN32)
 else()
   if(APPLE)
     set(MATH_LIB cblas)
-  else(APPLE)
+  elseif(WIN32)
+    set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas.lib)
+  else()
     set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas.a)
   endif(APPLE)
 endif()

python/CMakeLists.txt

@@ -48,18 +48,13 @@ configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in
 IF(WIN32)
     # Python would use the .pyd by default under Windows series platform
     set(FLUID_DST_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/)
-    get_filename_component(openblas_refpath ${CBLAS_LIBRARIES} DIRECTORY)
     set(FLUID_CORE ${FLUID_DST_DIR}/core.pyd)
-    add_custom_command(OUTPUT ${FLUID_CORE}
-            COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${FLUID_CORE}
-            COMMAND cmake -E copy ${openblas_refpath}/openblas.dll ${FLUID_DST_DIR}
-            DEPENDS paddle_pybind)
 ELSE()
     set(FLUID_CORE ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so)
-    add_custom_command(OUTPUT ${FLUID_CORE}
+ENDIF()
+add_custom_command(OUTPUT ${FLUID_CORE}
         COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${FLUID_CORE}
         DEPENDS paddle_pybind)
-ENDIF()
 add_custom_target(copy_paddle_pybind ALL DEPENDS ${FLUID_CORE})
 
 IF(WIN32)

python/paddle/fluid/data_feeder.py

@@ -44,8 +44,6 @@ class DataToLoDTensorConverter(object):
             self.dtype = 'int64'
         elif dtype == core.VarDesc.VarType.FP64:
             self.dtype = 'float64'
-        elif dtype == core.VarDesc.VarType.FP16:
-            self.dtype = 'float16'
         elif dtype == core.VarDesc.VarType.INT32:
             self.dtype = 'int32'
         elif dtype == core.VarDesc.VarType.UINT8:

python/paddle/fluid/initializer.py

@@ -18,7 +18,6 @@ from . import framework
 import numpy as np
 import contextlib
 from .core import VarDesc
-from . import unique_name
 
 __all__ = [
     'Constant', 'Uniform', 'Normal', 'TruncatedNormal', 'Xavier', 'Bilinear',
@@ -208,39 +207,16 @@ class UniformInitializer(Initializer):
         # Initialization Ops should be prepended and not appended
         if self._seed == 0:
             self._seed = block.program.random_seed
-
-        # to be compatible of fp16 initalizers
-        if var.dtype == VarDesc.VarType.FP16:
-            out_dtype = VarDesc.VarType.FP32
-            out_var = block.create_var(
-                name=unique_name.generate(".".join(['gaussian_random', 'tmp'])),
-                shape=var.shape,
-                dtype=out_dtype,
-                type=VarDesc.VarType.LOD_TENSOR,
-                persistable=False)
-        else:
-            out_dtype = var.dtype
-            out_var = var
-
         op = block._prepend_op(
             type="uniform_random",
-            outputs={"Out": out_var},
+            outputs={"Out": var},
             attrs={
                 "shape": var.shape,
-                "dtype": out_dtype,
+                "dtype": int(var.dtype),
                 "min": self._low,
                 "max": self._high,
                 "seed": self._seed
             })
-
-        if var.dtype == VarDesc.VarType.FP16:
-            block.append_op(
-                type="cast",
-                inputs={"X": out_var},
-                outputs={"Out": var},
-                attrs={"in_dtype": out_var.dtype,
-                       "out_dtype": var.dtype})
-
         var.op = op
         return op
 
@@ -285,39 +261,17 @@ class NormalInitializer(Initializer):
         # Initialization Ops should be prepended and not appended
         if self._seed == 0:
             self._seed = block.program.random_seed
-
-        # to be compatible of fp16 initalizers
-        if var.dtype == VarDesc.VarType.FP16:
-            out_dtype = VarDesc.VarType.FP32
-            out_var = block.create_var(
-                name=unique_name.generate(".".join(['gaussian_random', 'tmp'])),
-                shape=var.shape,
-                dtype=out_dtype,
-                type=VarDesc.VarType.LOD_TENSOR,
-                persistable=False)
-        else:
-            out_dtype = var.dtype
-            out_var = var
-
         op = block._prepend_op(
             type="gaussian_random",
-            outputs={"Out": out_var},
+            outputs={"Out": var},
             attrs={
                 "shape": var.shape,
-                "dtype": out_dtype,
+                "dtype": int(var.dtype),
                 "mean": self._mean,
                 "std": self._std_dev,
                 "seed": self._seed,
                 "use_mkldnn": False
             })
-
-        if var.dtype == VarDesc.VarType.FP16:
-            block.append_op(
-                type="cast",
-                inputs={"X": out_var},
-                outputs={"Out": var},
-                attrs={"in_dtype": out_var.dtype,
-                       "out_dtype": var.dtype})
         var.op = op
         return op
 

python/paddle/fluid/layers/learning_rate_scheduler.py

@@ -63,8 +63,6 @@ def noam_decay(d_model, warmup_steps):
     Returns:
         The decayed learning rate.
     """
-
-    def _lr_schedule(dtype):
     with default_main_program()._lr_schedule_guard():
         global_step = _decay_step_counter(1)
 
@@ -74,8 +72,6 @@ def noam_decay(d_model, warmup_steps):
 
     return lr_value
 
-    return _lr_schedule
-
 
 def exponential_decay(learning_rate, decay_steps, decay_rate, staircase=False):
     """
@@ -113,8 +109,6 @@ def exponential_decay(learning_rate, decay_steps, decay_rate, staircase=False):
           sgd_optimizer.minimize(avg_cost)
 
     """
-
-    def _lr_schedule(dtype):
     with default_main_program()._lr_schedule_guard():
         global_step = _decay_step_counter()
 
@@ -125,8 +119,6 @@ def exponential_decay(learning_rate, decay_steps, decay_rate, staircase=False):
 
         return decayed_lr
 
-    return _lr_schedule
-
 
 def natural_exp_decay(learning_rate, decay_steps, decay_rate, staircase=False):
     """Applies natural exponential decay to the initial learning rate.
@@ -146,8 +138,6 @@ def natural_exp_decay(learning_rate, decay_steps, decay_rate, staircase=False):
     Returns:
         The decayed learning rate
     """
-
-    def _lr_schedule(dtype):
     with default_main_program()._lr_schedule_guard():
         global_step = _decay_step_counter()
 
@@ -158,8 +148,6 @@ def natural_exp_decay(learning_rate, decay_steps, decay_rate, staircase=False):
 
         return decayed_lr
 
-    return _lr_schedule
-
 
 def inverse_time_decay(learning_rate, decay_steps, decay_rate, staircase=False):
     """
@@ -196,8 +184,6 @@ def inverse_time_decay(learning_rate, decay_steps, decay_rate, staircase=False):
                     staircase=True))
           sgd_optimizer.minimize(avg_cost)
     """
-
-    def _lr_schedule(dtype):
     with default_main_program()._lr_schedule_guard():
         global_step = _decay_step_counter()
 
@@ -209,8 +195,6 @@ def inverse_time_decay(learning_rate, decay_steps, decay_rate, staircase=False):
 
         return decayed_lr
 
-    return _lr_schedule
-
 
 def polynomial_decay(learning_rate,
                      decay_steps,
@@ -240,17 +224,15 @@ def polynomial_decay(learning_rate,
     Returns:
         Variable: The decayed learning rate
     """
-
-    def _lr_schedule(dtype, decay_steps=decay_steps):
     with default_main_program()._lr_schedule_guard():
         global_step = _decay_step_counter()
 
         if cycle:
             div_res = ops.ceil(global_step / decay_steps)
             zero_var = tensor.fill_constant(
-                    shape=[1], dtype=dtype, value=0.0)
+                shape=[1], dtype='float32', value=0.0)
             one_var = tensor.fill_constant(
-                    shape=[1], dtype=dtype, value=1.0)
+                shape=[1], dtype='float32', value=1.0)
 
             with control_flow.Switch() as switch:
                 with switch.case(global_step == zero_var):
@@ -258,16 +240,13 @@ def polynomial_decay(learning_rate,
             decay_steps = decay_steps * div_res
         else:
             decay_steps_var = tensor.fill_constant(
-                    shape=[1], dtype=dtype, value=float(decay_steps))
-                global_step = nn.elementwise_min(
-                    x=global_step, y=decay_steps_var)
+                shape=[1], dtype='float32', value=float(decay_steps))
+            global_step = nn.elementwise_min(x=global_step, y=decay_steps_var)
 
         decayed_lr = (learning_rate - end_learning_rate) * \
             ((1 - global_step / decay_steps) ** power) + end_learning_rate
         return decayed_lr
 
-    return _lr_schedule
-
 
 def piecewise_decay(boundaries, values):
     """Applies piecewise decay to the initial learning rate.
@@ -294,8 +273,6 @@ def piecewise_decay(boundaries, values):
 
 
     """
-
-    def _lr_schedule(dtype):
     with default_main_program()._lr_schedule_guard():
         if len(values) - len(boundaries) != 1:
             raise ValueError("len(values) - len(boundaries) should be 1")
@@ -329,8 +306,6 @@ def piecewise_decay(boundaries, values):
 
     return lr
 
-    return _lr_schedule
-
 
 def append_LARS(params_grads, learning_rate, weight_decay):
     """

python/paddle/fluid/layers/nn.py

@@ -18,7 +18,9 @@ All layers just related to the neural network.
 from __future__ import print_function
 
 import numpy as np
+import six
 import os
+import inspect
 from ..layer_helper import LayerHelper
 from ..initializer import Normal, Constant
 from ..framework import Variable, OpProtoHolder
@@ -176,6 +178,7 @@ __all__ = [
     'merge_selected_rows',
     'get_tensor_from_selected_rows',
     'lstm',
+    'py_func',
     'psroi_pool',
     'huber_loss',
 ]
@@ -2798,10 +2801,6 @@ def batch_norm(input,
     helper = LayerHelper('batch_norm', **locals())
     dtype = helper.input_dtype()
 
-    # use fp32 for bn parameter
-    if dtype == core.VarDesc.VarType.FP16:
-        dtype = core.VarDesc.VarType.FP32
-
     input_shape = input.shape
     if data_layout == 'NCHW':
         channel_num = input_shape[1]
@@ -2836,7 +2835,7 @@ def batch_norm(input,
             trainable=False,
             do_model_average=do_model_average_for_mean_and_var),
         shape=param_shape,
-        dtype=dtype)
+        dtype=input.dtype)
     mean.stop_gradient = True
 
     variance = helper.create_parameter(
@@ -2846,7 +2845,7 @@ def batch_norm(input,
             trainable=False,
             do_model_average=do_model_average_for_mean_and_var),
         shape=param_shape,
-        dtype=dtype)
+        dtype=input.dtype)
     variance.stop_gradient = True
 
     # create output
@@ -9331,6 +9330,224 @@ def get_tensor_from_selected_rows(x, name=None):
     return out
 
 
+class PyFuncRegistry(object):
+    _register_funcs = []
+
+    def __init__(self, func):
+        if func is None or not callable(func):
+            raise TypeError('func must be a Python function')
+
+        self._func = func
+        # find named args using reflection 
+        args = inspect.getargspec(self._func)
+        if len(args[0]) == 0 and args[1] is None and args[2] is None:
+            # Function with no inputs
+            self._named_args = None
+        else:
+            self._named_args = args[0]
+        self._id = core._append_python_callable_object_and_return_id(self)
+        '''
+        Why record self here?
+
+        1. For debug usage. Users can call 
+           :code:`py_func.registered_func(idx)` method 
+           to find the registered function corresponding
+           to :code:`idx`. 
+
+        2. For increasing reference count of self. 
+           It seems that to release Python object 
+           whose reference count is 1 would cause
+           segmentation fault error in C++ side. 
+           May be lack of Python GC in C++ side?
+        '''
+        PyFuncRegistry._register_funcs.append(self)
+
+    @classmethod
+    def registered_func(cls, idx):
+        return cls._register_funcs[idx]._func
+
+    @classmethod
+    def registered_func_num(cls):
+        return len(cls._register_funcs)
+
+    @property
+    def id(self):
+        return self._id
+
+    def __call__(self, *args):
+        if self._named_args is None:
+            func_ret = self._func()
+        else:
+            kwargs = dict()
+            idx = 0
+            for arg in self._named_args:
+                kwargs[arg] = args[idx]
+                idx += 1
+            func_ret = self._func(*args[idx:], **kwargs)
+
+        if not isinstance(func_ret, (list, tuple)):
+            func_ret = (func_ret, )
+
+        ret = []
+        for each_ret in func_ret:
+            if each_ret is None or isinstance(each_ret, core.LoDTensor):
+                ret.append(each_ret)
+                continue
+
+            if not isinstance(each_ret, np.ndarray):
+                each_ret = np.array(each_ret)
+
+            tensor = core.LoDTensor()
+            tensor.set(each_ret, core.CPUPlace())
+            ret.append(tensor)
+
+        return tuple(ret)
+
+
+@templatedoc()
+def py_func(func, x, out, backward_func=None, skip_vars_in_backward_input=None):
+    """
+    PyFunc Operator.
+    
+    User can use :code:`py_func` to register operators in Python side.
+    The inputs of :code:`func` is :code:`LoDTensor` and outputs can be
+    numpy array or :code:`LoDTensor`. Paddle would call the registered
+    :code:`func` in forward part, and call :code:`backward_func` in
+    backward part (if :code:`backward_func` is not None).
+
+    User should set the right data type and shape of :code:`out` before
+    calling this function. However, data types and shapes of gradients of
+    :code:`out` and :code:`x` would be inferred automatically.
+
+    Input orders of :code:`backward_func` would be: forward inputs
+    :code:`x`, forward outputs :code:`out` and backward input gradients of
+    :code:`out`. If some variables of :code:`out` have no gradient, the input
+    tensor would be None in Python side. If some variables of :code:`in` have
+    no gradient, users should return None.
+
+    This function can also be used to debug the running network. User can
+    add a :code:`py_func` operator without output, and print input 
+    :code:`x` inside :code:`func`.
+
+    Args:
+        func (callable): forward Python function.
+        x (Variable|list(Variable)|tuple(Variable)): inputs of :code:`func`.
+        out (Variable|list(Variable)|tuple(Variable)): outputs of :code:`func`.
+            Paddle cannot infer shapes and data types of :code:`out`. Users
+            should create :code:`out` beforehand. 
+        backward_func (callable|None): backward Python function.
+                                       None means no backward. Default None. 
+        skip_vars_in_backward_input (Variable|list(Variable)|tuple(Variable)):
+            Variables that are not needed in :code:`backward_func` inputs. 
+            These variables must be any of :code:`x` and :code:`out`.
+            If set, these vars would not be inputs of :code:`backward_func`,
+            Only useful when :code:`backward_func` is not None. Default None. 
+
+    Returns:
+        out (Variable|list(Variable)|tuple(Variable)): input :code:`out`
+
+    Examples:
+    
+        >>> import paddle.fluid as fluid
+        >>> import six
+        >>>
+        >>> def create_tmp_var(name, dtype, shape):
+        >>>     return fluid.default_main_program().current_block().create_var(
+        >>>         name=name, dtype=dtype, shape=shape) 
+        >>>
+        >>> # tanh activation has been provided by Paddle C++ op
+        >>> # Here, we only use tanh to be an example to show the usage 
+        >>> # of py_func
+        >>> def tanh(x):
+        >>>     return np.tanh(x)
+        >>> 
+        >>> # forward input x is skipped
+        >>> def tanh_grad(y, dy):
+        >>>     return np.array(dy) * (1 - np.square(np.array(y)))
+        >>>
+        >>> def debug_func(x):
+        >>>     print(x) 
+        >>>
+        >>> def simple_net(img, label):
+        >>>     hidden = img
+        >>>     for idx in six.moves.range(4):
+        >>>         hidden = fluid.layers.fc(hidden, size=200)
+        >>>         new_hidden = create_tmp_var(name='hidden_{}'.format(idx),
+        >>>             dtype=hidden.dtype, shape=hidden.shape)    
+        >>>
+        >>>         # user-defined layers with forward and backward
+        >>>         hidden = fluid.layers.py_func(func=tanh, x=hidden, 
+        >>>             out=new_hidden, backward_func=tanh_grad, 
+        >>>             skip_vars_in_backward_input=hidden)
+        >>>
+        >>>         # user-defined debug layers to print variables
+        >>>         fluid.layers.py_func(func=debug_func, x=hidden, out=None)
+        >>>
+        >>>     prediction = fluid.layers.fc(hidden, size=10, act='softmax')
+        >>>     loss = fluid.layers.cross_entropy(input=prediction, label=label)
+        >>>     return fluid.layers.mean(loss)
+    """
+    helper = LayerHelper('py_func', **locals())
+    if x is None:
+        x = []
+    elif isinstance(x, Variable):
+        x = [x]
+    elif not isinstance(x, (list, tuple)):
+        raise TypeError('Input must be Variable/list(Variable)/tuple(Variable)')
+
+    if out is None:
+        out_list = []
+    elif isinstance(out, Variable):
+        out_list = [out]
+    elif isinstance(out, (list, tuple)):
+        out_list = out
+    else:
+        raise TypeError(
+            'Output must be Variable/list(Variable)/tuple(Variable)')
+
+    fwd_func_id = PyFuncRegistry(func).id
+    bwd_func_id = PyFuncRegistry(
+        backward_func).id if backward_func is not None else -1
+
+    for each_out in out_list:
+        if len(each_out.shape) == 0:
+            raise ValueError(
+                'Output shapes of py_func op should be provided by users manually'
+            )
+
+    backward_skip_vars = set()
+    if backward_func is not None and skip_vars_in_backward_input is not None:
+        if isinstance(skip_vars_in_backward_input, Variable):
+            skip_vars_in_backward_input = [skip_vars_in_backward_input]
+
+        fwd_in_out = [v.name for v in x]
+        fwd_in_out.extend([v.name for v in out_list])
+        fwd_in_out = set(fwd_in_out)
+        backward_skip_vars = set()
+        for v in skip_vars_in_backward_input:
+            if not v.name in fwd_in_out:
+                raise ValueError(
+                    'Variable {} is not found in forward inputs and outputs'
+                    .format(v.name))
+            backward_skip_vars.add(v.name)
+
+    helper.append_op(
+        type='py_func',
+        inputs={'X': x},
+        outputs={'Out': out_list},
+        attrs={
+            'forward_callable_id': fwd_func_id,
+            'backward_callable_id': bwd_func_id,
+            'backward_skip_vars': list(backward_skip_vars)
+        })
+    return out
+
+
+# For debug usage
+py_func.registered_func = PyFuncRegistry.registered_func
+py_func.registered_func_num = PyFuncRegistry.registered_func_num
+
+
 @templatedoc()
 def psroi_pool(input,
                rois,

python/paddle/fluid/net_drawer.py

@@ -26,7 +26,7 @@ logger = logging.getLogger(__name__)
 logger.setLevel(logging.INFO)
 
 try:
-    from .graphviz import Digraph
+    from .graphviz import Graph
 except ImportError:
     logger.info(
         'Cannot import graphviz, which is required for drawing a network. This '
@@ -112,7 +112,7 @@ def draw_graph(startup_program, main_program, **kwargs):
     filename = kwargs.get("filename")
     if filename == None:
         filename = str(graph_id) + ".gv"
-    g = Digraph(
+    g = Graph(
         name=str(graph_id),
         filename=filename,
         graph_attr=GRAPH_STYLE,

python/paddle/fluid/optimizer.py

@@ -50,21 +50,17 @@ class Optimizer(object):
 
     def __init__(self, learning_rate, regularization=None, name=None):
         if not isinstance(learning_rate, float) and \
-                not isinstance(learning_rate, framework.Variable) and \
-                not callable(learning_rate):
-            raise TypeError(
-                "learning rate should be float or Variable or callable(dtype)")
+                not isinstance(learning_rate, framework.Variable):
+            raise TypeError("learning rate should be float or Variable")
         self._name = name
         self.regularization = regularization
         self._learning_rate = learning_rate
         # the learning rate type should be inferenced from loss
         self._dtype = None
         # each program should have a independent learning rate
-        # program -> Variable(learning_rate) or:
-        # program -> callable(return learning_rate Variable)
+        # program -> Variable(learning_rate)
         self._learning_rate_map = dict()
-        if isinstance(self._learning_rate, framework.Variable) or \
-            callable(self._learning_rate):
+        if isinstance(self._learning_rate, framework.Variable):
             self._learning_rate_map[framework.default_main_program(
             )] = self._learning_rate
         # Dictionary of accumulators. Some optimizer subclasses need to
@@ -79,11 +75,6 @@ class Optimizer(object):
 
         if isinstance(lr, framework.Variable):
             return
-        elif callable(lr):
-            dtype = 'float32' if self._dtype is None else self._dtype
-            self._learning_rate_map[framework.default_main_program()] = lr(
-                dtype)
-            return
         else:
             if not isinstance(self._learning_rate, float):
                 raise TypeError(

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

@@ -370,8 +370,6 @@ class OpTest(unittest.TestCase):
                     return [place]
             else:
                 return []
-            else:
-                return []
         places = [fluid.CPUPlace()]
         cpu_only = self._cpu_only if hasattr(self, '_cpu_only') else False
         if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type)\

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

@@ -22,10 +22,8 @@ from op_test import OpTest
 class TestAccuracyOp(OpTest):
     def setUp(self):
         self.op_type = "accuracy"
-        self.dtype = np.float32
-        self.init_dtype()
         n = 8192
-        infer = np.random.random((n, 1)).astype(self.dtype)
+        infer = np.random.random((n, 1)).astype("float32")
         indices = np.random.randint(0, 2, (n, 1))
         label = np.random.randint(0, 2, (n, 1))
         self.inputs = {'Out': infer, 'Indices': indices, "Label": label}
@@ -36,25 +34,14 @@ class TestAccuracyOp(OpTest):
                     num_correct += 1
                     break
         self.outputs = {
-            'Accuracy': np.array([num_correct / float(n)]).astype(self.dtype),
+            'Accuracy': np.array([num_correct / float(n)]).astype("float32"),
             'Correct': np.array([num_correct]).astype("int32"),
             'Total': np.array([n]).astype("int32")
         }
 
-    def init_dtype(self):
-        pass
-
     def test_check_output(self):
         self.check_output()
 
 
-class TestAccuracyOpFp16(TestAccuracyOp):
-    def init_dtype(self):
-        self.dtype = np.float16
-
-    def test_check_output(self):
-        self.check_output(atol=1e-3)
-
-
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import unittest
 
-from test_conv2d_op import TestConv2dOp, TestWithPad, TestWithStride
+from test_conv2d_op import TestConv2dOp, TestWithPad, TestWithStride, TestWithGroup, TestWith1x1, TestWithInput1x1Filter1x1
 
 
 class TestMKLDNN(TestConv2dOp):
@@ -37,5 +37,23 @@ class TestMKLDNNWithStride(TestWithStride):
         self.data_format = "NCHW"
 
 
+class TestMKLDNNWithGroup(TestWithGroup):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+        self.data_format = "NCHW"
+
+
+class TestMKLDNNWith1x1(TestWith1x1):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+        self.data_format = "NCHW"
+
+
+class TestMKLDNNWithInput1x1Filter1x1(TestWithInput1x1Filter1x1):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+        self.data_format = "NCHW"
+
+
 if __name__ == '__main__':
     unittest.main()

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

@@ -21,16 +21,14 @@ from op_test import OpTest
 class ElementwiseDivOp(OpTest):
     def setUp(self):
         self.op_type = "elementwise_div"
-        self.dtype = np.float32
-        self.init_dtype()
         """ Warning
         CPU gradient check error!
         'X': np.random.random((32,84)).astype("float32"),
         'Y': np.random.random((32,84)).astype("float32")
         """
         self.inputs = {
-            'X': np.random.uniform(0.1, 1, [13, 17]).astype(self.dtype),
-            'Y': np.random.uniform(0.1, 1, [13, 17]).astype(self.dtype)
+            'X': np.random.uniform(0.1, 1, [13, 17]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [13, 17]).astype("float32")
         }
         self.outputs = {'Out': np.divide(self.inputs['X'], self.inputs['Y'])}
 
@@ -48,9 +46,6 @@ class ElementwiseDivOp(OpTest):
         self.check_grad(
             ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Y'))
 
-    def init_dtype(self):
-        pass
-
 
 class TestElementwiseDivOp_scalar(ElementwiseDivOp):
     def setUp(self):
@@ -131,21 +126,5 @@ class TestElementwiseDivOp_broadcast_3(ElementwiseDivOp):
         }
 
 
-class TestElementwiseDivOpFp16(ElementwiseDivOp):
-    def init_dtype(self):
-        self.dtype = np.float16
-
-    def test_check_grad_normal(self):
-        self.check_grad(['X', 'Y'], 'Out', max_relative_error=1)
-
-    def test_check_grad_ingore_x(self):
-        self.check_grad(
-            ['Y'], 'Out', max_relative_error=1, no_grad_set=set("X"))
-
-    def test_check_grad_ingore_y(self):
-        self.check_grad(
-            ['X'], 'Out', max_relative_error=1, no_grad_set=set('Y'))
-
-
 if __name__ == '__main__':
     unittest.main()

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

@@ -135,10 +135,5 @@ class TestElementwiseMulOp_broadcast_3(ElementwiseMulOp):
         }
 
 
-class TestElementwiseMulOpFp16(ElementwiseMulOp):
-    def init_dtype(self):
-        self.dtype = np.float16
-
-
 if __name__ == '__main__':
     unittest.main()

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

@@ -22,22 +22,12 @@ from op_test import OpTest
 class TestFillZerosLikeOp(OpTest):
     def setUp(self):
         self.op_type = "fill_zeros_like"
-        self.dtype = np.float32
-        self.init_dtype()
-        self.inputs = {'X': np.random.random((219, 232)).astype(self.dtype)}
+        self.inputs = {'X': np.random.random((219, 232)).astype("float32")}
         self.outputs = {'Out': np.zeros_like(self.inputs["X"])}
 
-    def init_dtype(self):
-        pass
-
     def test_check_output(self):
         self.check_output()
 
 
-class TestFillZerosLikeOpFp16(TestFillZerosLikeOp):
-    def init_dtype(self):
-        self.dtype = np.float16
-
-
 if __name__ == "__main__":
     unittest.main()

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

@@ -29,7 +29,7 @@ class TestGetTensorFromSelectedRows(unittest.TestCase):
 
     def check_with_place(self, place):
         scope = core.Scope()
-        x_rows = [0, 5, 5, 4, 20]
+        x_rows = [0, 5, 5, 4, 19]
         height = 20
         row_numel = 2
 

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

@@ -97,7 +97,7 @@ class TestLearningRateDecay(unittest.TestCase):
         startup_prog = fluid.Program()
 
         with fluid.program_guard(main_prog, startup_prog):
-            decayed_lr = fluid_decay_fn(**kwargs)("float32")
+            decayed_lr = fluid_decay_fn(**kwargs)
 
         place = fluid.CPUPlace()
         exe = fluid.Executor(place)

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

@@ -29,8 +29,8 @@ class TestMergeSelectedRows(unittest.TestCase):
 
     def check_with_place(self, place):
         scope = core.Scope()
-        x_rows = [0, 5, 5, 4, 20]
-        out_rows = [0, 4, 5, 20]
+        x_rows = [0, 5, 5, 4, 19]
+        out_rows = [0, 4, 5, 19]
         height = 20
         row_numel = 2
 

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

@@ -24,13 +24,11 @@ from op_test import OpTest
 class TestMomentumOp1(OpTest):
     def setUp(self):
         self.op_type = "momentum"
-        self.dtype = np.float32
-        self.init_dtype()
 
-        param = np.random.random((123, 321)).astype(self.dtype)
-        grad = np.random.random((123, 321)).astype(self.dtype)
-        velocity = np.zeros((123, 321)).astype(self.dtype)
-        learning_rate = np.array([0.001]).astype(self.dtype)
+        param = np.random.random((123, 321)).astype("float32")
+        grad = np.random.random((123, 321)).astype("float32")
+        velocity = np.zeros((123, 321)).astype("float32")
+        learning_rate = np.array([0.001]).astype("float32")
         mu = 0.0001
         use_nesterov = False
 
@@ -52,21 +50,10 @@ class TestMomentumOp1(OpTest):
 
         self.outputs = {'ParamOut': param_out, 'VelocityOut': velocity_out}
 
-    def init_dtype(self):
-        pass
-
     def test_check_output(self):
         self.check_output()
 
 
-class TestMomentumOpFp16(TestMomentumOp1):
-    def init_dtype(self):
-        self.dtype = np.float16
-
-    def test_check_output(self):
-        self.check_output(atol=1e-3)
-
-
 class TestMomentumOp2(OpTest):
     '''Test Momentum with default values for attributes
     '''

python/paddle/fluid/tests/unittests/test_py_func_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 os
+import paddle.fluid as fluid
+import paddle
+import unittest
+import six
+import numpy as np
+
+dev_cnt = 2
+if fluid.core.is_compiled_with_cuda():
+    dev_cnt = fluid.core.get_cuda_device_count()
+os.environ['CPU_NUM'] = str(dev_cnt)
+
+
+def dummy_func_with_no_input():
+    return float(1.0)
+
+
+def dummy_func_with_no_output(x):
+    pass
+
+
+def tanh(x):
+    return np.tanh(x)
+
+
+def tanh_grad(y, dy):
+    return np.array(dy) * (1 - np.square(np.array(y)))
+
+
+def cross_entropy(logits, labels):
+    logits = np.array(logits)
+    labels = np.array(labels)
+    M = logits.shape[0]
+    N = logits.shape[1]
+    ret = np.ndarray([M, 1]).astype(logits.dtype)
+    for idx in six.moves.range(M):
+        ret[idx][0] = -np.log(logits[idx][labels[idx][0]])
+    return ret
+
+
+def cross_entropy_grad(logits, labels, bwd_dout):
+    logits = np.array(logits)
+    labels = np.array(labels)
+    bwd_dout = np.array(bwd_dout)
+    M = logits.shape[0]
+    N = logits.shape[1]
+    dlogits = np.zeros([M, N]).astype(logits.dtype)
+    for idx in six.moves.range(M):
+        dlogits[idx][labels[idx][0]] = -bwd_dout[idx] / logits[idx][labels[idx][
+            0]]
+    return dlogits, None
+
+
+def simple_fc_net(img, label, use_py_func_op):
+    hidden = img
+    for idx in range(4):
+        hidden = fluid.layers.fc(
+            hidden,
+            size=200,
+            bias_attr=fluid.ParamAttr(
+                initializer=fluid.initializer.Constant(value=1.0)))
+        if not use_py_func_op:
+            hidden = fluid.layers.tanh(hidden)
+        else:
+            new_hidden = fluid.default_main_program().current_block(
+            ).create_var(
+                name='hidden_{}'.format(idx),
+                dtype='float32',
+                shape=hidden.shape)
+            hidden = fluid.layers.py_func(
+                func=tanh,
+                x=hidden,
+                out=new_hidden,
+                backward_func=tanh_grad,
+                skip_vars_in_backward_input=hidden)
+
+    prediction = fluid.layers.fc(hidden, size=10, act='softmax')
+    if not use_py_func_op:
+        loss = fluid.layers.cross_entropy(input=prediction, label=label)
+    else:
+        loss = fluid.default_main_program().current_block().create_var(
+            name='loss', dtype='float32', shape=[-1, 1])
+        loss = fluid.layers.py_func(
+            func=cross_entropy,
+            x=[prediction, label],
+            out=loss,
+            backward_func=cross_entropy_grad,
+            skip_vars_in_backward_input=loss)
+
+        dummy_var = fluid.default_main_program().current_block().create_var(
+            name='test_tmp_var', dtype='float32', shape=[1])
+        fluid.layers.py_func(
+            func=dummy_func_with_no_input, x=None, out=dummy_var)
+
+        fluid.layers.py_func(func=dummy_func_with_no_output, x=loss, out=None)
+
+    loss = fluid.layers.mean(loss)
+    return loss
+
+
+def reader():
+    for _ in six.moves.range(dev_cnt * 100):
+        yield np.random.random([784]), np.random.random_integers(
+            size=[1], low=0, high=9)
+
+
+def test_main(use_cuda, use_py_func_op, use_parallel_executor):
+    if use_cuda and not fluid.core.is_compiled_with_cuda():
+        return None
+
+    with fluid.program_guard(fluid.Program(), fluid.Program()):
+        with fluid.scope_guard(fluid.core.Scope()):
+            fluid.default_main_program().random_seed = 1
+            fluid.default_startup_program().random_seed = 1
+            np.random.seed(1)
+
+            img = fluid.layers.data(name='image', shape=[784], dtype='float32')
+            label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+            loss = simple_fc_net(img, label, use_py_func_op)
+            optimizer = fluid.optimizer.SGD(learning_rate=1e-3)
+            optimizer.minimize(loss)
+
+            place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
+            feeder = fluid.DataFeeder(feed_list=[img, label], place=place)
+            r = paddle.batch(reader, batch_size=10)
+
+            exe = fluid.Executor(place)
+            exe.run(fluid.default_startup_program())
+            if use_parallel_executor:
+                exe = fluid.ParallelExecutor(
+                    use_cuda=use_cuda, loss_name=loss.name)
+                fetch_list = [loss.name]
+            else:
+                fetch_list = [loss]
+
+            ret = []
+            for epoch_id in six.moves.range(2):
+                for d in r():
+                    L, = exe.run(feed=feeder.feed(d), fetch_list=fetch_list)
+                    ret.append(L)
+
+            return np.array(ret)
+
+
+class TestPyFuncOpUseExecutor(unittest.TestCase):
+    def setUp(self):
+        self.use_parallel_executor = False
+
+    def test_loss_diff(self):
+        losses = []
+        for use_cuda in [True, False]:
+            for use_py_func_op in [True, False]:
+                L = test_main(use_cuda, use_py_func_op,
+                              self.use_parallel_executor)
+                if L is not None:
+                    losses.append(L)
+
+        for idx in six.moves.range(len(losses) - 1):
+            max_diff = np.max(np.abs(losses[idx] - losses[0]))
+            self.assertAlmostEqual(max_diff, 0, delta=1e-3)
+
+
+class TestPyFuncOpUseParallelExecutor(unittest.TestCase):
+    def setUp(self):
+        self.use_parallel_executor = True
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -23,11 +23,8 @@ class TestTopkOp(OpTest):
     def setUp(self):
         self.set_args()
         self.op_type = "top_k"
-        self.dtype = np.float32
-        self.init_dtype()
-
         k = self.top_k
-        input = np.random.random((self.row, k)).astype(self.dtype)
+        input = np.random.random((self.row, k)).astype("float32")
         output = np.ndarray((self.row, k))
         indices = np.ndarray((self.row, k)).astype("int64")
 
@@ -41,9 +38,6 @@ class TestTopkOp(OpTest):
 
         self.outputs = {'Out': output, 'Indices': indices}
 
-    def init_dtype(self):
-        pass
-
     def set_args(self):
         self.row = 32
         self.top_k = 1
@@ -52,11 +46,6 @@ class TestTopkOp(OpTest):
         self.check_output()
 
 
-class TestTopkOpFp16(TestTopkOp):
-    def init_dtype(self):
-        self.dtype = np.float16
-
-
 class TestTopkOp3d(OpTest):
     def setUp(self):
         self.op_type = "top_k"

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

@@ -23,16 +23,6 @@ class TestTransposeMKLDNN(TestTransposeOp):
     def init_op_type(self):
         self.op_type = "transpose2"
         self.use_mkldnn = True
-        self.is_test = True
-        return
-
-    def test_check_grad(self):
-        return
-
-    def test_check_grad_no_input(self):
-        return
-
-    def test_check_grad_no_filter(self):
         return
 
 

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

@@ -27,7 +27,6 @@ class TestTransposeOp(OpTest):
         self.attrs = {
             'axis': list(self.axis),
             'use_mkldnn': self.use_mkldnn,
-            'is_test': self.is_test,
         }
         self.outputs = {
             'XShape': np.random.random(self.shape).astype("float32"),
@@ -37,7 +36,6 @@ class TestTransposeOp(OpTest):
     def init_op_type(self):
         self.op_type = "transpose2"
         self.use_mkldnn = False
-        self.is_test = False
 
     def test_check_output(self):
         self.check_output(no_check_set=['XShape'])

python/setup.py.in

@@ -140,8 +140,6 @@ if '${WITH_FLUID_ONLY}'== 'OFF':
                    '${PADDLE_BINARY_DIR}/paddle/scripts/paddle']
 
 package_data={'paddle.fluid': ['core' + (ext_name if os.name != 'nt' else '.pyd')]}
-if os.name == 'nt':
-    package_data['paddle.fluid'] += ['openblas' + ext_name]
 
 if '${WITH_FLUID_ONLY}'== 'OFF':
     package_data['paddle.v2.master']=['libpaddle_master' + ext_name]
@@ -166,11 +164,17 @@ package_data['paddle.libs']=[('libwarpctc' if os.name != 'nt' else 'warpctc') +
 shutil.copy('${WARPCTC_LIBRARIES}', libs_path)
 
 if '${WITH_MKL}' == 'ON':
-    shutil.copy('${MKLML_LIB}', libs_path)
-    shutil.copy('${MKLML_IOMP_LIB}', libs_path)
-    package_data['paddle.libs']+=['libmklml_intel' + ext_name,'libiomp5' + ext_name]
+    shutil.copy('${MKLML_SHARED_LIB}', libs_path)
+    shutil.copy('${MKLML_SHARED_IOMP_LIB}', libs_path)
+    package_data['paddle.libs']+=[('libmklml_intel' if os.name != 'nt' else 'mklml') + ext_name, ('libiomp5' if os.name != 'nt' else 'libiomp5md') + ext_name]
+else:
+    if os.name == 'nt':
+        # copy the openblas.dll
+        shutil.copy(os.path.dirname('${CBLAS_LIBRARIES}') + '/openblas' + ext_name, libs_path)
+        package_data['paddle.libs'] += ['openblas' + ext_name]
+
 if '${WITH_MKLDNN}' == 'ON':
-    if '${CMAKE_BUILD_TYPE}' == 'Release':
+    if '${CMAKE_BUILD_TYPE}' == 'Release' and os.name != 'nt':
         # only change rpath in Release mode.
         # TODO(typhoonzero): use install_name_tool to patch mkl libs once
         # we can support mkl on mac.
@@ -181,7 +185,7 @@ if '${WITH_MKLDNN}' == 'ON':
         command = "patchelf --set-rpath '$ORIGIN/' ${MKLDNN_SHARED_LIB}"
         if os.system(command) != 0:
             raise Exception("patch libmkldnn.so failed, command: %s" % command)
-    package_data['paddle.libs']+=['libmkldnn.so.0']
+    package_data['paddle.libs']+=['libmkldnn.so.0' if os.name != 'nt' else ('mkldnn' + ext_name)]
     shutil.copy('${MKLDNN_SHARED_LIB}', libs_path)
 if '${WITH_NGRAPH}' == 'ON':
     # only change rpath in Release mode,