test=develop

paddle/fluid/framework/ir/CMakeLists.txt

@@ -42,12 +42,10 @@ if(WITH_MKLDNN)
     pass_library(mkldnn_placement_pass base)
     pass_library(conv_bias_mkldnn_fuse_pass inference)
     pass_library(conv_relu_mkldnn_fuse_pass inference)
+    pass_library(conv_elementwise_add_mkldnn_fuse_pass inference)
 endif()
 
 cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector )
-if(WITH_MKLDNN)
-  pass_library(conv_elementwise_add_mkldnn_fuse_pass inference)
-endif()
 
 set(GLOB_PASS_LIB ${PASS_LIBRARY} CACHE INTERNAL "Global PASS library")
 

paddle/fluid/framework/ir/graph_helper_test.cc

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

paddle/fluid/framework/ir/graph_test.cc

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

paddle/fluid/framework/op_desc.cc

@@ -515,20 +515,14 @@ void OpDesc::InferShape(const BlockDesc &block) const {
 }
 
 void OpDesc::InferVarType(BlockDesc *block) const {
+  // There are a few places that var type can be set.
+  // When VarDesc is created, default set to LOD_TENSOR.
+  // When output variable is created, default is defaut set to LOD_TENSOR.
+  // We limit here to be the only place that operator defines its customized
+  // var type inference. Hence, we don't do any "default" setting here.
   auto &info = OpInfoMap::Instance().Get(this->Type());
   if (info.infer_var_type_) {
     info.infer_var_type_(*this, block);
-  } else {
-    // all output type is LoDTensor by default
-    VLOG(10) << this->Type()
-             << " has not registered InferVarType. Set output variables to "
-                "LOD_TENSOR";
-    for (auto &out_pair : this->outputs_) {
-      for (auto &out_var_name : out_pair.second) {
-        block->FindRecursiveOrCreateVar(out_var_name)
-            .SetType(proto::VarType::LOD_TENSOR);
-      }
-    }
   }
 }
 

paddle/fluid/framework/program_desc_test.cc

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

paddle/fluid/framework/reader_test.cc

@@ -40,7 +40,7 @@ TEST(READER, decorate_chain) {
     auto endpoints = root->GetEndPoints();
     ASSERT_EQ(endpoints.size(), 2U);
     ASSERT_NE(endpoints.count(end_point1.get()), 0UL);
-    ASSERT_NE(endpoints.count(end_point2.get()), 0);
+    ASSERT_NE(endpoints.count(end_point2.get()), 0UL);
   }
 
   {

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

@@ -71,7 +71,7 @@ void profile(bool use_mkldnn = false) {
 }
 
 TEST(Analyzer_resnet50, profile) { profile(); }
-#ifndef PADDLE_WITH_MKLDNN
+#ifdef PADDLE_WITH_MKLDNN
 TEST(Analyzer_resnet50, profile_mkldnn) { profile(true /* use_mkldnn */); }
 #endif
 

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

@@ -50,7 +50,7 @@ void CompareResult(const std::vector<PaddleTensor> &outputs,
     auto &ref_out = ref_outputs[i];
     size_t size = VecReduceToInt(out.shape);
     size_t ref_size = VecReduceToInt(ref_out.shape);
-    EXPECT_GT(size, 0);
+    EXPECT_GT(size, 0UL);
     EXPECT_EQ(size, ref_size);
     EXPECT_EQ(out.dtype, ref_out.dtype);
     switch (out.dtype) {

paddle/fluid/operators/CMakeLists.txt

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

paddle/fluid/operators/array_to_lod_tensor_op.cc

@@ -11,7 +11,7 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
-#include <paddle/fluid/operators/math/concat.h>
+#include <paddle/fluid/operators/math/concat_and_split.h>
 #include <numeric>
 
 #include "paddle/fluid/framework/lod_rank_table.h"

paddle/fluid/operators/concat_op.h

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

paddle/fluid/operators/detection/generate_proposal_labels_op.cc

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

paddle/fluid/operators/lod_tensor_to_array_op.cc

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

paddle/fluid/operators/math/CMakeLists.txt

@@ -1,5 +1,5 @@
 if (NOT WIN32)
-add_subdirectory(detail)
+    add_subdirectory(detail)
 endif(NOT WIN32)
 
 function(math_library TARGET)
@@ -35,7 +35,7 @@ function(math_library TARGET)
 endfunction()
 
 # please add new math_library in alphabetical order
-math_library(concat)
+math_library(concat_and_split)
 math_library(context_project DEPS im2col math_function)
 math_library(cross_entropy)
 math_library(cos_sim_functor)
@@ -43,8 +43,8 @@ math_library(depthwise_conv)
 math_library(im2col)
 
 if (NOT WIN32) # windows do not support avx functions yet.
-math_library(gru_compute DEPS activation_functions math_function)
-math_library(lstm_compute DEPS activation_functions)
+    math_library(gru_compute DEPS activation_functions math_function)
+    math_library(lstm_compute DEPS activation_functions)
 endif (NOT WIN32)
 
 cc_library(blas SRCS blas.cc DEPS cblas framework_proto device_context)
@@ -58,7 +58,7 @@ math_library(sequence_pooling DEPS math_function)
 math_library(sequence_scale)
 math_library(softmax DEPS math_function)
 if (NOT WIN32)
-math_library(matrix_bit_code)
+    math_library(matrix_bit_code)
 endif (NOT WIN32)
 math_library(unpooling)
 math_library(vol2col)
@@ -72,7 +72,7 @@ if(WITH_GPU)
     nv_test(math_function_gpu_test SRCS math_function_test.cu DEPS math_function)
     nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor math_function)
 endif()
-cc_test(concat_test SRCS concat_test.cc DEPS concat)
+cc_test(concat_test SRCS concat_test.cc DEPS concat_and_split)
 cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info)
 cc_library(jit_kernel 
     SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_lstm.cc

paddle/fluid/operators/math/concat_and_split.cc

@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/operators/math/concat.h"
+#include "paddle/fluid/operators/math/concat_and_split.h"
 #include <vector>
 
 namespace paddle {
@@ -67,7 +67,7 @@ class ConcatFunctor<platform::CPUDeviceContext, T> {
  * each dimension must be the same, except the axis dimension.
  */
 template <typename T>
-class ConcatGradFunctor<platform::CPUDeviceContext, T> {
+class SplitFunctor<platform::CPUDeviceContext, T> {
  public:
   void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input,
@@ -111,7 +111,7 @@ class ConcatGradFunctor<platform::CPUDeviceContext, T> {
 };
 #define DEFINE_FUNCTOR(type)                                      \
   template class ConcatFunctor<platform::CPUDeviceContext, type>; \
-  template class ConcatGradFunctor<platform::CPUDeviceContext, type>;
+  template class SplitFunctor<platform::CPUDeviceContext, type>;
 
 FOR_ALL_TYPES(DEFINE_FUNCTOR);
 

paddle/fluid/operators/math/concat_and_split.cu

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

paddle/fluid/operators/math/concat_and_split.h

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

paddle/fluid/operators/math/concat_test.cc

@@ -12,10 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/operators/math/concat.h"
 #include <gtest/gtest.h>
 #include <vector>
 #include "paddle/fluid/framework/tensor_util.h"
+#include "paddle/fluid/operators/math/concat_and_split.h"
 
 template <typename DeviceContext, typename Place>
 void testConcat() {

paddle/fluid/operators/reader/reader_blocking_queue_test.cc

@@ -237,7 +237,7 @@ TEST(BlockingQueue, speed_test_mode) {
   }
   for (size_t i = 0; i < queue_size; ++i) {
     q2.Receive(&b);
-    EXPECT_EQ(b, 0);
+    EXPECT_EQ(b, 0UL);
   }
   EXPECT_EQ(q2.Size(), queue_size);
 }

paddle/fluid/operators/sequence_concat_op.h

@@ -17,7 +17,7 @@
 #include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
-#include "paddle/fluid/operators/math/concat.h"
+#include "paddle/fluid/operators/math/concat_and_split.h"
 
 namespace paddle {
 namespace operators {
@@ -106,7 +106,7 @@ class SeqConcatGradKernel : public framework::OpKernel<T> {
       }
     }
 
-    math::ConcatGradFunctor<DeviceContext, T> functor;
+    math::SplitFunctor<DeviceContext, T> functor;
     std::vector<const framework::Tensor *> sliced_x_ptr;
     std::vector<framework::Tensor *> sliced_dx_ptr;
     for (auto &x : sliced_x) {

paddle/fluid/operators/split_op.cc

@@ -111,11 +111,10 @@ Example:
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-USE_CPU_ONLY_OP(concat);
 
 REGISTER_OPERATOR(split, ops::SplitOp, ops::SplitOpMaker, ops::SplitGradMaker);
-REGISTER_OP_CPU_KERNEL(split,
-                       ops::SplitOpKernel<paddle::platform::CPUPlace, double>,
-                       ops::SplitOpKernel<paddle::platform::CPUPlace, float>,
-                       ops::SplitOpKernel<paddle::platform::CPUPlace, int64_t>,
-                       ops::SplitOpKernel<paddle::platform::CPUPlace, int>);
+REGISTER_OP_CPU_KERNEL(
+    split, ops::SplitOpKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::SplitOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SplitOpKernel<paddle::platform::CPUDeviceContext, int64_t>,
+    ops::SplitOpKernel<paddle::platform::CPUDeviceContext, int>);

paddle/fluid/operators/split_op.h

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <chrono>  // NOLINT
 #include <vector>
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/concat_and_split.h"
 #include "paddle/fluid/operators/strided_memcpy.h"
 
 namespace paddle {
@@ -28,18 +29,22 @@ class SplitOpKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto* in = ctx.Input<framework::Tensor>("X");
     auto outs = ctx.MultiOutput<framework::Tensor>("Out");
-    auto in_stride = framework::stride_numel(in->dims());
-    int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
+    int axis = ctx.Attr<int>("axis");
     auto place = ctx.GetPlace();
 
-    size_t input_offset = 0;
-    for (auto& out : outs) {
-      out->mutable_data<T>(ctx.GetPlace());
-      auto out_stride = framework::stride_numel(out->dims());
-      StridedNumelCopyWithAxis<T>(ctx.device_context(), axis, out->data<T>(),
-                                  out_stride, in->data<T>() + input_offset,
-                                  in_stride, out_stride[axis]);
-      input_offset += out_stride[axis];
+    std::vector<const framework::Tensor*> shape_refer;
+    for (size_t j = 0; j < outs.size(); ++j) {
+      outs[j]->mutable_data<T>(ctx.GetPlace());
+      shape_refer.emplace_back(outs[j]);
+    }
+
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
+    // Sometimes direct copies will be faster, this maybe need deeply analysis.
+    if (axis == 0 && outs.size() < 10) {
+      StridedMemcpyWithAxis0<T>(dev_ctx, *in, shape_refer, &outs);
+    } else {
+      math::SplitFunctor<DeviceContext, T> functor;
+      functor(dev_ctx, *in, shape_refer, axis, &outs);
     }
   }
 };

paddle/fluid/operators/strided_memcpy.h

@@ -13,8 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <vector>
+#include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/operators/detail/strided_memcpy.h"
-
 namespace paddle {
 namespace operators {
 
@@ -98,5 +99,26 @@ inline void StridedNumelCopyWithAxis(const platform::DeviceContext& ctx,
   }
 }
 
+template <typename T>
+inline void StridedMemcpyWithAxis0(
+    const platform::DeviceContext& dev_ctx, const framework::Tensor& input,
+    const std::vector<const framework::Tensor*>& shape_refer,
+    std::vector<framework::Tensor*>* outputs) {
+  const framework::DDim in_stride = stride_numel(input.dims());
+  const int axis = 0;
+  size_t input_offset = 0;
+
+  for (size_t i = 0; i < outputs->size(); ++i) {
+    auto out_stride = stride_numel(shape_refer[i]->dims());
+    auto out = outputs->at(i);
+    if (out != nullptr) {
+      StridedNumelCopyWithAxis<T>(dev_ctx, axis, out->data<T>(), out_stride,
+                                  input.data<T>() + input_offset, in_stride,
+                                  out_stride[axis]);
+    }
+    input_offset += out_stride[axis];
+  }
+}
+
 }  // namespace operators
 }  // namespace paddle

python/paddle/fluid/layer_helper.py

@@ -324,10 +324,19 @@ class LayerHelper(object):
             raise ValueError("no Parameter name %s found" % name)
         return param
 
-    def create_tmp_variable(self, dtype, stop_gradient=False):
+    def create_variable_for_type_inference(self, dtype, stop_gradient=False):
+        """Create a temporary variable that should be type inferred layer.
+
+        Note:
+            The default type will be set to LOD_TENSOR. However, when
+            the var is used as operator output, its type will be updated
+            based on operator's `VarTypeInference` implementation in
+            infer_var_type.
+        """
         return self.main_program.current_block().create_var(
             name=unique_name.generate(".".join([self.name, 'tmp'])),
             dtype=dtype,
+            type=core.VarDesc.VarType.LOD_TENSOR,
             persistable=False,
             stop_gradient=stop_gradient)
 
@@ -388,7 +397,7 @@ class LayerHelper(object):
 
         b = self.create_parameter(
             attr=bias_attr, shape=size, dtype=input_var.dtype, is_bias=True)
-        tmp = self.create_tmp_variable(dtype=input_var.dtype)
+        tmp = self.create_variable_for_type_inference(dtype=input_var.dtype)
         self.append_op(
             type='elementwise_add',
             inputs={'X': [input_var],
@@ -414,7 +423,7 @@ class LayerHelper(object):
         tmp = input_var
         # NOTE(dzhwinter): some activation support inplace compution.
         if not core.IsInplace(act_type):
-            tmp = self.create_tmp_variable(dtype=input_var.dtype)
+            tmp = self.create_variable_for_type_inference(dtype=input_var.dtype)
         self.append_op(
             type=act_type,
             inputs={"X": [input_var]},

python/paddle/fluid/layers/control_flow.py

@@ -80,8 +80,8 @@ def split_lod_tensor(input, mask, level=0):
 
     """
     helper = LayerHelper('split_lod_tensor', **locals())
-    out_true = helper.create_tmp_variable(dtype=input.dtype)
-    out_false = helper.create_tmp_variable(dtype=input.dtype)
+    out_true = helper.create_variable_for_type_inference(dtype=input.dtype)
+    out_false = helper.create_variable_for_type_inference(dtype=input.dtype)
     helper.append_op(
         type='split_lod_tensor',
         inputs={
@@ -131,7 +131,7 @@ def merge_lod_tensor(in_true, in_false, x, mask, level=0):
                 in_true=out_true, in_false=out_false, mask=y, x=x, level=level)
     """
     helper = LayerHelper('merge_lod_tensor', **locals())
-    out = helper.create_tmp_variable(dtype=in_true.dtype)
+    out = helper.create_variable_for_type_inference(dtype=in_true.dtype)
     helper.append_op(
         type='merge_lod_tensor',
         inputs={'X': x,
@@ -524,7 +524,7 @@ class StaticRNN(object):
         if not isinstance(o, Variable):
             raise TypeError("step output takes a Variable")
 
-        tmp_o = self.helper.create_tmp_variable(dtype=o.dtype)
+        tmp_o = self.helper.create_variable_for_type_inference(dtype=o.dtype)
         self.helper.append_op(
             type='rnn_memory_helper',
             inputs={'X': [o]},
@@ -606,7 +606,8 @@ class StaticRNN(object):
             pre_memories.append(mem.pre_mem.name)
             mem_var = rnn_block.var(mem.mem.name)
             assert isinstance(mem_var, Variable)
-            new_mem = self.helper.create_tmp_variable(dtype=mem_var.dtype)
+            new_mem = self.helper.create_variable_for_type_inference(
+                dtype=mem_var.dtype)
 
             rnn_block.append_op(
                 type='rnn_memory_helper',
@@ -813,7 +814,7 @@ def max_sequence_len(rank_table):
         ${out_comment}.
     """
     helper = LayerHelper("max_seqence_len", **locals())
-    res = helper.create_tmp_variable(dtype="int64")
+    res = helper.create_variable_for_type_inference(dtype="int64")
     helper.append_op(
         type="max_sequence_len",
         inputs={"RankTable": rank_table},
@@ -884,7 +885,7 @@ def array_to_lod_tensor(x, table):
           lod_tensor = fluid.layers.array_to_lod_tensor(array, table)
     """
     helper = LayerHelper("array_to_lod_tensor", **locals())
-    tmp = helper.create_tmp_variable(dtype=x.dtype)
+    tmp = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type="array_to_lod_tensor",
         inputs={'X': x,
@@ -915,7 +916,7 @@ def increment(x, value=1.0, in_place=True):
     """
     helper = LayerHelper("increment", **locals())
     if not in_place:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = x
     helper.append_op(
@@ -1012,7 +1013,7 @@ def less_than(x, y, force_cpu=None, cond=None, **ignored):
     """
     helper = LayerHelper("less_than", **locals())
     if cond is None:
-        cond = helper.create_tmp_variable(dtype='bool')
+        cond = helper.create_variable_for_type_inference(dtype='bool')
         cond.stop_gradient = True
 
     attrs = dict()
@@ -1051,7 +1052,7 @@ def equal(x, y, cond=None, **ignored):
     """
     helper = LayerHelper("equal", **locals())
     if cond is None:
-        cond = helper.create_tmp_variable(dtype='bool')
+        cond = helper.create_variable_for_type_inference(dtype='bool')
         cond.stop_gradient = True
 
     helper.append_op(
@@ -1098,7 +1099,7 @@ def array_read(array, i):
             array,
             Variable) or array.type != core.VarDesc.VarType.LOD_TENSOR_ARRAY:
         raise TypeError("array should be tensor array vairable")
-    out = helper.create_tmp_variable(dtype=array.dtype)
+    out = helper.create_variable_for_type_inference(dtype=array.dtype)
     helper.append_op(
         type='read_from_array',
         inputs={'X': [array],
@@ -1133,7 +1134,7 @@ def shrink_memory(x, i, table):
         usage.
     """
     helper = LayerHelper('shrink_memory', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='shrink_rnn_memory',
         inputs={'X': [x],
@@ -1170,7 +1171,7 @@ def array_length(array):
 
     """
     helper = LayerHelper('array_length', **locals())
-    tmp = helper.create_tmp_variable(dtype='int64')
+    tmp = helper.create_variable_for_type_inference(dtype='int64')
     tmp.stop_gradient = True
     helper.append_op(
         type='lod_array_length', inputs={'X': [array]}, outputs={'Out': [tmp]})
@@ -1590,7 +1591,7 @@ class DynamicRNN(object):
         self.mem_dict = dict()
         self.output_array = []
         self.outputs = []
-        self.cond = self.helper.create_tmp_variable(dtype='bool')
+        self.cond = self.helper.create_variable_for_type_inference(dtype='bool')
         self.cond.stop_gradient = False
         self.while_op = While(self.cond)
         self.input_array = []
@@ -1924,7 +1925,7 @@ def reorder_lod_tensor_by_rank(x, rank_table):
     helper.is_instance('x', Variable)
     helper.is_instance('rank_table', Variable)
 
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='reorder_lod_tensor_by_rank',
         inputs={'X': [x],
@@ -1958,7 +1959,7 @@ def is_empty(x, cond=None, **ignored):
     """
     helper = LayerHelper("is_empty", **locals())
     if cond is None:
-        cond = helper.create_tmp_variable(dtype='bool')
+        cond = helper.create_variable_for_type_inference(dtype='bool')
         cond.stop_gradient = True
     elif not isinstance(cond, Variable):
         raise TypeError("cond takes a variable")