Merge branch 'develop' into tr_convert_init

.travis.yml

@@ -12,7 +12,7 @@ services:
 os:
   - linux
 env:
-  - JOB=build_doc
+  - JOB=doc
   - JOB=check_style
   - JOB=build_android
 addons:
@@ -36,21 +36,18 @@ addons:
       - ccache
   ssh_known_hosts: 13.229.163.131
 before_install:
-  - if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi
-  # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python
-  # protobuf version.
   - sudo pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
-  - sudo pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
+  - sudo pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit
   - |
     function timeout() { perl -e 'alarm shift; exec @ARGV' "$@"; }
 script:
   - |
     # 43min timeout
-    if [[ "$JOB" == "build_android" ]]; then timeout 2580 docker run -it --rm -v "$TRAVIS_BUILD_DIR:/paddle" paddlepaddle/paddle:latest-dev-android;
-    else timeout 2580 paddle/scripts/travis/${JOB}.sh; fi;
-    RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true; else exit 1; fi;
+    if [[ "$JOB" != "doc" ]]; then timeout 2580 paddle/scripts/paddle_docker_build.sh ${JOB}; else paddle/scripts/paddle_build.sh ${JOB}; fi;
+    if [ $? -eq 0 ] || [ $? -eq 142 ]; then true; else exit 1; fi;
   - |
-    if [[ "$JOB" != "build_doc" ]]; then exit 0; fi;
+    if [[ "$JOB" != "doc" ]]; then exit 0; fi;
+    # For document only
     if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
     if [[ "$TRAVIS_BRANCH" != "develop"  && ! "$TRAVIS_BRANCH" =~ ^v[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then exit 0; fi;
     export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/master/scripts/deploy/deploy_docs.sh

AUTHORS.md

@@ -2,12 +2,14 @@
 |---|---|
 | abhinavarora | Abhinav Arora |
 | backyes | Yan-Fei Wang |
+| baiyfbupt | Yi-Fan Bai |
 | beckett1124 | Bin Qi |
-| JiayiFeng | Jia-Yi Feng |
 | chengxiaohua1105 | Xiao-Hua Cheng |
 | cxwangyi, yiwangbaidu, wangkuiyi | Yi Wang |
 | cxysteven | Xing-Yi Cheng |
 | dzhwinter | Zhi-Hong Dong |
+| dragonwarrior | Long Wang |
+| dyning | Yuning Du |
 | emailweixu | Wei Xu |
 | gangliao | Gang Liao |
 | gongweibao | Wei-Bao Gong |
@@ -16,6 +18,9 @@
 | hedaoyuan | Dao-Yuan He |
 | helinwang | He-Lin Wang |
 | jacquesqiao | Long-Fei Qiao |
+| jczaja | Jacek Czaja |
+| JiayiFeng | Jia-Yi Feng |
+| kbinias | Krzysztof Binias |
 | kuke | Yi-Bing Liu |
 | lcy-seso | Ying Cao |
 | lipeng-unisound | Peng Li |
@@ -24,15 +29,20 @@
 | llxxxll | Yong-Feng Liu |
 | luotao01 | Tao Luo |
 | lzhao4ever | Liang Zhao |
+| mozga-intel | Mateusz Ozga |
 | NHZlX | Zhao-Long Xing |
+| Noplz | Yuan Gao |
 | pakchoi | Chuan-Jiang Song |
+| panyx0718 | Xin Pan |
 | pengli09 | Peng Li |
 | pkuyym | Ya-Ming Yang |
+| pzelazko-intel | Pawel Zelazko |
 | QiJune | Jun Qi |
 | qingqing01 | Qing-Qing Dang |
 | reyoung | Yang Yu |
 | Superjom | Chun-Wei Yan |
 | tianbingsz | Tian-Bing Xu |
+| tpatejko | Tomasz Patejko |
 | typhoonzero | Yi Wu |
 | wanghaoshuang | Hao-Shuang Wang |
 | wangyang59 | Yang Wang |

Dockerfile

 # A image for building paddle binaries
 # Use cuda devel base image for both cpu and gpu environment
-
-# When you modify it, please be aware of cudnn-runtime version 
+# When you modify it, please be aware of cudnn-runtime version
 # and libcudnn.so.x in paddle/scripts/docker/build.sh
 FROM nvidia/cuda:8.0-cudnn7-devel-ubuntu16.04
 MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
@@ -24,7 +23,7 @@ ENV HOME /root
 COPY ./paddle/scripts/docker/root/ /root/
 
 RUN apt-get update && \
-    apt-get install -y \
+    apt-get install -y --allow-downgrades \
     git python-pip python-dev openssh-server bison \
     libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 \
     wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
@@ -33,7 +32,7 @@ RUN apt-get update && \
     automake locales clang-format swig doxygen cmake  \
     liblapack-dev liblapacke-dev \
     clang-3.8 llvm-3.8 libclang-3.8-dev \
-    net-tools libtool && \
+    net-tools libtool ccache && \
     apt-get clean -y
 
 # Install Go and glide

benchmark/fluid/machine_translation.py

@@ -21,7 +21,7 @@ import argparse
 import time
 import distutils.util
 
-import paddle.v2 as paddle
+import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 import paddle.fluid.framework as framework

benchmark/fluid/mnist.py

@@ -20,7 +20,7 @@ import numpy as np
 import argparse
 import time
 
-import paddle.v2 as paddle
+import paddle
 import paddle.fluid as fluid
 import paddle.fluid.profiler as profiler
 

benchmark/fluid/resnet.py

@@ -23,7 +23,7 @@ import time
 
 import cProfile, pstats, StringIO
 
-import paddle.v2 as paddle
+import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 import paddle.fluid.profiler as profiler

benchmark/fluid/stacked_dynamic_lstm.py

@@ -23,10 +23,10 @@ import random
 import time
 
 import numpy
-import paddle.v2 as paddle
-import paddle.v2.dataset.imdb as imdb
+import paddle
+import paddle.dataset.imdb as imdb
 import paddle.fluid as fluid
-from paddle.v2 import batch
+import paddle.batch as batch
 import paddle.fluid.profiler as profiler
 
 

benchmark/fluid/vgg.py

@@ -17,7 +17,7 @@ from __future__ import print_function
 import sys
 import time
 import numpy as np
-import paddle.v2 as paddle
+import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 import argparse

cmake/cuda.cmake

@@ -172,6 +172,8 @@ set(CUDA_PROPAGATE_HOST_FLAGS OFF)
 list(APPEND CUDA_NVCC_FLAGS "-std=c++11")
 list(APPEND CUDA_NVCC_FLAGS "--use_fast_math")
 list(APPEND CUDA_NVCC_FLAGS "-Xcompiler -fPIC")
+# in cuda9, suppress cuda warning on eigen 
+list(APPEND CUDA_NVCC_FLAGS "-w")
 # Set :expt-relaxed-constexpr to suppress Eigen warnings
 list(APPEND CUDA_NVCC_FLAGS "--expt-relaxed-constexpr")
 

cmake/external/eigen.cmake

@@ -22,7 +22,9 @@ else()
         extern_eigen3
         ${EXTERNAL_PROJECT_LOG_ARGS}
         GIT_REPOSITORY  "https://github.com/RLovelett/eigen.git"
-        GIT_TAG         70661066beef694cadf6c304d0d07e0758825c10
+        # eigen on cuda9.1 missing header of math_funtions.hpp
+        # https://stackoverflow.com/questions/43113508/math-functions-hpp-not-found-when-using-cuda-with-eigen
+        GIT_TAG         917060c364181f33a735dc023818d5a54f60e54c
         PREFIX          ${EIGEN_SOURCE_DIR}
         UPDATE_COMMAND  ""
         CONFIGURE_COMMAND ""

cmake/external/warpctc.cmake

@@ -38,8 +38,7 @@ ENDIF()
 ExternalProject_Add(
     extern_warpctc
     ${EXTERNAL_PROJECT_LOG_ARGS}
-    GIT_REPOSITORY  "https://github.com/gangliao/warp-ctc.git"
-    GIT_TAG         b63a0644654a3e0ed624c85a1767bc8193aead09
+    GIT_REPOSITORY  "https://github.com/dzhwinter/warp-ctc.git"
     PREFIX          ${WARPCTC_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}

doc/fluid/api/data/data_reader.rst

@@ -56,11 +56,11 @@ DataFeeder
 Reader
 ======
 
-..  automodule:: paddle.v2.reader
+..  automodule:: paddle.reader
     :members:
     :noindex:
 
-..  automodule:: paddle.v2.reader.creator
+..  automodule:: paddle.reader.creator
     :members:
     :noindex:
 

doc/fluid/api/layers.rst

@@ -479,6 +479,13 @@ label_smooth
 ..  autofunction:: paddle.fluid.layers.label_smooth
     :noindex:
 
+roi_pool
+---------
+
+..  autofunction:: paddle.fluid.layers.roi_pool
+    :noindex:
+
+    
 ops
 ===
 
@@ -820,3 +827,5 @@ topk
 
 ..  autofunction:: paddle.fluid.layers.topk
     :noindex:
+
+

doc/fluid/design/algorithm/parameter_average.md

 # Averaging Parameter in PaddlePaddle
 
 ## Why Averaging
-In a large scale machine learning setup where the size of the training data is huge, it could take us a large number of iterations over the training data before we can achieve the optimal values of parameters of our model. Looking at the problem setup, it is desirable if we can obtain the optimal values of parameters by going through the data in as few passes as we can.
+In a large scale machine learning setup where the size of the training data is huge, it could take us a large number of iterations over the training data before we can achieve the optimal values of parameters of our model. Looking at the problem setup, it is desirable to obtain the optimal values of parameters by going through the data in as few passes as possible.
 
 Polyak and Juditsky (1992) showed that the test performance of simple average of parameters obtained by Stochastic Gradient Descent (SGD) is as good as that of parameter values that are obtained by training the model over and over again, over the training dataset.
 
@@ -16,16 +16,16 @@ We propose averaging for any optimizer similar to how ASGD performs it, as menti
 ### How to perform Parameter Averaging in PaddlePaddle
 
 Parameter Averaging in PaddlePaddle works in the following way during training :
-1. It will take in an instance of a normal optimizer as an input, e.g. RMSPropOptimizer
+1. It will take in an instance of an optimizer as an input, e.g. RMSPropOptimizer
 2. The optimizer itself is responsible for updating the parameters.
 3. The ParameterAverageOptimizer maintains a separate copy of the parameters for itself:
-    1. In concept, the values of this copy are the average of the values of the parameters in the most recent N batches.
-    2. However, saving all the N instances of the parameters in memory is not feasible.
+    1. In theory, the values of this copy are the average of the values of the parameters in the most recent N batches.
+    2. However, saving all N instances of the parameters in memory is not feasible.
     3. Therefore, an approximation algorithm is used.
 
 Hence, overall we have have two copies of the parameters: one for the optimizer itself, and one for the ParameterAverageOptimizer. The former should be used in back propagation, while the latter should be used during testing and should be saved.
 
-During the testing/ saving the model phase, we perform the following steps:
+During the testing/saving the model phase, we perform the following steps:
 1. Perform the delayed operations.
 2. Save current values of the parameters to a temporary variable.
 3. Replace the values of the parameters with the averaged values.

doc/fluid/design/data_type/float16.md

@@ -3,7 +3,7 @@
 ## Why float16
 Half precision (float16) is a binary floating-point format that occupies 16 bits in memory. float16 is half the size of traditional 32-bit single precision format (float) and has lower precision and smaller range. 
 
-When high precision computation is not required, using float16 data type could potentially 
+When high precision computation is not required (which is usually the case at least in the deep learning inference stage), using float16 data type could potentially 
 
 - reduce storage space, memory bandwidth, and power usages; 
 - increase the chance of data fitting into a smaller cache of lower latency; 
@@ -12,7 +12,7 @@ When high precision computation is not required, using float16 data type could p
 ## Survey of current float16 support
 A brief survey of float16 support on different compilers, hardwares, and libraries can be found below. Interested readers can refer to [link1](https://github.com/PaddlePaddle/Paddle/issues/4853) and [link2](https://github.com/Xreki/Xreki.github.io/blob/master/multi_data_types_in_dl_framework/ppt/float16_and_quantized_type.md) for more info.
 
-The goal of float16 is to serve as a key for the executor to find and run the correct version of compute method specialized for float16 in operator kernel. It should be compatible with various natively supported float16 implementations including `__half` for cuda, `float16_t` for ARM, and `Eigen::half` for Eigen to make writing customized float16 kernels easier. 
+The goal of float16 is to serve as a key for the executor to find and run the correct version of compute method specialized for float16 in operator kernels. It should be compatible with various natively supported float16 implementations including `__half` for cuda, `float16_t` for ARM, and `Eigen::half` for Eigen to make writing customized float16 kernels easier. 
 
 ### Compiler
 - nvcc supports `__half` data type after CUDA 7.5.
@@ -95,11 +95,89 @@ float half_to_float(float16 h);
 ```
 which provides one-to-one conversion between float32 and float16. These twos functions will do different conversion routines based on the current hardware. CUDA/ARM instrinsics will be used when the corresonding hardware is available. If the hardware or compiler level does not support float32 to float16 conversion, software emulation will be performed to do the conversion.
 
-## To do
-After float16 class is available, some of the future items are below:
+## float16 inference
+In Fluid, a neural network is represented as a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md), whose Python wrapper is a [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#program). The basic structure of a program is some nested [blocks](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program desc by executing the sequence of operators in the entrance block of the program one by one.  
 
-- Update pybind/tensor_py.h to bind c++ float16 with numpy float16. 
+### Operator level requirement
+Each operator has many kernels for different data types, devices, and library types. The operator will select the appropriate kernel to run based on, among other things, the data type of the input variables. By default, every Fluid operator has a float data type kernel that takes float variables as input and generates float output. 
 
-- Modify `GetKernelType()` method in `framework/operator.h` to make it compatible with float16.
+This means that if we provide float input to the first operator in a program, then each opeartor will use float kernel to compute float output and send it as input to the next operator to trigger the float kernel. Overall, the program will run in float mode and give us a final output of float data type.
 
-- Create a type-casting operator that can convert the data type in tensor between float16 and other types.
+The same principle applies if we want a program to run in float16 mode. We provide input variable of float16 data type to the first operator, and then one by one, each operator in the program will run the float16 kernel (provided that each operator in this program has float16 kernels registered) until we finally obtain a float16 output variable.
+
+So the preliminary requirement for float16 inference is to add float16 kernel to operators that are needed in a specific kind of program. For example, float16 inference on an image classification neural network like Vgg or Resnet, typically requires the following operators to have float16 kernels: convolution, pooling, multiplication, addition, batch norm, dropout, relu, and softmax. Please refer to [new_op_en](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/dev/new_op_en.md) for details of how to add new kernels to an operator.
+
+### Variable level requirement
+Operators including convolution and multiplication (used in fully-connected layers) takes as input not only the variables generated by the preceding operators but also [parameter](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#parameter) variables, which contains the trained weights to apply to the input data. These weights are obtained in the Fluid training process and are by default of float data type.
+
+When these operators are running in float16 mode, the float16 kernel requires those parameter variables to contain weights of Fluid float16 data type. Thus, we need a convenient way to convert the original float weights to float16 weights. 
+
+In Fluid, we use tensor to hold actual data for a variable on the c++ end. [Pybind](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/pybind/tensor_py.h) is used to bind c++ tensors of certain data type with numpy array of the correponding numpy data type on the Python end. Each common c++ built-in data type has a corresponding numpy data type of the same name. However, since there is no built-in float16 type in c++, we cannot directly bind numpy float16 data type with the Fluid float16 class. Since both Fluid float16 and numpy float16 use uint16 as the internal data storage type, we use c++ built-in type `uint16_t` and the corresponding numpy uint16 data type to bridge the gap via [Pybind](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/pybind/tensor_py.h). 
+
+The following code demonstrates how to do the tensor conversion.
+```Python
+# var is the variable of float weights
+# tensor is a numpy array of data copied from the tensor data in var 
+# fp16_var is the variable that will contain float16 weights converted from var  
+tensor = numpy.array(var.get_tensor())
+fp16_tensor = fp16_var.get_tensor()
+
+# After the original tensor data is converted to numpy float16 data type, 
+# view(numpy.uint16) is used so that the internal memory of the numpy array 
+# will be reinterpreted to be of uint16 data type, which is binded to 
+# Fluid float16 class via pybind with the help of uint16_t built-in c++ type
+fp16_tensor.set(tensor.astype(numpy.float16).view(numpy.uint16), GPUPlace)  
+```
+
+### Consistent API requirement
+The basic inference in float16 mode requires users to feed input and obtain output both of float16 data type. However, in this way, the inference APIs are not consistent between float16 mode and float mode, and users may find it confusing and diffcult to use float16 inference since they need to do extra steps to provide float16 input data and convert float16 output data back to float. To have consistent API for different inference modes, we need to transpile the program desc in some way so that we can run float16 inference by feeding and fetching variables of float data type.
+
+This problem can be solved by introducing a type-casting operator which takes an input variable of certain data type, cast it to another specified data type, and put the casted data into the output variable. Insert cast operator where needed can make a program internally run in float16 mode.   
+
+### float16 transpiler
+Put all the above requirements in mind, we designed a float16 inference transpiler that can tranpile a float32 mode inference program desc to a float16 mode one.
+
+Given a float inference program and the corresponding variables of float32 weights in the [scope](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/scope.md),
+this transpiler mainly does the following modifications:
+
+1. Insert cast operators at the beginning of the program so that the input float data will be converted to float16 data type before feeding to subsequent operators to invoke the float16 kernel. 
+
+2. Insert cast operators at the end of the program so that the output float16 data will be converted back to float data type before users obtain the result.
+
+3. For each parameter variable of float weights, create in the scope a corresponding variable of float16 weights which are converted from the corresponding float weights and add this new float16 variable to the program.
+
+4. Update the operator information in the program so that each relevant operator use the newly created float16 variable instead of its float counterpart.
+
+Below is an example of usage:
+```Python
+# Get the float inference program
+[float_inference_program, feed_target_names,
+ fetch_targets] = fluid.io.load_inference_model(save_dirname, exe)
+
+# Prepare the float input data
+tensor_img = numpy.random.rand(1, 3, 32, 32).astype(numpy.float32)
+
+# Running inference_program in float mode
+float_results = exe.run(float_inference_program,
+                        feed={feed_target_names[0]: tensor_img},
+                        fetch_list=fetch_targets)
+
+# Use float16 transpiler to speedup
+float16_inference_program = float_inference_program.clone()
+t = fluid.InferenceTranspiler()
+t.float16_transpile(float16_inference_program, GPUPlace)
+
+# Running 
+float16_results = exe.run(float16_inference_program,
+                          feed={feed_target_names[0]: tensor_img},
+                          fetch_list=fetch_targets)
+```
+
+As we can see from the example above, users can simply use the `float16_transpile` method provided by the infernece transpiler class on an existing float inference program to run inference in float16 mode.
+
+### Speedup on GPU
+Currently, Fluid inference in float16 mode is only supported on Nvidia GPU device. There is no motivation to support float16 inference on non-ARM CPUs because float16 is not natively supported there and float16 calculation will only be slower than its float counterpart. 
+
+Nvidia started to support its native float16 data type (which has the same internal memory representation as Fluid float16 class) on CUDA 7.5. Moreover, float16 speedups on common computational intensive tasks including GEMM (general matrix-matrix multiplication) and convolution are supported since cublas 7.5 and cuDNN 5.0.
+
+Recently, the introduction of [tensor core](https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/) in volta architecture GPUs and the support of tensor core calculation in CUDA 9.0 and cuDNN 7.0 make float16 truly superior to float in certain deep learning applications. Please refer to this [benchmark report](https://github.com/kexinzhao/Paddle_benchmark/blob/master/float16_benchmark.md) for more details.

doc/v2/api/data/data_reader.rst

@@ -56,11 +56,11 @@ DataFeeder
 Reader
 ======
 
-..  automodule:: paddle.v2.reader
+..  automodule:: paddle.reader
     :members:
     :noindex:
 
-..  automodule:: paddle.v2.reader.creator
+..  automodule:: paddle.reader.creator
     :members:
     :noindex:
 

doc/v2/api/data/dataset.rst

 Dataset
 =======
 
-..  automodule:: paddle.v2.dataset
+..  automodule:: paddle.dataset
     :members:
     :noindex:
 
 mnist
 +++++
 
-..  automodule:: paddle.v2.dataset.mnist
+..  automodule:: paddle.dataset.mnist
     :members:
     :noindex:
 
 cifar
 +++++
 
-..  automodule:: paddle.v2.dataset.cifar
+..  automodule:: paddle.dataset.cifar
     :members:
     :noindex:
 
 conll05
 +++++++
 
-..  automodule:: paddle.v2.dataset.conll05
+..  automodule:: paddle.dataset.conll05
     :members: get_dict,get_embedding,test
     :noindex:
 
 imdb
 ++++
 
-..  automodule:: paddle.v2.dataset.imdb
+..  automodule:: paddle.dataset.imdb
     :members:
     :noindex:
 
 imikolov
 ++++++++
 
-..  automodule:: paddle.v2.dataset.imikolov
+..  automodule:: paddle.dataset.imikolov
     :members:
     :noindex:
 
 movielens
 +++++++++
 
-..  automodule:: paddle.v2.dataset.movielens
+..  automodule:: paddle.dataset.movielens
     :members:
     :noindex:
 
-..  autoclass:: paddle.v2.dataset.movielens.MovieInfo
+..  autoclass:: paddle.dataset.movielens.MovieInfo
     :noindex:
-    
-..  autoclass:: paddle.v2.dataset.movielens.UserInfo
+
+..  autoclass:: paddle.dataset.movielens.UserInfo
     :noindex:
 
 sentiment
 +++++++++
 
-..  automodule:: paddle.v2.dataset.sentiment
+..  automodule:: paddle.dataset.sentiment
     :members:
     :noindex:
 
 uci_housing
 +++++++++++
 
-..  automodule:: paddle.v2.dataset.uci_housing
+..  automodule:: paddle.dataset.uci_housing
     :members:
     :noindex:
 
 wmt14
 +++++
 
-..  automodule:: paddle.v2.dataset.wmt14
+..  automodule:: paddle.dataset.wmt14
     :members:
     :noindex:
 
 wmt16
 +++++
 
-..  automodule:: paddle.v2.dataset.wmt16
+..  automodule:: paddle.dataset.wmt16
     :members:
     :noindex:

paddle/cuda/include/hl_base.h

@@ -228,6 +228,21 @@ extern __thread cudaStream_t default_stream;
         << "CUDA error: " << hl_get_device_error_string((size_t)err); \
   }
 
+// __shfl has been deprecated as of CUDA 9.0.
+#if CUDA_VERSION < 9000
+template <typename T>
+__forceinline__ __device__ T
+__shfl_sync(unsigned, T val, int src_line, int width) {
+  return __shfl(val, src_line, width);
+}
+
+#define CREATE_SHFL_MASK(mask, predicate) mask = 0u;
+#else
+#define FULL_WARP_MASK 0xFFFFFFFF
+#define CREATE_SHFL_MASK(mask, predicate) \
+  mask = __ballot_sync(FULL_WARP_MASK, (predicate))
+#endif
+
 #endif /* __NVCC__ */
 
 #endif /* HL_BASE_H_ */

paddle/cuda/src/hl_cuda_lstm.cu

@@ -341,12 +341,15 @@ void hl_lstm_parallel_forward(real *gateValue,
 }
 
 __device__ __forceinline__ void transpose_32x32(real a[], const int idx) {
-  int addr = idx % 32;
+  const int warp_size = 32;
+  int addr = idx % warp_size;
+  unsigned mask = 0u;
+  CREATE_SHFL_MASK(mask, addr < warp_size);
 #pragma unroll
   for (int k = 1; k < 32; k++) {
-    // rSrc[k] = __shfl(rSrc[k], (threadIdx.x + k) % 32, 32);
-    addr = __shfl(addr, (idx + 1) % 32, 32);
-    a[k] = __shfl(a[k], addr, 32);
+    // rSrc[k] = __shfl_sync(rSrc[k], (threadIdx.x + k) % 32, 32);
+    addr = __shfl_sync(mask, addr, (idx + 1) % 32, 32);
+    a[k] = __shfl_sync(mask, a[k], addr, 32);
   }
 
 #pragma unroll
@@ -360,10 +363,11 @@ __device__ __forceinline__ void transpose_32x32(real a[], const int idx) {
   }
 
   addr = (32 - idx) % 32;
+  CREATE_SHFL_MASK(mask, idx % 32 < warp_size);
 #pragma unroll
   for (int k = 0; k < 32; k++) {
-    a[k] = __shfl(a[k], addr, 32);
-    addr = __shfl(addr, (idx + 31) % 32, 32);
+    a[k] = __shfl_sync(mask, a[k], addr, 32);
+    addr = __shfl_sync(mask, addr, (idx + 31) % 32, 32);
   }
 }
 

paddle/cuda/src/hl_top_k.cu

@@ -244,13 +244,16 @@ __device__ __forceinline__ void blockReduce(Pair* shTopK,
     if (--beamSize == 0) break;
     __syncthreads();
 
+    unsigned mask = 0u;
+    // CREATE_SHFL_MASK(mask, tid < len);
+
     if (tid == maxId[0]) {
       if (beam < maxLength) {
         shTopK[tid] = topK[beam];
       }
     }
     if (maxId[0] / 32 == warp) {
-      if (__shfl(beam, (maxId[0]) % 32, 32) == maxLength) break;
+      if (__shfl_sync(mask, beam, (maxId[0]) % 32, 32) == maxLength) break;
     }
   }
 }

paddle/fluid/framework/details/broadcast_op_handle_test.cc

@@ -139,7 +139,7 @@ struct TestBroadcastOpHandle {
       PADDLE_ENFORCE_EQ(out_tensor.lod(), lod, "lod is not equal.");
 
       f::Tensor result_tensor;
-      f::TensorCopy(out_tensor, cpu_place, *(ctxs_[j]), &result_tensor);
+      f::TensorCopySync(out_tensor, cpu_place, &result_tensor);
       float* ct = result_tensor.mutable_data<float>(cpu_place);
 
       for (int64_t i = 0; i < f::product(kDims); ++i) {
@@ -185,7 +185,7 @@ struct TestBroadcastOpHandle {
       }
 
       f::Tensor result_tensor;
-      f::TensorCopy(rt, cpu_place, *(ctxs_[j]), &result_tensor);
+      f::TensorCopySync(rt, cpu_place, &result_tensor);
       float* ct = result_tensor.data<float>();
 
       for (int64_t i = 0; i < f::product(kDims); ++i) {

paddle/fluid/framework/details/fetch_op_handle.cc

@@ -66,8 +66,7 @@ void FetchOpHandle::RunImpl() {
     auto &t = var->Get<framework::LoDTensor>();
     if (platform::is_gpu_place(t.place())) {
 #ifdef PADDLE_WITH_CUDA
-      TensorCopy(t, cpu, *dev_ctxes_[t.place()], &tensors_[i], true);
-      dev_ctxes_.at(t.place())->Wait();
+      TensorCopySync(t, cpu, &tensors_[i]);
 #endif
     } else {
       tensors_[i].ShareDataWith(t);

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -34,7 +34,7 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::vector<platform::Place> &places,
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
-    const std::vector<Scope *> &local_scopes, bool skip_scale_loss,
+    const std::vector<Scope *> &local_scopes, bool use_default_grad_scale,
     platform::NCCLContextMap *nccl_ctxs)
     : loss_var_name_(loss_var_name),
       places_(places),
@@ -45,7 +45,7 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::vector<platform::Place> &places,
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
-    const std::vector<Scope *> &local_scopes, bool skip_scale_loss)
+    const std::vector<Scope *> &local_scopes, bool use_default_grad_scale)
     : loss_var_name_(loss_var_name),
       places_(places),
       local_scopes_(local_scopes) {
@@ -53,28 +53,25 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
   for (auto &p : params) {
     grad_names_.insert(GradVarName(p));
   }
-  skip_scale_loss_ = skip_scale_loss;
+  use_default_grad_scale_ = use_default_grad_scale;
 }
 
 void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result,
                                                 const OpDesc &op,
-                                                const platform::Place &p,
-                                                const size_t &i) const {
+                                                size_t place_id) const {
+  auto p = places_[place_id];
   auto *op_handle = result->ops_.back().get();
   op_handle->SetDeviceContext(p,
                               platform::DeviceContextPool::Instance().Get(p));
 
-  auto var_names = op.InputArgumentNames();
-
-  for (auto &each_var_name : var_names) {
-    VarHandle *var = CreateOrGetLatestVarHandle(result, each_var_name, p, i);
+  for (auto &each_var_name : op.InputArgumentNames()) {
+    VarHandle *var =
+        CreateOrGetLatestVarHandle(result, each_var_name, p, place_id);
     op_handle->AddInput(var);
   }
 
-  var_names = op.OutputArgumentNames();
-
-  for (auto &each_var_name : var_names) {
-    CreateOpOutput(result, op_handle, each_var_name, p, i);
+  for (auto &each_var_name : op.OutputArgumentNames()) {
+    CreateOpOutput(result, op_handle, each_var_name, p, place_id);
   }
 }
 
@@ -84,17 +81,18 @@ bool MultiDevSSAGraphBuilder::IsDistTrainOp(const OpDesc &op,
     return false;
   }
 
-  auto checker = [&](const std::vector<std::string> opvars,
-                     const std::vector<std::string> sendvars) -> bool {
-    bool is_dist_train_op = false;
+  /**
+   * Check any of opvars contains `.block` and in sendvars
+   */
+  auto checker = [](const std::vector<std::string> &opvars,
+                    const std::vector<std::string> &sendvars) -> bool {
     for (auto &var : opvars) {
       if (var.find(".block") != std::string::npos &&
           std::find(sendvars.begin(), sendvars.end(), var) != sendvars.end()) {
-        is_dist_train_op = true;
-        break;
+        return true;
       }
     }
-    return is_dist_train_op;
+    return false;
   };
 
   if (op.Type() == "split") {
@@ -117,13 +115,7 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
       places_.size());
 
   // Find "send" op first for split is in front of send.
-  OpDesc *send_op = nullptr;
-  for (auto *op : program.Block(0).AllOps()) {
-    if (op->Type() == "send") {
-      send_op = op;
-      break;
-    }
-  }
+  OpDesc *send_op = GetSendOpDesc(program);
 
   bool is_forwarding = true;
   for (auto *op : program.Block(0).AllOps()) {
@@ -134,7 +126,8 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
     } else if (IsDistTrainOp(*op, send_op)) {
       CreateComputationalOps(&result, *op, 1);
     } else if (IsScaleLossOp(*op)) {
-      if (!skip_scale_loss_) {
+      // user can customize loss@grad if not use_default_grad_scale_
+      if (use_default_grad_scale_) {
         CreateScaleLossGradOp(&result);
       }
       is_forwarding = false;
@@ -142,10 +135,7 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
       CreateComputationalOps(&result, *op, places_.size());
       if (!is_forwarding) {
         // Currently, we assume that once gradient is generated, it can be
-        // broadcast, and each gradient is only broadcast once. But there are no
-        // other cases, for example, we need to adjust the gradient according to
-        // the input when we get the gradient, which is not considered at
-        // present.
+        // broadcast, and each gradient is only broadcast once.
         for (auto &og : op->OutputArgumentNames()) {
           if (IsParameterGradientOnce(og, &og_has_been_broadcast)) {
             InsertNCCLAllReduceOp(&result, og);
@@ -175,6 +165,16 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
   return std::unique_ptr<SSAGraph>(graph);
 }
 
+OpDesc *MultiDevSSAGraphBuilder::GetSendOpDesc(
+    const ProgramDesc &program) const {
+  for (auto *op : program.Block(0).AllOps()) {
+    if (op->Type() == "send") {
+      return op;
+    }
+  }
+  return nullptr;
+}
+
 void MultiDevSSAGraphBuilder::InsertNCCLAllReduceOp(
     SSAGraph *result, const std::string &og) const {
 #ifdef PADDLE_WITH_CUDA
@@ -243,7 +243,7 @@ void MultiDevSSAGraphBuilder::CreateComputationalOps(SSAGraph *result,
     auto p = places_[scope_idx];
     auto s = local_scopes_[scope_idx];
     result->ops_.emplace_back(new ComputationOpHandle(op, s, p));
-    CreateOpHandleIOs(result, op, p, scope_idx);
+    CreateOpHandleIOs(result, op, scope_idx);
   }
 }
 
@@ -255,7 +255,7 @@ void MultiDevSSAGraphBuilder::CreateSendOp(SSAGraph *result,
   result->ops_.emplace_back(new SendOpHandle(op, s, p));
   // Create inputs for output on original place and no ssa output
   // is created for send op.
-  CreateOpHandleIOs(result, op, p, 0);
+  CreateOpHandleIOs(result, op, 0);
 }
 
 bool MultiDevSSAGraphBuilder::IsScaleLossOp(const OpDesc &op) const {

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -41,14 +41,14 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
                           const std::string &loss_var_name,
                           const std::unordered_set<std::string> &params,
                           const std::vector<Scope *> &local_scopes,
-                          bool skip_scale_loss);
+                          bool use_default_grad_scale);
 #endif
 
   std::unique_ptr<SSAGraph> Build(const ProgramDesc &program) const override;
 
  private:
   void CreateOpHandleIOs(SSAGraph *result, const OpDesc &op,
-                         const platform::Place &p, const size_t &i) const;
+                         size_t place_id) const;
 
  private:
   std::string loss_var_name_;
@@ -59,12 +59,15 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
 #ifdef PADDLE_WITH_CUDA
   platform::NCCLContextMap *nccl_ctxs_;
 #endif
-  bool skip_scale_loss_;
+  bool use_default_grad_scale_;
 
   bool IsScaleLossOp(const OpDesc &op) const;
 
   void CreateSendOp(SSAGraph *result, const OpDesc &op) const;
 
+  /**
+   * Is this operator as the end-point operator before/after send operator.
+   */
   bool IsDistTrainOp(const OpDesc &op, OpDesc *send_op) const;
 
   void CreateComputationalOps(SSAGraph *result, const OpDesc &op,
@@ -77,6 +80,12 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
       std::unordered_set<std::string> *og_has_been_broadcast) const;
 
   void InsertNCCLAllReduceOp(SSAGraph *result, const std::string &og) const;
+
+  /**
+   * Get send op in the global block of program.
+   * nullptr if not found.
+   */
+  OpDesc *GetSendOpDesc(const ProgramDesc &program) const;
 };
 }  // namespace details
 }  // namespace framework

paddle/fluid/framework/details/reduce_op_handle_test.cc

@@ -194,7 +194,7 @@ struct TestReduceOpHandle {
     }
 
     f::Tensor result_tensor;
-    f::TensorCopy(rt, cpu_place, *(ctxs_[output_scope_idx]), &result_tensor);
+    f::TensorCopySync(rt, cpu_place, &result_tensor);
     float *ct = result_tensor.data<float>();
 
     for (int64_t j = 0; j < f::product(result_tensor.dims()); ++j) {
@@ -239,7 +239,7 @@ struct TestReduceOpHandle {
     auto &rt = out_var->Get<f::LoDTensor>();
 
     f::Tensor result_tensor;
-    f::TensorCopy(rt, cpu_place, *(ctxs_[output_scope_idx]), &result_tensor);
+    f::TensorCopySync(rt, cpu_place, &result_tensor);
     float *ct = result_tensor.data<float>();
 
     for (int64_t j = 0; j < f::product(result_tensor.dims()); ++j) {

paddle/fluid/framework/details/scale_loss_grad_op_handle.cc

@@ -46,6 +46,7 @@ void ScaleLossGradOpHandle::RunImpl() {
               ->stream();
       memory::Copy(boost::get<platform::CUDAPlace>(place_), tmp,
                    platform::CPUPlace(), &coeff_, sizeof(float), stream);
+      VLOG(1) << place_ << "RUN Scale loss grad op";
     });
 #endif
   }

paddle/fluid/framework/details/ssa_graph.h

@@ -25,12 +25,22 @@ namespace paddle {
 namespace framework {
 namespace details {
 
+// A SSA graph used by parallel executor.
 struct SSAGraph {
+  // all variable in each devices.
+  // The outside vector is the device vector. Each element of this vector is a
+  // map from variable name to variables. The variables, who have the same name,
+  // will have a different version. The offset in the
+  // `std::vector<std::unique_ptr<VarHandle>>` is the version of varaibles.
   std::vector<
       std::unordered_map<std::string, std::vector<std::unique_ptr<VarHandle>>>>
       vars_;
+
   // aux variables to represent dependency. Useful to resolve data hazard.
   std::unordered_set<std::unique_ptr<VarHandleBase>> dep_vars_;
+
+  // all operators. NOTE that even we use a vector here, the operators is
+  // unordered.
   std::vector<std::unique_ptr<OpHandleBase>> ops_;
 };
 

paddle/fluid/framework/details/ssa_graph_builder.h

@@ -48,6 +48,8 @@ class SSAGraphBuilder {
                                                const platform::Place &place,
                                                size_t place_offset);
 
+  // Add an output variable (each_var_name, place, place_offset) to op_handle,
+  // which belongs to graph
   static void CreateOpOutput(SSAGraph *graph, OpHandleBase *op_handle,
                              const std::string &each_var_name,
                              const platform::Place &place, size_t place_offset);

paddle/fluid/framework/init.cc

@@ -15,7 +15,6 @@ limitations under the License. */
 #include <algorithm>
 #include <stdexcept>
 #include <string>
-#include <vector>
 
 #include "paddle/fluid/framework/init.h"
 #include "paddle/fluid/framework/operator.h"
@@ -31,6 +30,7 @@ std::once_flag p2p_init_flag;
 
 void InitGflags(std::vector<std::string> argv) {
   std::call_once(gflags_init_flag, [&]() {
+    argv.insert(argv.begin(), "dummy");
     int argc = argv.size();
     char **arr = new char *[argv.size()];
     std::string line;
@@ -44,20 +44,23 @@ void InitGflags(std::vector<std::string> argv) {
   });
 }
 
-void InitP2P(int count) {
+void InitP2P(std::vector<int> devices) {
 #ifdef PADDLE_WITH_CUDA
   std::call_once(p2p_init_flag, [&]() {
+    int count = devices.size();
     for (int i = 0; i < count; ++i) {
       for (int j = 0; j < count; ++j) {
-        if (i == j) continue;
+        if (devices[i] == devices[j]) continue;
         int can_acess = -1;
-        PADDLE_ENFORCE(cudaDeviceCanAccessPeer(&can_acess, i, j),
-                       "Failed to test P2P access.");
+        PADDLE_ENFORCE(
+            cudaDeviceCanAccessPeer(&can_acess, devices[i], devices[j]),
+            "Failed to test P2P access.");
         if (can_acess != 1) {
-          LOG(WARNING) << "Cannot enable P2P access from " << i << " to " << j;
+          LOG(WARNING) << "Cannot enable P2P access from " << devices[i]
+                       << " to " << devices[j];
         } else {
-          cudaSetDevice(i);
-          cudaDeviceEnablePeerAccess(j, 0);
+          cudaSetDevice(devices[i]);
+          cudaDeviceEnablePeerAccess(devices[j], 0);
         }
       }
     }
@@ -67,11 +70,26 @@ void InitP2P(int count) {
 
 void InitDevices(bool init_p2p) {
   /*Init all available devices by default */
+  std::vector<int> devices;
+#ifdef PADDLE_WITH_CUDA
+  try {
+    int count = platform::GetCUDADeviceCount();
+    for (int i = 0; i < count; ++i) {
+      devices.push_back(i);
+    }
+  } catch (const std::exception &exp) {
+    LOG(WARNING) << "Compiled with WITH_GPU, but no GPU found in runtime.";
+  }
+#else
+  LOG(WARNING)
+      << "'CUDA' is not supported, Please re-compile with WITH_GPU option";
+#endif
+  InitDevices(init_p2p, devices);
+}
 
+void InitDevices(bool init_p2p, const std::vector<int> devices) {
   std::vector<platform::Place> places;
-  places.emplace_back(platform::CPUPlace());
   int count = 0;
-
 #ifdef PADDLE_WITH_CUDA
   try {
     count = platform::GetCUDADeviceCount();
@@ -83,12 +101,17 @@ void InitDevices(bool init_p2p) {
       << "'CUDA' is not supported, Please re-compile with WITH_GPU option";
 #endif
 
-  for (int i = 0; i < count; ++i) {
-    places.emplace_back(platform::CUDAPlace(i));
+  for (size_t i = 0; i < devices.size(); ++i) {
+    if (devices[i] >= count || devices[i] < 0) {
+      LOG(WARNING) << "Invalid devices id.";
+      continue;
+    }
+    places.emplace_back(platform::CUDAPlace(devices[i]));
   }
   if (init_p2p) {
-    InitP2P(count);
+    InitP2P(devices);
   }
+  places.emplace_back(platform::CPUPlace());
   platform::DeviceContextPool::Init(places);
 }
 

paddle/fluid/framework/init.h

@@ -28,5 +28,7 @@ void InitGLOG(const std::string &prog_name);
 
 void InitDevices(bool init_p2p);
 
+void InitDevices(bool init_p2p, const std::vector<int> devices);
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/parallel_executor.cc

@@ -58,7 +58,7 @@ ParallelExecutor::ParallelExecutor(
     const std::unordered_set<std::string> &bcast_vars,
     const ProgramDesc &main_program, const std::string &loss_var_name,
     Scope *scope, const std::vector<Scope *> &local_scopes, bool allow_op_delay,
-    bool customize_scale_loss)
+    bool use_default_grad_scale)
     : member_(new ParallelExecutorPrivate(places)) {
   member_->global_scope_ = scope;
 
@@ -93,11 +93,11 @@ ParallelExecutor::ParallelExecutor(
 #ifdef PADDLE_WITH_CUDA
   details::MultiDevSSAGraphBuilder builder(
       member_->places_, loss_var_name, params, member_->local_scopes_,
-      customize_scale_loss, member_->nccl_ctxs_.get());
+      use_default_grad_scale, member_->nccl_ctxs_.get());
 #else
   details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name,
                                            params, member_->local_scopes_,
-                                           customize_scale_loss);
+                                           use_default_grad_scale);
 #endif
   auto graph = builder.Build(main_program);
 

paddle/fluid/framework/parallel_executor.h

@@ -40,7 +40,7 @@ class ParallelExecutor {
                             const ProgramDesc& main_program,
                             const std::string& loss_var_name, Scope* scope,
                             const std::vector<Scope*>& local_scopes,
-                            bool allow_op_delay, bool customize_scale_loss);
+                            bool allow_op_delay, bool use_default_grad_scale);
 
   ~ParallelExecutor();
 

paddle/fluid/framework/tensor_util.cc

@@ -20,7 +20,7 @@ namespace paddle {
 namespace framework {
 
 void TensorCopy(const Tensor& src, const platform::Place& dst_place,
-                const platform::DeviceContext& ctx, Tensor* dst, bool sync) {
+                const platform::DeviceContext& ctx, Tensor* dst) {
   VLOG(3) << "TensorCopy " << src.dims() << " from " << src.place() << " to "
           << dst_place;
   src.check_memory_size();
@@ -48,9 +48,7 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
     auto ctx_gpu_place = boost::get<platform::CUDAPlace>(ctx_place);
     PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place);
     auto stream =
-        sync ? nullptr
-             : reinterpret_cast<const platform::CUDADeviceContext&>(ctx)
-                   .stream();
+        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
     memory::Copy(dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, stream);
   } else if (platform::is_cpu_place(src_place) &&
              platform::is_gpu_place(dst_place)) {
@@ -61,9 +59,7 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
     auto ctx_gpu_place = boost::get<platform::CUDAPlace>(ctx_place);
     PADDLE_ENFORCE_EQ(dst_gpu_place, ctx_gpu_place);
     auto stream =
-        sync ? nullptr
-             : reinterpret_cast<const platform::CUDADeviceContext&>(ctx)
-                   .stream();
+        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
     memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, stream);
   } else if (platform::is_gpu_place(src_place) &&
              platform::is_gpu_place(dst_place)) {
@@ -72,9 +68,7 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
     auto ctx_place = ctx.GetPlace();
     PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
     auto stream =
-        sync ? nullptr
-             : reinterpret_cast<const platform::CUDADeviceContext&>(ctx)
-                   .stream();
+        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
     memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, stream);
   }
 #endif
@@ -92,6 +86,41 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
   TensorCopy(src, dst_place, *dev_ctx, dst);
 }
 
+void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
+                    Tensor* dst) {
+  VLOG(3) << "TensorCopySync " << src.dims() << " from " << src.place()
+          << " to " << dst_place;
+  src.check_memory_size();
+  dst->Resize(src.dims());
+  dst->set_layout(src.layout());
+  auto src_place = src.place();
+  auto src_ptr = src.data<void>();
+  auto dst_ptr = dst->mutable_data(dst_place, src.type());
+  auto size = src.numel() * SizeOfType(src.type());
+  if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
+    memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
+                 boost::get<platform::CPUPlace>(src_place), src_ptr, size);
+  }
+#ifdef PADDLE_WITH_CUDA
+  else if (platform::is_gpu_place(src_place) &&  // NOLINT
+           platform::is_cpu_place(dst_place)) {
+    auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
+    auto dst_cpu_place = boost::get<platform::CPUPlace>(dst_place);
+    memory::Copy(dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr);
+  } else if (platform::is_cpu_place(src_place) &&
+             platform::is_gpu_place(dst_place)) {
+    auto src_cpu_place = boost::get<platform::CPUPlace>(src_place);
+    auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
+    memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, nullptr);
+  } else if (platform::is_gpu_place(src_place) &&
+             platform::is_gpu_place(dst_place)) {
+    auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
+    auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
+    memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr);
+  }
+#endif
+}
+
 template <typename Predicate, typename DevCtx>
 struct AnyDTypeVisitor {
   Predicate predicate_;

paddle/fluid/framework/tensor_util.h

@@ -24,10 +24,11 @@ namespace paddle {
 namespace framework {
 
 void TensorCopy(const Tensor& src, const platform::Place& dst_place,
-                const platform::DeviceContext& ctx, Tensor* dst,
-                bool sync = false);
+                const platform::DeviceContext& ctx, Tensor* dst);
 void TensorCopy(const Tensor& src, const platform::Place& dst_place,
                 Tensor* dst);
+void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
+                    Tensor* dst);
 
 template <typename T>
 void TensorFromVector(const std::vector<T>& src,

paddle/fluid/inference/engine.h

@@ -46,7 +46,6 @@ class EngineBase {
   virtual void Execute(int batch_size) = 0;
 
   virtual ~EngineBase() {}
-
 };  // class EngineBase
 
 }  // namespace inference

paddle/fluid/inference/io.cc

@@ -16,17 +16,29 @@ limitations under the License. */
 
 #include <algorithm>
 #include <fstream>
+#include <vector>
 #include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/feed_fetch_type.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/pybind/pybind.h"
 
+DEFINE_string(devices, "", "The devices to be used which is joined by comma.");
+DEFINE_bool(init_p2p, false, "Whether to init p2p.");
+
 namespace paddle {
 namespace inference {
 
-// Temporarily add this function for exposing framework::InitDevices() when
-// linking the inference shared library.
-void Init(bool init_p2p) { framework::InitDevices(init_p2p); }
+void Init(const std::vector<std::string> argv) {
+  framework::InitGflags(argv);
+  // init devices
+  std::vector<int> devices;
+  std::string token;
+  std::istringstream tokenStream(FLAGS_devices);
+  while (std::getline(tokenStream, token, ',')) {
+    devices.push_back(std::stoi(token));
+  }
+  framework::InitDevices(FLAGS_init_p2p, devices);
+}
 
 void ReadBinaryFile(const std::string& filename, std::string* contents) {
   std::ifstream fin(filename, std::ios::in | std::ios::binary);

paddle/fluid/inference/io.h

@@ -25,7 +25,7 @@ limitations under the License. */
 namespace paddle {
 namespace inference {
 
-void Init(bool init_p2p);
+void Init(const std::vector<std::string> argv);
 
 void LoadPersistables(framework::Executor* executor, framework::Scope* scope,
                       const framework::ProgramDesc& main_program,

paddle/fluid/inference/tensorrt/engine.cc

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <NvInfer.h>
 #include <cuda.h>
 #include <glog/logging.h>
+#include <string>
 #include "paddle/fluid/inference/tensorrt/helper.h"
 #include "paddle/fluid/platform/enforce.h"
 

paddle/fluid/inference/tensorrt/engine.h

@@ -16,7 +16,9 @@ limitations under the License. */
 
 #include <NvInfer.h>
 #include <memory>
+#include <string>
 #include <unordered_map>
+#include <vector>
 #include "paddle/fluid/inference/engine.h"
 #include "paddle/fluid/inference/tensorrt/helper.h"
 
@@ -56,9 +58,9 @@ class TensorRTEngine : public EngineBase {
   virtual ~TensorRTEngine();
 
   // TODO(Superjomn) implement it later when graph segmentation is supported.
-  virtual void Build(const DescType& paddle_model) override;
+  void Build(const DescType& paddle_model) override;
 
-  virtual void Execute(int batch_size) override;
+  void Execute(int batch_size) override;
 
   // Initialize the inference network, so that TensorRT layers can add to this
   // network.

paddle/fluid/inference/tensorrt/test_engine.cc

@@ -12,13 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/inference/tensorrt/engine.h"
-
 #include <cuda.h>
 #include <cuda_runtime_api.h>
 #include <glog/logging.h>
 #include <gtest/gtest.h>
 
+#include "paddle/fluid/inference/tensorrt/engine.h"
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {
@@ -65,7 +64,8 @@ TEST_F(TensorRTEngineTest, add_layer) {
 
   // fill in real data
   float x_v = 1234;
-  engine_->SetInputFromCPU("x", (void*)&x_v, 1 * sizeof(float));
+  engine_->SetInputFromCPU("x", reinterpret_cast<void*>(&x_v),
+                           1 * sizeof(float));
   LOG(INFO) << "to execute";
   engine_->Execute(1);
 

paddle/fluid/inference/tests/book/test_inference_image_classification.cc

@@ -62,5 +62,21 @@ TEST(inference, image_classification) {
   LOG(INFO) << output2.dims();
 
   CheckError<float>(output1, output2);
+
+  // float16 inference requires cuda GPUs with >= 5.3 compute capability
+  if (paddle::platform::GetCUDAComputeCapability(0) >= 53) {
+    paddle::framework::LoDTensor output3;
+    std::vector<paddle::framework::LoDTensor*> cpu_fetchs3;
+    cpu_fetchs3.push_back(&output3);
+
+    LOG(INFO) << "--- GPU Runs in float16 mode: ---";
+    std::string fp16_dirname = dirname;
+    fp16_dirname.replace(fp16_dirname.find("book/"),
+                         std::string("book/").size(), "book/float16_");
+    TestInference<paddle::platform::CUDAPlace, false, true>(
+        fp16_dirname, cpu_feeds, cpu_fetchs3, FLAGS_repeat);
+
+    CheckError<float>(output2, output3);
+  }
 #endif
 }

paddle/fluid/operators/accuracy_op.cu

@@ -15,7 +15,7 @@ limitations under the License. */
 #include <thrust/execution_policy.h>
 #include <thrust/reduce.h>
 #include "paddle/fluid/operators/accuracy_op.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 #include "paddle/fluid/platform/gpu_info.h"
 
 namespace paddle {

paddle/fluid/operators/adagrad_op.cu

@@ -16,7 +16,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/adagrad_op.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/adam_op.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 #include <math.h>  // for sqrt in CPU and CUDA
+#include <Eigen/Dense>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
@@ -24,8 +25,14 @@ namespace operators {
 
 namespace scatter = paddle::operators::math::scatter;
 
+struct GPUAdam;
+struct CPUAdam;
+
+template <typename T, typename Flavour>
+struct AdamFunctor;
+
 template <typename T>
-struct AdamFunctor {
+struct AdamFunctor<T, GPUAdam> {
   T beta1_;
   T beta2_;
   T epsilon_;
@@ -71,6 +78,7 @@ struct AdamFunctor {
 
     // Calculation
     lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow);
+
     mom1 = beta1_ * mom1 + (1 - beta1_) * g;
     mom2 = beta2_ * mom2 + (1 - beta2_) * g * g;
     p -= lr * (mom1 / (sqrt(mom2) + epsilon_));
@@ -82,6 +90,71 @@ struct AdamFunctor {
   }
 };
 
+template <typename T>
+struct AdamFunctor<T, CPUAdam> {
+  T beta1_;
+  T beta2_;
+  T epsilon_;
+
+  const T* beta1_pow_;
+  const T* beta2_pow_;
+  const T* moment1_;
+  T* moment1_out_;
+  const T* moment2_;
+  T* moment2_out_;
+  const T* lr_;
+  const T* grad_;
+  const T* param_;
+  T* param_out_;
+
+  AdamFunctor(T beta1, T beta2, T epsilon, const T* beta1_pow,
+              const T* beta2_pow, const T* mom1, T* mom1_out, const T* mom2,
+              T* mom2_out, const T* lr, const T* grad, const T* param,
+              T* param_out)
+      : beta1_(beta1),
+        beta2_(beta2),
+        epsilon_(epsilon),
+        beta1_pow_(beta1_pow),
+        beta2_pow_(beta2_pow),
+        moment1_(mom1),
+        moment1_out_(mom1_out),
+        moment2_(mom2),
+        moment2_out_(mom2_out),
+        lr_(lr),
+        grad_(grad),
+        param_(param),
+        param_out_(param_out) {}
+
+  void operator()(size_t numel) const {
+    Eigen::Map<const Eigen::Array<T, 1, Eigen::Dynamic>> g{
+        grad_, static_cast<Eigen::Index>(numel)};
+    Eigen::Map<const Eigen::Array<T, 1, Eigen::Dynamic>> mom1{
+        moment1_, static_cast<Eigen::Index>(numel)};
+    Eigen::Map<const Eigen::Array<T, 1, Eigen::Dynamic>> mom2{
+        moment2_, static_cast<Eigen::Index>(numel)};
+    Eigen::Map<const Eigen::Array<T, 1, Eigen::Dynamic>> param{
+        param_, static_cast<Eigen::Index>(numel)};
+
+    Eigen::Map<Eigen::Array<T, 1, Eigen::Dynamic>> param_out{
+        param_out_, static_cast<Eigen::Index>(numel)};
+    Eigen::Map<Eigen::Array<T, 1, Eigen::Dynamic>> moment1_out{
+        moment1_out_, static_cast<Eigen::Index>(numel)};
+    Eigen::Map<Eigen::Array<T, 1, Eigen::Dynamic>> moment2_out{
+        moment2_out_, static_cast<Eigen::Index>(numel)};
+
+    T lr = *lr_;
+    T beta1_pow = *beta1_pow_;
+    T beta2_pow = *beta2_pow_;
+
+    // Calculation
+    lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow);
+
+    moment1_out = beta1_ * mom1 + (1 - beta1_) * g;
+    moment2_out = beta2_ * mom2 + (1 - beta2_) * g * g;
+    param_out = param - lr * (moment1_out / (moment2_out.sqrt() + epsilon_));
+  }
+};
+
 template <typename T>
 struct SparseAdamFunctor {
   T beta1_;
@@ -134,6 +207,7 @@ struct SparseAdamFunctor {
       T p = param_[rows_[i] * row_numel_ + j];
 
       lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow);
+
       mom1 = beta1_ * mom1 + (1 - beta1_) * g;
       mom2 = beta2_ * mom2 + (1 - beta2_) * g * g;
       p -= lr * (mom1 / (sqrt(mom2) + epsilon_));
@@ -177,19 +251,34 @@ class AdamOpKernel : public framework::OpKernel<T> {
 
     if (grad_var->IsType<framework::LoDTensor>()) {
       auto& grad = Ref(ctx.Input<LoDTensor>("Grad"), "Must set Grad");
-      AdamFunctor<T> functor(
-          beta1, beta2, epsilon, beta1_pow.template data<T>(),
-          beta2_pow.template data<T>(), mom1.template data<T>(),
-          mom1_out.template mutable_data<T>(ctx.GetPlace()),
-          mom2.template data<T>(),
-          mom2_out.template mutable_data<T>(ctx.GetPlace()),
-          lr.template data<T>(), grad.template data<T>(),
-          param.template data<T>(),
-          param_out.template mutable_data<T>(ctx.GetPlace()));
-      platform::ForRange<DeviceContext> for_range(
-          static_cast<const DeviceContext&>(ctx.device_context()),
-          param.numel());
-      for_range(functor);
+
+      if (platform::is_cpu_place(ctx.GetPlace())) {
+        AdamFunctor<T, CPUAdam> functor(
+            beta1, beta2, epsilon, beta1_pow.template data<T>(),
+            beta2_pow.template data<T>(), mom1.template data<T>(),
+            mom1_out.template mutable_data<T>(ctx.GetPlace()),
+            mom2.template data<T>(),
+            mom2_out.template mutable_data<T>(ctx.GetPlace()),
+            lr.template data<T>(), grad.template data<T>(),
+            param.template data<T>(),
+            param_out.template mutable_data<T>(ctx.GetPlace()));
+        functor(param.numel());
+      } else if (platform::is_gpu_place(ctx.GetPlace())) {
+        AdamFunctor<T, GPUAdam> functor(
+            beta1, beta2, epsilon, beta1_pow.template data<T>(),
+            beta2_pow.template data<T>(), mom1.template data<T>(),
+            mom1_out.template mutable_data<T>(ctx.GetPlace()),
+            mom2.template data<T>(),
+            mom2_out.template mutable_data<T>(ctx.GetPlace()),
+            lr.template data<T>(), grad.template data<T>(),
+            param.template data<T>(),
+            param_out.template mutable_data<T>(ctx.GetPlace()));
+
+        platform::ForRange<DeviceContext> for_range(
+            static_cast<const DeviceContext&>(ctx.device_context()),
+            param.numel());
+        for_range(functor);
+      }
     } else if (grad_var->IsType<framework::SelectedRows>()) {
       auto& grad =
           Ref(ctx.Input<framework::SelectedRows>("Grad"), "Must set Grad");

paddle/fluid/operators/beam_search_op.cc

@@ -195,10 +195,9 @@ std::string ItemToString(const BeamSearch::Item &item) {
   return stream.str();
 }
 
-class BeamSearchProtoAndCheckerMaker
-    : public framework::OpProtoAndCheckerMaker {
+class BeamSearchOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BeamSearchProtoAndCheckerMaker(OpProto *proto, OpAttrChecker *op_checker)
+  BeamSearchOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     // inputs and outputs stored in proto
     AddInput("pre_ids", "ids in previous step");
@@ -222,20 +221,32 @@ class BeamSearchProtoAndCheckerMaker
   }
 };
 
-class BeamSearchInferShape : public framework::InferShapeBase {
+class BeamSearchOp : public framework::OperatorWithKernel {
  public:
-  void operator()(framework::InferShapeContext *context) const override {
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
     for (const std::string &arg :
          std::vector<std::string>({"pre_ids", "ids", "scores"})) {
-      PADDLE_ENFORCE(context->HasInput(arg),
-                     "BeamSearch need input argument '%s'", arg);
+      PADDLE_ENFORCE(ctx->HasInput(arg), "BeamSearch need input argument '%s'",
+                     arg);
     }
     for (const std::string &arg :
          std::vector<std::string>({"selected_ids", "selected_scores"})) {
-      PADDLE_ENFORCE(context->HasOutput(arg),
+      PADDLE_ENFORCE(ctx->HasOutput(arg),
                      "BeamSearch need output argument '%s'", arg);
     }
   }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    framework::OpKernelType kt = framework::OpKernelType(
+        framework::ToDataType(
+            ctx.Input<framework::LoDTensor>("pre_ids")->type()),
+        platform::CPUPlace());
+    return kt;
+  }
 };
 
 class BeamSearchInferVarType : public framework::VarTypeInference {
@@ -254,8 +265,13 @@ class BeamSearchInferVarType : public framework::VarTypeInference {
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OPERATOR(beam_search, paddle::operators::BeamSearchOp,
-                  paddle::operators::BeamSearchProtoAndCheckerMaker,
-                  paddle::operators::BeamSearchInferShape,
-                  paddle::operators::BeamSearchInferVarType,
-                  paddle::framework::EmptyGradOpMaker);
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(beam_search, ops::BeamSearchOp, ops::BeamSearchOpMaker,
+                  ops::BeamSearchInferVarType);
+REGISTER_OP_CPU_KERNEL(
+    beam_search,
+    ops::BeamSearchOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::BeamSearchOpKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::BeamSearchOpKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::BeamSearchOpKernel<paddle::platform::CPUDeviceContext, int64_t>);

paddle/fluid/operators/beam_search_op.h

@@ -192,49 +192,29 @@ std::ostream& operator<<(std::ostream& os, const BeamSearch::Item& item);
 
 std::string ItemToString(const BeamSearch::Item& item);
 
-class BeamSearchOp : public framework::OperatorBase {
+template <typename DeviceContext, typename T>
+class BeamSearchOpKernel : public framework::OpKernel<T> {
  public:
-  BeamSearchOp(const std::string& type,
-               const framework::VariableNameMap& inputs,
-               const framework::VariableNameMap& outputs,
-               const framework::AttributeMap& attrs)
-      : OperatorBase(type, inputs, outputs, attrs) {}
-
-  BeamSearchOp(const BeamSearchOp& o)
-      : framework::OperatorBase(
-            static_cast<const framework::OperatorBase&>(o)) {
-    PADDLE_THROW("Not Implemented");
-  }
-
- private:
-  void RunImpl(const framework::Scope& scope,
-               const platform::Place& dev_place) const override {
-    auto ids_var = scope.FindVar(Input("ids"));
-    auto scores_var = scope.FindVar(Input("scores"));
-    auto pre_ids_var = scope.FindVar(Input("pre_ids"));
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* ids_var = context.Input<framework::LoDTensor>("ids");
+    auto* scores_var = context.Input<framework::LoDTensor>("scores");
+    auto* pre_ids_var = context.Input<framework::LoDTensor>("pre_ids");
     PADDLE_ENFORCE_NOT_NULL(ids_var);
     PADDLE_ENFORCE_NOT_NULL(scores_var);
     PADDLE_ENFORCE_NOT_NULL(pre_ids_var);
 
-    auto& ids = ids_var->Get<framework::LoDTensor>();
-    auto& scores = scores_var->Get<framework::LoDTensor>();
-    auto& pre_ids = pre_ids_var->Get<framework::LoDTensor>();
-    size_t level = Attr<int>("level");
-    size_t beam_size = Attr<int>("beam_size");
-    int end_id = Attr<int>("end_id");
-    BeamSearch alg(ids, scores, level, beam_size, end_id);
-
-    auto selected_ids_var = scope.FindVar(Output("selected_ids"));
-    auto selected_scores_var = scope.FindVar(Output("selected_scores"));
+    size_t level = context.Attr<int>("level");
+    size_t beam_size = context.Attr<int>("beam_size");
+    int end_id = context.Attr<int>("end_id");
+    BeamSearch alg(*ids_var, *scores_var, level, beam_size, end_id);
+    auto selected_ids_var =
+        context.Output<framework::LoDTensor>("selected_ids");
+    auto selected_scores_var =
+        context.Output<framework::LoDTensor>("selected_scores");
     PADDLE_ENFORCE_NOT_NULL(selected_ids_var);
     PADDLE_ENFORCE_NOT_NULL(selected_scores_var);
-    auto& selected_ids_tensor =
-        *selected_ids_var->GetMutable<framework::LoDTensor>();
-    auto& selected_scores_tensor =
-        *selected_scores_var->GetMutable<framework::LoDTensor>();
-    alg(pre_ids, &selected_ids_tensor, &selected_scores_tensor);
+    alg(*pre_ids_var, selected_ids_var, selected_scores_var);
   }
 };
-
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/bilinear_interp_op.cu

@@ -10,7 +10,7 @@
    limitations under the License. */
 
 #include "paddle/fluid/operators/bilinear_interp_op.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/box_coder_op.cu

@@ -10,7 +10,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/box_coder_op.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/concat_op.h

@@ -87,7 +87,7 @@ class ConcatGradKernel : public framework::OpKernel<T> {
       auto& dev_ctx = ctx.template device_context<DeviceContext>();
       paddle::operators::math::ConcatGradFunctor<DeviceContext, T>
           concat_grad_functor;
-      concat_grad_functor(dev_ctx, *in, static_cast<int>(axis), outputs);
+      concat_grad_functor(dev_ctx, *in, static_cast<int>(axis), &outputs);
     }
   }
 };

paddle/fluid/operators/concurrency/channel_util.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 "channel_util.h"
+#include "paddle/fluid/operators/concurrency/channel_util.h"
 #include "paddle/fluid/framework/var_type.h"
 
 namespace poc = paddle::operators::concurrency;

paddle/fluid/operators/conv_cudnn_op.cu.cc

@@ -20,6 +20,11 @@ limitations under the License. */
 #include "paddle/fluid/platform/cudnn_helper.h"
 #include "paddle/fluid/platform/float16.h"
 
+DEFINE_bool(cudnn_algo_use_autotune, true,
+            "Whether allow using an autotuning algorithm for convolution "
+            "operator. The autotuning algorithm may be non-deterministic. If "
+            "false, the algorithm is deterministic.");
+
 namespace paddle {
 namespace operators {
 
@@ -267,17 +272,23 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
     auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
     auto handle = dev_ctx.cudnn_handle();
     if (input_grad) {
-      PADDLE_ENFORCE(
-          platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
-              handle, cudnn_filter_desc,
-              // dyDesc: Handle to the previously initialized input differential
-              // tensor descriptor.
-              cudnn_output_grad_desc, cudnn_conv_desc,
-              // dxDesc: Handle to the previously initialized output tensor
-              // descriptor.
-              cudnn_input_desc,
-              CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
-              workspace_size_limit, &data_algo));
+      if (FLAGS_cudnn_algo_use_autotune) {
+        PADDLE_ENFORCE(
+            platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
+                handle, cudnn_filter_desc,
+                // dyDesc: Handle to the previously initialized input
+                // differential
+                // tensor descriptor.
+                cudnn_output_grad_desc, cudnn_conv_desc,
+                // dxDesc: Handle to the previously initialized output tensor
+                // descriptor.
+                cudnn_input_desc,
+                CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
+                workspace_size_limit, &data_algo));
+      } else {
+        data_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1;
+      }
+
       PADDLE_ENFORCE(
           platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
               handle, cudnn_filter_desc, cudnn_output_grad_desc,
@@ -286,12 +297,16 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
     }
 
     if (filter_grad) {
-      PADDLE_ENFORCE(
-          platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
-              handle, cudnn_input_desc, cudnn_output_grad_desc, cudnn_conv_desc,
-              cudnn_filter_desc,
-              CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
-              workspace_size_limit, &filter_algo));
+      if (FLAGS_cudnn_algo_use_autotune) {
+        PADDLE_ENFORCE(
+            platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
+                handle, cudnn_input_desc, cudnn_output_grad_desc,
+                cudnn_conv_desc, cudnn_filter_desc,
+                CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
+                workspace_size_limit, &filter_algo));
+      } else {
+        filter_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1;
+      }
 
       PADDLE_ENFORCE(
           platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize(

paddle/fluid/operators/conv_shift_op.cu

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/conv_shift_op.h"
 #include "paddle/fluid/operators/math/math_function.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/detail/grpc_server.cc

@@ -30,9 +30,13 @@ enum CallStatus { PROCESS = 0, FINISH };
 class RequestBase {
  public:
   explicit RequestBase(GrpcService::AsyncService* service,
-                       ::grpc::ServerCompletionQueue* cq,
+                       ::grpc::ServerCompletionQueue* cq, bool sync_mode,
                        const platform::DeviceContext* dev_ctx)
-      : service_(service), cq_(cq), status_(PROCESS), dev_ctx_(dev_ctx) {
+      : service_(service),
+        cq_(cq),
+        sync_mode_(sync_mode),
+        status_(PROCESS),
+        dev_ctx_(dev_ctx) {
     PADDLE_ENFORCE(cq_);
   }
   virtual ~RequestBase() {}
@@ -49,6 +53,7 @@ class RequestBase {
   ::grpc::ServerContext ctx_;
   GrpcService::AsyncService* service_;
   ::grpc::ServerCompletionQueue* cq_;
+  const bool sync_mode_;
   CallStatus status_;
   const platform::DeviceContext* dev_ctx_;
 };
@@ -56,11 +61,17 @@ class RequestBase {
 class RequestSend final : public RequestBase {
  public:
   explicit RequestSend(GrpcService::AsyncService* service,
-                       ::grpc::ServerCompletionQueue* cq,
+                       ::grpc::ServerCompletionQueue* cq, bool sync_mode,
                        framework::Scope* scope, ReceivedQueue* queue,
                        const platform::DeviceContext* dev_ctx)
-      : RequestBase(service, cq, dev_ctx), queue_(queue), responder_(&ctx_) {
-    request_.reset(new VariableResponse(scope, dev_ctx_));
+      : RequestBase(service, cq, sync_mode, dev_ctx),
+        queue_(queue),
+        responder_(&ctx_) {
+    if (sync_mode_) {
+      request_.reset(new VariableResponse(scope, dev_ctx_, false));
+    } else {
+      request_.reset(new VariableResponse(scope, dev_ctx_, true));
+    }
     int method_id = static_cast<int>(detail::GrpcMethod::kSendVariable);
     service_->RequestAsyncUnary(method_id, &ctx_, request_.get(), &responder_,
                                 cq_, cq_, this);
@@ -87,11 +98,11 @@ class RequestSend final : public RequestBase {
 class RequestGet final : public RequestBase {
  public:
   explicit RequestGet(GrpcService::AsyncService* service,
-                      ::grpc::ServerCompletionQueue* cq,
+                      ::grpc::ServerCompletionQueue* cq, bool sync_mode,
                       framework::Scope* scope,
                       const platform::DeviceContext* dev_ctx,
                       framework::BlockingQueue<MessageWithName>* queue)
-      : RequestBase(service, cq, dev_ctx),
+      : RequestBase(service, cq, sync_mode, dev_ctx),
         responder_(&ctx_),
         scope_(scope),
         queue_(queue) {
@@ -134,19 +145,23 @@ class RequestGet final : public RequestBase {
 class RequestPrefetch final : public RequestBase {
  public:
   explicit RequestPrefetch(GrpcService::AsyncService* service,
-                           ::grpc::ServerCompletionQueue* cq,
+                           ::grpc::ServerCompletionQueue* cq, bool sync_mode,
                            framework::Scope* scope,
                            const platform::DeviceContext* dev_ctx,
                            framework::Executor* executor,
                            framework::ProgramDesc* program,
                            framework::ExecutorPrepareContext* prefetch_ctx)
-      : RequestBase(service, cq, dev_ctx),
+      : RequestBase(service, cq, sync_mode, dev_ctx),
         responder_(&ctx_),
         scope_(scope),
         executor_(executor),
         program_(program),
         prefetch_ctx_(prefetch_ctx) {
-    request_.reset(new VariableResponse(scope, dev_ctx_));
+    if (sync_mode_) {
+      request_.reset(new VariableResponse(scope, dev_ctx_, false));
+    } else {
+      request_.reset(new VariableResponse(scope, dev_ctx_, true));
+    }
     int method_id = static_cast<int>(detail::GrpcMethod::kPrefetchVariable);
     service_->RequestAsyncUnary(method_id, &ctx_, request_.get(), &responder_,
                                 cq_, cq_, this);
@@ -181,7 +196,6 @@ class RequestPrefetch final : public RequestBase {
   framework::Executor* executor_;
   framework::ProgramDesc* program_;
   framework::ExecutorPrepareContext* prefetch_ctx_;
-  int blkid_;
 };
 
 void AsyncGRPCServer::WaitClientGet(int count) {
@@ -254,8 +268,8 @@ void AsyncGRPCServer::TryToRegisterNewSendOne() {
     VLOG(3) << "shutdown, do not TryToRegisterNewSendOne";
     return;
   }
-  RequestSend* send = new RequestSend(&service_, cq_send_.get(), scope_,
-                                      &var_recv_queue_, dev_ctx_);
+  RequestSend* send = new RequestSend(&service_, cq_send_.get(), sync_mode_,
+                                      scope_, &var_recv_queue_, dev_ctx_);
   VLOG(4) << "Create RequestSend status:" << send->Status();
 }
 
@@ -265,8 +279,8 @@ void AsyncGRPCServer::TryToRegisterNewGetOne() {
     VLOG(3) << "shutdown, do not TryToRegisterNewGetOne";
     return;
   }
-  RequestGet* get = new RequestGet(&service_, cq_get_.get(), scope_, dev_ctx_,
-                                   &var_get_queue_);
+  RequestGet* get = new RequestGet(&service_, cq_get_.get(), sync_mode_, scope_,
+                                   dev_ctx_, &var_get_queue_);
   VLOG(4) << "Create RequestGet status:" << get->Status();
 }
 
@@ -277,8 +291,8 @@ void AsyncGRPCServer::TryToRegisterNewPrefetchOne() {
     return;
   }
   RequestPrefetch* prefetch =
-      new RequestPrefetch(&service_, cq_prefetch_.get(), scope_, dev_ctx_,
-                          executor_, program_, prefetch_ctx_);
+      new RequestPrefetch(&service_, cq_prefetch_.get(), sync_mode_, scope_,
+                          dev_ctx_, executor_, program_, prefetch_ctx_);
 
   VLOG(4) << "Create RequestPrefetch status:" << prefetch->Status();
 }
@@ -301,9 +315,11 @@ void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq,
     VLOG(3) << "HandleRequest for " << cq_name << " while after Next";
 
     PADDLE_ENFORCE(tag);
-    // FIXME(typhoonzero): de-couple the barriers with recv_op
-    if (!is_shut_down_ && cq_name == "cq_get") WaitCond(1);
-    if (!is_shut_down_ && cq_name == "cq_send") WaitCond(0);
+    if (sync_mode_) {
+      // FIXME(typhoonzero): de-couple the barriers with recv_op
+      if (!is_shut_down_ && cq_name == "cq_get") WaitCond(1);
+      if (!is_shut_down_ && cq_name == "cq_send") WaitCond(0);
+    }
 
     RequestBase* base = reinterpret_cast<RequestBase*>(tag);
     // reference:
@@ -320,13 +336,13 @@ void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq,
 
     switch (base->Status()) {
       case PROCESS: {
-        VLOG(4) << cq_name << " status:" << base->Status();
+        VLOG(4) << cq_name << " PROCESS status:" << base->Status();
         TryToRegisterNewOne();
         base->Process();
         break;
       }
       case FINISH: {
-        VLOG(4) << cq_name << " status:" << base->Status();
+        VLOG(4) << cq_name << " FINISH status:" << base->Status();
         delete base;
         break;
       }

paddle/fluid/operators/detail/grpc_server.h

@@ -44,7 +44,8 @@ class RequestBase;
 
 class AsyncGRPCServer final {
  public:
-  explicit AsyncGRPCServer(const std::string &address) : address_(address) {}
+  explicit AsyncGRPCServer(const std::string &address, bool sync_mode)
+      : address_(address), sync_mode_(sync_mode) {}
 
   void RunSyncUpdate();
 
@@ -95,6 +96,7 @@ class AsyncGRPCServer final {
   std::unique_ptr<::grpc::Server> server_;
 
   std::string address_;
+  const bool sync_mode_;
   framework::Scope *scope_;
   const platform::DeviceContext *dev_ctx_;
 

paddle/fluid/operators/detail/grpc_server_test.cc

@@ -89,7 +89,7 @@ void InitTensorsOnServer(framework::Scope* scope, platform::CPUPlace* place,
 }
 
 void StartServer(const std::string& endpoint) {
-  rpc_service_.reset(new detail::AsyncGRPCServer(endpoint));
+  rpc_service_.reset(new detail::AsyncGRPCServer(endpoint, true));
   framework::ProgramDesc program;
   framework::Scope scope;
   platform::CPUPlace place;

paddle/fluid/operators/detail/variable_response.h

@@ -46,7 +46,9 @@ class VariableResponse {
   }
 
   virtual ~VariableResponse() {
-    if (create_scope_) scope_->DeleteScope(local_scope_);
+    if (create_scope_) {
+      scope_->DeleteScope(local_scope_);
+    }
   }
 
   // return:
@@ -63,6 +65,8 @@ class VariableResponse {
 
   const framework::Scope& GetLocalScope() const { return *local_scope_; }
 
+  framework::Scope* GetMutableLocalScope() const { return local_scope_; }
+
   inline std::string Varname() { return meta_.varname(); }
   inline std::string OutVarname() { return meta_.out_varname(); }
 

paddle/fluid/operators/edit_distance_op.cu

@@ -16,7 +16,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/edit_distance_op.h"
 #include "paddle/fluid/operators/math/math_function.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 #include "paddle/fluid/platform/gpu_info.h"
 
 namespace paddle {

paddle/fluid/operators/elementwise_op_function.h

@@ -22,6 +22,7 @@ limitations under the License. */
 #ifdef __NVCC__
 #include <cuda.h>
 #include <thrust/iterator/iterator_adaptor.h>
+#include "paddle/fluid/platform/cuda_primitives.h"
 constexpr int ELEMWISE_MAX_BLOCK_DIM = 1024;
 #endif
 
@@ -333,24 +334,12 @@ static void ElemwiseGradBroadcast1CPU(const T* x, const T* y, const T* out,
     }
   }
 }
-#ifdef __NVCC__
 
-// __shfl_down has been deprecated as of CUDA 9.0.
-#if CUDA_VERSION < 9000
-template <typename T>
-__forceinline__ __device__ T __shfl_down_sync(unsigned, T val, int delta) {
-  return __shfl_down(val, delta);
-}
-#define CREATE_SHFL_MASK(mask, predicate) mask = 0u;
-#else
-#define FULL_WARP_MASK 0xFFFFFFFF
-#define CREATE_SHFL_MASK(mask, predicate) \
-  mask = __ballot_sync(FULL_WARP_MASK, (predicate))
-#endif
+#ifdef __NVCC__
 
 template <typename T>
 __device__ T reduceSum(T val, int tid, int len) {
-  // TODO(zcd): The warp size should be taken from the
+  // NOTE(zcd): The warp size should be taken from the
   // parameters of the GPU but not specified as 32 simply.
   // To make the reduceSum more efficiently,
   // I use Warp-Level Parallelism and assume the Warp size
@@ -362,7 +351,7 @@ __device__ T reduceSum(T val, int tid, int len) {
   CREATE_SHFL_MASK(mask, tid < len);
 
   for (int offset = warpSize / 2; offset > 0; offset /= 2)
-    val += __shfl_down_sync(mask, val, offset);
+    val += platform::__shfl_down_sync(mask, val, offset);
 
   if (tid < warpSize) shm[tid] = 0;
 
@@ -378,7 +367,7 @@ __device__ T reduceSum(T val, int tid, int len) {
   if (tid < warpSize) {
     val = shm[tid];
     for (int offset = warpSize / 2; offset > 0; offset /= 2)
-      val += __shfl_down_sync(mask, val, offset);
+      val += platform::__shfl_down_sync(mask, val, offset);
   }
 
   return val;

paddle/fluid/operators/fetch_op.cc

@@ -57,10 +57,7 @@ class FetchOp : public framework::OperatorBase {
 
     // FIXME(yuyang18): Should we assume the fetch operator always generate
     // CPU outputs?
-    auto &dev_ctx = *pool.Get(src_item.place());
-
-    TensorCopy(src_item, platform::CPUPlace(), dev_ctx, &dst_item);
-    dev_ctx.Wait();
+    TensorCopySync(src_item, platform::CPUPlace(), &dst_item);
     dst_item.set_lod(src_item.lod());
 
     VLOG(3) << "Fetch variable " << fetch_var_name << " to " << out_name;

paddle/fluid/operators/gru_op.h

@@ -34,7 +34,7 @@ inline void ReorderInitState(const DeviceContext& ctx,
                              framework::Tensor* dst, bool indexed_src) {
   math::CopyMatrixRowsFunctor<DeviceContext, T> row_shuffle;
   dst->mutable_data<T>(src.dims(), ctx.GetPlace());
-  row_shuffle(ctx, src, index_lod, *dst, indexed_src);
+  row_shuffle(ctx, src, index_lod, dst, indexed_src);
 }
 
 template <typename DeviceContext, typename T>
@@ -61,7 +61,7 @@ class GRUKernel : public framework::OpKernel<T> {
     bool is_reverse = context.Attr<bool>("is_reverse");
     math::LoDTensor2BatchFunctor<DeviceContext, T> to_batch;
     auto& dev_ctx = context.template device_context<DeviceContext>();
-    to_batch(dev_ctx, *input, *batch_gate, true, is_reverse);
+    to_batch(dev_ctx, *input, batch_gate, true, is_reverse);
 
     if (bias) {
       math::RowwiseAdd<DeviceContext, T> add_bias;
@@ -113,7 +113,7 @@ class GRUKernel : public framework::OpKernel<T> {
 
     math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
     batch_hidden->set_lod(batch_gate->lod());
-    to_seq(dev_ctx, *batch_hidden, *hidden);
+    to_seq(dev_ctx, *batch_hidden, hidden);
   }
 
   void Compute(const framework::ExecutionContext& context) const override {
@@ -174,7 +174,7 @@ class GRUGradKernel : public framework::OpKernel<T> {
 
     bool is_reverse = context.Attr<bool>("is_reverse");
     batch_hidden_grad.set_lod(batch_hidden->lod());
-    to_batch(dev_ctx, *hidden_grad, batch_hidden_grad, false, is_reverse);
+    to_batch(dev_ctx, *hidden_grad, &batch_hidden_grad, false, is_reverse);
 
     math::GRUMetaValue<T> gru_value;
     gru_value.gate_weight = const_cast<T*>(weight_data);
@@ -236,7 +236,7 @@ class GRUGradKernel : public framework::OpKernel<T> {
       input_grad->mutable_data<T>(context.GetPlace());
       math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
       batch_gate_grad.set_lod(batch_gate->lod());
-      to_seq(dev_ctx, batch_gate_grad, *input_grad);
+      to_seq(dev_ctx, batch_gate_grad, input_grad);
     }
     if (bias_grad) {
       bias_grad->mutable_data<T>(context.GetPlace());

paddle/fluid/operators/iou_similarity_op.h

@@ -41,22 +41,24 @@ struct IOUSimilarityFunctor {
   IOUSimilarityFunctor(const T* x, const T* y, T* z, int cols)
       : x_(x), y_(y), z_(z), cols_(static_cast<size_t>(cols)) {}
 
-  inline HOSTDEVICE void operator()(size_t row_id) const {
+  inline HOSTDEVICE void operator()(size_t tid) const {
+    size_t row_id = tid / cols_;
+    size_t col_id = tid % cols_;
+
     T x_min1 = x_[row_id * 4];
     T y_min1 = x_[row_id * 4 + 1];
     T x_max1 = x_[row_id * 4 + 2];
     T y_max1 = x_[row_id * 4 + 3];
-    for (size_t i = 0; i < cols_; ++i) {
-      T x_min2 = y_[i * 4];
-      T y_min2 = y_[i * 4 + 1];
-      T x_max2 = y_[i * 4 + 2];
-      T y_max2 = y_[i * 4 + 3];
 
-      T sim = IOUSimilarity(x_min1, y_min1, x_max1, y_max1, x_min2, y_min2,
-                            x_max2, y_max2);
+    T x_min2 = y_[col_id * 4];
+    T y_min2 = y_[col_id * 4 + 1];
+    T x_max2 = y_[col_id * 4 + 2];
+    T y_max2 = y_[col_id * 4 + 3];
+
+    T sim = IOUSimilarity(x_min1, y_min1, x_max1, y_max1, x_min2, y_min2,
+                          x_max2, y_max2);
 
-      z_[row_id * cols_ + i] = sim;
-    }
+    z_[row_id * cols_ + col_id] = sim;
   }
   const T* x_;
   const T* y_;
@@ -81,7 +83,7 @@ class IOUSimilarityKernel : public framework::OpKernel<T> {
                                     out->mutable_data<T>(ctx.GetPlace()), y_n);
 
     platform::ForRange<DeviceContext> for_range(
-        static_cast<const DeviceContext&>(ctx.device_context()), x_n);
+        static_cast<const DeviceContext&>(ctx.device_context()), x_n * y_n);
     for_range(functor);
   }
 };  // namespace operators

paddle/fluid/operators/listen_and_serv_op.cc

@@ -27,6 +27,38 @@ void RunServer(std::shared_ptr<detail::AsyncGRPCServer> service) {
   VLOG(4) << "RunServer thread end";
 }
 
+static void split(const std::string &str, char sep,
+                  std::vector<std::string> *pieces) {
+  pieces->clear();
+  if (str.empty()) {
+    return;
+  }
+  size_t pos = 0;
+  size_t next = str.find(sep, pos);
+  while (next != std::string::npos) {
+    pieces->push_back(str.substr(pos, next - pos));
+    pos = next + 1;
+    next = str.find(sep, pos);
+  }
+  if (!str.substr(pos).empty()) {
+    pieces->push_back(str.substr(pos));
+  }
+}
+
+static void AsyncExecuteBlock(framework::Executor *executor,
+                              framework::ExecutorPrepareContext *prepared,
+                              framework::Scope *scope) {
+  std::future<void> future = framework::Async([&executor, &prepared, &scope]() {
+    try {
+      executor->RunPreparedContext(prepared, scope, false, false);
+    } catch (std::exception &e) {
+      LOG(ERROR) << "run sub program error " << e.what();
+    }
+  });
+  // TODO(qiao) maybe we can remove this
+  future.wait();
+}
+
 static void ParallelExecuteBlocks(
     const std::vector<size_t> &parallel_blkids, framework::Executor *executor,
     const std::vector<std::shared_ptr<framework::ExecutorPrepareContext>>
@@ -169,15 +201,82 @@ void ListenAndServOp::RunSyncLoop(framework::Executor *executor,
   }  // while(true)
 }
 
+void ListenAndServOp::RunAsyncLoop(framework::Executor *executor,
+                                   framework::ProgramDesc *program,
+                                   framework::Scope *recv_scope,
+                                   framework::BlockDesc *prefetch_block) const {
+  VLOG(3) << "RunAsyncLoop in";
+  // grad name to block id
+  std::unordered_map<std::string, int32_t> grad_to_block_id;
+  std::unordered_map<int32_t, std::string> id_to_grad;
+
+  auto grad_to_block_id_str =
+      Attr<std::vector<std::string>>("grad_to_block_id");
+  for (auto &grad_and_id : grad_to_block_id_str) {
+    std::vector<std::string> pieces;
+    split(grad_and_id, ':', &pieces);
+    VLOG(3) << "after split, grad = " << pieces[0] << ", id=" << pieces[1];
+    PADDLE_ENFORCE_EQ(pieces.size(), 2);
+    PADDLE_ENFORCE_EQ(grad_to_block_id.count(pieces[0]), 0);
+    int block_id = std::stoi(pieces[1]);
+    grad_to_block_id[pieces[0]] = block_id;
+    id_to_grad[block_id] = pieces[0];
+  }
+  size_t num_blocks = program->Size();
+  PADDLE_ENFORCE_GE(num_blocks, 2,
+                    "server program should have at least 2 blocks");
+
+  std::vector<int> block_list;
+  for (size_t blkid = 1; blkid < num_blocks; ++blkid) {
+    block_list.push_back(blkid);
+  }
+  auto optimize_prepared = executor->Prepare(*program, block_list);
+  std::unordered_map<std::string,
+                     std::shared_ptr<framework::ExecutorPrepareContext>>
+      grad_to_prepared_ctx;
+  for (size_t i = 0; i < block_list.size(); ++i) {
+    grad_to_prepared_ctx[id_to_grad[block_list[i]]] = optimize_prepared[i];
+  }
+
+  VLOG(3) << "RunAsyncLoop into while";
+  bool exit_flag = false;
+  while (!exit_flag) {
+    const detail::ReceivedMessage v = rpc_service_->Get();
+    auto recv_var_name = v.first;
+    if (recv_var_name == LISTEN_TERMINATE_MESSAGE) {
+      LOG(INFO) << "received terminate message and exit";
+      exit_flag = true;
+      break;
+    } else {
+      VLOG(3) << "received grad: " << recv_var_name;
+      auto var = v.second->GetVar();
+      if (var == nullptr) {
+        LOG(ERROR) << "Can not find server side var: " << recv_var_name;
+        PADDLE_THROW("Can not find server side var");
+      }
+      AsyncExecuteBlock(executor, grad_to_prepared_ctx[recv_var_name].get(),
+                        v.second->GetMutableLocalScope());
+    }
+
+    if (exit_flag) {
+      rpc_service_->ShutDown();
+      break;
+    }
+  }  // while(true)
+}
+
 void ListenAndServOp::RunImpl(const framework::Scope &scope,
                               const platform::Place &dev_place) const {
   platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
   auto &dev_ctx = *pool.Get(dev_place);
   framework::Scope &recv_scope = scope.NewScope();
 
+  bool sync_mode = Attr<bool>("sync_mode");
+
   PADDLE_ENFORCE(!rpc_service_);
   std::string endpoint = Attr<std::string>("endpoint");
-  rpc_service_.reset(new detail::AsyncGRPCServer(endpoint));
+
+  rpc_service_.reset(new detail::AsyncGRPCServer(endpoint, sync_mode));
 
   auto *optimize_block = Attr<framework::BlockDesc *>(kOptimizeBlock);
   auto *prefetch_block = Attr<framework::BlockDesc *>(kPrefetchBlock);
@@ -202,7 +301,11 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
   sleep(5);
   // Write to a file of server selected port for python use.
   SavePort(rpc_service_);
-  RunSyncLoop(&executor, program, &recv_scope, prefetch_block);
+  if (sync_mode) {
+    RunSyncLoop(&executor, program, &recv_scope, prefetch_block);
+  } else {
+    RunAsyncLoop(&executor, program, &recv_scope, prefetch_block);
+  }
 }
 
 class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -221,6 +324,12 @@ from send_op and send back variables to recv_op.
                          "IP address to listen on.")
         .SetDefault("127.0.0.1:6164")
         .AddCustomChecker([](const std::string &ip) { return !ip.empty(); });
+    AddAttr<std::vector<std::string>>(
+        "grad_to_block_id",
+        "['param1@GRAD.block0:1', 'param2@GRAD.blockn:2'] "
+        "a map from grad name to it's optimize block id")
+        .SetDefault({});
+    AddAttr<bool>("sync_mode", "if works at sync_mode or not").SetDefault(true);
     AddAttr<framework::BlockDesc *>(kOptimizeBlock,
                                     "BlockID to run on server side.");
     AddAttr<framework::BlockDesc *>(kPrefetchBlock,

paddle/fluid/operators/listen_and_serv_op.h

@@ -46,6 +46,11 @@ class ListenAndServOp : public framework::OperatorBase {
                    framework::Scope* recv_scope,
                    framework::BlockDesc* prefetch_block) const;
 
+  void RunAsyncLoop(framework::Executor* executor,
+                    framework::ProgramDesc* program,
+                    framework::Scope* recv_scope,
+                    framework::BlockDesc* prefetch_block) const;
+
   void Stop() override;
 
   void RunImpl(const framework::Scope& scope,

paddle/fluid/operators/lookup_table_op.cu

@@ -16,7 +16,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/lookup_table_op.h"
 #include "paddle/fluid/platform/assert.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/lstm_op.h

@@ -33,7 +33,7 @@ inline void ReorderInitState(const DeviceContext& ctx,
                              framework::Tensor* dst, bool indexed_src) {
   math::CopyMatrixRowsFunctor<DeviceContext, T> row_shuffle;
   dst->mutable_data<T>(src.dims(), ctx.GetPlace());
-  row_shuffle(ctx, src, index_lod, *dst, indexed_src);
+  row_shuffle(ctx, src, index_lod, dst, indexed_src);
 }
 
 template <typename DeviceContext, typename T>
@@ -57,7 +57,7 @@ class LSTMKernel : public framework::OpKernel<T> {
     bool is_reverse = ctx.Attr<bool>("is_reverse");
     math::LoDTensor2BatchFunctor<DeviceContext, T> to_batch;
     auto& device_ctx = ctx.template device_context<DeviceContext>();
-    to_batch(device_ctx, *input, *batch_gate, true, is_reverse);
+    to_batch(device_ctx, *input, batch_gate, true, is_reverse);
 
     auto in_dims = input->dims();
     int frame_size = static_cast<int>(in_dims[1] / 4);
@@ -161,11 +161,11 @@ class LSTMKernel : public framework::OpKernel<T> {
     math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
     batch_hidden.set_lod(batch_gate->lod());
     // restore the output hidden in LoDTensor from the batch hidden
-    to_seq(device_ctx, batch_hidden, *hidden_out);
+    to_seq(device_ctx, batch_hidden, hidden_out);
 
     batch_cell.set_lod(batch_gate->lod());
     // restore the output cell state in LoDTensor from the batch cell
-    to_seq(device_ctx, batch_cell, *cell_out);
+    to_seq(device_ctx, batch_cell, cell_out);
   }
 };
 
@@ -257,7 +257,7 @@ class LSTMGradKernel : public framework::OpKernel<T> {
         const framework::DDim& dims, framework::LoDTensor& dst) {
       dst.mutable_data<T>(dims, ctx.GetPlace());
       dst.set_lod(batch_gate->lod());
-      to_batch(ctx, src, dst, false);
+      to_batch(ctx, src, &dst, false);
     };
 
     LoDTensor batch_hidden, batch_hidden_g, batch_cell;
@@ -351,7 +351,7 @@ class LSTMGradKernel : public framework::OpKernel<T> {
     if (in_g) {
       /* backward data */
       in_g->mutable_data<T>(ctx.GetPlace());
-      to_seq(device_ctx, batch_gate_g, *in_g);
+      to_seq(device_ctx, batch_gate_g, in_g);
     }
     if (bias && bias_g) {
       /* backward bias */

paddle/fluid/operators/lstmp_op.h

@@ -40,7 +40,7 @@ inline void ReorderInitState(const DeviceContext& ctx,
                              framework::Tensor* dst, bool indexed_src) {
   math::CopyMatrixRowsFunctor<DeviceContext, T> row_shuffle;
   dst->mutable_data<T>(src.dims(), ctx.GetPlace());
-  row_shuffle(ctx, src, index, *dst, indexed_src);
+  row_shuffle(ctx, src, index, dst, indexed_src);
 }
 
 template <typename DeviceContext, typename T>
@@ -81,7 +81,7 @@ class LSTMPKernel : public framework::OpKernel<T> {
     bool is_reverse = ctx.Attr<bool>("is_reverse");
     math::LoDTensor2BatchFunctor<DeviceContext, T> to_batch;
     auto& device_ctx = ctx.template device_context<DeviceContext>();
-    to_batch(device_ctx, *input, *batch_gate, true, is_reverse);
+    to_batch(device_ctx, *input, batch_gate, true, is_reverse);
 
     auto in_dims = input->dims();
     int frame_size = static_cast<int>(in_dims[1] / 4);
@@ -208,11 +208,11 @@ class LSTMPKernel : public framework::OpKernel<T> {
     math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
     batch_proj.set_lod(batch_gate->lod());
     // restore the output hidden in LoDTensor from the batch hidden
-    to_seq(device_ctx, batch_proj, *proj_out);
+    to_seq(device_ctx, batch_proj, proj_out);
 
     batch_cell.set_lod(batch_gate->lod());
     // restore the output cell state in LoDTensor from the batch cell
-    to_seq(device_ctx, batch_cell, *cell_out);
+    to_seq(device_ctx, batch_cell, cell_out);
   }
 };
 
@@ -332,7 +332,7 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
         const framework::DDim& dims, framework::LoDTensor& dst) {
       dst.mutable_data<T>(dims, ctx.GetPlace());
       dst.set_lod(batch_gate->lod());
-      to_batch(ctx, src, dst, false);
+      to_batch(ctx, src, &dst, false);
     };
 
     LoDTensor batch_hidden_g, batch_proj, batch_proj_g, batch_cell;
@@ -471,7 +471,7 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
     if (in_g) {
       /* backward data */
       in_g->mutable_data<T>(ctx.GetPlace());
-      to_seq(device_ctx, batch_gate_g, *in_g);
+      to_seq(device_ctx, batch_gate_g, in_g);
     }
     if (bias && bias_g) {
       /* backward bias */

paddle/fluid/operators/math/concat.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/math/concat.h"
+#include <vector>
 
 namespace paddle {
 namespace operators {
@@ -70,20 +71,20 @@ class ConcatGradFunctor<platform::CPUDeviceContext, T> {
  public:
   void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input, const int axis,
-                  std::vector<framework::Tensor>& outputs) {
+                  std::vector<framework::Tensor>* outputs) {
     // TODO(zcd): Add input data validity checking
-    int num = outputs.size();
+    int num = outputs->size();
 
     int input_rows = 1;
-    auto dim_0 = outputs[0].dims();
+    auto dim_0 = outputs->at(0).dims();
     for (int i = 0; i < axis; ++i) {
       input_rows *= dim_0[i];
     }
     int input_cols = 0;
 
-    std::vector<int64_t> output_cols(outputs.size());
+    std::vector<int64_t> output_cols(outputs->size());
     for (int i = 0; i < num; ++i) {
-      int t_cols = outputs[i].numel() / input_rows;
+      int t_cols = outputs->at(i).numel() / input_rows;
       input_cols += t_cols;
       output_cols[i] = t_cols;
     }
@@ -95,7 +96,7 @@ class ConcatGradFunctor<platform::CPUDeviceContext, T> {
       int col_idx = 0;
       for (int j = 0; j < num; ++j) {
         int col_len = output_cols[j];
-        T* dst_ptr = outputs[j].data<T>() + k * col_len;
+        T* dst_ptr = outputs->at(j).data<T>() + k * col_len;
         memory::Copy(cpu_place, dst_ptr, cpu_place, src_ptr + col_idx,
                      sizeof(T) * col_len);
         col_idx += col_len;

paddle/fluid/operators/math/concat.cu

@@ -12,9 +12,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
+#include <algorithm>
+#include <vector>
 #include "paddle/fluid/framework/mixed_vector.h"
 #include "paddle/fluid/operators/math/concat.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {
 namespace operators {
@@ -202,16 +204,16 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
  public:
   void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input, const int axis,
-                  std::vector<framework::Tensor>& outputs) {
+                  std::vector<framework::Tensor>* outputs) {
     // TODO(zcd): Add input data validity checking
-    int o_num = outputs.size();
+    int o_num = outputs->size();
     int out_row = 1;
-    auto dim_0 = outputs[0].dims();
+    auto dim_0 = outputs->at(0).dims();
     for (int i = 0; i < axis; ++i) {
       out_row *= dim_0[i];
     }
 
-    int out_col = outputs[0].numel() / out_row;
+    int out_col = outputs->at(0).numel() / out_row;
     int in_col = 0, in_row = out_row;
     bool sameShape = true;
 
@@ -221,13 +223,13 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
 
     outputs_cols[0] = 0;
     for (int i = 0; i < o_num; ++i) {
-      int t_col = outputs[i].numel() / out_row;
+      int t_col = outputs->at(i).numel() / out_row;
       if (sameShape) {
         if (t_col != out_col) sameShape = false;
       }
       in_col += t_col;
       outputs_cols[i + 1] = in_col;
-      outputs_ptr[i] = outputs[i].data<T>();
+      outputs_ptr[i] = outputs->at(i).data<T>();
     }
 
     T** dev_out_gpu_data =

paddle/fluid/operators/math/concat.h

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <vector>
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/framework/tensor.h"
 
@@ -56,7 +57,7 @@ template <typename DeviceContext, typename T>
 class ConcatGradFunctor {
  public:
   void operator()(const DeviceContext& context, const framework::Tensor& input,
-                  const int axis, std::vector<framework::Tensor>& outputs);
+                  const int axis, std::vector<framework::Tensor>* outputs);
 };
 
 }  // namespace math

paddle/fluid/operators/math/concat_test.cc

@@ -17,17 +17,14 @@ limitations under the License. */
 #include <vector>
 #include "paddle/fluid/framework/tensor_util.h"
 
-using namespace paddle::framework;
-using namespace paddle::platform;
-
 template <typename DeviceContext, typename Place>
 void testConcat() {
-  Tensor input_a_cpu;
-  Tensor input_b_cpu;
-  Tensor out_cpu;
-  Tensor input_a;
-  Tensor input_b;
-  Tensor out;
+  paddle::framework::Tensor input_a_cpu;
+  paddle::framework::Tensor input_b_cpu;
+  paddle::framework::Tensor out_cpu;
+  paddle::framework::Tensor input_a;
+  paddle::framework::Tensor input_b;
+  paddle::framework::Tensor out;
 
   DeviceContext* context = new DeviceContext(Place());
   //  DeviceContext context(Place());
@@ -40,18 +37,18 @@ void testConcat() {
    *    output:
    *        out.shape: [5, 3, 4]
    */
-  auto dim_a = make_ddim({2, 3, 4});
-  auto dim_b = make_ddim({3, 3, 4});
-  auto dim_out = make_ddim({5, 3, 4});
+  auto dim_a = paddle::framework::make_ddim({2, 3, 4});
+  auto dim_b = paddle::framework::make_ddim({3, 3, 4});
+  auto dim_out = paddle::framework::make_ddim({5, 3, 4});
 
   input_a.mutable_data<int>(dim_a, Place());
   input_b.mutable_data<int>(dim_b, Place());
   out.mutable_data<int>(dim_out, Place());
 
   if (paddle::platform::is_gpu_place(Place())) {
-    input_a_cpu.mutable_data<int>(dim_a, CPUPlace());
-    input_b_cpu.mutable_data<int>(dim_b, CPUPlace());
-    out_cpu.mutable_data<int>(dim_out, CPUPlace());
+    input_a_cpu.mutable_data<int>(dim_a, paddle::platform::CPUPlace());
+    input_b_cpu.mutable_data<int>(dim_b, paddle::platform::CPUPlace());
+    out_cpu.mutable_data<int>(dim_out, paddle::platform::CPUPlace());
   }
 
   int* a_ptr;
@@ -72,11 +69,11 @@ void testConcat() {
   }
 
   if (paddle::platform::is_gpu_place(Place())) {
-    TensorCopy(input_a_cpu, Place(), *context, &input_a);
-    TensorCopy(input_b_cpu, Place(), *context, &input_b);
+    paddle::framework::TensorCopy(input_a_cpu, Place(), *context, &input_a);
+    paddle::framework::TensorCopy(input_b_cpu, Place(), *context, &input_b);
   }
 
-  std::vector<Tensor> input;
+  std::vector<paddle::framework::Tensor> input;
   input.push_back(input_a);
   input.push_back(input_b);
 
@@ -89,7 +86,8 @@ void testConcat() {
 
   int* out_ptr;
   if (paddle::platform::is_gpu_place(Place())) {
-    TensorCopy(out, CPUPlace(), *context, &out_cpu);
+    paddle::framework::TensorCopy(out, paddle::platform::CPUPlace(), *context,
+                                  &out_cpu);
     out_ptr = out_cpu.data<int>();
   } else {
     out_ptr = out.data<int>();
@@ -115,9 +113,9 @@ void testConcat() {
     *    output:
     *        out.shape: [2, 7, 4]
     */
-  dim_a = make_ddim({2, 3, 4});
-  dim_b = make_ddim({2, 4, 4});
-  dim_out = make_ddim({2, 7, 4});
+  dim_a = paddle::framework::make_ddim({2, 3, 4});
+  dim_b = paddle::framework::make_ddim({2, 4, 4});
+  dim_out = paddle::framework::make_ddim({2, 7, 4});
 
   input_a.Resize(dim_a);
   input_b.Resize(dim_b);
@@ -144,8 +142,8 @@ void testConcat() {
   }
 
   if (paddle::platform::is_gpu_place(Place())) {
-    TensorCopy(input_a_cpu, Place(), *context, &input_a);
-    TensorCopy(input_b_cpu, Place(), *context, &input_b);
+    paddle::framework::TensorCopy(input_a_cpu, Place(), *context, &input_a);
+    paddle::framework::TensorCopy(input_b_cpu, Place(), *context, &input_b);
   }
 
   input.clear();
@@ -159,7 +157,8 @@ void testConcat() {
   PADDLE_ENFORCE_EQ(input_b.dims(), dim_b);
 
   if (paddle::platform::is_gpu_place(Place())) {
-    TensorCopy(out, CPUPlace(), *context, &out_cpu);
+    paddle::framework::TensorCopy(out, paddle::platform::CPUPlace(), *context,
+                                  &out_cpu);
     out_ptr = out_cpu.data<int>();
   } else {
     out_ptr = out.data<int>();
@@ -187,9 +186,9 @@ void testConcat() {
     *    output:
     *        out.shape: [2, 3, 9]
     */
-  dim_a = make_ddim({2, 3, 4});
-  dim_b = make_ddim({2, 3, 5});
-  dim_out = make_ddim({2, 3, 9});
+  dim_a = paddle::framework::make_ddim({2, 3, 4});
+  dim_b = paddle::framework::make_ddim({2, 3, 5});
+  dim_out = paddle::framework::make_ddim({2, 3, 9});
 
   input_a.Resize(dim_a);
   input_b.Resize(dim_b);
@@ -216,8 +215,8 @@ void testConcat() {
   }
 
   if (paddle::platform::is_gpu_place(Place())) {
-    TensorCopy(input_a_cpu, Place(), *context, &input_a);
-    TensorCopy(input_b_cpu, Place(), *context, &input_b);
+    paddle::framework::TensorCopy(input_a_cpu, Place(), *context, &input_a);
+    paddle::framework::TensorCopy(input_b_cpu, Place(), *context, &input_b);
   }
 
   input.clear();
@@ -231,7 +230,8 @@ void testConcat() {
   PADDLE_ENFORCE_EQ(input_b.dims(), dim_b);
 
   if (paddle::platform::is_gpu_place(Place())) {
-    TensorCopy(out, CPUPlace(), *context, &out_cpu);
+    paddle::framework::TensorCopy(out, paddle::platform::CPUPlace(), *context,
+                                  &out_cpu);
     out_ptr = out_cpu.data<int>();
   } else {
     out_ptr = out.data<int>();
@@ -261,9 +261,9 @@ void testConcat() {
     *    output:
     *        out.shape: [2, 6, 4]
     */
-  dim_a = make_ddim({2, 3, 4});
-  dim_b = make_ddim({2, 3, 4});
-  dim_out = make_ddim({2, 6, 4});
+  dim_a = paddle::framework::make_ddim({2, 3, 4});
+  dim_b = paddle::framework::make_ddim({2, 3, 4});
+  dim_out = paddle::framework::make_ddim({2, 6, 4});
 
   input_a.Resize(dim_a);
   input_b.Resize(dim_b);
@@ -290,8 +290,8 @@ void testConcat() {
   }
 
   if (paddle::platform::is_gpu_place(Place())) {
-    TensorCopy(input_a_cpu, Place(), *context, &input_a);
-    TensorCopy(input_b_cpu, Place(), *context, &input_b);
+    paddle::framework::TensorCopy(input_a_cpu, Place(), *context, &input_a);
+    paddle::framework::TensorCopy(input_b_cpu, Place(), *context, &input_b);
   }
 
   input.clear();
@@ -305,7 +305,8 @@ void testConcat() {
   PADDLE_ENFORCE_EQ(input_b.dims(), dim_b);
 
   if (paddle::platform::is_gpu_place(Place())) {
-    TensorCopy(out, CPUPlace(), *context, &out_cpu);
+    paddle::framework::TensorCopy(out, paddle::platform::CPUPlace(), *context,
+                                  &out_cpu);
     out_ptr = out_cpu.data<int>();
   } else {
     out_ptr = out.data<int>();

paddle/fluid/operators/math/context_project.h

@@ -14,6 +14,8 @@ limitations under the License. */
 
 #pragma once
 
+#include <algorithm>
+#include <vector>
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/operators/math/im2col.h"
 #include "paddle/fluid/operators/math/math_function.h"

paddle/fluid/operators/math/cos_sim_functor.cu

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/math/cos_sim_functor.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/math/cross_entropy.cu

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/math/cross_entropy.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {
 namespace operators {
@@ -31,11 +32,11 @@ __global__ void CrossEntropyKernel(T* Y, const T* X, const int64_t* label,
 
 template <typename T>
 __device__ __forceinline__ T sum_single_warp(T val) {
-  val += __shfl_down(val, 16);
-  val += __shfl_down(val, 8);
-  val += __shfl_down(val, 4);
-  val += __shfl_down(val, 2);
-  val += __shfl_down(val, 1);
+  val += platform::__shfl_down_sync(0, val, 16);
+  val += platform::__shfl_down_sync(0, val, 8);
+  val += platform::__shfl_down_sync(0, val, 4);
+  val += platform::__shfl_down_sync(0, val, 2);
+  val += platform::__shfl_down_sync(0, val, 1);
   return val;
 }
 
@@ -108,7 +109,9 @@ class CrossEntropyFunctor<platform::CUDADeviceContext, T> {
 
     if (softLabel) {
       const T* label_data = labels->data<T>();
-      int block = class_num > 512 ? 512 : pow(2, int(std::log2(class_num)));
+      int block = class_num > 512
+                      ? 512
+                      : pow(2, static_cast<int>(std::log2(class_num)));
 
       SoftCrossEntropyKernel<T><<<
           batch_size, block, block * sizeof(T),

paddle/fluid/operators/math/depthwise_conv.cu

@@ -12,8 +12,9 @@ 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 <vector>
 #include "paddle/fluid/operators/math/depthwise_conv.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/math/depthwise_conv.h

@@ -13,6 +13,7 @@ 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/platform/device_context.h"
 #include "paddle/fluid/platform/hostdevice.h"

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

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 #include <math.h>
+#include <string>
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/hostdevice.h"
 

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

@@ -89,14 +89,14 @@ void hl_avx_gru_forward_reset_output(OpResetOutput op_reset_output,
   __m256 r_value_reset_gate;
   __m256 r_value_reset_output;
   __m256 r_prev_out = _mm256_set1_ps(0.0f);
-  __m256 *update_gate = (__m256 *)gate_value;
-  __m256 *reset_gate = (__m256 *)(gate_value + frame_size);
+  __m256 *update_gate = reinterpret_cast<__m256 *>(gate_value);
+  __m256 *reset_gate = reinterpret_cast<__m256 *>(gate_value + frame_size);
 
   for (int i = 0; i < frame_size / 8; i++) {
     r_value_update_gate = update_gate[i];
     r_value_reset_gate = reset_gate[i];
     if (prev_output_value) {
-      r_prev_out = ((__m256 *)prev_output_value)[i];
+      r_prev_out = (reinterpret_cast<__m256 *>(prev_output_value))[i];
     }
 
     op_reset_output(r_value_update_gate, r_value_reset_gate, r_prev_out,
@@ -104,7 +104,7 @@ void hl_avx_gru_forward_reset_output(OpResetOutput op_reset_output,
 
     update_gate[i] = r_value_update_gate;
     reset_gate[i] = r_value_reset_gate;
-    ((__m256 *)reset_output_value)[i] = r_value_reset_output;
+    (reinterpret_cast<__m256 *>(reset_output_value))[i] = r_value_reset_output;
   }
 #endif
 }
@@ -119,21 +119,21 @@ void hl_avx_gru_forward_final_output(OpFinalOutput op_final_output,
   __m256 r_value_frame_state;
   __m256 r_prev_out = _mm256_set1_ps(0.0f);
   __m256 r_output;
-  __m256 *update_gate = (__m256 *)gate_value;
-  __m256 *frame_state = (__m256 *)(gate_value + frame_size * 2);
+  __m256 *update_gate = reinterpret_cast<__m256 *>(gate_value);
+  __m256 *frame_state = reinterpret_cast<__m256 *>(gate_value + frame_size * 2);
 
   for (int i = 0; i < frame_size / 8; i++) {
     r_value_update_gate = update_gate[i];
     r_value_frame_state = frame_state[i];
     if (prev_output_value) {
-      r_prev_out = ((__m256 *)prev_output_value)[i];
+      r_prev_out = (reinterpret_cast<__m256 *>(prev_output_value))[i];
     }
 
     op_final_output(r_value_update_gate, r_value_frame_state, r_prev_out,
                     r_output, active_node);
 
     frame_state[i] = r_value_frame_state;
-    ((__m256 *)output_value)[i] = r_output;
+    (reinterpret_cast<__m256 *>(output_value))[i] = r_output;
   }
 #endif
 }
@@ -284,20 +284,22 @@ void hl_avx_gru_backward_state_grad(OpStateGrad op_state_grad, T *gate_value,
   __m256 r_out_grad;
   __m256 r_prev_out_value = _mm256_set1_ps(0.0f);
   __m256 r_prev_out_grad = _mm256_set1_ps(0.0f);
-  __m256 *update_gate_value = (__m256 *)gate_value;
-  __m256 *update_gate_grad = (__m256 *)gate_grad;
-  __m256 *frame_state_value = (__m256 *)(gate_value + frame_size * 2);
-  __m256 *frame_state_grad = (__m256 *)(gate_grad + frame_size * 2);
+  __m256 *update_gate_value = reinterpret_cast<__m256 *>(gate_value);
+  __m256 *update_gate_grad = reinterpret_cast<__m256 *>(gate_grad);
+  __m256 *frame_state_value =
+      reinterpret_cast<__m256 *>(gate_value + frame_size * 2);
+  __m256 *frame_state_grad =
+      reinterpret_cast<__m256 *>(gate_grad + frame_size * 2);
 
   for (int i = 0; i < frame_size / 8; i++) {
     r_update_gate_value = update_gate_value[i];
     r_frame_state_value = frame_state_value[i];
-    r_out_grad = ((__m256 *)output_grad)[i];
+    r_out_grad = (reinterpret_cast<__m256 *>(output_grad))[i];
     if (prev_out_value) {
-      r_prev_out_value = ((__m256 *)prev_out_value)[i];
+      r_prev_out_value = (reinterpret_cast<__m256 *>(prev_out_value))[i];
     }
     if (prev_out_grad) {
-      r_prev_out_grad = ((__m256 *)prev_out_grad)[i];
+      r_prev_out_grad = (reinterpret_cast<__m256 *>(prev_out_grad))[i];
     }
 
     op_state_grad(r_update_gate_value, r_update_gate_grad, r_frame_state_value,
@@ -307,7 +309,7 @@ void hl_avx_gru_backward_state_grad(OpStateGrad op_state_grad, T *gate_value,
     update_gate_grad[i] = r_update_gate_grad;
     frame_state_grad[i] = r_frame_state_grad;
     if (prev_out_grad) {
-      ((__m256 *)prev_out_grad)[i] = r_prev_out_grad;
+      (reinterpret_cast<__m256 *>(prev_out_grad))[i] = r_prev_out_grad;
     }
   }
 #endif
@@ -327,10 +329,11 @@ void hl_avx_gru_backward_reset_grad(OpResetGrad op_reset_grad, T *gate_value,
   __m256 r_reset_output_grad = _mm256_set1_ps(0.0f);
   __m256 r_prev_out_value = _mm256_set1_ps(0.0f);
   __m256 r_prev_out_grad = _mm256_set1_ps(0.0f);
-  __m256 *update_gate_value = (__m256 *)gate_value;
-  __m256 *update_gate_grad = (__m256 *)gate_grad;
-  __m256 *reset_gate_value = (__m256 *)(gate_value + frame_size);
-  __m256 *reset_gate_grad = (__m256 *)(gate_grad + frame_size);
+  __m256 *update_gate_value = reinterpret_cast<__m256 *>(gate_value);
+  __m256 *update_gate_grad = reinterpret_cast<__m256 *>(gate_grad);
+  __m256 *reset_gate_value =
+      reinterpret_cast<__m256 *>(gate_value + frame_size);
+  __m256 *reset_gate_grad = reinterpret_cast<__m256 *>(gate_grad + frame_size);
 
   for (int i = 0; i < frame_size / 8; i++) {
     r_update_gate_value = update_gate_value[i];
@@ -338,13 +341,13 @@ void hl_avx_gru_backward_reset_grad(OpResetGrad op_reset_grad, T *gate_value,
     r_reset_gate_value = reset_gate_value[i];
 
     if (prev_out_value && prev_out_grad) {
-      r_reset_output_grad = ((__m256 *)reset_output_grad)[i];
+      r_reset_output_grad = (reinterpret_cast<__m256 *>(reset_output_grad))[i];
     }
     if (prev_out_value) {
-      r_prev_out_value = ((__m256 *)prev_out_value)[i];
+      r_prev_out_value = (reinterpret_cast<__m256 *>(prev_out_value))[i];
     }
     if (prev_out_grad) {
-      r_prev_out_grad = ((__m256 *)prev_out_grad)[i];
+      r_prev_out_grad = (reinterpret_cast<__m256 *>(prev_out_grad))[i];
     }
 
     op_reset_grad(r_update_gate_value, r_update_gate_grad, r_reset_gate_value,
@@ -354,7 +357,7 @@ void hl_avx_gru_backward_reset_grad(OpResetGrad op_reset_grad, T *gate_value,
     update_gate_grad[i] = r_update_gate_grad;
     reset_gate_grad[i] = r_reset_gate_grad;
     if (prev_out_grad) {
-      ((__m256 *)prev_out_grad)[i] = r_prev_out_grad;
+      (reinterpret_cast<__m256 *>(prev_out_grad))[i] = r_prev_out_grad;
     }
   }
 #endif

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

@@ -16,7 +16,7 @@ limitations under the License. */
 #include <type_traits>
 #include "paddle/fluid/operators/math/detail/activation_functions.h"
 #include "paddle/fluid/operators/math/gru_compute.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 #include "paddle/fluid/platform/device_context.h"
 
 namespace paddle {

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

@@ -164,10 +164,12 @@ void avx_lstm_forward_one_sequence(Op op, LstmMetaValue<T> value,
   __m256 r_state_atv;
   __m256 r_out;
 
-  __m256 *value_in = (__m256 *)value.gate_value;
-  __m256 *value_ig = (__m256 *)(value.gate_value + frame_size);
-  __m256 *value_fg = (__m256 *)(value.gate_value + frame_size * 2);
-  __m256 *value_og = (__m256 *)(value.gate_value + frame_size * 3);
+  __m256 *value_in = reinterpret_cast<__m256 *>(value.gate_value);
+  __m256 *value_ig = reinterpret_cast<__m256 *>(value.gate_value + frame_size);
+  __m256 *value_fg =
+      reinterpret_cast<__m256 *>(value.gate_value + frame_size * 2);
+  __m256 *value_og =
+      reinterpret_cast<__m256 *>(value.gate_value + frame_size * 3);
 
   for (int i = 0; i < frame_size / 8; i++) {
     r_value_in = value_in[i];
@@ -175,13 +177,13 @@ void avx_lstm_forward_one_sequence(Op op, LstmMetaValue<T> value,
     r_value_fg = value_fg[i];
     r_value_og = value_og[i];
     if (value.check_ig) {
-      r_checkI = ((__m256 *)value.check_ig)[i];
-      r_checkF = ((__m256 *)value.check_fg)[i];
-      r_checkO = ((__m256 *)value.check_og)[i];
+      r_checkI = (reinterpret_cast<__m256 *>(value.check_ig))[i];
+      r_checkF = (reinterpret_cast<__m256 *>(value.check_fg))[i];
+      r_checkO = (reinterpret_cast<__m256 *>(value.check_og))[i];
     }
 
     if (value.prev_state_value) {
-      r_prev_state = ((__m256 *)value.prev_state_value)[i];
+      r_prev_state = (reinterpret_cast<__m256 *>(value.prev_state_value))[i];
     }
 
     op(r_value_in, r_value_ig, r_value_fg, r_value_og, r_prev_state, r_state,
@@ -192,9 +194,9 @@ void avx_lstm_forward_one_sequence(Op op, LstmMetaValue<T> value,
     value_ig[i] = r_value_ig;
     value_fg[i] = r_value_fg;
     value_og[i] = r_value_og;
-    ((__m256 *)value.state_value)[i] = r_state;
-    ((__m256 *)value.state_active_value)[i] = r_state_atv;
-    ((__m256 *)value.output_value)[i] = r_out;
+    (reinterpret_cast<__m256 *>(value.state_value))[i] = r_state;
+    (reinterpret_cast<__m256 *>(value.state_active_value))[i] = r_state_atv;
+    (reinterpret_cast<__m256 *>(value.output_value))[i] = r_out;
   }
 #endif
 }
@@ -227,14 +229,16 @@ void avx_lstm_backward_one_sequence(Op op, LstmMetaValue<T> value,
   __m256 r_checkFGrad;
   __m256 r_checkOGrad;
 
-  __m256 *value_in = (__m256 *)value.gate_value;
-  __m256 *value_ig = (__m256 *)(value.gate_value + frame_size);
-  __m256 *value_fg = (__m256 *)(value.gate_value + frame_size * 2);
-  __m256 *value_og = (__m256 *)(value.gate_value + frame_size * 3);
-  __m256 *grad_in = (__m256 *)grad.gate_grad;
-  __m256 *grad_ig = (__m256 *)(grad.gate_grad + frame_size);
-  __m256 *grad_fg = (__m256 *)(grad.gate_grad + frame_size * 2);
-  __m256 *grad_og = (__m256 *)(grad.gate_grad + frame_size * 3);
+  __m256 *value_in = reinterpret_cast<__m256 *>(value.gate_value);
+  __m256 *value_ig = reinterpret_cast<__m256 *>(value.gate_value + frame_size);
+  __m256 *value_fg =
+      reinterpret_cast<__m256 *>(value.gate_value + frame_size * 2);
+  __m256 *value_og =
+      reinterpret_cast<__m256 *>(value.gate_value + frame_size * 3);
+  __m256 *grad_in = reinterpret_cast<__m256 *>(grad.gate_grad);
+  __m256 *grad_ig = reinterpret_cast<__m256 *>(grad.gate_grad + frame_size);
+  __m256 *grad_fg = reinterpret_cast<__m256 *>(grad.gate_grad + frame_size * 2);
+  __m256 *grad_og = reinterpret_cast<__m256 *>(grad.gate_grad + frame_size * 3);
 
   for (int i = 0; i < frame_size / 8; i++) {
     r_value_in = value_in[i];
@@ -242,16 +246,16 @@ void avx_lstm_backward_one_sequence(Op op, LstmMetaValue<T> value,
     r_value_fg = value_fg[i];
     r_value_og = value_og[i];
     if (value.check_ig) {
-      r_checkI = ((__m256 *)value.check_ig)[i];
-      r_checkF = ((__m256 *)value.check_fg)[i];
-      r_checkO = ((__m256 *)value.check_og)[i];
+      r_checkI = (reinterpret_cast<__m256 *>(value.check_ig))[i];
+      r_checkF = (reinterpret_cast<__m256 *>(value.check_fg))[i];
+      r_checkO = (reinterpret_cast<__m256 *>(value.check_og))[i];
     }
-    r_state = ((__m256 *)value.state_value)[i];
-    r_state_atv = ((__m256 *)value.state_active_value)[i];
-    r_output_grad = ((__m256 *)grad.output_grad)[i];
-    r_state_grad = ((__m256 *)grad.state_grad)[i];
+    r_state = (reinterpret_cast<__m256 *>(value.state_value))[i];
+    r_state_atv = (reinterpret_cast<__m256 *>(value.state_active_value))[i];
+    r_output_grad = (reinterpret_cast<__m256 *>(grad.output_grad))[i];
+    r_state_grad = (reinterpret_cast<__m256 *>(grad.state_grad))[i];
     if (value.prev_state_value) {
-      r_prev_state = ((__m256 *)value.prev_state_value)[i];
+      r_prev_state = (reinterpret_cast<__m256 *>(value.prev_state_value))[i];
     }
 
     op(r_value_in, r_value_ig, r_value_fg, r_value_og, r_grad_in, r_grad_ig,
@@ -264,15 +268,18 @@ void avx_lstm_backward_one_sequence(Op op, LstmMetaValue<T> value,
     grad_ig[i] = r_grad_ig;
     grad_fg[i] = r_grad_fg;
     grad_og[i] = r_grad_og;
-    ((__m256 *)grad.state_grad)[i] = r_state_grad;
+    (reinterpret_cast<__m256 *>(grad.state_grad))[i] = r_state_grad;
 
     if (grad.prev_state_grad)
-      ((__m256 *)grad.prev_state_grad)[i] = r_prev_state_grad;
+      (reinterpret_cast<__m256 *>(grad.prev_state_grad))[i] = r_prev_state_grad;
     if (value.prev_state_value) {
-      if (grad.check_ig_grad) ((__m256 *)grad.check_ig_grad)[i] += r_checkIGrad;
-      if (grad.check_fg_grad) ((__m256 *)grad.check_fg_grad)[i] += r_checkFGrad;
+      if (grad.check_ig_grad)
+        (reinterpret_cast<__m256 *>(grad.check_ig_grad))[i] += r_checkIGrad;
+      if (grad.check_fg_grad)
+        (reinterpret_cast<__m256 *>(grad.check_fg_grad))[i] += r_checkFGrad;
     }
-    if (grad.check_og_grad) ((__m256 *)grad.check_og_grad)[i] += r_checkOGrad;
+    if (grad.check_og_grad)
+      (reinterpret_cast<__m256 *>(grad.check_og_grad))[i] += r_checkOGrad;
   }
 #endif
 }

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

@@ -13,13 +13,13 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <type_traits>
+
 #include "paddle/fluid/operators/math/detail/activation_functions.h"
 #include "paddle/fluid/operators/math/lstm_compute.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 #include "paddle/fluid/platform/device_context.h"
 
-#include <type_traits>
-
 namespace paddle {
 namespace operators {
 namespace math {

paddle/fluid/operators/math/im2col.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/math/im2col.h"
+#include <vector>
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/math/im2col.cu

@@ -12,8 +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 <algorithm>
+#include <vector>
 #include "paddle/fluid/operators/math/im2col.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/math/im2col.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <vector>
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/framework/tensor_util.h"
 #include "paddle/fluid/platform/device_context.h"

paddle/fluid/operators/math/im2col_test.cc

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/math/im2col.h"
 #include <gtest/gtest.h>
+#include <vector>
 
 template <typename DeviceContext, typename Place>
 void testIm2col() {
@@ -62,7 +63,7 @@ void testIm2col() {
   if (paddle::platform::is_cpu_place(*place)) {
     input = input_tmp;
   } else {
-    TensorCopy(input_tmp, *place, *context, &input);
+    TensorCopySync(input_tmp, *place, &input);
   }
   output_cfo.mutable_data<float>(
       {1, filter_size, filter_size, output_height, output_width}, *place);
@@ -87,7 +88,7 @@ void testIm2col() {
   if (paddle::platform::is_cpu_place(*place)) {
     out_cfo_ptr = output_cfo.data<float>();
   } else {
-    TensorCopy(output_cfo, paddle::platform::CPUPlace(), *context, &output_tmp);
+    TensorCopySync(output_cfo, paddle::platform::CPUPlace(), &output_tmp);
     out_cfo_ptr = output_tmp.data<float>();
   }
   for (int i = 0; i < 6; ++i) {
@@ -98,7 +99,7 @@ void testIm2col() {
   if (paddle::platform::is_cpu_place(*place)) {
     out_ocf_ptr = output_ocf.data<float>();
   } else {
-    TensorCopy(output_ocf, paddle::platform::CPUPlace(), *context, &output_tmp);
+    TensorCopySync(output_ocf, paddle::platform::CPUPlace(), &output_tmp);
     out_ocf_ptr = output_tmp.data<float>();
   }
 
@@ -119,7 +120,7 @@ void testIm2col() {
   if (paddle::platform::is_cpu_place(*place)) {
     input = input_tmp;
   } else {
-    TensorCopy(input_tmp, *place, *context, &input);
+    TensorCopySync(input_tmp, *place, &input);
   }
 
   col2im(*context, output_cfo, dilation, stride, padding, &input);
@@ -128,7 +129,7 @@ void testIm2col() {
   if (paddle::platform::is_cpu_place(*place)) {
     in_ptr = input.data<float>();
   } else {
-    TensorCopy(input, paddle::platform::CPUPlace(), *context, &input_tmp);
+    TensorCopySync(input, paddle::platform::CPUPlace(), &input_tmp);
     in_ptr = input_tmp.data<float>();
   }
   for (int i = 0; i < 6; ++i) {
@@ -140,7 +141,7 @@ void testIm2col() {
   if (paddle::platform::is_cpu_place(*place)) {
     input = input_tmp;
   } else {
-    TensorCopy(input_tmp, *place, *context, &input);
+    TensorCopySync(input_tmp, *place, &input);
   }
 
   col2im_ocf(*context, output_ocf, dilation, stride, padding, &input);
@@ -148,7 +149,7 @@ void testIm2col() {
   if (paddle::platform::is_cpu_place(*place)) {
     in_ptr = input.data<float>();
   } else {
-    TensorCopy(input, paddle::platform::CPUPlace(), *context, &input_tmp);
+    TensorCopySync(input, paddle::platform::CPUPlace(), &input_tmp);
     in_ptr = input_tmp.data<float>();
   }
   for (int i = 0; i < 6; ++i) {

paddle/fluid/operators/math/math_function_impl.h

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <vector>
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/operators/math/math_function.h"
 

paddle/fluid/operators/math/math_function_test.cu

@@ -23,32 +23,29 @@ void fill_fp16_data(paddle::platform::float16* in_ptr, size_t size,
 }
 
 TEST(math_function, notrans_mul_trans_fp32) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
+  paddle::framework::Tensor input1;
+  paddle::framework::Tensor input1_gpu;
+  paddle::framework::Tensor input2_gpu;
+  paddle::framework::Tensor out_gpu;
+  paddle::framework::Tensor out;
 
-  Tensor input1;
-  Tensor input1_gpu;
-  Tensor input2_gpu;
-  Tensor out_gpu;
-  Tensor out;
-
-  CPUPlace cpu_place;
-  CUDAPlace gpu_place(0);
-  CUDADeviceContext context(gpu_place);
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CUDAPlace gpu_place(0);
+  paddle::platform::CUDADeviceContext context(gpu_place);
 
   float* input1_ptr = input1.mutable_data<float>({2, 3}, cpu_place);
   float arr[6] = {0, 1, 2, 3, 4, 5};
   memcpy(input1_ptr, arr, 6 * sizeof(float));
 
-  TensorCopy(input1, gpu_place, context, &input1_gpu);
-  TensorCopy(input1, gpu_place, context, &input2_gpu);
+  paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu);
+  paddle::framework::TensorCopySync(input1, gpu_place, &input2_gpu);
 
   out_gpu.mutable_data<float>({2, 2}, gpu_place);
 
-  paddle::operators::math::matmul<CUDADeviceContext, float>(
+  paddle::operators::math::matmul<paddle::platform::CUDADeviceContext, float>(
       context, input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0);
 
-  TensorCopy(out_gpu, cpu_place, context, &out);
+  paddle::framework::TensorCopySync(out_gpu, cpu_place, &out);
 
   float* out_ptr = out.data<float>();
   context.Wait();
@@ -59,39 +56,38 @@ TEST(math_function, notrans_mul_trans_fp32) {
 }
 
 TEST(math_function, notrans_mul_trans_fp16) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-
-  Tensor input1;
-  Tensor input1_gpu;
-  Tensor input2_gpu;
-  Tensor out_gpu;
-  Tensor out;
+  paddle::framework::Tensor input1;
+  paddle::framework::Tensor input1_gpu;
+  paddle::framework::Tensor input2_gpu;
+  paddle::framework::Tensor out_gpu;
+  paddle::framework::Tensor out;
 
-  CPUPlace cpu_place;
-  CUDAPlace gpu_place(0);
-  CUDADeviceContext context(gpu_place);
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CUDAPlace gpu_place(0);
+  paddle::platform::CUDADeviceContext context(gpu_place);
 
   // fp16 GEMM in cublas requires GPU compute capability >= 53
   if (context.GetComputeCapability() < 53) {
     return;
   }
 
-  float16* input1_ptr = input1.mutable_data<float16>({2, 3}, cpu_place);
+  paddle::platform::float16* input1_ptr =
+      input1.mutable_data<paddle::platform::float16>({2, 3}, cpu_place);
   fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5});
 
-  TensorCopy(input1, gpu_place, context, &input1_gpu);
-  TensorCopy(input1, gpu_place, context, &input2_gpu);
+  paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu);
+  paddle::framework::TensorCopySync(input1, gpu_place, &input2_gpu);
 
-  out_gpu.mutable_data<float16>({2, 2}, gpu_place);
+  out_gpu.mutable_data<paddle::platform::float16>({2, 2}, gpu_place);
 
-  paddle::operators::math::matmul<CUDADeviceContext, float16>(
-      context, input1_gpu, false, input2_gpu, true, float16(1), &out_gpu,
-      float16(0));
+  paddle::operators::math::matmul<paddle::platform::CUDADeviceContext,
+                                  paddle::platform::float16>(
+      context, input1_gpu, false, input2_gpu, true,
+      paddle::platform::float16(1), &out_gpu, paddle::platform::float16(0));
 
-  TensorCopy(out_gpu, cpu_place, context, &out);
+  paddle::framework::TensorCopySync(out_gpu, cpu_place, &out);
 
-  float16* out_ptr = out.data<float16>();
+  paddle::platform::float16* out_ptr = out.data<paddle::platform::float16>();
   context.Wait();
   EXPECT_EQ(static_cast<float>(out_ptr[0]), 5);
   EXPECT_EQ(static_cast<float>(out_ptr[1]), 14);
@@ -100,32 +96,29 @@ TEST(math_function, notrans_mul_trans_fp16) {
 }
 
 TEST(math_function, trans_mul_notrans_fp32) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
+  paddle::framework::Tensor input1;
+  paddle::framework::Tensor input1_gpu;
+  paddle::framework::Tensor input2_gpu;
+  paddle::framework::Tensor out_gpu;
+  paddle::framework::Tensor out;
 
-  Tensor input1;
-  Tensor input1_gpu;
-  Tensor input2_gpu;
-  Tensor out_gpu;
-  Tensor out;
-
-  CPUPlace cpu_place;
-  CUDAPlace gpu_place(0);
-  CUDADeviceContext context(gpu_place);
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CUDAPlace gpu_place(0);
+  paddle::platform::CUDADeviceContext context(gpu_place);
 
   float* input1_ptr = input1.mutable_data<float>({2, 3}, cpu_place);
   float arr[6] = {0, 1, 2, 3, 4, 5};
   memcpy(input1_ptr, arr, 6 * sizeof(float));
 
-  TensorCopy(input1, gpu_place, context, &input1_gpu);
-  TensorCopy(input1, gpu_place, context, &input2_gpu);
+  paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu);
+  paddle::framework::TensorCopySync(input1, gpu_place, &input2_gpu);
 
   out_gpu.mutable_data<float>({3, 3}, gpu_place);
 
   paddle::operators::math::matmul<paddle::platform::CUDADeviceContext, float>(
       context, input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0);
 
-  TensorCopy(out_gpu, cpu_place, context, &out);
+  paddle::framework::TensorCopySync(out_gpu, cpu_place, &out);
 
   float* out_ptr = out.data<float>();
   context.Wait();
@@ -141,39 +134,38 @@ TEST(math_function, trans_mul_notrans_fp32) {
 }
 
 TEST(math_function, trans_mul_notrans_fp16) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-
-  Tensor input1;
-  Tensor input1_gpu;
-  Tensor input2_gpu;
-  Tensor out_gpu;
-  Tensor out;
+  paddle::framework::Tensor input1;
+  paddle::framework::Tensor input1_gpu;
+  paddle::framework::Tensor input2_gpu;
+  paddle::framework::Tensor out_gpu;
+  paddle::framework::Tensor out;
 
-  CPUPlace cpu_place;
-  CUDAPlace gpu_place(0);
-  CUDADeviceContext context(gpu_place);
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CUDAPlace gpu_place(0);
+  paddle::platform::CUDADeviceContext context(gpu_place);
 
   // fp16 GEMM in cublas requires GPU compute capability >= 53
   if (context.GetComputeCapability() < 53) {
     return;
   }
 
-  float16* input1_ptr = input1.mutable_data<float16>({2, 3}, cpu_place);
+  paddle::platform::float16* input1_ptr =
+      input1.mutable_data<paddle::platform::float16>({2, 3}, cpu_place);
   fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5});
 
-  TensorCopy(input1, gpu_place, context, &input1_gpu);
-  TensorCopy(input1, gpu_place, context, &input2_gpu);
+  paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu);
+  paddle::framework::TensorCopySync(input1, gpu_place, &input2_gpu);
 
-  out_gpu.mutable_data<float16>({3, 3}, gpu_place);
+  out_gpu.mutable_data<paddle::platform::float16>({3, 3}, gpu_place);
 
-  paddle::operators::math::matmul<paddle::platform::CUDADeviceContext, float16>(
-      context, input1_gpu, true, input2_gpu, false, float16(1), &out_gpu,
-      float16(0));
+  paddle::operators::math::matmul<paddle::platform::CUDADeviceContext,
+                                  paddle::platform::float16>(
+      context, input1_gpu, true, input2_gpu, false,
+      paddle::platform::float16(1), &out_gpu, paddle::platform::float16(0));
 
-  TensorCopy(out_gpu, cpu_place, context, &out);
+  paddle::framework::TensorCopySync(out_gpu, cpu_place, &out);
 
-  float16* out_ptr = out.data<float16>();
+  paddle::platform::float16* out_ptr = out.data<paddle::platform::float16>();
   context.Wait();
   EXPECT_EQ(static_cast<float>(out_ptr[0]), 9);
   EXPECT_EQ(static_cast<float>(out_ptr[1]), 12);
@@ -187,19 +179,16 @@ TEST(math_function, trans_mul_notrans_fp16) {
 }
 
 TEST(math_function, gemm_notrans_cublas_fp32) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
+  paddle::framework::Tensor input1;
+  paddle::framework::Tensor input2;
+  paddle::framework::Tensor input3;
+  paddle::framework::Tensor input1_gpu;
+  paddle::framework::Tensor input2_gpu;
+  paddle::framework::Tensor input3_gpu;
 
-  Tensor input1;
-  Tensor input2;
-  Tensor input3;
-  Tensor input1_gpu;
-  Tensor input2_gpu;
-  Tensor input3_gpu;
-
-  CPUPlace cpu_place;
-  CUDAPlace gpu_place(0);
-  CUDADeviceContext context(gpu_place);
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CUDAPlace gpu_place(0);
+  paddle::platform::CUDADeviceContext context(gpu_place);
 
   int m = 2;
   int n = 3;
@@ -214,9 +203,9 @@ TEST(math_function, gemm_notrans_cublas_fp32) {
   float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7};
   memcpy(input3_ptr, arr3, 8 * sizeof(float));
 
-  TensorCopy(input1, gpu_place, context, &input1_gpu);
-  TensorCopy(input2, gpu_place, context, &input2_gpu);
-  TensorCopy(input3, gpu_place, context, &input3_gpu);
+  paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu);
+  paddle::framework::TensorCopySync(input2, gpu_place, &input2_gpu);
+  paddle::framework::TensorCopySync(input3, gpu_place, &input3_gpu);
   float* a = input1_gpu.data<float>();
   float* b = input2_gpu.data<float>();
   float* c = input3_gpu.mutable_data<float>(gpu_place);
@@ -224,7 +213,7 @@ TEST(math_function, gemm_notrans_cublas_fp32) {
   paddle::operators::math::gemm<paddle::platform::CUDADeviceContext, float>(
       context, false, false, m, n, k, 1, a, 3, b + 1, 4, 1, c + 1, 4);
 
-  TensorCopy(input3_gpu, cpu_place, context, &input3);
+  paddle::framework::TensorCopySync(input3_gpu, cpu_place, &input3);
 
   // numpy code:
   // a = np.arange(6).reshape(2, 3)
@@ -244,19 +233,16 @@ TEST(math_function, gemm_notrans_cublas_fp32) {
 }
 
 TEST(math_function, gemm_notrans_cublas_fp16) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-
-  Tensor input1;
-  Tensor input2;
-  Tensor input3;
-  Tensor input1_gpu;
-  Tensor input2_gpu;
-  Tensor input3_gpu;
+  paddle::framework::Tensor input1;
+  paddle::framework::Tensor input2;
+  paddle::framework::Tensor input3;
+  paddle::framework::Tensor input1_gpu;
+  paddle::framework::Tensor input2_gpu;
+  paddle::framework::Tensor input3_gpu;
 
-  CPUPlace cpu_place;
-  CUDAPlace gpu_place(0);
-  CUDADeviceContext context(gpu_place);
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CUDAPlace gpu_place(0);
+  paddle::platform::CUDADeviceContext context(gpu_place);
 
   // fp16 GEMM in cublas requires GPU compute capability >= 53
   if (context.GetComputeCapability() < 53) {
@@ -266,26 +252,31 @@ TEST(math_function, gemm_notrans_cublas_fp16) {
   int m = 2;
   int n = 3;
   int k = 3;
-  float16* input1_ptr = input1.mutable_data<float16>({2, 3}, cpu_place);
+  paddle::platform::float16* input1_ptr =
+      input1.mutable_data<paddle::platform::float16>({2, 3}, cpu_place);
   fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5});
-  float16* input2_ptr = input2.mutable_data<float16>({3, 4}, cpu_place);
+  paddle::platform::float16* input2_ptr =
+      input2.mutable_data<paddle::platform::float16>({3, 4}, cpu_place);
   fill_fp16_data(input2_ptr, input2.numel(),
                  {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11});
-  float16* input3_ptr = input3.mutable_data<float16>({2, 4}, cpu_place);
+  paddle::platform::float16* input3_ptr =
+      input3.mutable_data<paddle::platform::float16>({2, 4}, cpu_place);
   fill_fp16_data(input3_ptr, input3.numel(), {0, 1, 2, 3, 4, 5, 6, 7});
 
-  TensorCopy(input1, gpu_place, context, &input1_gpu);
-  TensorCopy(input2, gpu_place, context, &input2_gpu);
-  TensorCopy(input3, gpu_place, context, &input3_gpu);
-  float16* a = input1_gpu.data<float16>();
-  float16* b = input2_gpu.data<float16>();
-  float16* c = input3_gpu.mutable_data<float16>(gpu_place);
+  paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu);
+  paddle::framework::TensorCopySync(input2, gpu_place, &input2_gpu);
+  paddle::framework::TensorCopySync(input3, gpu_place, &input3_gpu);
+  paddle::platform::float16* a = input1_gpu.data<paddle::platform::float16>();
+  paddle::platform::float16* b = input2_gpu.data<paddle::platform::float16>();
+  paddle::platform::float16* c =
+      input3_gpu.mutable_data<paddle::platform::float16>(gpu_place);
 
-  paddle::operators::math::gemm<paddle::platform::CUDADeviceContext, float16>(
-      context, false, false, m, n, k, float16(1), a, 3, b + 1, 4, float16(1),
-      c + 1, 4);
+  paddle::operators::math::gemm<paddle::platform::CUDADeviceContext,
+                                paddle::platform::float16>(
+      context, false, false, m, n, k, paddle::platform::float16(1), a, 3, b + 1,
+      4, paddle::platform::float16(1), c + 1, 4);
 
-  TensorCopy(input3_gpu, cpu_place, context, &input3);
+  paddle::framework::TensorCopySync(input3_gpu, cpu_place, &input3);
 
   // numpy code:
   // a = np.arange(6).reshape(2, 3)
@@ -305,19 +296,16 @@ TEST(math_function, gemm_notrans_cublas_fp16) {
 }
 
 TEST(math_function, gemm_trans_cublas_fp32) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-
-  Tensor input1;
-  Tensor input2;
-  Tensor input3;
-  Tensor input1_gpu;
-  Tensor input2_gpu;
-  Tensor input3_gpu;
+  paddle::framework::Tensor input1;
+  paddle::framework::Tensor input2;
+  paddle::framework::Tensor input3;
+  paddle::framework::Tensor input1_gpu;
+  paddle::framework::Tensor input2_gpu;
+  paddle::framework::Tensor input3_gpu;
 
-  CPUPlace cpu_place;
-  CUDAPlace gpu_place(0);
-  CUDADeviceContext context(gpu_place);
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CUDAPlace gpu_place(0);
+  paddle::platform::CUDADeviceContext context(gpu_place);
 
   int m = 2;
   int n = 3;
@@ -332,9 +320,9 @@ TEST(math_function, gemm_trans_cublas_fp32) {
   float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7};
   memcpy(input3_ptr, arr3, 8 * sizeof(float));
 
-  TensorCopy(input1, gpu_place, context, &input1_gpu);
-  TensorCopy(input2, gpu_place, context, &input2_gpu);
-  TensorCopy(input3, gpu_place, context, &input3_gpu);
+  paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu);
+  paddle::framework::TensorCopySync(input2, gpu_place, &input2_gpu);
+  paddle::framework::TensorCopySync(input3, gpu_place, &input3_gpu);
   float* a = input1_gpu.data<float>();
   float* b = input2_gpu.data<float>();
   float* c = input3_gpu.mutable_data<float>(gpu_place);
@@ -342,7 +330,7 @@ TEST(math_function, gemm_trans_cublas_fp32) {
   paddle::operators::math::gemm<paddle::platform::CUDADeviceContext, float>(
       context, false, true, m, n, k, 1, a, 3, b + 3, 3, 1, c + 1, 4);
 
-  TensorCopy(input3_gpu, cpu_place, context, &input3);
+  paddle::framework::TensorCopySync(input3_gpu, cpu_place, &input3);
 
   context.Wait();
   EXPECT_EQ(input3_ptr[0], 0);
@@ -356,19 +344,16 @@ TEST(math_function, gemm_trans_cublas_fp32) {
 }
 
 TEST(math_function, gemm_trans_cublas_fp16) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
+  paddle::framework::Tensor input1;
+  paddle::framework::Tensor input2;
+  paddle::framework::Tensor input3;
+  paddle::framework::Tensor input1_gpu;
+  paddle::framework::Tensor input2_gpu;
+  paddle::framework::Tensor input3_gpu;
 
-  Tensor input1;
-  Tensor input2;
-  Tensor input3;
-  Tensor input1_gpu;
-  Tensor input2_gpu;
-  Tensor input3_gpu;
-
-  CPUPlace cpu_place;
-  CUDAPlace gpu_place(0);
-  CUDADeviceContext context(gpu_place);
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CUDAPlace gpu_place(0);
+  paddle::platform::CUDADeviceContext context(gpu_place);
 
   // fp16 GEMM in cublas requires GPU compute capability >= 53
   if (context.GetComputeCapability() < 53) {
@@ -378,26 +363,31 @@ TEST(math_function, gemm_trans_cublas_fp16) {
   int m = 2;
   int n = 3;
   int k = 3;
-  float16* input1_ptr = input1.mutable_data<float16>({2, 3}, cpu_place);
+  paddle::platform::float16* input1_ptr =
+      input1.mutable_data<paddle::platform::float16>({2, 3}, cpu_place);
   fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5});
-  float16* input2_ptr = input2.mutable_data<float16>({4, 3}, cpu_place);
+  paddle::platform::float16* input2_ptr =
+      input2.mutable_data<paddle::platform::float16>({4, 3}, cpu_place);
   fill_fp16_data(input2_ptr, input2.numel(),
                  {0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11});
-  float16* input3_ptr = input3.mutable_data<float16>({2, 4}, cpu_place);
+  paddle::platform::float16* input3_ptr =
+      input3.mutable_data<paddle::platform::float16>({2, 4}, cpu_place);
   fill_fp16_data(input3_ptr, input3.numel(), {0, 1, 2, 3, 4, 5, 6, 7});
 
-  TensorCopy(input1, gpu_place, context, &input1_gpu);
-  TensorCopy(input2, gpu_place, context, &input2_gpu);
-  TensorCopy(input3, gpu_place, context, &input3_gpu);
-  float16* a = input1_gpu.data<float16>();
-  float16* b = input2_gpu.data<float16>();
-  float16* c = input3_gpu.mutable_data<float16>(gpu_place);
+  paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu);
+  paddle::framework::TensorCopySync(input2, gpu_place, &input2_gpu);
+  paddle::framework::TensorCopySync(input3, gpu_place, &input3_gpu);
+  paddle::platform::float16* a = input1_gpu.data<paddle::platform::float16>();
+  paddle::platform::float16* b = input2_gpu.data<paddle::platform::float16>();
+  paddle::platform::float16* c =
+      input3_gpu.mutable_data<paddle::platform::float16>(gpu_place);
 
-  paddle::operators::math::gemm<paddle::platform::CUDADeviceContext, float16>(
-      context, false, true, m, n, k, float16(1), a, 3, b + 3, 3, float16(1),
-      c + 1, 4);
+  paddle::operators::math::gemm<paddle::platform::CUDADeviceContext,
+                                paddle::platform::float16>(
+      context, false, true, m, n, k, paddle::platform::float16(1), a, 3, b + 3,
+      3, paddle::platform::float16(1), c + 1, 4);
 
-  TensorCopy(input3_gpu, cpu_place, context, &input3);
+  paddle::framework::TensorCopySync(input3_gpu, cpu_place, &input3);
 
   context.Wait();
   EXPECT_EQ(static_cast<float>(input3_ptr[0]), 0);
@@ -412,24 +402,21 @@ TEST(math_function, gemm_trans_cublas_fp16) {
 
 template <typename T>
 void GemvTest(int m, int n, bool trans) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-
-  Tensor mat_a;
-  Tensor vec_b;
-  Tensor vec_c;
+  paddle::framework::Tensor mat_a;
+  paddle::framework::Tensor vec_b;
+  paddle::framework::Tensor vec_c;
 
-  CPUPlace cpu_place;
-  CUDAPlace gpu_place(0);
-  CUDADeviceContext context(gpu_place);
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CUDAPlace gpu_place(0);
+  paddle::platform::CUDADeviceContext context(gpu_place);
 
   T* data_a = mat_a.mutable_data<T>({m, n}, cpu_place);
   T* data_b = vec_b.mutable_data<T>({trans ? m : n}, cpu_place);
   T* data_c = vec_c.mutable_data<T>({trans ? n : m}, cpu_place);
 
-  Tensor g_mat_a;
-  Tensor g_vec_b;
-  Tensor g_vec_c;
+  paddle::framework::Tensor g_mat_a;
+  paddle::framework::Tensor g_vec_b;
+  paddle::framework::Tensor g_vec_c;
   T* g_data_a = g_mat_a.mutable_data<T>(mat_a.dims(), gpu_place);
   T* g_data_b = g_vec_b.mutable_data<T>(vec_b.dims(), gpu_place);
   T* g_data_c = g_vec_c.mutable_data<T>(vec_c.dims(), gpu_place);
@@ -441,14 +428,14 @@ void GemvTest(int m, int n, bool trans) {
     data_b[i] = static_cast<T>(i);
   }
 
-  TensorCopy(mat_a, gpu_place, context, &g_mat_a);
-  TensorCopy(vec_b, gpu_place, context, &g_vec_b);
+  paddle::framework::TensorCopySync(mat_a, gpu_place, &g_mat_a);
+  paddle::framework::TensorCopySync(vec_b, gpu_place, &g_vec_b);
 
-  paddle::operators::math::gemv<CUDADeviceContext, T>(
+  paddle::operators::math::gemv<paddle::platform::CUDADeviceContext, T>(
       context, trans, static_cast<int>(m), static_cast<int>(n), 1., g_data_a,
       g_data_b, 0., g_data_c);
 
-  TensorCopy(g_vec_c, cpu_place, context, &vec_c);
+  paddle::framework::TensorCopySync(g_vec_c, cpu_place, &vec_c);
 
   if (!trans) {
     for (int i = 0; i < m; ++i) {

paddle/fluid/operators/math/matmul.h

@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <algorithm>
+#include <vector>
 #include "paddle/fluid/operators/math/math_function.h"
 
 namespace paddle {

paddle/fluid/operators/math/maxouting.cu

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/math/maxouting.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/math/pooling.cu

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/math/pooling.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/math/sampler.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 "sampler.h"
+#include "paddle/fluid/operators/math/sampler.h"
 
 namespace paddle {
 namespace random {

paddle/fluid/operators/math/sampler.h

@@ -13,9 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <cstdint>
 #include <memory>
 #include <random>
-typedef long int64;
 namespace paddle {
 namespace operators {
 namespace math {
@@ -27,25 +27,25 @@ namespace math {
 */
 class Sampler {
  public:
-  explicit Sampler(int64 range) : range_(range) {
+  explicit Sampler(int64_t range) : range_(range) {
     PADDLE_ENFORCE_GT(range, 0);
     std::random_device r;
     seed_ = r();
   }
-  explicit Sampler(int64 range, unsigned int seed)
+  explicit Sampler(int64_t range, unsigned int seed)
       : range_(range), seed_(seed) {
     PADDLE_ENFORCE_GT(range, 0);
   }
   virtual ~Sampler();
   // Sample a single value
-  virtual int64 Sample() const = 0;
+  virtual int64_t Sample() const = 0;
   // The probability that a single call to Sample() returns the given value.
-  virtual float Probability(int64 value) const = 0;
+  virtual float Probability(int64_t value) const = 0;
 
-  int64 range() { return range_; };
+  int64 range() { return range_; }
 
  protected:
-  const int64 range_;
+  const int64_t range_;
   unsigned int seed_;
 };
 
@@ -56,15 +56,15 @@ class Sampler {
  */
 class UniformSampler : public Sampler {
  public:
-  explicit UniformSampler(int64 range);
+  explicit UniformSampler(int64_t range);
 
-  explicit UniformSampler(int64 range, unsigned int seed);
+  explicit UniformSampler(int64_t range, unsigned int seed);
 
   ~UniformSampler() override {}
 
   int64 Sample() const override;
 
-  float Probability(int64 value) const override;
+  float Probability(int64_t value) const override;
 
  private:
   const float inv_range_;
@@ -79,15 +79,15 @@ class UniformSampler : public Sampler {
  */
 class LogUniformSampler : public Sampler {
  public:
-  explicit LogUniformSampler(int64 range);
+  explicit LogUniformSampler(int64_t range);
 
-  explicit LogUniformSampler(int64 range, unsigned int seed);
+  explicit LogUniformSampler(int64_t range, unsigned int seed);
 
   ~LogUniformSampler() override {}
 
   int64 Sample() const override;
 
-  float Probability(int64 value) const override;
+  float Probability(int64_t value) const override;
 
  private:
   const float log_range_;
@@ -95,6 +95,6 @@ class LogUniformSampler : public Sampler {
   std::shared_ptr<std::uniform_real_distribution<>> dist_;
 };
 
-}  // math
+}  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/math/selected_rows_functor.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <set>
+#include <vector>
 
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"

paddle/fluid/operators/math/selected_rows_functor.cu

@@ -13,10 +13,11 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <set>
+#include <vector>
 
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
-#include "paddle/fluid/platform/cuda_helper.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/math/selected_rows_functor_test.cc

@@ -13,41 +13,50 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
+#include <vector>
 #include "gtest/gtest.h"
 #include "paddle/fluid/operators/math/math_function.h"
 
 TEST(selected_rows_functor, cpu_add) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-  using namespace paddle::operators::math;
-
-  CPUPlace cpu_place;
-  CPUDeviceContext ctx(cpu_place);
-  SetConstant<CPUDeviceContext, float> functor;
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CPUDeviceContext ctx(cpu_place);
+  paddle::operators::math::SetConstant<paddle::platform::CPUDeviceContext,
+                                       float>
+      functor;
   int64_t height = 10;
   int64_t row_numel = 10;
 
   std::vector<int64_t> rows1{0, 4, 7};
-  std::unique_ptr<SelectedRows> selected_rows1{new SelectedRows(rows1, height)};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows1{
+      new paddle::framework::SelectedRows(rows1, height)};
   auto* in1_value = selected_rows1->mutable_value();
   in1_value->mutable_data<float>(
-      make_ddim({static_cast<int64_t>(rows1.size()), row_numel}), cpu_place);
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows1.size()), row_numel}),
+      cpu_place);
   functor(ctx, in1_value, 1.0);
 
   std::vector<int64_t> rows2{0, 5, 7, 9};
-  std::unique_ptr<SelectedRows> selected_rows2{new SelectedRows(rows2, height)};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows2{
+      new paddle::framework::SelectedRows(rows2, height)};
   auto* in2_value = selected_rows2->mutable_value();
   in2_value->mutable_data<float>(
-      make_ddim({static_cast<int64_t>(rows2.size()), row_numel}), cpu_place);
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows2.size()), row_numel}),
+      cpu_place);
   functor(ctx, in2_value, 2.0);
 
-  std::unique_ptr<SelectedRows> output{new SelectedRows()};
+  std::unique_ptr<paddle::framework::SelectedRows> output{
+      new paddle::framework::SelectedRows()};
   auto* out_value = output->mutable_value();
 
   // simplely concat two SelectedRows
-  out_value->mutable_data<float>(make_ddim({7, 10}), cpu_place);
+  out_value->mutable_data<float>(paddle::framework::make_ddim({7, 10}),
+                                 cpu_place);
 
-  SelectedRowsAdd<CPUDeviceContext, float> add_functor;
+  paddle::operators::math::SelectedRowsAdd<paddle::platform::CPUDeviceContext,
+                                           float>
+      add_functor;
   add_functor(ctx, *selected_rows1, *selected_rows2, output.get());
 
   auto out_height = output->height();
@@ -78,14 +87,20 @@ TEST(selected_rows_functor, cpu_add) {
   EXPECT_EQ(out_data[5 * row_numel + 7], 2.0);
   EXPECT_EQ(out_data[6 * row_numel + 9], 2.0);
 
-  std::unique_ptr<Tensor> tensor1{new Tensor()};
-  tensor1->mutable_data<float>(make_ddim({height, row_numel}), cpu_place);
+  std::unique_ptr<paddle::framework::Tensor> tensor1{
+      new paddle::framework::Tensor()};
+  tensor1->mutable_data<float>(
+      paddle::framework::make_ddim({height, row_numel}), cpu_place);
   functor(ctx, tensor1.get(), 3.0);
 
-  std::unique_ptr<Tensor> tensor2{new Tensor()};
-  tensor2->mutable_data<float>(make_ddim({height, row_numel}), cpu_place);
+  std::unique_ptr<paddle::framework::Tensor> tensor2{
+      new paddle::framework::Tensor()};
+  tensor2->mutable_data<float>(
+      paddle::framework::make_ddim({height, row_numel}), cpu_place);
 
-  SelectedRowsAddTensor<CPUDeviceContext, float> add_tensor_functor;
+  paddle::operators::math::SelectedRowsAddTensor<
+      paddle::platform::CPUDeviceContext, float>
+      add_tensor_functor;
   add_tensor_functor(ctx, *output, *tensor1, tensor2.get());
 
   auto* tensor2_data = tensor2->data<float>();
@@ -106,38 +121,46 @@ TEST(selected_rows_functor, cpu_add) {
 }
 
 TEST(selected_rows_functor, cpu_add_to) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-  using namespace paddle::operators::math;
-
-  CPUPlace cpu_place;
-  CPUDeviceContext ctx(cpu_place);
-  SetConstant<CPUDeviceContext, float> functor;
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CPUDeviceContext ctx(cpu_place);
+  paddle::operators::math::SetConstant<paddle::platform::CPUDeviceContext,
+                                       float>
+      functor;
   int64_t height = 10;
   int64_t row_numel = 10;
 
   std::vector<int64_t> rows1{0, 4, 7};
-  std::unique_ptr<SelectedRows> selected_rows1{new SelectedRows(rows1, height)};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows1{
+      new paddle::framework::SelectedRows(rows1, height)};
   auto* in1_value = selected_rows1->mutable_value();
   in1_value->mutable_data<float>(
-      make_ddim({static_cast<int64_t>(rows1.size()), row_numel}), cpu_place);
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows1.size()), row_numel}),
+      cpu_place);
   functor(ctx, in1_value, 1.0);
 
   std::vector<int64_t> rows2{0, 5, 7, 9};
-  std::unique_ptr<SelectedRows> selected_rows2{new SelectedRows(rows2, height)};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows2{
+      new paddle::framework::SelectedRows(rows2, height)};
   auto* in2_value = selected_rows2->mutable_value();
   in2_value->mutable_data<float>(
-      make_ddim({static_cast<int64_t>(rows2.size()), row_numel}), cpu_place);
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows2.size()), row_numel}),
+      cpu_place);
   functor(ctx, in2_value, 2.0);
 
-  std::unique_ptr<SelectedRows> output{new SelectedRows()};
+  std::unique_ptr<paddle::framework::SelectedRows> output{
+      new paddle::framework::SelectedRows()};
   output->set_height(height);
   auto* out_value = output->mutable_value();
 
   // simplely concat two SelectedRows
-  out_value->mutable_data<float>(make_ddim({7, 10}), cpu_place);
+  out_value->mutable_data<float>(paddle::framework::make_ddim({7, 10}),
+                                 cpu_place);
 
-  SelectedRowsAddTo<CPUDeviceContext, float> add_to_functor;
+  paddle::operators::math::SelectedRowsAddTo<paddle::platform::CPUDeviceContext,
+                                             float>
+      add_to_functor;
   add_to_functor(ctx, *selected_rows1, 0, output.get());
   add_to_functor(ctx, *selected_rows2, in1_value->numel(), output.get());
 
@@ -169,11 +192,15 @@ TEST(selected_rows_functor, cpu_add_to) {
   EXPECT_EQ(out_data[5 * row_numel + 7], 2.0);
   EXPECT_EQ(out_data[6 * row_numel + 9], 2.0);
 
-  std::unique_ptr<Tensor> tensor1{new Tensor()};
-  tensor1->mutable_data<float>(make_ddim({height, row_numel}), cpu_place);
+  std::unique_ptr<paddle::framework::Tensor> tensor1{
+      new paddle::framework::Tensor()};
+  tensor1->mutable_data<float>(
+      paddle::framework::make_ddim({height, row_numel}), cpu_place);
   functor(ctx, tensor1.get(), 3.0);
 
-  SelectedRowsAddToTensor<CPUDeviceContext, float> add_to_tensor_functor;
+  paddle::operators::math::SelectedRowsAddToTensor<
+      paddle::platform::CPUDeviceContext, float>
+      add_to_tensor_functor;
   add_to_tensor_functor(ctx, *output, tensor1.get());
 
   auto* tensor1_data = tensor1->data<float>();

paddle/fluid/operators/math/selected_rows_functor_test.cu

@@ -12,43 +12,52 @@ 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 <vector>
 #include "gtest/gtest.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
 
 TEST(selected_rows_functor, gpu_add) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-  using namespace paddle::operators::math;
-
-  CUDAPlace gpu_place(0);
-  CPUPlace cpu_place;
-  CUDADeviceContext ctx(gpu_place);
-  SetConstant<CUDADeviceContext, float> functor;
+  paddle::platform::CUDAPlace gpu_place(0);
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CUDADeviceContext ctx(gpu_place);
+  paddle::operators::math::SetConstant<paddle::platform::CUDADeviceContext,
+                                       float>
+      functor;
   int64_t height = 10;
   int64_t row_numel = 10;
 
   std::vector<int64_t> rows1{0, 4, 7};
-  std::unique_ptr<SelectedRows> selected_rows1{new SelectedRows(rows1, height)};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows1{
+      new paddle::framework::SelectedRows(rows1, height)};
   auto* in1_value = selected_rows1->mutable_value();
   in1_value->mutable_data<float>(
-      make_ddim({static_cast<int64_t>(rows1.size()), row_numel}), gpu_place);
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows1.size()), row_numel}),
+      gpu_place);
   functor(ctx, in1_value, 1.0);
 
   std::vector<int64_t> rows2{0, 5, 7, 9};
-  std::unique_ptr<SelectedRows> selected_rows2{new SelectedRows(rows2, height)};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows2{
+      new paddle::framework::SelectedRows(rows2, height)};
   auto* in2_value = selected_rows2->mutable_value();
   in2_value->mutable_data<float>(
-      make_ddim({static_cast<int64_t>(rows2.size()), row_numel}), gpu_place);
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows2.size()), row_numel}),
+      gpu_place);
   functor(ctx, in2_value, 2.0);
 
-  std::unique_ptr<SelectedRows> output{new SelectedRows()};
+  std::unique_ptr<paddle::framework::SelectedRows> output{
+      new paddle::framework::SelectedRows()};
   auto* out_value = output->mutable_value();
 
-  // simplely concat two SelectedRows
-  out_value->mutable_data<float>(make_ddim({7, 10}), gpu_place);
+  // simply concat two SelectedRows
+  out_value->mutable_data<float>(paddle::framework::make_ddim({7, 10}),
+                                 gpu_place);
 
-  SelectedRowsAdd<CUDADeviceContext, float> add_functor;
+  paddle::operators::math::SelectedRowsAdd<paddle::platform::CUDADeviceContext,
+                                           float>
+      add_functor;
   add_functor(ctx, *selected_rows1, *selected_rows2, output.get());
 
   auto out_height = output->height();
@@ -66,8 +75,8 @@ TEST(selected_rows_functor, gpu_add) {
   EXPECT_EQ(out_rows[5], 7);
   EXPECT_EQ(out_rows[6], 9);
 
-  Tensor out_cpu;
-  TensorCopy(*out_value, cpu_place, ctx, &out_cpu);
+  paddle::framework::Tensor out_cpu;
+  paddle::framework::TensorCopy(*out_value, cpu_place, ctx, &out_cpu);
   ctx.Wait();
 
   auto* out_cpu_data = out_cpu.data<float>();
@@ -83,18 +92,24 @@ TEST(selected_rows_functor, gpu_add) {
   EXPECT_EQ(out_cpu_data[5 * row_numel + 7], 2.0);
   EXPECT_EQ(out_cpu_data[6 * row_numel + 9], 2.0);
 
-  std::unique_ptr<Tensor> tensor1{new Tensor()};
-  tensor1->mutable_data<float>(make_ddim({height, row_numel}), gpu_place);
+  std::unique_ptr<paddle::framework::Tensor> tensor1{
+      new paddle::framework::Tensor()};
+  tensor1->mutable_data<float>(
+      paddle::framework::make_ddim({height, row_numel}), gpu_place);
   functor(ctx, tensor1.get(), 3.0);
 
-  std::unique_ptr<Tensor> tensor2{new Tensor()};
-  tensor2->mutable_data<float>(make_ddim({height, row_numel}), gpu_place);
+  std::unique_ptr<paddle::framework::Tensor> tensor2{
+      new paddle::framework::Tensor()};
+  tensor2->mutable_data<float>(
+      paddle::framework::make_ddim({height, row_numel}), gpu_place);
 
-  SelectedRowsAddTensor<CUDADeviceContext, float> add_tensor_functor;
+  paddle::operators::math::SelectedRowsAddTensor<
+      paddle::platform::CUDADeviceContext, float>
+      add_tensor_functor;
   add_tensor_functor(ctx, *output, *tensor1, tensor2.get());
 
-  Tensor tensor2_cpu;
-  TensorCopy(*tensor2, cpu_place, ctx, &tensor2_cpu);
+  paddle::framework::Tensor tensor2_cpu;
+  paddle::framework::TensorCopy(*tensor2, cpu_place, ctx, &tensor2_cpu);
   ctx.Wait();
 
   auto* tensor2_cpu_data = tensor2_cpu.data<float>();
@@ -115,39 +130,47 @@ TEST(selected_rows_functor, gpu_add) {
 }
 
 TEST(selected_rows_functor, gpu_add_to) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-  using namespace paddle::operators::math;
-
-  CUDAPlace gpu_place(0);
-  CPUPlace cpu_place;
-  CUDADeviceContext ctx(gpu_place);
-  SetConstant<CUDADeviceContext, float> functor;
+  paddle::platform::CUDAPlace gpu_place(0);
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CUDADeviceContext ctx(gpu_place);
+  paddle::operators::math::SetConstant<paddle::platform::CUDADeviceContext,
+                                       float>
+      functor;
   int64_t height = 10;
   int64_t row_numel = 10;
 
   std::vector<int64_t> rows1{0, 4, 7};
-  std::unique_ptr<SelectedRows> selected_rows1{new SelectedRows(rows1, height)};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows1{
+      new paddle::framework::SelectedRows(rows1, height)};
   auto* in1_value = selected_rows1->mutable_value();
   in1_value->mutable_data<float>(
-      make_ddim({static_cast<int64_t>(rows1.size()), row_numel}), gpu_place);
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows1.size()), row_numel}),
+      gpu_place);
   functor(ctx, in1_value, 1.0);
 
   std::vector<int64_t> rows2{0, 5, 7, 9};
-  std::unique_ptr<SelectedRows> selected_rows2{new SelectedRows(rows2, height)};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows2{
+      new paddle::framework::SelectedRows(rows2, height)};
   auto* in2_value = selected_rows2->mutable_value();
   in2_value->mutable_data<float>(
-      make_ddim({static_cast<int64_t>(rows2.size()), row_numel}), gpu_place);
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows2.size()), row_numel}),
+      gpu_place);
   functor(ctx, in2_value, 2.0);
 
-  std::unique_ptr<SelectedRows> output{new SelectedRows()};
+  std::unique_ptr<paddle::framework::SelectedRows> output{
+      new paddle::framework::SelectedRows()};
   output->set_height(height);
   auto* out_value = output->mutable_value();
 
-  // simplely concat two SelectedRows
-  out_value->mutable_data<float>(make_ddim({7, 10}), gpu_place);
+  // simply concat two SelectedRows
+  out_value->mutable_data<float>(paddle::framework::make_ddim({7, 10}),
+                                 gpu_place);
 
-  SelectedRowsAddTo<CUDADeviceContext, float> add_to_functor;
+  paddle::operators::math::SelectedRowsAddTo<
+      paddle::platform::CUDADeviceContext, float>
+      add_to_functor;
   add_to_functor(ctx, *selected_rows1, 0, output.get());
   add_to_functor(ctx, *selected_rows2, in1_value->numel(), output.get());
 
@@ -166,8 +189,8 @@ TEST(selected_rows_functor, gpu_add_to) {
   EXPECT_EQ(out_rows[5], 7);
   EXPECT_EQ(out_rows[6], 9);
 
-  Tensor out_cpu;
-  TensorCopy(*out_value, cpu_place, ctx, &out_cpu);
+  paddle::framework::Tensor out_cpu;
+  paddle::framework::TensorCopy(*out_value, cpu_place, ctx, &out_cpu);
   ctx.Wait();
 
   auto* out_cpu_data = out_cpu.data<float>();
@@ -183,15 +206,19 @@ TEST(selected_rows_functor, gpu_add_to) {
   EXPECT_EQ(out_cpu_data[5 * row_numel + 7], 2.0);
   EXPECT_EQ(out_cpu_data[6 * row_numel + 9], 2.0);
 
-  std::unique_ptr<Tensor> tensor1{new Tensor()};
-  tensor1->mutable_data<float>(make_ddim({height, row_numel}), gpu_place);
+  std::unique_ptr<paddle::framework::Tensor> tensor1{
+      new paddle::framework::Tensor()};
+  tensor1->mutable_data<float>(
+      paddle::framework::make_ddim({height, row_numel}), gpu_place);
   functor(ctx, tensor1.get(), 3.0);
 
-  SelectedRowsAddToTensor<CUDADeviceContext, float> add_to_tensor_functor;
+  paddle::operators::math::SelectedRowsAddToTensor<
+      paddle::platform::CUDADeviceContext, float>
+      add_to_tensor_functor;
   add_to_tensor_functor(ctx, *output, tensor1.get());
 
-  Tensor tensor1_cpu;
-  TensorCopy(*tensor1, cpu_place, ctx, &tensor1_cpu);
+  paddle::framework::Tensor tensor1_cpu;
+  paddle::framework::TensorCopy(*tensor1, cpu_place, ctx, &tensor1_cpu);
   ctx.Wait();
 
   auto* tensor1_cpu_data = tensor1_cpu.data<float>();

paddle/fluid/operators/math/sequence2batch.cc

@@ -23,11 +23,11 @@ class CopyMatrixRowsFunctor<platform::CPUDeviceContext, T> {
  public:
   void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& src,
-                  framework::Vector<size_t> index_lod, framework::Tensor& dst,
+                  framework::Vector<size_t> index_lod, framework::Tensor* dst,
                   bool is_src_index) {
     size_t* index = index_lod.data();
     auto src_dims = src.dims();
-    auto dst_dims = dst.dims();
+    auto dst_dims = dst->dims();
     PADDLE_ENFORCE_EQ(src_dims.size(), 2UL,
                       "The src must be matrix with rank 2.");
     PADDLE_ENFORCE_EQ(dst_dims.size(), 2UL,
@@ -37,7 +37,7 @@ class CopyMatrixRowsFunctor<platform::CPUDeviceContext, T> {
     auto height = dst_dims[0];
     auto width = dst_dims[1];
     auto* src_data = src.data<T>();
-    auto* dst_data = dst.data<T>();
+    auto* dst_data = dst->data<T>();
     for (int i = 0; i < height; ++i) {
       if (is_src_index) {
         memcpy(dst_data + i * width, src_data + index[i] * width,

paddle/fluid/operators/math/sequence2batch.cu

@@ -43,10 +43,10 @@ class CopyMatrixRowsFunctor<platform::CUDADeviceContext, T> {
  public:
   void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& src,
-                  framework::Vector<size_t> index_lod, framework::Tensor& dst,
+                  framework::Vector<size_t> index_lod, framework::Tensor* dst,
                   bool is_src_index) {
     auto src_dims = src.dims();
-    auto dst_dims = dst.dims();
+    auto dst_dims = dst->dims();
     PADDLE_ENFORCE_EQ(src_dims.size(), 2,
                       "The src must be matrix with rank 2.");
     PADDLE_ENFORCE_EQ(dst_dims.size(), 2,
@@ -56,7 +56,7 @@ class CopyMatrixRowsFunctor<platform::CUDADeviceContext, T> {
     auto height = dst_dims[0];
     auto width = dst_dims[1];
     auto* src_data = src.data<T>();
-    auto* dst_data = dst.data<T>();
+    auto* dst_data = dst->data<T>();
 
     dim3 threads(128, 8);
     dim3 grid(8, 1);

paddle/fluid/operators/math/sequence2batch.h

@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <algorithm>
+#include <vector>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/tensor.h"
@@ -35,7 +37,7 @@ class CopyMatrixRowsFunctor {
   // copy the input src to the indexed rows of output dst.
   // The indexed rows are based on the input index.
   void operator()(const DeviceContext& context, const framework::Tensor& src,
-                  framework::Vector<size_t> index_lod, framework::Tensor& dst,
+                  framework::Vector<size_t> index_lod, framework::Tensor* dst,
                   bool is_src_index);
 };
 
@@ -58,10 +60,10 @@ class LoDTensor2BatchFunctor {
  public:
   void operator()(const DeviceContext& context,
                   const framework::LoDTensor& lod_tensor,
-                  framework::LoDTensor& batch, bool is_cal_batch_lod,
+                  framework::LoDTensor* batch, bool is_cal_batch_lod,
                   bool is_reverse = false) const {
     if (!is_cal_batch_lod) {
-      auto lods = batch.lod();
+      auto lods = batch->lod();
       PADDLE_ENFORCE_GT(lods.size(), 2UL);
       PADDLE_ENFORCE_EQ(lods[1].size(),
                         static_cast<size_t>(lod_tensor.dims()[0]));
@@ -141,7 +143,7 @@ class LoDTensor2BatchFunctor {
     for (size_t i = 0; i < seq_info.size(); ++i) {
       seq_order[i] = seq_info[i].seq_idx;
     }
-    batch.set_lod(batch_lods);
+    batch->set_lod(batch_lods);
 
     CopyMatrixRowsFunctor<DeviceContext, T> to_batch;
     to_batch(context, lod_tensor, batch_lods[1], batch, true);
@@ -153,11 +155,11 @@ class Batch2LoDTensorFunctor {
  public:
   void operator()(const DeviceContext& context,
                   const framework::LoDTensor& batch,
-                  framework::LoDTensor& lod_tensor) const {
+                  framework::LoDTensor* lod_tensor) const {
     auto in_lod = batch.lod();
     PADDLE_ENFORCE_GT(in_lod.size(), 2UL);
     PADDLE_ENFORCE_EQ(in_lod[1].size(),
-                      static_cast<size_t>(lod_tensor.dims()[0]));
+                      static_cast<size_t>(lod_tensor->dims()[0]));
     CopyMatrixRowsFunctor<DeviceContext, T> to_seq;
     to_seq(context, batch, in_lod[1], lod_tensor, false);
   }

paddle/fluid/operators/math/sequence_padding_test.cc

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/math/sequence_padding.h"
 #include <gtest/gtest.h>
+#include <vector>
 
 template <typename DeviceContext, typename Place, typename T>
 void TestSequencePadding(const paddle::framework::LoD& lod,
@@ -75,7 +76,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod,
 
   delete place;
   delete context;
-};
+}
 
 TEST(Seq2BatchPadding, CPU) {
   paddle::framework::LoD lod1;

paddle/fluid/operators/math/sequence_pooling.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/math/sequence_pooling.h"
+#include <string>
 #include "paddle/fluid/operators/math/math_function.h"
 
 namespace paddle {