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

test=develop

CMakeLists.txt

@@ -25,12 +25,18 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: "
 message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
         "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
 if(WIN32)
+    set(CMAKE_SUPPRESS_REGENERATION ON)
     set(CMAKE_STATIC_LIBRARY_PREFIX lib)
     add_definitions("/DGOOGLE_GLOG_DLL_DECL=")
     set(CMAKE_C_FLAGS_DEBUG   "${CMAKE_C_FLAGS_DEBUG} /bigobj /MTd")
     set(CMAKE_C_FLAGS_RELEASE  "${CMAKE_C_FLAGS_RELEASE} /bigobj /MT")
     set(CMAKE_CXX_FLAGS_DEBUG  "${CMAKE_CXX_FLAGS_DEBUG} /bigobj /MTd")
     set(CMAKE_CXX_FLAGS_RELEASE   "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT")
+    add_compile_options(/wd4068 /wd4129 /wd4244 /wd4267 /wd4297 /wd4530 /wd4577 /wd4819 /wd4838)
+    set(PADDLE_LINK_FLAGS "/IGNORE:4006 /IGNORE:4098 /IGNORE:4217 /IGNORE:4221")
+    set(CMAKE_STATIC_LINKER_FLAGS  "${CMAKE_STATIC_LINKER_FLAGS} ${PADDLE_LINK_FLAGS}")
+    set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} ${PADDLE_LINK_FLAGS}")
+    set(CMAKE_EXE_LINKER_FLAGS  "${CMAKE_EXE_LINKER_FLAGS} ${PADDLE_LINK_FLAGS}")
 endif(WIN32)
 
 find_package(CUDA QUIET)

README.md

 # PaddlePaddle
 
+English | [简体中文](./README_cn.md)
 
 [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
 [![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
@@ -7,7 +8,6 @@
 [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
 
-
 Welcome to the PaddlePaddle GitHub.
 
 PaddlePaddle (PArallel Distributed Deep LEarning) is an easy-to-use,
@@ -18,16 +18,6 @@ learning to many products at Baidu.
 Our vision is to enable deep learning for everyone via PaddlePaddle.
 Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
 
-
-欢迎来到 PaddlePaddle GitHub
-
-PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效灵活、可扩展的深度学习平台，最初由百度科学家和工程师共同开发，目的是将深度学习技术应用到百度的众多产品中。
-
-我们的愿景是让每个人都能通过PaddlePaddle接触深度学习
-
-跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
-
-
 ### Latest PaddlePaddle Release: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
 ### Install Latest Stable Release:
 ```
@@ -43,23 +33,6 @@ pip install paddlepaddle-gpu==1.2.0.post85
 # For installation on other platform, refer to http://paddlepaddle.org/
 ```
 
-
-### PaddlePaddle最新版本: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
-### 安装最新稳定版本:
-```
-# Linux CPU
-pip install paddlepaddle
-# Linux GPU cuda9cudnn7
-pip install paddlepaddle-gpu
-# Linux GPU cuda8cudnn7
-pip install paddlepaddle-gpu==1.2.0.post87
-# Linux GPU cuda8cudnn5
-pip install paddlepaddle-gpu==1.2.0.post85
-
-# 其他平台上的安装指引请参考 http://paddlepaddle.org/
-```
-
-
 ## Features
 
 - **Flexibility**
@@ -100,38 +73,10 @@ pip install paddlepaddle-gpu==1.2.0.post85
     Baidu and it has achieved a significant impact. We hope you can also explore
     the capability of PaddlePaddle to make an impact on your product.
 
-## 特点
-
-- **灵活性**
-
-    PaddlePaddle支持丰富的神经网络架构和优化算法。易于配置复杂模型，例如带有注意力机制或复杂记忆连接的神经网络机器翻译模型。
-
--  **高效性**
-
-    为了高效使用异步计算资源，PaddlePaddle对框架的不同层进行优化，包括计算、存储、架构和通信。下面是一些样例：
-    
-    - 通过SSE/AVX 内置函数、BLAS库(例如MKL、OpenBLAS、cuBLAS)或定制的CPU/GPU内核优化数学操作。
-    - 通过MKL-DNN库优化CNN网络
-    - 高度优化循环网络，无需执行 `padding` 操作即可处理 **变长** 序列
-    - 针对高维稀疏数据模型，优化了局部和分布式训练。
-     
-
-- **稳定性**
-
-    有了 PaddlePaddle，使得利用各种CPU/GPU和机器来加速训练变得简单。PaddlePaddle 通过优化通信可以实现巨大吞吐量和快速执行。
-
-- **连接产品**
-
-    另外，PaddlePaddle 的设计也易于部署。在百度，PaddlePaddle 已经部署到含有巨大用户量的产品和服务上，包括广告点击率（CTR）预测、大规模图像分类、光学字符识别（OCR）、搜索排序，计算机病毒检测、推荐系统等等。PaddlePaddle广泛应用于百度产品中，产生了非常重要的影响。我们希望您也能探索 PaddlePaddle 的能力，为您的产品创造新的影响力和效果。
-
 ## Installation
 
 It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) on our website.
 
-## 安装
-
-推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) 
-
 ## Documentation
 
 We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and
@@ -153,37 +98,9 @@ We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarte
 
    We appreciate your contributions!
 
-## 文档
-
-我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)和
-[中文](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) 文档
-
-- [深度学习101](https://github.com/PaddlePaddle/book)
-
-  或许您想从这个在线交互式书籍开始，可以在Jupyter Notebook中运行
-
-- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
-
-  可以在MPI集群上运行分布式训练任务
-
-- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
-
-   新的API支持代码更少更简洁的程序
-
-- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
-
-   欢迎您的贡献!
-
 ## Ask Questions
 
 You are welcome to submit questions and bug reports as [Github Issues](https://github.com/PaddlePaddle/Paddle/issues).
 
-## 答疑
-
-欢迎您将问题和bug报告以[Github Issues](https://github.com/PaddlePaddle/Paddle/issues)的形式提交
-
 ## Copyright and License
 PaddlePaddle is provided under the [Apache-2.0 license](LICENSE).
-
-## 版权和许可证
-PaddlePaddle由[Apache-2.0 license](LICENSE)提供

README_cn.md

+# PaddlePaddle
+
+[English](./README.md) | 简体中文
+
+[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
+[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
+[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html)
+[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
+[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
+
+欢迎来到 PaddlePaddle GitHub
+
+PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效灵活、可扩展的深度学习平台，最初由百度科学家和工程师共同开发，目的是将深度学习技术应用到百度的众多产品中。
+
+我们的愿景是让每个人都能通过PaddlePaddle接触深度学习
+
+跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
+
+### PaddlePaddle最新版本: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
+### 安装最新稳定版本:
+```
+# Linux CPU
+pip install paddlepaddle
+# Linux GPU cuda9cudnn7
+pip install paddlepaddle-gpu
+# Linux GPU cuda8cudnn7
+pip install paddlepaddle-gpu==1.2.0.post87
+# Linux GPU cuda8cudnn5
+pip install paddlepaddle-gpu==1.2.0.post85
+
+# 其他平台上的安装指引请参考 http://paddlepaddle.org/
+```
+
+## 特性
+
+- **灵活性**
+
+    PaddlePaddle支持丰富的神经网络架构和优化算法。易于配置复杂模型，例如带有注意力机制或复杂记忆连接的神经网络机器翻译模型。
+
+-  **高效性**
+
+    为了高效使用异步计算资源，PaddlePaddle对框架的不同层进行优化，包括计算、存储、架构和通信。下面是一些样例：
+
+    - 通过SSE/AVX 内置函数、BLAS库(例如MKL、OpenBLAS、cuBLAS)或定制的CPU/GPU内核优化数学操作。
+    - 通过MKL-DNN库优化CNN网络
+    - 高度优化循环网络，无需执行 `padding` 操作即可处理 **变长** 序列
+    - 针对高维稀疏数据模型，优化了局部和分布式训练。
+
+
+- **稳定性**
+
+    有了 PaddlePaddle，使得利用各种CPU/GPU和机器来加速训练变得简单。PaddlePaddle 通过优化通信可以实现巨大吞吐量和快速执行。
+
+- **与产品相连**
+
+    另外，PaddlePaddle 的设计也易于部署。在百度，PaddlePaddle 已经部署到含有巨大用户量的产品和服务上，包括广告点击率（CTR）预测、大规模图像分类、光学字符识别（OCR）、搜索排序，计算机病毒检测、推荐系统等等。PaddlePaddle广泛应用于百度产品中，产生了非常重要的影响。我们希望您也能探索 PaddlePaddle 的能力，为您的产品创造新的影响力和效果。
+
+## 安装
+
+推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html)
+
+## 文档
+
+我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)和
+[中文](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) 文档
+
+- [深度学习101](https://github.com/PaddlePaddle/book)
+
+  或许您想从这个在线交互式书籍开始，可以在Jupyter Notebook中运行
+
+- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
+
+  可以在MPI集群上运行分布式训练任务
+
+- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
+
+   新的API支持代码更少更简洁的程序
+
+- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
+
+   欢迎您的贡献!
+
+## 答疑
+
+欢迎您将问题和bug报告以[Github Issues](https://github.com/PaddlePaddle/Paddle/issues)的形式提交
+
+## 版权和许可证
+PaddlePaddle由[Apache-2.0 license](LICENSE)提供

cmake/configure.cmake

@@ -152,7 +152,12 @@ endif()
 
 if (WITH_MKLML AND MKLML_IOMP_LIB)
     message(STATUS "Enable Intel OpenMP with ${MKLML_IOMP_LIB}")
-    set(OPENMP_FLAGS "-fopenmp")
+    if(WIN32)
+        # openmp not support well for now on windows
+        set(OPENMP_FLAGS "")
+    else(WIN32)
+        set(OPENMP_FLAGS "-fopenmp")
+    endif(WIN32)
     set(CMAKE_C_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
     set(CMAKE_CXX_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OPENMP_FLAGS}")

cmake/cuda.cmake

@@ -203,25 +203,26 @@ list(APPEND CUDA_NVCC_FLAGS "-w")
 list(APPEND CUDA_NVCC_FLAGS "--expt-relaxed-constexpr")
 
 if (NOT WIN32)
-if(CMAKE_BUILD_TYPE  STREQUAL "Debug")
-    list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_DEBUG})
-elseif(CMAKE_BUILD_TYPE  STREQUAL "Release")
-    list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELEASE})
-elseif(CMAKE_BUILD_TYPE  STREQUAL "RelWithDebInfo")
-    list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELWITHDEBINFO})
-elseif(CMAKE_BUILD_TYPE  STREQUAL "MinSizeRel")
-    # nvcc 9 does not support -Os. Use Release flags instead
-    list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELEASE})
-endif()
+  if(CMAKE_BUILD_TYPE  STREQUAL "Debug")
+      list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_DEBUG})
+  elseif(CMAKE_BUILD_TYPE  STREQUAL "Release")
+      list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELEASE})
+  elseif(CMAKE_BUILD_TYPE  STREQUAL "RelWithDebInfo")
+      list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELWITHDEBINFO})
+  elseif(CMAKE_BUILD_TYPE  STREQUAL "MinSizeRel")
+      # nvcc 9 does not support -Os. Use Release flags instead
+      list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELEASE})
+  endif()
 else(NOT WIN32)
-list(APPEND CUDA_NVCC_FLAGS  "--compiler-options;/bigobj")
-if(CMAKE_BUILD_TYPE  STREQUAL "Debug")
-  list(APPEND CUDA_NVCC_FLAGS  "-g -G")
-  # match the cl's _ITERATOR_DEBUG_LEVEL
-  list(APPEND CUDA_NVCC_FLAGS  "-D_DEBUG")
-elseif(CMAKE_BUILD_TYPE STREQUAL "Release")
-  list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG")
-else()
+  list(APPEND CUDA_NVCC_FLAGS  "-Xcompiler \"/wd 4244 /wd 4267 /wd 4819\"")
+  list(APPEND CUDA_NVCC_FLAGS  "--compiler-options;/bigobj")
+  if(CMAKE_BUILD_TYPE  STREQUAL "Debug")
+    list(APPEND CUDA_NVCC_FLAGS  "-g -G")
+    # match the cl's _ITERATOR_DEBUG_LEVEL
+    list(APPEND CUDA_NVCC_FLAGS  "-D_DEBUG")
+  elseif(CMAKE_BUILD_TYPE STREQUAL "Release")
+    list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG")
+  else()
   message(FATAL "Windows only support Release or Debug build now. Please set visual studio build type to Release/Debug, x64 build.")
 endif()
 endif(NOT WIN32)

cmake/external/glog.cmake

@@ -20,8 +20,10 @@ SET(GLOG_INCLUDE_DIR "${GLOG_INSTALL_DIR}/include" CACHE PATH "glog include dire
 
 IF(WIN32)
   SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.lib" CACHE FILEPATH "glog library." FORCE)
+  SET(GLOG_CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd4267 /wd4530")
 ELSE(WIN32)
   SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.a" CACHE FILEPATH "glog library." FORCE)
+  SET(GLOG_CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS})
 ENDIF(WIN32)
 
 INCLUDE_DIRECTORIES(${GLOG_INCLUDE_DIR})
@@ -39,7 +41,7 @@ ExternalProject_Add(
     UPDATE_COMMAND  ""
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
                     -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-                    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                    -DCMAKE_CXX_FLAGS=${GLOG_CMAKE_CXX_FLAGS}
                     -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
                     -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
                     -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}

cmake/external/mkldnn.cmake

@@ -49,6 +49,8 @@ IF(NOT WIN32)
     SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value")
     SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
     SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
+ELSE()
+    SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} /EHsc")
 ENDIF(NOT WIN32)
 
 ExternalProject_Add(
@@ -61,7 +63,6 @@ ExternalProject_Add(
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
     CMAKE_ARGS          -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-    CMAKE_ARGS          -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
     CMAKE_ARGS          -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
     CMAKE_ARGS          -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
     CMAKE_ARGS          -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}

cmake/external/snappy.cmake

@@ -20,6 +20,12 @@ set(SNAPPY_SOURCES_DIR ${THIRD_PARTY_PATH}/snappy)
 set(SNAPPY_INSTALL_DIR ${THIRD_PARTY_PATH}/install/snappy)
 set(SNAPPY_INCLUDE_DIR "${SNAPPY_INSTALL_DIR}/include" CACHE PATH "snappy include directory." FORCE)
 
+if(WIN32)
+    SET(SNAPPY_CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd4244 /wd4267")
+else()
+    SET(SNAPPY_CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS})
+endif()
+
 ExternalProject_Add(
     extern_snappy
     GIT_REPOSITORY "https://github.com/google/snappy"
@@ -31,7 +37,7 @@ ExternalProject_Add(
                     -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
                     -DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}
                     -DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}
-                    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                    -DCMAKE_CXX_FLAGS=${SNAPPY_CMAKE_CXX_FLAGS}
                     -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
                     -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
                     -DCMAKE_INSTALL_PREFIX=${SNAPPY_INSTALL_DIR}

cmake/flags.cmake

@@ -147,12 +147,6 @@ set(GPU_COMMON_FLAGS
     -Wno-error=unused-function  # Warnings in Numpy Header.
     -Wno-error=array-bounds # Warnings in Eigen::array
 )
-
-else(NOT WIN32)
-set(COMMON_FLAGS
-    "/w") #disable all warnings.
-set(GPU_COMMON_FLAGS
-    "/w") #disable all warnings
 endif(NOT WIN32)
 
 if (APPLE)
@@ -193,8 +187,7 @@ safe_set_static_flag()
         CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO
         CMAKE_C_FLAGS CMAKE_C_FLAGS_DEBUG CMAKE_C_FLAGS_RELEASE
         CMAKE_C_FLAGS_MINSIZEREL CMAKE_C_FLAGS_RELWITHDEBINFO)
-      if(${flag_var} MATCHES "/W3")
-        string(REGEX REPLACE "/W3" "/w" ${flag_var} "${${flag_var}}")
-      endif(${flag_var} MATCHES "/W3")
+        string(REGEX REPLACE "(^| )/W[0-9]( |$)" " " ${flag_var} "${${flag_var}}")
+        set(flag_var "${flag_var} /w")
     endforeach(flag_var)
 endif(WIN32)

cmake/version.cmake

@@ -31,8 +31,23 @@ while ("${PADDLE_VERSION}" STREQUAL "")
           set(tmp_version "${GIT_TAG_NAME}~1")
         endif()
       else()
-        # otherwise, we always set PADDLE_VERSION to 0.0.0 to represent latest
-        set(PADDLE_VERSION "0.0.0")
+        execute_process(
+          COMMAND ${GIT_EXECUTABLE} describe --exact-match --tags ${tmp_version}
+          WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
+          OUTPUT_VARIABLE GIT_EXACT_TAG_NAME
+          RESULT_VARIABLE GIT_EXACT_TAG_RESULT
+          ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+        if (NOT ${GIT_EXACT_TAG_NAME})
+          # Check if current branch is tag branch
+          if (${GIT_EXACT_TAG_NAME} MATCHES "v${TAG_VERSION_REGEX}")
+            string(REPLACE "v" "" PADDLE_VERSION ${GIT_EXACT_TAG_NAME})
+          else()
+            set(PADDLE_VERSION "0.0.0")
+          endif()
+        else()
+          # otherwise, we always set PADDLE_VERSION to 0.0.0 to represent latest
+          set(PADDLE_VERSION "0.0.0")
+        endif()
       endif()
     else()
       set(PADDLE_VERSION "0.0.0")

paddle/fluid/API.spec

@@ -325,6 +325,7 @@ paddle.fluid.layers.iou_similarity ArgSpec(args=['x', 'y', 'name'], varargs=None
 paddle.fluid.layers.box_coder ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name', 'axis'], varargs=None, keywords=None, defaults=('encode_center_size', True, None, 0))
 paddle.fluid.layers.polygon_box_transform ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.yolov3_loss ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.box_clip ArgSpec(args=['input', 'im_info', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.multiclass_nms ArgSpec(args=['bboxes', 'scores', 'score_threshold', 'nms_top_k', 'keep_top_k', 'nms_threshold', 'normalized', 'nms_eta', 'background_label', 'name'], varargs=None, keywords=None, defaults=(0.3, True, 1.0, 0, None))
 paddle.fluid.layers.accuracy ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None))
 paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk', 'slide_steps'], varargs=None, keywords=None, defaults=('ROC', 4095, 1, 1))

paddle/fluid/framework/CMakeLists.txt

@@ -128,7 +128,7 @@ cc_test(version_test SRCS version_test.cc DEPS version)
 
 cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version)
 
-cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
+cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc memory_optimize_helper)
 nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
 
 py_proto_compile(framework_py_proto SRCS framework.proto data_feed.proto)
@@ -192,6 +192,7 @@ cc_library(prune SRCS prune.cc DEPS framework_proto)
 cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)
 cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
         proto_desc)
+cc_test(inplace_op_inference_test SRCS inplace_op_inference_test.cc DEPS op_registry proto_desc op_info memory_optimize_helper)
 cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
 cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
 

paddle/fluid/framework/details/CMakeLists.txt

@@ -50,7 +50,9 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_
 cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
 cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
 
-cc_library(memory_optimize_pass SRCS analysis_var_pass.cc memory_reuse_types.cc DEPS graph graph_helper pass)
+cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper)
+cc_library(memory_optimize_pass SRCS memory_optimize_pass.cc DEPS memory_optimize_helper pass)
+cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_info)
 cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper)
 cc_library(memory_early_delete_pass SRCS memory_early_delete_pass.cc DEPS memory_optimize_pass computation_op_handle scale_loss_grad_op_handle rpc_op_handle
         all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass)
@@ -65,12 +67,12 @@ cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_he
 cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
         scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
 
-set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass memory_early_delete_pass)
+set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass memory_early_delete_pass inplace_op_pass)
 if (WITH_GPU)
   list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass)
 endif()
-cc_test(memory_reuse_types_test SRCS memory_reuse_types_test.cc memory_reuse_types.cc DEPS framework_proto graph)
-cc_test(analysis_var_pass_test SRCS analysis_var_pass_test.cc analysis_var_pass.cc memory_reuse_types.cc DEPS framework_proto graph graph_helper op_registry pass)
+cc_test(memory_optimize_helper_test SRCS memory_optimize_helper_test.cc memory_optimize_helper.cc DEPS framework_proto graph)
+cc_test(memory_optimize_pass_test SRCS memory_optimize_pass_test.cc memory_optimize_pass.cc memory_optimize_helper.cc DEPS framework_proto graph graph_helper op_registry pass)
 
 cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS})
 

paddle/fluid/framework/details/build_strategy.cc

@@ -17,7 +17,7 @@ limitations under the License. */
 #include <glog/logging.h>
 #include <memory>
 
-#include "paddle/fluid/framework/details/memory_reuse_types.h"
+#include "paddle/fluid/framework/details/memory_optimize_helper.h"
 #include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
 #include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h"
 #include "paddle/fluid/framework/details/reduce_op_handle.h"
@@ -47,6 +47,22 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
       AppendPass("sequential_execution_pass");
     }
 
+    // Add op fusion.
+    if (strategy.fuse_relu_depthwise_conv_) {
+      AppendPass("fuse_relu_depthwise_conv_pass");
+    }
+
+    // NOTE(dzhwinter): A note for automatical inplace.
+    // 1. modify program desc passes should put
+    // before inplace pass.
+    // 2. manually configured inplace should put
+    // before inplace_pass
+
+    // Add automatically inplace.
+    if (strategy_.enable_inplace_) {
+      AppendPass("inplace_pass");
+    }
+
     // Add a graph viz pass to record a graph.
     if (!strategy_.debug_graphviz_path_.empty()) {
       auto viz_pass = AppendPass("graph_viz_pass");
@@ -55,10 +71,6 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
       viz_pass->Set<std::string>("graph_viz_path", new std::string(graph_path));
     }
 
-    // Add op fusion.
-    if (strategy.fuse_relu_depthwise_conv_) {
-      AppendPass("fuse_relu_depthwise_conv_pass");
-    }
     if (strategy.fuse_elewise_add_act_ops_) {
       auto fuse_elewise_add_act_pass = AppendPass("fuse_elewise_add_act_pass");
       // Add a graph viz pass to record a graph.
@@ -88,7 +100,7 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
     // A side-effect of that, memory optimize cannot forsee the fetched vars
     // , so fetchlist should be set persistable before call the Run interface.
     if (strategy.memory_optimize_) {
-      auto analysis_var_pass = AppendPass("analysis_var_pass");
+      auto memory_optimize_pass = AppendPass("memory_optimize_pass");
     }
 
     AppendMultiDevPass(strategy);
@@ -186,8 +198,10 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
       pass->Erase("nccl_ctxs");
       pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
 #endif
-
-    } else if (pass->Type() == "analysis_var_pass") {
+    } else if (pass->Type() == "memory_optimize_pass") {
+      if (graph->Has(kAllOpDescs)) {
+        graph->Erase(kAllOpDescs);
+      }
       const std::vector<OpDesc *> *all_op_descs =
           new std::vector<OpDesc *>(main_program.Block(0).AllOps());
       graph->Set<const std::vector<OpDesc *>>(kAllOpDescs,
@@ -214,6 +228,13 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
       pass->Set<const std::vector<OpDesc *>>(
           kAllOpDescs,
           new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
+    } else if (pass->Type() == "inplace_pass") {
+      if (graph->Has(kAllOpDescs)) {
+        graph->Erase(kAllOpDescs);
+      }
+      graph->Set<const std::vector<OpDesc *>>(
+          kAllOpDescs,
+          new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
     } else if (pass->Type() == "fuse_relu_depthwise_conv_pass") {
       if (!use_cuda) {
         LOG(WARNING) << "fuse_relu_depthwise_conv_pass is only supported on "
@@ -239,9 +260,10 @@ USE_PASS(allreduce_mode_multi_devices_pass);
 USE_PASS(dist_multi_devices_pass);
 USE_PASS(multi_devices_check_pass);
 USE_PASS(multi_devices_print_pass);
-USE_PASS(analysis_var_pass);
+USE_PASS(memory_optimize_pass);
 USE_PASS(sequential_execution_pass);
 USE_PASS(all_reduce_deps_pass);
 USE_PASS(modify_op_lock_and_record_event_pass);
+USE_PASS(inplace_pass);
 USE_PASS(lock_free_optimize_pass);
 USE_PASS(graph_to_program_pass);

paddle/fluid/framework/details/build_strategy.h

@@ -80,6 +80,11 @@ struct BuildStrategy {
 
   bool memory_early_delete_{false};
 
+  // TODO(dzhwinter):
+  // make enable_inplace, memory_optimize_
+  // memory_early_delete_ true by default
+  bool enable_inplace_{false};
+
   bool enable_sequential_execution_{false};
 
   bool fuse_broadcast_op_{false};

paddle/fluid/framework/details/graph_test_base.h

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <algorithm>
+#include <iostream>
+#include <iterator>
+#include <string>
+#include "glog/logging.h"
+#include "gtest/gtest.h"
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/program_desc.h"
+
+namespace paddle {
+namespace framework {
+
+class DummyOp : public OperatorBase {
+ public:
+  DummyOp(const std::string& type, const VariableNameMap& inputs,
+          const VariableNameMap& outputs, const AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+
+ private:
+  void RunImpl(const Scope& scope,
+               const platform::Place& place) const override {}
+};
+
+class SumOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  void Make() {
+    AddInput("X", "").AsDuplicable();
+    AddOutput("Out", "");
+    AddComment("");
+  }
+};
+
+class AssignOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  void Make() {
+    AddInput("X", "").AsDuplicable();
+    AddOutput("Out", "");
+    AddComment("");
+  }
+};
+
+class SplitOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  void Make() {
+    AddInput("X", "");
+    AddOutput("Out", "").AsDuplicable();
+    AddComment("");
+  }
+};
+
+class DummyVarTypeInference : public VarTypeInference {
+ public:
+  void operator()(const OpDesc& op_desc, BlockDesc* block) const override {
+    auto& inputs = op_desc.Input("X");
+    auto type = block->Var(inputs.front())->GetType();
+    auto out_var_name = op_desc.Output("Out").front();
+    block->Var(out_var_name)->SetType(type);
+  }
+};
+
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/inplace_op_pass.cc

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/framework/details/inplace_op_pass.h"
+#include <algorithm>
+#include <deque>
+#include <iterator>
+#include <stack>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+#include "paddle/fluid/framework/details/memory_optimize_pass.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+#include "paddle/fluid/framework/op_info.h"
+
+// NOTE(dzhwinter): inplace means one op output variable reuse the input space.
+// By our design, one operator only can read its input(const Variable),
+// write its output(non-const Variable). If one operator is inplaced, means
+// user have chance to write the space before reading happens.
+// Especially when some optimize code writing style is applied.
+//
+//
+// /* wrong case in operator */
+// /*In this case, a larger allocation is allocated, input content is lost*/
+// const Tensor* in = ctx.Input<Tensor>("In")
+// Tensor* out = ctx.Output<Tensor>("Out");
+// auto* out_ptr = out->mutable_data<T>(ctx.GetPlace());
+// out_ptr[0] = 0;  // input contect is overwrited.
+
+// NOTE(dzhwinter):
+// Only for backward compacity and stable. if enable_inplace_whitelist is turn
+// on.
+// only the ops in whitelist will be use inplace strategy.
+// if not, all the op will be inplaced if it registered with InplaceClass
+DEFINE_bool(
+    enable_inplace_whitelist, false,
+    "If this option turns on, only these op in whitelist can be inplaced."
+    "If it turns off, all of the running op can be candidate of inplaced op."
+    "Such as scale, elementwise_add"
+    "By default, it's turned on");
+
+DECLARE_string(memory_optimize_debug);
+
+// clang-format off
+const std::string kInplacedOpWhiteList[] = { // NOLINT
+    "sigmoid",
+    "exp",
+    "relu",
+    "tanh",
+    "sqrt",
+    "ceil",
+    "floor",
+    "reciprocal",
+    "relu6",
+    "soft_relu",
+    "hard_sigmoid",
+    "batch_norm",
+    "batch_norm_grad",
+    "sum",
+    "sum_grad",
+    "scale",
+    "reshape",
+    "elementwise_add",
+    "elementwise_add_grad",
+};
+// clang-format on
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+static inline ir::Node* GetNextCascadeInplacedVar(ir::Node* var) {
+  // if next op is inplaced, then return the output var
+  // otherwise return nullptr
+  PADDLE_ENFORCE(var && var->IsVar() && !var->IsCtrlVar());
+  ir::Node* inplaced_var = nullptr;
+  for (auto* next_op : var->outputs) {
+    for (auto* output : next_op->outputs) {
+      if (output->IsVar() && !output->IsCtrlVar() &&
+          output->Name() == var->Name()) {
+        inplaced_var = output;
+      }
+    }
+  }
+  return inplaced_var;
+}
+
+static inline ir::Node* GetPrevCascadeInplacedVar(ir::Node* var) {
+  PADDLE_ENFORCE(var && var->IsVar() && !var->IsCtrlVar());
+  if (var->inputs.empty()) return nullptr;
+  auto* prev_op = var->inputs.at(0);
+  auto input_it = std::find_if(prev_op->inputs.begin(), prev_op->inputs.end(),
+                               [&](ir::Node* node) {
+                                 if (node->IsVar() && !node->IsCtrlVar() &&
+                                     node->Name() == var->Name()) {
+                                   return true;
+                                 } else {
+                                   return false;
+                                 }
+                               });
+  return input_it == prev_op->inputs.end() ? nullptr : *input_it;
+}
+
+InplacePass::InplacePass() : Pass() {
+  if (FLAGS_enable_inplace_whitelist) {
+    for (auto& s : kInplacedOpWhiteList) {
+      whitelist_.emplace(s);
+    }
+  }
+}
+
+void InplacePass::InitSSAGraphNodes() const {
+  std::unordered_map<std::string, std::unordered_set<ir::Node*>> all_vars;
+  for (auto* op : view_.AllOps()) {
+    for (auto* node : op->inputs) {
+      if (!node->IsVar() || node->IsCtrlVar()) continue;
+      if (all_vars[node->Name()].count(node) == 0) {
+        all_vars[node->Name()].emplace(node);
+        var_nodes_[node->Name()].emplace_back(node);
+      }
+    }
+    for (auto* node : op->outputs) {
+      if (!node->IsVar() || node->IsCtrlVar()) continue;
+      if (all_vars[node->Name()].count(node) == 0) {
+        all_vars[node->Name()].emplace(node);
+        var_nodes_[node->Name()].emplace_back(node);
+      }
+    }
+  }
+}
+
+std::unique_ptr<ir::Graph> InplacePass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  var_nodes_.clear();
+  view_.Build(graph.get());
+  InitSSAGraphNodes();
+
+  for (auto* op : view_.AllOps()) {
+    if (FLAGS_enable_inplace_whitelist && !whitelist_.count(op->Name()))
+      continue;
+    TryInplaceOpInputOutput(op, graph.get());
+  }
+  graph->ResolveHazard(var_nodes_);
+
+  return graph;
+}
+
+void InplacePass::InplaceModifyDesc(const std::string& var,
+                                    const std::string& cache_var,
+                                    const size_t& idx) const {
+  for (size_t i = idx; i < view_.AllOps().size(); ++i) {
+    ir::Node* op = view_.AllOps()[i];
+    PADDLE_ENFORCE(op->IsOp() && op->Op());
+    auto* op_desc = op->Op();
+    op_desc->RenameInput(var, cache_var);
+    op_desc->RenameOutput(var, cache_var);
+    if (op_desc->Block()->HasVar(var)) op_desc->Block()->RemoveVar(var);
+    op_desc->Flush();
+  }
+}
+
+const SSANodePair InplacePass::TryInplaceModifyVar(const std::string& var,
+                                                   const std::string& cache_var,
+                                                   const size_t& idx,
+                                                   ir::Graph* graph) const {
+  PADDLE_ENFORCE(var_nodes_[var].size() >= 1 &&
+                 var_nodes_[var].at(0)->Var() != nullptr);
+  std::unique_ptr<VarDesc> var_desc(new VarDesc(*var_nodes_[var].at(0)->Var()));
+  var_desc->SetName(cache_var);
+
+  SSANodePair swap_nodes;
+
+  for (size_t i = idx; i < view_.AllOps().size(); ++i) {
+    auto* op = view_.AllOps()[i];
+
+    // redirect the input to the latest version of cache_var
+    for (auto* node : op->inputs) {
+      if (node->Name() == var) {
+        ir::Node* cache_node = graph->CreateVarNode(var_desc.get());
+
+        // swap node to cache_node
+        cache_node->outputs.insert(cache_node->outputs.end(),
+                                   node->outputs.begin(), node->outputs.end());
+        PADDLE_ENFORCE(node->inputs.size() == 1 && node->inputs[0]->IsOp());
+        auto* prev_op = node->inputs[0];
+        std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), node,
+                     cache_node);
+        cache_node->inputs.emplace_back(prev_op);
+        for (auto* next_op : node->outputs) {
+          std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
+                       cache_node);
+        }
+
+        swap_nodes.emplace_back(std::make_pair(node, cache_node));
+      }
+    }
+
+    // if we need to rename the output,
+    // always create a newer version of cache_var
+    for (auto* node : op->outputs) {
+      if (node->Name() == var) {
+        ir::Node* cache_node = graph->CreateVarNode(var_desc.get());
+        // swap node to cache node
+        cache_node->outputs.insert(cache_node->outputs.end(),
+                                   node->outputs.begin(), node->outputs.end());
+        cache_node->inputs.emplace_back(op);
+        std::replace(op->outputs.begin(), op->outputs.end(), node, cache_node);
+        for (auto* next_op : node->outputs) {
+          std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
+                       cache_node);
+        }
+
+        swap_nodes.emplace_back(std::make_pair(node, cache_node));
+      }
+    }
+  }
+
+  return swap_nodes;
+}
+
+void InplacePass::CommitModify(const SSANodePair& swap_nodes,
+                               ir::Graph* graph) const {
+  for (auto& pair : swap_nodes) {
+    auto *node = pair.first, *cache_node = pair.second;
+    const std::string var = node->Name(), cache_var = cache_node->Name();
+    var_nodes_[cache_var].emplace_back(cache_node);
+    graph->RemoveNode(node);
+    auto& nodes = var_nodes_.at(var);
+    // release unused var in graph. Because python side memory optimize
+    // may reused the var in same name, so we only clear the var node
+    // after current inplaced index.
+    nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end());
+  }
+}
+
+void InplacePass::WithdrawModify(const SSANodePair& nodes,
+                                 ir::Graph* graph) const {
+  for (auto& pair : nodes) {
+    auto *node = pair.first, *cache_node = pair.second;
+    const std::string var = node->Name(), cache_var = cache_node->Name();
+    auto* prev_op = node->inputs[0];
+    std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), cache_node,
+                 node);
+    for (auto* next_op : node->outputs) {
+      std::replace(next_op->inputs.begin(), next_op->inputs.end(), cache_node,
+                   node);
+    }
+    graph->RemoveNode(cache_node);
+  }
+}
+
+void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
+                                          ir::Graph* graph) const {
+  VLOG(4) << "Try to inplace op " << op->Name();
+  PADDLE_ENFORCE(op->Op() != nullptr && op->Op()->Block() != nullptr,
+                 "op_desc is nullptr");
+  // some pre-requirments need to meet if the op want to inplaced.
+
+  auto* op_desc = op->Op();
+  auto& infer_inplace =
+      OpInfoMap::Instance().Get(op_desc->Type()).infer_inplace_;
+
+  // 1. infer_inplace_ is registered.
+  if (!static_cast<bool>(infer_inplace)) return;
+  PADDLE_ENFORCE(static_cast<bool>(infer_inplace),
+                 "%s's infer_inplace has not been registered", op_desc->Type());
+
+  auto* block = op_desc->Block();
+  auto in_to_outs = infer_inplace(*op_desc, block);
+
+  auto& all_ops = view_.AllOps();
+  auto cursor = std::find(all_ops.begin(), all_ops.end(), op);
+  size_t idx = std::distance(all_ops.begin(), cursor);
+
+  for (auto& pair : in_to_outs) {
+    auto& in_var_name = pair.first;
+    auto& out_var_name = pair.second;
+    auto* in_node = view_.GetNodeByName(in_var_name, op->inputs);
+    auto* out_node = view_.GetNodeByName(out_var_name, op->outputs);
+
+    // 2. there is no external pending op on the input node
+    if (view_.PendingOpsOnVar(in_node).size() > 1) {
+      VLOG(4) << string::Sprintf(
+          "Skiped pair %s => %s. %s input has external dependency."
+          "inplace such pair will overwrite the memory.",
+          out_var_name, in_var_name, op->Name());
+      continue;
+    }
+
+    // 3. if output has been memory optimize by python(fluid.memory_optmize()).
+    // this candidate  can not be inplaced. Will be deprecated in the future.
+    if (view_.InSkipSet(out_node->Name())) {
+      VLOG(4) << string::Sprintf(
+          "Skiped %s => %s reused previous memory block in python memory "
+          "optmize,"
+          "it inplace may generate a circle",
+          out_var_name, in_var_name, op->Name());
+      continue;
+    }
+
+    // Debug Interface. Which would be skipped by the pass.
+    if (out_node->Name() == FLAGS_memory_optimize_debug) {
+      VLOG(3) << "Skiped var by force. FLAGS_memory_optimize_debug="
+              << out_node->Name();
+      continue;
+    }
+
+    // NOTE(dzhwinter):
+    // two stage commit of inplaced process. if after inplace happens generate a
+    // circle,
+    // then withdraw the changes. Otherwise, safely add the node.
+    auto swap_nodes =
+        TryInplaceModifyVar(out_var_name, in_var_name, idx, graph);
+
+    if (!ir::HasCircle(*graph)) {
+      VLOG(3) << string::Sprintf("!!! %s,  %s => %s inplaced", op->Name(),
+                                 out_var_name, in_var_name);
+      InplaceModifyDesc(out_var_name, in_var_name, idx);
+      CommitModify(swap_nodes, graph);
+    } else {
+      VLOG(3) << string::Sprintf(
+          "Skiped pair %s => %s, inplace will generate a circle. withdraw %s",
+          out_var_name, in_var_name, op->Name());
+      WithdrawModify(swap_nodes, graph);
+    }
+  }
+}
+
+ir::Node* GraphView::GetNodeByName(const std::string& name,
+                                   const std::vector<ir::Node*>& nodes) const {
+  // nodes should be op->inputs/outputs
+  // node in same node do have different name.
+  std::unordered_set<std::string> nodes_in_op;
+  bool has_dup_node =
+      std::all_of(nodes.begin(), nodes.end(), [&nodes_in_op](ir::Node* node) {
+        if (!node->IsVar() || node->IsCtrlVar() || node->Var() == nullptr) {
+          if (nodes_in_op.count(node->Name())) return true;
+          nodes_in_op.emplace(node->Name());
+        }
+        return false;
+      });
+  PADDLE_ENFORCE(has_dup_node == false, "nodes has same name!");
+  ir::Node* node = nullptr;
+  for (auto* it : nodes) {
+    if (!it->IsVar() || it->IsCtrlVar() || it->Var() == nullptr) continue;
+    if (it->Name() == name) {
+      node = it;
+      break;
+    }
+  }
+  PADDLE_ENFORCE(node != nullptr,
+                 string::Sprintf("Not found var %s in nodes!", name));
+  return node;
+}
+
+std::vector<ir::Node*> GraphView::PendingOpsOnVar(ir::Node* node) {
+  // get the pending ops depends on same var node.
+  // because node also maybe a inplaced variable, so need to backtrack all the
+  // previous inplaced vars.
+  std::vector<ir::Node*> pending_ops;
+  ir::Node* p = node;
+  while (p != nullptr) {
+    pending_ops.insert(pending_ops.end(), p->outputs.begin(), p->outputs.end());
+    p = GetPrevCascadeInplacedVar(p);
+  }
+  return pending_ops;
+}
+
+void GraphView::Build(ir::Graph* g) {
+  // track the var nodes in correct order.
+  // Because we insert some new created node. Which may have data race between
+  // nodes.
+  // resolve data harzards depends on the var nodes in right order.
+  ops_ = SortOpLikeDescOrder(*g);
+
+  // 1. track the nodes which reused previous node in Python memory optimize.
+  // these node can not be inplaced, otherwise may generate a circle in graph.
+  std::unordered_set<std::string> all_vars;
+  for (auto& node : g->Nodes()) {
+    if (node->IsVar()) continue;
+    for (auto& out : node->outputs) {
+      if (out->IsCtrlVar() || out->Var() == nullptr) continue;
+      if (all_vars.count(out->Name())) {
+        dup_nodes_.emplace(out->Name());
+      } else {
+        all_vars.emplace(out->Name());
+      }
+    }
+  }
+
+  // 2. track the nodes which used by parameter server.
+  // these node can not be inplaced, otherwise trainer
+  // pserver can not find each other name.
+  auto update_skip_set = [&](ir::Node* node) {
+    for (auto& in : node->inputs) {
+      if (in->IsVar() && in->Var() != nullptr) dup_nodes_.emplace(in->Name());
+    }
+    for (auto& out : node->outputs) {
+      if (out->IsVar() && out->Var() != nullptr)
+        dup_nodes_.emplace(out->Name());
+    }
+  };
+  for (auto& node : g->Nodes()) {
+    if (!node->IsOp()) continue;
+    if (node->Name() == "send") update_skip_set(node);
+    if (node->Name() == "recv") update_skip_set(node);
+    if (node->Name() == "prefetch") update_skip_set(node);
+  }
+}
+
+const std::vector<ir::Node*>& GraphView::AllOps() { return ops_; }
+
+bool GraphView::InSkipSet(const std::string& var) const {
+  return dup_nodes_.count(var);
+}
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(inplace_pass, paddle::framework::details::InplacePass);

paddle/fluid/framework/details/inplace_op_pass.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may abtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+#include <map>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+#include "paddle/fluid/framework/details/memory_optimize_helper.h"
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/pass.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+class GraphView {
+ public:
+  GraphView() = default;
+
+  void Build(ir::Graph* g);
+
+  const std::vector<ir::Node*>& AllOps();
+
+  ir::Node* GetNodeByName(const std::string& name,
+                          const std::vector<ir::Node*>& nodes) const;
+
+  std::vector<ir::Node*> PendingOpsOnVar(ir::Node* var);
+
+  // Will Deperated in the future.
+  // NOTE(dzhwinter) :
+  // 1. Python memory optimize will reuse
+  // memory based var name, so different op output may
+  // have the same variable name. enable inplace on such node
+  // will generate a circle in ssa graph.
+  // 2. DistributeTranspiler will use unique name to
+  // map the parameter and gradient, must be skipped.
+  bool InSkipSet(const std::string& var) const;
+
+ private:
+  std::vector<ir::Node*> ops_;
+  std::unordered_set<std::string> dup_nodes_;  // mem opt affect nodes
+  std::map<ir::Node*, std::unordered_set<ir::Node*>> adj_list_;
+};
+
+typedef std::vector<std::pair<ir::Node*, ir::Node*>> SSANodePair;
+class InplacePass : public ir::Pass {
+ public:
+  InplacePass();
+
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override;
+
+  void InitSSAGraphNodes() const;
+
+ private:
+  const SSANodePair TryInplaceModifyVar(const std::string& var,
+                                        const std::string& cache_var,
+                                        const size_t& idx,
+                                        ir::Graph* graph) const;
+
+  void CommitModify(const SSANodePair&, ir::Graph* graph) const;
+
+  void WithdrawModify(const SSANodePair& nodes, ir::Graph* graph) const;
+
+  void InplaceModifyDesc(const std::string& in_var, const std::string& out_var,
+                         const size_t& idx) const;
+
+  void TryInplaceOpInputOutput(ir::Node* op, ir::Graph* graph) const;
+
+  mutable std::map<std::string, std::vector<ir::Node*>> var_nodes_;
+
+  mutable std::unordered_set<std::string> whitelist_;
+  mutable GraphView view_;
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/memory_early_delete_pass.cc

@@ -16,7 +16,7 @@
 #include <queue>
 #include <string>
 #include <vector>
-#include "paddle/fluid/framework/details/memory_reuse_types.h"
+#include "paddle/fluid/framework/details/memory_optimize_helper.h"
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
 #include "paddle/fluid/framework/details/reference_count_pass_helper.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"

paddle/fluid/framework/details/memory_reuse_types.cc → paddle/fluid/framework/details/memory_optimize_helper.cc

@@ -12,8 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/details/memory_reuse_types.h"
+#include "paddle/fluid/framework/details/memory_optimize_helper.h"
+#include <functional>
 #include <iostream>
+#include <numeric>
 #include <sstream>
 #include <string>
 
@@ -21,15 +23,17 @@ namespace paddle {
 namespace framework {
 namespace details {
 
+size_t NodeSizeInBytes(const VarDesc& node) {
+  auto shape = node.GetShape();
+  int size =
+      std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>());
+  size_t type_size = SizeOfType(node.GetDataType());
+  return type_size * std::abs(size);
+}
+
 size_t NodeSizeInBytes(ir::Node* n) {
   auto* desc = FindVarDescInBlock(n);
-  auto shape = desc->GetShape();
-  size_t type_size = SizeOfType(desc->GetDataType());
-  int size = 1;
-  for (auto& s : shape) {
-    size *= s;
-  }
-  return type_size * std::abs(size);
+  return NodeSizeInBytes(*desc);
 }
 
 std::string DebugStringImpl(VarDesc* var) {
@@ -83,7 +87,7 @@ struct NodeComparator {
   }
 };
 
-void OrderedNodePairPool::Insert(ir::Node* var, ir::Node* op) {
+void OrderedNodeList::Insert(ir::Node* var, ir::Node* op) {
   PADDLE_ENFORCE(var->IsVar() && !var->IsCtrlVar());
   PADDLE_ENFORCE(op->IsOp());
   if (mark_table_.count(var->Name()) != 0) {
@@ -119,11 +123,11 @@ void OrderedNodePairPool::Insert(ir::Node* var, ir::Node* op) {
   mark_table_[var->Name()] = it;
 }
 
-int OrderedNodePairPool::GetIndex(ir::Node* var) {
+int OrderedNodeList::GetIndex(ir::Node* var) {
   return std::distance(nodes_.begin(), mark_table_[var->Name()]);
 }
 
-ir::Node* OrderedNodePairPool::NodeMatch(ir::Node* var) const {
+ir::Node* OrderedNodeList::NodeMatch(ir::Node* var) const {
   ir::Node* found_node = nullptr;
   NodeComparator compare_node;
 
@@ -136,13 +140,15 @@ ir::Node* OrderedNodePairPool::NodeMatch(ir::Node* var) const {
   return found_node;
 }
 
-void OrderedNodePairPool::Erase(ir::Node* var) {
-  PADDLE_ENFORCE(mark_table_.count(var->Name()));
-  nodes_.erase(mark_table_[var->Name()]);
-  mark_table_.erase(var->Name());
+void OrderedNodeList::Erase(ir::Node* var) { Erase(var->Name()); }
+
+void OrderedNodeList::Erase(const std::string& var) {
+  PADDLE_ENFORCE(mark_table_.count(var));
+  nodes_.erase(mark_table_[var]);
+  mark_table_.erase(var);
 }
 
-std::string OrderedNodePairPool::ToString() const {
+std::string OrderedNodeList::ToString() const {
   std::stringstream ss;
   for (auto it = nodes_.begin(); it != nodes_.end(); ++it) {
     ss << DebugString(it->first) << " ";
@@ -150,6 +156,43 @@ std::string OrderedNodePairPool::ToString() const {
   return ss.str();
 }
 
+bool NodeCanReused(ir::Node* node) {
+  if (node == nullptr || !node->IsVar() || node->IsCtrlVar()) return false;
+  // auto* desc = node->Var();
+  bool flag = NodeCanReused(*node->Var());
+  for (auto* op : node->inputs) {
+    if (op->Op()->HasAttr("force_cpu")) {
+      // op output force generated in cpu, can not be reused.
+      flag &= framework::AttrReader(op->Op()->GetAttrMap())
+                  .Get<bool>("force_cpu") == 0;
+    }
+  }
+  return flag;
+}
+
+bool NodeCanReused(const VarDesc& node) {
+  auto type = node.GetType();
+  if (node.Persistable() || type != proto::VarType::LOD_TENSOR ||
+      node.GetShape().empty()) {
+    return false;
+  }
+  // vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad
+  std::string name = node.Name();
+  if (!name.empty() && name[0] == '@' && name[name.size() - 1] == '@')
+    return false;
+  return true;
+}
+
+bool OpHasSubBlock(OpDesc* desc) {
+  const AttributeMap& attrs = desc->GetAttrMap();
+  for (auto& attr : attrs) {
+    if (attr.second.type() == typeid(BlockDesc*) ||             // NOLINT
+        attr.second.type() == typeid(std::vector<BlockDesc*>))  // NOLINT
+      return true;
+  }
+  return false;
+}
+
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/memory_reuse_types.h → paddle/fluid/framework/details/memory_optimize_helper.h

@@ -43,7 +43,7 @@ using GraphNodePool = std::vector<
 // For example,
 // node0[-1, 1] node1[-1, 1, 1], node2[1,1], node3[1,1024], ..
 // O(1) insert, delete
-class OrderedNodePairPool {
+class OrderedNodeList {
  public:
   using NodePair = std::pair<ir::Node*, std::unordered_set<ir::Node*>>;
   using Iter = typename std::list<NodePair>::iterator;
@@ -53,8 +53,12 @@ class OrderedNodePairPool {
 
   void Erase(ir::Node* var);
 
+  void Erase(const std::string& var);
+
   bool Has(ir::Node* var) { return mark_table_.count(var->Name()); }
 
+  bool Has(const std::string& var) { return mark_table_.count(var); }
+
   ir::Node* NodeMatch(ir::Node* var) const;
   // map store non-const iterator, can not promise const
   int GetIndex(ir::Node* var);
@@ -67,6 +71,11 @@ class OrderedNodePairPool {
   ConstIter end() const { return nodes_.end(); }
   size_t size() const { return nodes_.size(); }
 
+  void Clear() {
+    mark_table_.clear();
+    nodes_.clear();
+  }
+
  private:
   // for searching.
   std::unordered_map<std::string, Iter> mark_table_;
@@ -74,14 +83,53 @@ class OrderedNodePairPool {
   std::list<NodePair> nodes_;
 };
 
+// valid a tensor can be reuse or not
+bool NodeCanReused(ir::Node* node);
+
+// valid a tensor can be reuse or not.
+bool NodeCanReused(const VarDesc& node);
+
+// check op has subblock or not
+bool OpHasSubBlock(OpDesc* desc);
+
 // node memory size in bytes
 size_t NodeSizeInBytes(ir::Node* n);
 
+// node memory size in bytes
+size_t NodeSizeInBytes(const VarDesc&);
+
 std::string DebugString(ir::Node* var);
 
-// std::string DebugString(VarDesc* var);
 VarDesc* FindVarDescInBlock(ir::Node* n);
 
+template <typename Container, typename Callback>
+class FilterVariableImpl {
+ public:
+  void operator()(const Container& nodes, Callback callback) {
+    for (auto* node : nodes) {
+      callback(node);
+    }
+  }
+};
+
+// filter var node for op->inputs/outputs
+template <typename Callback>
+class FilterVariableImpl<std::vector<ir::Node*>, Callback> {
+ public:
+  void operator()(const std::vector<ir::Node*>& nodes, Callback callback) {
+    for (auto* var : nodes) {
+      if (var->IsVar() && !var->IsCtrlVar()) {
+        callback(var);
+      }
+    }
+  }
+};
+
+template <typename Container, typename Callback>
+void FilterVariables(const Container& nodes, Callback callback) {
+  FilterVariableImpl<Container, Callback>()(nodes, callback);
+}
+
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/memory_reuse_types_test.cc → paddle/fluid/framework/details/memory_optimize_helper_test.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/details/memory_reuse_types.h"
+#include "paddle/fluid/framework/details/memory_optimize_helper.h"
 #include <algorithm>
 #include <iostream>
 #include <memory>
@@ -27,8 +27,8 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-TEST(OrderedNodePairPool, Normal) {
-  OrderedNodePairPool pool;
+TEST(OrderedNodeList, Normal) {
+  OrderedNodeList pool;
   std::vector<std::unique_ptr<ir::Node>> nodes;
 
   // clang-format off

paddle/fluid/framework/details/analysis_var_pass.cc → paddle/fluid/framework/details/memory_optimize_pass.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/details/analysis_var_pass.h"
+#include "paddle/fluid/framework/details/memory_optimize_pass.h"
 #include <algorithm>
 #include <atomic>
 #include <deque>
@@ -48,39 +48,10 @@ static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) {
          op1->Outputs() == op2->Outputs();
 }
 
-template <typename Container, typename Callback>
-class FilterVariableImpl {
- public:
-  void operator()(const Container& nodes, Callback callback) {
-    for (auto* node : nodes) {
-      callback(node);
-    }
-  }
-};
-
-// filter var node for op->inputs/outputs
-template <typename Callback>
-class FilterVariableImpl<std::vector<ir::Node*>, Callback> {
- public:
-  void operator()(const std::vector<ir::Node*>& nodes, Callback callback) {
-    for (auto* var : nodes) {
-      if (var->IsVar() && !var->IsCtrlVar()) {
-        callback(var);
-      }
-    }
-  }
-};
-
-template <typename Container, typename Callback>
-void FilterVariables(const Container& nodes, Callback callback) {
-  FilterVariableImpl<Container, Callback>()(nodes, callback);
-}
-
-std::unique_ptr<ir::Graph> AnalysisVarPass::ApplyImpl(
+std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
     std::unique_ptr<ir::Graph> graph) const {
   auto nodes = graph->Nodes();
-  auto subblock_vars = GetSubBlockVars(nodes);
-  skip_set_.insert(subblock_vars.begin(), subblock_vars.end());
+  CollectSkipVarsSet(nodes);
 
   cfg_.reset(new details::ControlFlowGraph(*graph));
   cfg_->LiveVariableAnalysis();
@@ -103,48 +74,53 @@ std::unique_ptr<ir::Graph> AnalysisVarPass::ApplyImpl(
     }
 
     for (auto& var : op->outputs) {
-      if (NodeCanReused(var) && cfg_->Use(op).count(var->Name()) == 0) {
-        ir::Node* cache = pool_.NodeMatch(var);
-        if (var->Name() == FLAGS_memory_optimize_debug) {
-          VLOG(3) << "start match var " << DebugString(var) << " of op "
-                  << op->Name();
-          VLOG(3) << pool_.ToString();
-          VLOG(3) << "matched in pool : "
-                  << ((cache == nullptr) ? "False" : "True");
-        }
-        if (cache != nullptr) {
-          if (var->Name() == cache->Name()) {
-            VLOG(3) << "The same cache variable is cascade reused."
-                    << var->Name() << " is re-filled to the pool after"
-                    << "the reused op is finished. Current op can not "
-                    << "replace it again. Skip this candidate.";
-            continue;
-          }
+      if (!NodeCanReused(var) || cfg_->Use(op).count(var->Name()) == 0 ||
+          skip_set_.count(var->Name()))
+        continue;
+      ir::Node* cache = pool_.NodeMatch(var);
+
+      if (var->Name() == FLAGS_memory_optimize_debug) {
+        VLOG(3) << "start match var " << DebugString(var) << " of op "
+                << op->Name();
+        VLOG(3) << pool_.ToString();
+        VLOG(3) << "matched in pool : "
+                << ((cache == nullptr) ? "False" : "True");
+      }
 
-          int node_idx_in_pool = pool_.GetIndex(cache);
-          VLOG(3) << string::Sprintf(
-              "!!! %s,  %s => %s, cache idx %d, pool size %d",
-              std::to_string(reuse_id++), DebugString(var), DebugString(cache),
-              node_idx_in_pool, static_cast<int>(pool_.size()));
-          // update CFG Graph on the fly.
-          // reused var maybe re-fill into the pool
-          cfg_->RenameVarInCFGGraph(var->Name(), cache->Name(), idx);
-          // NOTE(dzhwinter): we need to both update the ProgramDesc
-          // and IR Graph. because op_desc/var_desc is used in CreateOp,
-          // CreateVar when running happens. But IR Graph
-          // define the dependence relationship between nodes.
-          RenameVarInGraphDesc(var->Name(), cache->Name(), idx);
-          RenameVarInGraphNode(var->Name(), cache->Name(), idx, graph.get());
-
-          pool_.Erase(cache);
+      if (cache == nullptr) continue;
+      if (var->Name() == cache->Name()) {
+        VLOG(3) << "The same cache variable is cascade reused." << var->Name()
+                << " is re-filled to the pool after"
+                << "the reused op is finished. Current op can not "
+                << "replace it again. Skip this candidate.";
+        continue;
+
+        int node_idx_in_pool = pool_.GetIndex(cache);
+        VLOG(3) << string::Sprintf(
+            "!!! %s,  %s => %s, cache idx %d, pool size %d",
+            std::to_string(reuse_id++), DebugString(var), DebugString(cache),
+            node_idx_in_pool, static_cast<int>(pool_.size()));
+        // update CFG Graph on the fly.
+        // reused var maybe re-fill into the pool
+        cfg_->RenameVarInCFGGraph(var->Name(), cache->Name(), idx);
+        // NOTE(dzhwinter): we need to both update the ProgramDesc
+        // and IR Graph. because op_desc/var_desc is used in CreateOp,
+        // CreateVar when running happens. But IR Graph
+        // define the dependence relationship between nodes.
+        RenameVarInGraphDesc(var->Name(), cache->Name(), idx);
+        RenameVarInGraphNode(var->Name(), cache->Name(), idx, graph.get());
+
+        pool_.Erase(cache);
+      }
+      // fill the pool
+      std::unordered_set<std::string> unlived_vars;
+      for (auto var : cfg_->LiveIn(op)) {
+        if (cfg_->LiveOut(op).count(var) == 0) {
+          unlived_vars.emplace(var);
         }
       }
-    }
-    // fill the pool
-    for (auto var : cfg_->LiveIn(op)) {
-      if (cfg_->LiveOut(op).count(var) == 0) {
+      for (auto var : unlived_vars) {
         ir::Node* var_node = cfg_->GetNodeFromVarName(var, op);
-        if (var_node == nullptr) continue;
         if (NodeCanReused(var_node) && !pool_.Has(var_node)) {
           pool_.Insert(var_node, op);
         }
@@ -177,7 +153,7 @@ std::unique_ptr<ir::Graph> AnalysisVarPass::ApplyImpl(
   return graph;
 }
 
-void AnalysisVarPass::SubGraphOptimize(OpDesc* op_desc) const {
+void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const {
   // conditional block, while op and their grad op
   auto* sub_block_desc =
       AttrReader(op_desc->GetAttrMap()).Get<BlockDesc*>("sub_block");
@@ -247,25 +223,32 @@ void AnalysisVarPass::SubGraphOptimize(OpDesc* op_desc) const {
   }
 }
 
-std::unordered_set<std::string> AnalysisVarPass::GetSubBlockVars(
+void MemoryOptimizePass::CollectSkipVarsSet(
     const std::unordered_set<ir::Node*>& nodes) const {
-  std::unordered_set<std::string> vars;
+  auto update_skip_set = [&](OpDesc* op_desc) {
+    auto inputs = op_desc->InputArgumentNames();
+    auto outputs = op_desc->OutputArgumentNames();
+    skip_set_.insert(inputs.begin(), inputs.end());
+    skip_set_.insert(outputs.begin(), outputs.end());
+  };
   for (auto& op : nodes) {
     if (!op->IsOp() || op->Op() == nullptr) continue;
     auto* op_desc = op->Op();
-    if (OpHasSubBlock(op_desc)) {
-      auto inputs = op_desc->InputArgumentNames();
-      auto outputs = op_desc->OutputArgumentNames();
-      vars.insert(inputs.begin(), inputs.end());
-      vars.insert(outputs.begin(), outputs.end());
-    }
+    // NOTE(dzhwinter):
+    // current block can not reuse next level block vars.
+    if (OpHasSubBlock(op_desc)) update_skip_set(op_desc);
+    // NOTE(dzhwinter):
+    // distributed ops input/output name need to
+    // keep same bettwen trainer/pserver
+    if (op_desc->Type() == "send") update_skip_set(op_desc);
+    if (op_desc->Type() == "recv") update_skip_set(op_desc);
+    if (op_desc->Type() == "prefetch") update_skip_set(op_desc);
   }
-  return vars;
 }
 
-void AnalysisVarPass::RenameVarInGraphDesc(const std::string& var,
-                                           const std::string& cache_var,
-                                           size_t idx) const {
+void MemoryOptimizePass::RenameVarInGraphDesc(const std::string& var,
+                                              const std::string& cache_var,
+                                              size_t idx) const {
   for (size_t i = idx; i < cfg_->Ops().size(); ++i) {
     auto* op = cfg_->Ops()[i];
     PADDLE_ENFORCE(op->IsOp() && op->Op());
@@ -277,7 +260,7 @@ void AnalysisVarPass::RenameVarInGraphDesc(const std::string& var,
   }
 }
 
-void AnalysisVarPass::InitSSAGraphNodes() const {
+void MemoryOptimizePass::InitSSAGraphNodes() const {
   std::unordered_map<std::string, std::unordered_set<ir::Node*>> all_vars;
   if (var_nodes_.empty()) {
     for (auto* op : cfg_->Ops()) {
@@ -297,9 +280,10 @@ void AnalysisVarPass::InitSSAGraphNodes() const {
   }
 }
 
-void AnalysisVarPass::RenameVarInGraphNode(const std::string& var,
-                                           const std::string& cache_var,
-                                           size_t idx, ir::Graph* graph) const {
+void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var,
+                                              const std::string& cache_var,
+                                              size_t idx,
+                                              ir::Graph* graph) const {
   // if replace happens, we need to create a newer version cache_var
   // but use the same dims/data_type with var.
   PADDLE_ENFORCE(var_nodes_[var].size() >= 1 &&
@@ -358,39 +342,6 @@ void AnalysisVarPass::RenameVarInGraphNode(const std::string& var,
   var_nodes_.at(var).clear();
 }
 
-bool AnalysisVarPass::NodeCanReused(ir::Node* node) const {
-  if (!node->IsVar() || node->IsCtrlVar()) return false;
-  auto* desc = node->Var();
-  auto type = desc->GetType();
-  if (desc->Persistable() || type != proto::VarType::LOD_TENSOR ||
-      desc->GetShape().empty()) {
-    return false;
-  }
-  // vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad
-  std::string name = node->Name();
-  if (!name.empty() && name[0] == '@' && name[name.size() - 1] == '@')
-    return false;
-  if (skip_set_.count(name)) return false;
-  for (auto* op : node->inputs) {
-    if (op->Op()->HasAttr("force_cpu")) {
-      // op output force generated in cpu, can not be reused.
-      return framework::AttrReader(op->Op()->GetAttrMap())
-                 .Get<bool>("force_cpu") == 0;
-    }
-  }
-  return true;
-}
-
-bool AnalysisVarPass::OpHasSubBlock(OpDesc* desc) const {
-  const AttributeMap& attrs = desc->GetAttrMap();
-  for (auto& attr : attrs) {
-    if (attr.second.type() == typeid(BlockDesc*) ||             // NOLINT
-        attr.second.type() == typeid(std::vector<BlockDesc*>))  // NOLINT
-      return true;
-  }
-  return false;
-}
-
 std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph) {
   PADDLE_ENFORCE(graph.Has(kAllOpDescs),
                  "Graph has no attribute of kAllOpDescs.");
@@ -651,6 +602,7 @@ ir::Node* ControlFlowGraph::GetNodeFromVarName(const std::string& name,
 }  // namespace framework
 }  // namespace paddle
 
-REGISTER_PASS(analysis_var_pass, paddle::framework::details::AnalysisVarPass)
+REGISTER_PASS(memory_optimize_pass,
+              paddle::framework::details::MemoryOptimizePass)
     .RequireGraphAttr(paddle::framework::details::kGraphNodePool)
     .RequireGraphAttr(paddle::framework::details::kAllOpDescs);

paddle/fluid/framework/details/analysis_var_pass.h → paddle/fluid/framework/details/memory_optimize_pass.h

@@ -25,7 +25,7 @@
 #include <vector>
 
 #include "paddle/fluid/framework/data_type.h"
-#include "paddle/fluid/framework/details/memory_reuse_types.h"
+#include "paddle/fluid/framework/details/memory_optimize_helper.h"
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/pass.h"
 
@@ -35,12 +35,10 @@ namespace details {
 constexpr char kAllOpDescs[] = "all_op_descs";
 
 std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph);
-// sort op in bfs order
-std::vector<ir::Node*> BFSSortGraphOps(const ir::Graph& graph);
 
 class ControlFlowGraph;
 
-class AnalysisVarPass : public ir::Pass {
+class MemoryOptimizePass : public ir::Pass {
  protected:
   std::unique_ptr<ir::Graph> ApplyImpl(
       std::unique_ptr<ir::Graph> graph) const override;
@@ -57,17 +55,14 @@ class AnalysisVarPass : public ir::Pass {
                             ir::Graph* graph) const;
 
   void SubGraphOptimize(OpDesc* op_desc) const;
-  // valid a tensor can be reuse or not
-  bool NodeCanReused(ir::Node* node) const;
-  // scan subblock and collect the output/input variables.
-  std::unordered_set<std::string> GetSubBlockVars(
-      const std::unordered_set<ir::Node*>&) const;
-  // check op has subblock or not
-  bool OpHasSubBlock(OpDesc* desc) const;
+  // 1. scan op with subblock and collect the output/input vars.
+  // while, while_grad, conditional_block
+  // 2. scan distributed ops and collect the output/input vars
+  void CollectSkipVarsSet(const std::unordered_set<ir::Node*>&) const;
 
  private:
   // Reuse Node Pool, Owned.
-  mutable OrderedNodePairPool pool_;
+  mutable OrderedNodeList pool_;
   // controlflow Graph
   mutable std::unique_ptr<ControlFlowGraph> cfg_;
   // skip set

paddle/fluid/framework/details/analysis_var_pass_test.cc → paddle/fluid/framework/details/memory_optimize_pass_test.cc

@@ -12,63 +12,19 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/details/analysis_var_pass.h"
+#include "paddle/fluid/framework/details/memory_optimize_pass.h"
 #include <algorithm>
 #include <iostream>
 #include <iterator>
 #include "glog/logging.h"
 #include "gtest/gtest.h"
+#include "paddle/fluid/framework/details/graph_test_base.h"
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/program_desc.h"
 
-namespace paddle {
-namespace framework {
-
-class DummyOp : public OperatorBase {
- public:
-  DummyOp(const std::string& type, const VariableNameMap& inputs,
-          const VariableNameMap& outputs, const AttributeMap& attrs)
-      : OperatorBase(type, inputs, outputs, attrs) {}
-
- private:
-  void RunImpl(const Scope& scope,
-               const platform::Place& place) const override {}
-};
-
-class SumOpMaker : public OpProtoAndCheckerMaker {
- public:
-  void Make() {
-    AddInput("X", "").AsDuplicable();
-    AddOutput("Out", "");
-    AddComment("");
-  }
-};
-
-class AssignOpMaker : public OpProtoAndCheckerMaker {
- public:
-  void Make() {
-    AddInput("X", "").AsDuplicable();
-    AddOutput("Out", "");
-    AddComment("");
-  }
-};
-
-class DummyVarTypeInference : public VarTypeInference {
- public:
-  void operator()(const OpDesc& op_desc, BlockDesc* block) const override {
-    auto& inputs = op_desc.Input("X");
-    auto type = block->Var(inputs.front())->GetType();
-    auto out_var_name = op_desc.Output("Out").front();
-    block->Var(out_var_name)->SetType(type);
-  }
-};
-
-}  // namespace framework
-}  // namespace paddle
-
 REGISTER_OPERATOR(sum, paddle::framework::DummyOp,
                   paddle::framework::SumOpMaker,
                   paddle::framework::DummyVarTypeInference);
@@ -141,15 +97,6 @@ inline static ProgramDesc FillProgramDesc() {
   return prog;
 }
 
-template <typename Container>
-inline static std::string DebugString(const Container& c) {
-  std::stringstream ss;
-  for (auto& item : c) {
-    ss << item << " ";
-  }
-  return ss.str();
-}
-
 TEST(CFGGraph, IRGraph) {
   // prepare ir graph
   auto prog = FillProgramDesc();

paddle/fluid/framework/details/op_registry.h

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <tuple>
 #include <vector>
 #include "paddle/fluid/framework/grad_op_desc_maker.h"
+#include "paddle/fluid/framework/inplace_op_inference.h"
 #include "paddle/fluid/framework/op_info.h"
 #include "paddle/fluid/framework/op_proto_maker.h"
 #include "paddle/fluid/framework/operator.h"
@@ -32,7 +33,8 @@ enum OpInfoFillType {
   kOpProtoAndCheckerMaker = 1,
   kGradOpDescMaker = 2,
   kVarTypeInference = 3,
-  kShapeInference = 4
+  kShapeInference = 4,
+  kInplaceOpInference = 5
 };
 
 template <typename T>
@@ -48,8 +50,11 @@ struct OpInfoFillTypeID {
                                     ? kVarTypeInference
                                     : (std::is_base_of<InferShapeBase, T>::value
                                            ? kShapeInference
-                                           : static_cast<OpInfoFillType>(
-                                                 -1)))));
+                                           : (std::is_base_of<
+                                                  InplaceOpInference, T>::value
+                                                  ? kInplaceOpInference
+                                                  : static_cast<OpInfoFillType>(
+                                                        -1))))));
   }
 };
 
@@ -139,6 +144,16 @@ struct OpInfoFiller<T, kShapeInference> {
   }
 };
 
+template <typename T>
+struct OpInfoFiller<T, kInplaceOpInference> {
+  void operator()(const char* op_type, OpInfo* info) const {
+    info->infer_inplace_ = [](const OpDesc& op_desc, BlockDesc* block) {
+      T infer;
+      return infer(op_desc, block);
+    };
+  }
+};
+
 }  // namespace details
 
 }  // namespace framework

paddle/fluid/framework/inplace_op_inference.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+#include <functional>
+#include <numeric>
+#include <string>
+#include <unordered_map>
+#include "glog/logging.h"
+#include "paddle/fluid/framework/block_desc.h"
+#include "paddle/fluid/framework/details/memory_optimize_helper.h"
+#include "paddle/fluid/framework/op_desc.h"
+#include "paddle/fluid/framework/type_defs.h"
+
+namespace paddle {
+namespace framework {
+
+/*
+  Inplace Inference for create In->Out pairs for inplaced operator.
+  If we specify a pair of corresponding names. For example, X->Out.
+  then Out will inplaced use X's memory. The base class will do
+  legality validation for both variables.
+*/
+class InplaceOpInference {
+ public:
+  virtual ~InplaceOpInference() {}
+  virtual std::unordered_map<std::string, std::string> operator()(
+      const OpDesc& op_desc, BlockDesc* block) const = 0;
+};
+
+class InplaceInToOut : public InplaceOpInference {
+ public:
+  std::unordered_map<std::string, std::string> operator()(
+      const OpDesc& op_desc, BlockDesc* block) const {
+    std::unordered_map<std::string, std::string> ret;
+    auto in_out_var_names_pair = this->Apply(op_desc, block);
+    for (auto& pair : in_out_var_names_pair) {
+      PADDLE_ENFORCE(!op_desc.Input(pair.first).empty(),
+                     string::Sprintf("op %s do not have input of %s!",
+                                     op_desc.Type(), pair.first));
+      PADDLE_ENFORCE(!op_desc.Output(pair.second).empty(),
+                     string::Sprintf("op %s do not have output of %s!",
+                                     op_desc.Type(), pair.second));
+      auto& in_name = op_desc.Input(pair.first).at(0);
+      auto& out_name = op_desc.Output(pair.second).at(0);
+
+      auto in = block->FindRecursiveOrCreateVar(in_name);
+      auto out = block->FindRecursiveOrCreateVar(out_name);
+      if (TryInplaceInputOutput(in, out)) ret.insert({in_name, out_name});
+    }
+    return ret;
+  }
+
+ protected:
+  virtual std::unordered_map<std::string, std::string> Apply(
+      const OpDesc& op_desc, BlockDesc* block) const = 0;
+
+  bool TryInplaceInputOutput(const VarDesc& in, const VarDesc& out) const {
+    return in.Name() != out.Name() && details::NodeCanReused(in) &&
+           details::NodeCanReused(out) &&
+           details::NodeSizeInBytes(out) <= details::NodeSizeInBytes(in);
+  }
+};
+
+/*
+  Inplace In and Out for operator only have an Input and an Output.
+  For example, activation op.
+ */
+class SingleOpInplaceInToOut : public InplaceInToOut {
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const OpDesc& op_desc, BlockDesc* block) const override {
+    PADDLE_ENFORCE(!op_desc.InputNames().empty(),
+                   "Op inputs must not be empty");
+    PADDLE_ENFORCE(!op_desc.OutputNames().empty(),
+                   "Op outputs must not be empty");
+    auto x_name = op_desc.InputNames().at(0);
+    auto out_name = op_desc.OutputNames().at(0);
+    return std::unordered_map<std::string, std::string>{{x_name, out_name}};
+  }
+};
+
+/*
+  Gradient op. Inplace output use it's Input.
+  For example, Input@Grad->Input reuse strategy.
+ */
+class GradOpInplaceInToOut : public InplaceInToOut {
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const OpDesc& op_desc, BlockDesc* block) const override {
+    std::unordered_map<std::string, std::string> ret;
+    std::unordered_set<std::string> output_names(op_desc.OutputNames().begin(),
+                                                 op_desc.OutputNames().end());
+    for (auto& input_name : op_desc.InputNames()) {
+      if (output_names.count(GradVarName(input_name))) {
+        ret.insert({input_name, GradVarName(input_name)});
+      }
+    }
+    return ret;
+  }
+};
+
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/inplace_op_inference_test.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include <iterator>
+#include <string>
+#include "gtest/gtest.h"
+#include "paddle/fluid/framework/op_info.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/framework/var_type_inference.h"
+
+namespace paddle {
+namespace framework {
+
+class NOP : public OperatorBase {
+ public:
+  NOP(const std::string& type, const VariableNameMap& inputs,
+      const VariableNameMap& outputs, const AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+
+ private:
+  void RunImpl(const Scope& scope,
+               const platform::Place& place) const override {}
+};
+
+class SingleOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  void Make() {
+    AddInput("X", "").AsDuplicable();
+    AddOutput("Out", "");
+    AddComment("");
+  }
+};
+
+class SingleGradOpMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto* op = new framework::OpDesc();
+    op->SetType("single_op_grad");
+    op->SetInput("Out", OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
+    return std::unique_ptr<OpDesc>(op);
+  }
+};
+
+class SingleOpShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext* ctx) const override {
+    ctx->HasInput("X");
+    ctx->HasOutput("Out");
+    ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
+  }
+};
+
+class SingleGradOpShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext* ctx) const override {
+    ctx->HasInput(framework::GradVarName("Out"));
+    ctx->HasOutput(framework::GradVarName("X"));
+    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out"));
+  }
+};
+
+class MultiOutOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  void Make() {
+    AddInput("X", "").AsDuplicable();
+    AddInput("Y", "").AsDuplicable();
+    AddInput("Z", "").AsDuplicable();
+    AddOutput("Out", "");
+    AddOutput("YOut", "");
+    AddOutput("ZOut", "");
+    AddOutput("NotReuseOut", "");
+    AddComment("");
+  }
+};
+
+class MultiOutShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext* ctx) const override {
+    ctx->ShareDim("X", "Out");
+    ctx->ShareDim("Y", "YOut");
+    ctx->ShareDim("Z", "ZOut");
+  }
+};
+
+class MultiGradOpMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto* op = new framework::OpDesc();
+    op->SetType("multi_out_grad");
+    op->SetInput("X", Input("X"));
+    op->SetOutput(framework::GradVarName("Y"), OutputGrad("YOut"));
+    op->SetOutput(framework::GradVarName("X"), OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("Z"), OutputGrad("ZOut"));
+    return std::unique_ptr<framework::OpDesc>(op);
+  }
+};
+
+class MultiOutGradShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext* ctx) const override {
+    ctx->SetOutputDim(framework::GradVarName("Y"),
+                      ctx->GetInputDim(framework::GradVarName("YOut")));
+    ctx->SetOutputDim(framework::GradVarName("X"),
+                      ctx->GetInputDim(framework::GradVarName("Out")));
+    ctx->SetOutputDim(framework::GradVarName("Z"),
+                      ctx->GetInputDim(framework::GradVarName("ZOut")));
+  }
+};
+
+class MultiOutInplaceInToOut : public framework::InplaceInToOut {
+ public:
+  using framework::InplaceInToOut::InplaceInToOut;
+
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const OpDesc& op_desc, BlockDesc* block) const override {
+    return std::unordered_map<std::string, std::string>{
+        {"X", "Out"}, {"Y", "YOut"}, {"Z", "ZOut"},
+    };
+  }
+};
+
+class MultiOutGradInplaceInToOut : public framework::InplaceInToOut {
+ public:
+  using framework::InplaceInToOut::InplaceInToOut;
+
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const OpDesc& op_desc, BlockDesc* block) const override {
+    return std::unordered_map<std::string, std::string>{
+        {framework::GradVarName("YOut"), framework::GradVarName("Y")},
+        {framework::GradVarName("Out"), framework::GradVarName("X")},
+        {framework::GradVarName("ZOut"), framework::GradVarName("Z")},
+    };
+  }
+};
+
+}  // namespace framework
+}  // namespace paddle
+
+namespace f = paddle::framework;
+REGISTER_OPERATOR(single_op, f::NOP, f::SingleOpMaker, f::SingleGradOpMaker,
+                  f::SingleOpInplaceInToOut, f::SingleOpShapeInference);
+REGISTER_OPERATOR(single_op_grad, f::NOP, f::SingleOpInplaceInToOut,
+                  f::SingleGradOpShapeInference);
+REGISTER_OPERATOR(multi_out_op, f::NOP, f::MultiOutOpMaker, f::MultiGradOpMaker,
+                  f::MultiOutInplaceInToOut, f::MultiOutShapeInference);
+REGISTER_OPERATOR(multi_out_grad, f::NOP, f::MultiOutGradInplaceInToOut,
+                  f::MultiOutGradShapeInference);
+
+namespace paddle {
+namespace framework {
+
+TEST(InferInplace, SingleOpInplaceInToOut) {
+  ProgramDesc prog;
+  auto* op = prog.MutableBlock(0)->AppendOp();
+  op->SetType("single_op");
+  op->SetInput("X", {"test2_a", "test2_b", "test2_c"});
+  op->SetOutput("Out", {"test2_out"});
+
+  prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64});
+  prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("test2_out");
+  prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16});
+
+  auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
+  auto in_to_outs = infer_inplace(*op, op->Block());
+  EXPECT_EQ(in_to_outs.size(), 1ul);
+  auto it = in_to_outs.begin();
+  EXPECT_EQ(it->first, "test2_a");
+  EXPECT_EQ(it->second, "test2_out");
+}
+
+TEST(InferInplace, SingleGradOpInplaceInToOut) {
+  ProgramDesc prog;
+  auto* op = prog.MutableBlock(0)->AppendOp();
+  op->SetType("single_op_grad");
+  op->SetInput(GradVarName("Out"), {"test2_out"});
+  op->SetOutput(GradVarName("X"), {"test2_a", "test2_b", "test2_c"});
+
+  prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("test2_out");
+  prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16});
+
+  auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
+  auto in_to_outs = infer_inplace(*op, op->Block());
+  EXPECT_EQ(in_to_outs.size(), 1ul);
+  auto it = in_to_outs.begin();
+  EXPECT_EQ(it->first, "test2_out");
+  EXPECT_EQ(it->second, "test2_a");
+}
+
+TEST(InferInplace, MultiOutInplaceInToOut) {
+  ProgramDesc prog;
+  auto* op = prog.MutableBlock(0)->AppendOp();
+  op->SetType("multi_out_op");
+  op->SetInput("X", {"a0", "a1"});
+  op->SetInput("Y", {"b0"});
+  op->SetInput("Z", {"c0", "c1"});
+  op->SetOutput("Out", {"o0"});
+  op->SetOutput("YOut", {"y0"});
+  op->SetOutput("ZOut", {"z0"});
+
+  prog.MutableBlock(0)->Var("a0")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("b0")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("c0")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("c1")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("o0");
+  prog.MutableBlock(0)->Var("y0");
+  prog.MutableBlock(0)->Var("z0");
+  prog.MutableBlock(0)->Var("a0")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("b0")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("c0")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("o0")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("y0")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("z0")->SetShape({32, 16});
+
+  auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
+  auto in_to_outs = infer_inplace(*op, op->Block());
+  EXPECT_EQ(in_to_outs.size(), 3ul);
+  std::unordered_map<std::string, std::string> expects = {
+      {"a0", "o0"}, {"b0", "y0"}, {"c0", "z0"},
+  };
+  EXPECT_TRUE(expects == in_to_outs);
+}
+
+TEST(InferInplace, MultiGradInplaceInToOut) {
+  ProgramDesc prog;
+  auto* op = prog.MutableBlock(0)->AppendOp();
+  op->SetType("multi_out_grad");
+  op->SetInput(GradVarName("Out"), {"o0"});
+  op->SetInput(GradVarName("YOut"), {"y0"});
+  op->SetInput(GradVarName("ZOut"), {"z0"});
+  op->SetOutput(GradVarName("X"), {"a0", "a1"});
+  op->SetOutput(GradVarName("Y"), {"b0"});
+  op->SetOutput(GradVarName("Z"), {"c0", "c1"});
+
+  prog.MutableBlock(0)->Var("a0")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("b0")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("c0")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("c1")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("o0");
+  prog.MutableBlock(0)->Var("y0");
+  prog.MutableBlock(0)->Var("z0");
+  prog.MutableBlock(0)->Var("a0")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("b0")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("c0")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("o0")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("y0")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("z0")->SetShape({32, 16});
+
+  auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
+  auto in_to_outs = infer_inplace(*op, op->Block());
+
+  EXPECT_EQ(in_to_outs.size(), 3ul);
+  std::unordered_map<std::string, std::string> expects = {
+      {"o0", "a0"}, {"y0", "b0"}, {"z0", "c0"},
+  };
+  EXPECT_TRUE(expects == in_to_outs);
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/CMakeLists.txt

@@ -65,6 +65,7 @@ pass_library(conv_elementwise_add2_act_fuse_pass inference)
 pass_library(conv_elementwise_add_fuse_pass inference)
 pass_library(conv_affine_channel_fuse_pass inference)
 pass_library(transpose_flatten_concat_fuse_pass inference)
+pass_library(identity_scale_op_clean_pass base)
 
 # There may be many transpose-flatten structures in a model, and the output of
 # these structures will be used as inputs to the concat Op. This pattern will

paddle/fluid/framework/ir/graph.h

@@ -141,7 +141,8 @@ class Graph {
   ir::Node *CreateControlDepVar() {
     // TODO(panyx0718): control var name should be really unique.
     const std::string name = string::Sprintf(
-        "%s@%llu", ir::Node::kControlDepVarName, node_set_.size());
+        "%s@%llu", static_cast<const char *>(ir::Node::kControlDepVarName),
+        node_set_.size());
     auto *x = AddNode(new ir::Node(name, ir::Node::Type::kVariable));
     x->SetId(num_node_created_++);
     return x;

paddle/fluid/framework/ir/graph_helper.cc

@@ -52,16 +52,29 @@ bool HasCircleHelper(
     ir::Node *node,
     const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
     std::unordered_set<ir::Node *> *visited,
-    std::unordered_set<ir::Node *> *in_trace) {
+    std::unordered_set<ir::Node *> *in_trace,
+    std::vector<std::vector<ir::Node *>> *circles) {
   if (visited->find(node) == visited->end()) {
     visited->insert(node);
     in_trace->insert(node);
 
     for (ir::Node *in : adj_list.at(node)) {
       if (visited->find(in) == visited->end() &&
-          HasCircleHelper(in, adj_list, visited, in_trace)) {
+          HasCircleHelper(in, adj_list, visited, in_trace, circles)) {
         return true;
       } else if (in_trace->find(in) != in_trace->end()) {
+        if (circles != nullptr) {
+          std::vector<ir::Node *> circle;
+          circle.emplace_back(in);
+          ir::Node *p = in;
+          for (auto &adj : adj_list.at(p)) {
+            if (in_trace->count(adj)) {
+              circle.emplace_back(adj);
+              p = adj;
+            }
+          }
+          circles->emplace_back(circle);
+        }
         return true;
       }
     }
@@ -71,11 +84,12 @@ bool HasCircleHelper(
 }
 
 bool HasCircleInternal(
-    const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list) {
+    const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
+    std::vector<std::vector<ir::Node *>> *circles) {
   std::unordered_set<ir::Node *> visited;
   std::unordered_set<ir::Node *> in_trace;
   for (auto &adj : adj_list) {
-    if (HasCircleHelper(adj.first, adj_list, &visited, &in_trace)) {
+    if (HasCircleHelper(adj.first, adj_list, &visited, &in_trace, circles)) {
       return true;
     }
   }
@@ -84,13 +98,18 @@ bool HasCircleInternal(
 }  // namespace
 
 bool HasCircle(const Graph &graph) {
-  return HasCircleInternal(BuildOperationAdjList(graph));
+  return HasCircleInternal(BuildOperationAdjList(graph), nullptr);
+}
+
+bool FindCircleSubGraph(const Graph &graph,
+                        std::vector<std::vector<ir::Node *>> *circles) {
+  return HasCircleInternal(BuildOperationAdjList(graph), circles);
 }
 
 std::vector<ir::Node *> TopologySortOperations(const Graph &graph) {
   std::map<ir::Node *, std::unordered_set<ir::Node *>> adj_list =
       BuildOperationAdjList(graph);
-  PADDLE_ENFORCE(!HasCircleInternal(adj_list));
+  PADDLE_ENFORCE(!HasCircleInternal(adj_list, nullptr));
   std::unordered_set<ir::Node *> visited;
   std::vector<ir::Node *> ret;
   for (auto adj : adj_list) {

paddle/fluid/framework/ir/graph_helper.h

@@ -28,6 +28,11 @@ namespace ir {
 // Test if the graph contains circle.
 bool HasCircle(const Graph &graph);
 
+// Find All Circles for debugging,
+// store all subgraph in circles.
+bool FindCircleSubGraph(const Graph &graph,
+                        std::vector<std::vector<ir::Node *>> *circles);
+
 size_t GraphNum(const Graph &graph);
 
 // Topology Sort the operations in the graph from inputs to outputs.

paddle/fluid/framework/ir/graph_helper_test.cc

@@ -195,6 +195,17 @@ void BuildTwoGraphs(Graph* g) {
   //  v4->outputs.push_back(o5);
 }
 
+TEST(GraphHelperTest, Circles) {
+  ProgramDesc prog;
+
+  Graph g(prog);
+  BuildCircleGraph(&g);
+
+  std::vector<std::vector<ir::Node*>> circles;
+  ASSERT_TRUE(FindCircleSubGraph(g, &circles));
+  ASSERT_EQ(circles.size(), 1UL);
+}
+
 TEST(GraphHelperTest, GraphNum) {
   ProgramDesc prog;
 

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -117,11 +117,6 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) {
       // return false;
     }
   }
-  for (auto &item : pdnodes2nodes_) {
-    for (auto &n : item.second) {
-      GetMarkedNodes(const_cast<Graph *>(&graph)).insert(n);
-    }
-  }
   VLOG(3) << pdnodes2nodes_.size() << " nodes marked";
 
   return !pdnodes2nodes_.empty();

paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/framework/ir/identity_scale_op_clean_pass.h"
+#include <string>
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+std::unique_ptr<ir::Graph> IdentityScaleOpCleanPass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  FusePassBase::Init("identity_scale_op_clean", graph.get());
+
+  // pre_op -> scale_in -> scale_op -> scale_out
+  // ->
+  // pre_op -> scale_out
+  GraphPatternDetector detector;
+  auto pre_op = detector.mutable_pattern()->NewNode("pre_op")->assert_is_op();
+  auto scale_in = detector.mutable_pattern()
+                      ->NewNode("scale_in")
+                      ->assert_is_op_input("scale")
+                      ->AsIntermediate();
+  auto scale_op = detector.mutable_pattern()
+                      ->NewNode("scale_fuse")
+                      ->assert_is_op("scale")
+                      ->assert_op_attr<float>("scale", 1.)
+                      ->assert_op_attr<float>("bias", 0.);
+  auto scale_out = detector.mutable_pattern()
+                       ->NewNode("scale_out")
+                       ->assert_is_op_output("scale");
+
+  pre_op->LinksTo({scale_in});
+  scale_op->LinksFrom({scale_in}).LinksTo({scale_out});
+
+  GraphPatternDetector::handle_t handler = [&](
+      const GraphPatternDetector::subgraph_t& subgraph, Graph* graph) {
+    Node* scale_op_var = subgraph.at(scale_op);
+    Node* scale_in_var = subgraph.at(scale_in);
+    Node* scale_out_var = subgraph.at(scale_out);
+    Node* pre_op_var = subgraph.at(pre_op);
+    // Link pre_op directly to scale_out
+    const std::string scale_in_name = scale_in_var->Name();
+    const std::string scale_out_name = scale_out_var->Name();
+    // Remove links in graph
+    GraphSafeRemoveNodes(graph, {scale_in_var, scale_op_var});
+    // Modify proto message
+    auto* pre_op_desc = pre_op_var->Op();
+    for (auto& parameter : *pre_op_desc->Proto()->mutable_outputs()) {
+      auto* arguments = parameter.mutable_arguments();
+      auto it = std::find(arguments->begin(), arguments->end(), scale_in_name);
+      PADDLE_ENFORCE(it != arguments->end());
+      *it = scale_out_name;
+    }
+
+    IR_NODE_LINK_TO(pre_op_var, scale_out_var);
+  };
+
+  detector(graph.get(), handler);
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(identity_scale_op_clean_pass,
+              paddle::framework::ir::IdentityScaleOpCleanPass);

paddle/fluid/framework/ir/identity_scale_op_clean_pass.h

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class IdentityScaleOpCleanPass : public FusePassBase {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
+
+ private:
+  virtual ~IdentityScaleOpCleanPass() = default;
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/op_info.h

@@ -38,6 +38,7 @@ struct OpInfo {
   OpAttrChecker* checker_{nullptr};
   InferVarTypeFN infer_var_type_;
   InferShapeFN infer_shape_;
+  InferInplaceOpFN infer_inplace_;
 
   bool HasOpProtoAndChecker() const {
     return proto_ != nullptr && checker_ != nullptr;

paddle/fluid/framework/type_defs.h

@@ -57,5 +57,8 @@ using InferVarTypeFN =
 
 using InferShapeFN = std::function<void(InferShapeContext*)>;
 
+using InplacePair = std::unordered_map<std::string, std::string>;
+using InferInplaceOpFN = std::function<InplacePair(const OpDesc&, BlockDesc*)>;
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/imperative/CMakeLists.txt

 if(WITH_PYTHON)
-cc_library(layer SRCS layer.cc DEPS proto_desc operator device_context blas)
-cc_library(tracer SRCS tracer.cc DEPS proto_desc device_context)
+cc_library(layer SRCS layer.cc DEPS proto_desc operator device_context blas pybind)
+cc_library(tracer SRCS tracer.cc DEPS proto_desc device_context pybind)
 cc_library(engine SRCS engine.cc)
 endif()

paddle/fluid/inference/CMakeLists.txt

@@ -58,12 +58,13 @@ if(WIN32)
   sep_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS}
               DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array
                    analysis_config paddle_pass_builder)
-  target_link_libraries(paddle_fluid_shared shlwapi)
 else(WIN32)
   cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS}
              DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array
                   analysis_config paddle_pass_builder)
 endif()
+get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
+target_link_libraries(paddle_fluid_shared ${os_dependency_modules})
 
 set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
 if(NOT APPLE AND NOT WIN32)

paddle/fluid/inference/analysis/ir_pass_manager.cc

@@ -83,7 +83,6 @@ void IRPassManager::CreatePasses(Argument *argument,
           new std::string(GetOrCreateModelOptCacheDir(model_opt_cache_dir)));
     }
 
-    // graph_ = pass->Apply(std::move(graph_));
     pre_pass = pass_name;
 
     passes_.emplace_back(std::move(pass));
@@ -97,8 +96,9 @@ std::unique_ptr<Graph> IRPassManager::Apply(std::unique_ptr<Graph> graph) {
   PADDLE_ENFORCE(graph.get());
   // Apply all the passes
   for (const auto &pass : passes_) {
-    if (pass->Type() == "graph_viz_pass") continue;
-    PrettyLogEndl(Style::H2(), "--- Running IR pass [%s]", pass->Type());
+    if (pass->Type() != "graph_viz_pass") {
+      PrettyLogEndl(Style::H2(), "--- Running IR pass [%s]", pass->Type());
+    }
     graph = pass->Apply(std::move(graph));
   }
   return std::move(graph);

paddle/fluid/inference/analysis/ir_passes/CMakeLists.txt

 cc_library(subgraph_detector SRCS subgraph_detector.cc DEPS proto_desc)
+if(WITH_TESTING)
+  add_dependencies(subgraph_detector gtest)
+endif()
 
 if (WITH_GPU AND TENSORRT_FOUND)
   cc_library(tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass.cc DEPS subgraph_detector tensorrt_op_teller)

paddle/fluid/inference/analysis/passes/memory_optimize_pass.cc

@@ -18,6 +18,7 @@
 #include <limits>
 #include <map>
 #include <string>
+#include <type_traits>
 #include <utility>
 #include <vector>
 #include "paddle/fluid/framework/ir/graph_helper.h"
@@ -168,7 +169,11 @@ bool FindSuitableTensorToReuse(
     if (!cluster->count(candidate)) continue;
 
     size_t space = space_table.at(candidate);
-    size_t space_diff = std::abs<size_t>(space - space_required);
+    PADDLE_ENFORCE(
+        space <= std::numeric_limits<std::make_signed<size_t>::type>::max(),
+        "space overload");
+    size_t space_diff =
+        std::abs((std::make_signed<size_t>::type)space - space_required);
     if (space_diff < best_fit.second) {
       best_fit.first = candidate;
       best_fit.second = space_diff;

paddle/fluid/inference/api/analysis_config.cc

@@ -318,4 +318,9 @@ NativeConfig AnalysisConfig::ToNativeConfig() const {
   return config;
 }
 
+void AnalysisConfig::SwitchIrDebug(int x) {
+  ir_debug_ = x;
+  Update();
+}
+
 }  // namespace paddle

paddle/fluid/inference/api/analysis_predictor.cc

@@ -58,7 +58,8 @@ namespace {
 bool IsPersistable(const framework::VarDesc *var) {
   if (var->Persistable() &&
       var->GetType() != framework::proto::VarType::FEED_MINIBATCH &&
-      var->GetType() != framework::proto::VarType::FETCH_LIST) {
+      var->GetType() != framework::proto::VarType::FETCH_LIST &&
+      var->GetType() != framework::proto::VarType::RAW) {
     return true;
   }
   return false;

paddle/fluid/inference/api/analysis_predictor_tester.cc

@@ -196,7 +196,7 @@ TEST(AnalysisPredictor, memory_optim) {
   AnalysisConfig config(FLAGS_dirname);
   config.DisableGpu();
   config.EnableMemoryOptim(true);
-  config.pass_builder()->TurnOnDebug();
+  config.SwitchIrDebug();
 
   auto native_predictor =
       CreatePaddlePredictor<NativeConfig>(config.ToNativeConfig());

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -140,9 +140,12 @@ struct AnalysisConfig {
    */
   bool tensorrt_engine_enabled() const { return use_tensorrt_; }
 
-  /** Control whther to debug IR graph analysis phase.
+  /** \brief Control whether to debug IR graph analysis phase.
+   *
+   * This will generate DOT files for visualizing the computation graph after
+   * each analysis pass applied.
    */
-  void SwitchIrDebug(int x = true) { ir_debug_ = x; }
+  void SwitchIrDebug(int x = true);
 
   /** Turn on MKLDNN.
    */

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -117,6 +117,7 @@ class CpuPassStrategy : public PassStrategy {
         "conv_bn_fuse_pass",             //
         "conv_eltwiseadd_bn_fuse_pass",  //
         "is_test_pass",                  //
+        "identity_scale_op_clean_pass",  //
     });
     use_gpu_ = false;
   }
@@ -155,6 +156,7 @@ class GpuPassStrategy : public PassStrategy {
   GpuPassStrategy() : PassStrategy({}) {
     passes_.assign({
       "infer_clean_graph_pass",                        //
+          "identity_scale_op_clean_pass",              //
           "conv_affine_channel_fuse_pass",             //
           "conv_eltwiseadd_affine_channel_fuse_pass",  //
           "conv_bn_fuse_pass",                         //

paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc

@@ -142,7 +142,7 @@ void SetConfig(AnalysisConfig *cfg, bool use_mkldnn = false) {
   cfg->SetModel(FLAGS_infer_model + "/model", FLAGS_infer_model + "/params");
   cfg->DisableGpu();
   cfg->SwitchSpecifyInputNames();
-  cfg->pass_builder()->TurnOnDebug();
+  cfg->SwitchIrDebug();
   cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
   if (use_mkldnn) {
     cfg->EnableMKLDNN();

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

@@ -69,7 +69,7 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
 TEST(Analyzer_Text_Classification, profile) {
   AnalysisConfig cfg;
   SetConfig(&cfg);
-  cfg.pass_builder()->TurnOnDebug();
+  cfg.SwitchIrDebug();
   std::vector<PaddleTensor> outputs;
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;

paddle/fluid/inference/utils/benchmark_tester.cc

@@ -34,6 +34,6 @@ TEST(Benchmark, PersistToFile) {
   benchmark.SetLatency(220);
 
   benchmark.PersistToFile("1.log");
-  benchmark.PersistToFile("1.log");
-  benchmark.PersistToFile("1.log");
+  benchmark.PersistToFile("2.log");
+  benchmark.PersistToFile("3.log");
 }

paddle/fluid/memory/allocation/legacy_allocator.cc

@@ -59,11 +59,6 @@ size_t memory_usage(const platform::Place &p);
 
 using BuddyAllocator = detail::BuddyAllocator;
 
-std::unordered_map</*device id*/ int,
-                   std::pair</*current memory usage*/ uint64_t,
-                             /*peak memory usage*/ uint64_t>>
-    gpu_mem_info;
-
 BuddyAllocator *GetCPUBuddyAllocator() {
   // We tried thread_local for inference::RNN1 model, but that not works much
   // for multi-thread test.
@@ -144,6 +139,8 @@ BuddyAllocator *GetGPUBuddyAllocator(int gpu_id) {
     devices = platform::GetSelectedDevices();
     int gpu_num = devices.size();
 
+    allocation::GPUMemMonitor.Initialize(devices.size());
+
     a_arr = new BuddyAllocator *[gpu_num];
     for (size_t i = 0; i < devices.size(); ++i) {
       int dev_id = devices[i];
@@ -190,25 +187,19 @@ void *Alloc<platform::CUDAPlace>(const platform::CUDAPlace &place,
     platform::SetDeviceId(place.device);
     size_t avail, total;
     platform::GpuMemoryUsage(&avail, &total);
-    LOG(WARNING) << "Cannot allocate " << string::HumanReadableSize(size)
-                 << " in GPU " << place.device << ", available "
-                 << string::HumanReadableSize(avail);
-    LOG(WARNING) << "total " << total;
-    LOG(WARNING) << "GpuMinChunkSize "
-                 << string::HumanReadableSize(
-                        buddy_allocator->GetMinChunkSize());
-    LOG(WARNING) << "GpuMaxChunkSize "
-                 << string::HumanReadableSize(
-                        buddy_allocator->GetMaxChunkSize());
-    LOG(WARNING) << "GPU memory used: "
-                 << string::HumanReadableSize(Used<platform::CUDAPlace>(place));
+    LOG(FATAL) << "Cannot allocate " << string::HumanReadableSize(size)
+               << " in GPU " << place.device << ", available "
+               << string::HumanReadableSize(avail) << "total " << total
+               << "GpuMinChunkSize "
+               << string::HumanReadableSize(buddy_allocator->GetMinChunkSize())
+               << "GpuMaxChunkSize "
+               << string::HumanReadableSize(buddy_allocator->GetMaxChunkSize())
+               << "GPU memory used: "
+               << string::HumanReadableSize(Used<platform::CUDAPlace>(place));
     platform::SetDeviceId(cur_dev);
   } else {
-    gpu_mem_info[place.device].first += size;
-    if (gpu_mem_info[place.device].first > gpu_mem_info[place.device].second) {
-      gpu_mem_info[place.device].second = gpu_mem_info[place.device].first;
-      VLOG(3) << "device: " << place.device << " peak memory usage : "
-              << (gpu_mem_info[place.device].second >> 20) << " MiB";
+    if (VLOG_IS_ON(3)) {
+      allocation::GPUMemMonitor.Add(place.device, size);
     }
     if (FLAGS_init_allocated_mem) {
       cudaMemset(ptr, 0xEF, size);
@@ -225,7 +216,9 @@ void Free<platform::CUDAPlace>(const platform::CUDAPlace &place, void *p,
                                size_t size) {
 #ifdef PADDLE_WITH_CUDA
   GetGPUBuddyAllocator(place.device)->Free(p);
-  gpu_mem_info[place.device].first -= size;
+  if (VLOG_IS_ON(3)) {
+    allocation::GPUMemMonitor.Minus(place.device, size);
+  }
 #else
   PADDLE_THROW("'CUDAPlace' is not supported in CPU only device.");
 #endif
@@ -264,7 +257,7 @@ void *Alloc<platform::CUDAPinnedPlace>(const platform::CUDAPinnedPlace &place,
   void *ptr = buddy_allocator->Alloc(size);
 
   if (ptr == nullptr) {
-    LOG(WARNING) << "cudaMallocHost Cannot allocate " << size
+    LOG(WARNING) << "cudaHostAlloc Cannot allocate " << size
                  << " bytes in CUDAPinnedPlace";
   }
   if (FLAGS_init_allocated_mem) {
@@ -335,6 +328,8 @@ size_t Usage::operator()(const platform::CUDAPinnedPlace &cuda_pinned) const {
 
 namespace allocation {
 
+LegacyMemMonitor GPUMemMonitor;
+
 Allocation *LegacyAllocator::AllocateImpl(size_t size, Allocator::Attr attr) {
   void *ptr = boost::apply_visitor(legacy::AllocVisitor(size), place_);
   return new Allocation(ptr, size, place_);
@@ -346,6 +341,63 @@ void LegacyAllocator::Free(Allocation *allocation) {
       allocation->place());
   delete allocation;
 }
+
+bool MemInfo::Add(const size_t &size) {
+  std::lock_guard<std::mutex> lock(mutex_);
+  usage_ += size;
+  bool peak_point = usage_ > peak_usage_;
+  if (peak_point) peak_usage_ = usage_;
+  return peak_point;
+}
+
+void MemInfo::Minus(const size_t &size) {
+  std::lock_guard<std::mutex> lock(mutex_);
+  usage_ -= size;
+}
+
+uint64_t MemInfo::GetPeakUsage() { return peak_usage_; }
+
+LegacyMemMonitor::~LegacyMemMonitor() {
+  for (auto &item : gpu_mem_info_) delete item.second;
+}
+
+void LegacyMemMonitor::Initialize(const int &device_num) {
+  for (auto i = 0; i < device_num; ++i) {
+    gpu_mem_info_[i] = new MemInfo();
+  }
+}
+
+void LegacyMemMonitor::Add(const int &device, const size_t &size) {
+  if (gpu_mem_info_[device]->Add(size)) {
+    VLOG(3) << "#LegacyMemMonitor# device: " << device
+            << " peak memory usage : "
+            << (gpu_mem_info_[device]->GetPeakUsage() >> 20) << " MiB";
+  }
+}
+
+void LegacyMemMonitor::Minus(const int &device, const size_t &size) {
+  gpu_mem_info_[device]->Minus(size);
+}
+
+uint64_t LegacyMemMonitor::GetMemUsage(const int &device) {
+  return gpu_mem_info_.find(device) == gpu_mem_info_.end()
+             ? 0
+             : gpu_mem_info_[device]->GetPeakUsage();
+}
+
+void LegacyMemMonitor::PrintMemUsage() {
+  std::vector<int> devices;
+  for (const auto &item : gpu_mem_info_) {
+    devices.emplace_back(item.first);
+  }
+  std::sort(devices.begin(), devices.end());
+  for (const auto &device : devices) {
+    std::cout << "Device : " << device << " Peak Memory Usage : "
+              << (gpu_mem_info_[device]->GetPeakUsage() >> 20) << " MiB"
+              << std::endl;
+  }
+}
+
 }  // namespace allocation
 }  // namespace memory
 }  // namespace paddle

paddle/fluid/memory/allocation/legacy_allocator.h

@@ -13,12 +13,59 @@
 // limitations under the License.
 
 #pragma once
+#include <algorithm>
+#include <mutex>  // NOLINT
+#include <unordered_map>
+#include <utility>
+#include <vector>
 #include "paddle/fluid/memory/allocation/allocator.h"
 #include "paddle/fluid/platform/place.h"
 namespace paddle {
 namespace memory {
 namespace allocation {
 
+class MemInfo {
+ public:
+  MemInfo() : usage_(0), peak_usage_(0) {}
+  MemInfo(const MemInfo &) = delete;
+  MemInfo &operator=(const MemInfo &) = delete;
+
+  // return a flag to indicate current operation will create a peak point or not
+  bool Add(const size_t &);
+  void Minus(const size_t &);
+
+  uint64_t GetPeakUsage();
+
+ private:
+  /* current memory usage*/
+  uint64_t usage_;
+  uint64_t peak_usage_;
+  std::mutex mutex_;
+};
+
+class LegacyMemMonitor {
+ public:
+  // used to store the GPU memory usage of each devices
+  using MemUsage = std::unordered_map</*device id*/ int,
+                                      /*mem usage info node*/ MemInfo *>;
+
+  MemUsage GetMemUsageInfo() { return gpu_mem_info_; }
+  ~LegacyMemMonitor();
+
+  void Initialize(const int &);
+  void Add(const int &, const size_t &);
+  void Minus(const int &, const size_t &);
+
+  uint64_t GetMemUsage(const int &);
+
+  void PrintMemUsage();
+
+ protected:
+  MemUsage gpu_mem_info_;
+};
+
+extern LegacyMemMonitor GPUMemMonitor;
+
 class LegacyAllocatorPrivate;
 class LegacyAllocator : public Allocator {
  public:

paddle/fluid/memory/allocation/pinned_allocator.cc

@@ -32,7 +32,7 @@ Allocation *CPUPinnedAllocator::AllocateImpl(size_t size,
   //    "CPUPinnedAllocator should be used for Cross-Device Communication");
 
   void *ptr;
-  PADDLE_ENFORCE(cudaMallocHost(&ptr, size));
+  PADDLE_ENFORCE(cudaHostAlloc(&ptr, size, cudaHostAllocPortable));
   return new CPUPinnedAllocation(ptr, size);
 }
 }  // namespace allocation

paddle/fluid/memory/allocation/pinned_allocator.h

@@ -19,7 +19,7 @@ namespace paddle {
 namespace memory {
 namespace allocation {
 
-// Allocator uses `cudaMallocHost`
+// Allocator uses `cudaHostAlloc`
 class CPUPinnedAllocation : public Allocation {
  public:
   CPUPinnedAllocation(void *ptr, size_t size)

paddle/fluid/memory/detail/system_allocator.cc

@@ -173,14 +173,14 @@ void* CUDAPinnedAllocator::Alloc(size_t* index, size_t size) {
 
   void* p;
   // PINNED memory is visible to all CUDA contexts.
-  cudaError_t result = cudaMallocHost(&p, size);
+  cudaError_t result = cudaHostAlloc(&p, size, cudaHostAllocPortable);
 
   if (result == cudaSuccess) {
     *index = 1;  // PINNED memory
     cuda_pinnd_alloc_size_ += size;
     return p;
   } else {
-    LOG(WARNING) << "cudaMallocHost failed.";
+    LOG(WARNING) << "cudaHostAlloc failed.";
     return nullptr;
   }
 

paddle/fluid/operators/activation_op.cc

@@ -547,12 +547,14 @@ namespace ops = paddle::operators;
   __macro(Swish, swish);             \
   __macro(ThresholdedRelu, thresholded_relu);
 
-#define REGISTER_INPLACE_ACTIVATION_OP(OP_NAME, KERNEL_TYPE)        \
-  REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp, \
-                    ::paddle::operators::OP_NAME##OpMaker,          \
-                    ::paddle::operators::ActivationOpInferVarType,  \
-                    ::paddle::operators::OP_NAME##GradMaker);       \
-  REGISTER_OPERATOR(KERNEL_TYPE##_grad, ::paddle::operators::ActivationOpGrad)
+#define REGISTER_INPLACE_ACTIVATION_OP(OP_NAME, KERNEL_TYPE)                   \
+  REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp,            \
+                    ::paddle::operators::OP_NAME##OpMaker,                     \
+                    ::paddle::operators::ActivationOpInferVarType,             \
+                    ::paddle::operators::OP_NAME##GradMaker,                   \
+                    ::paddle::framework::SingleOpInplaceInToOut);              \
+  REGISTER_OPERATOR(KERNEL_TYPE##_grad, ::paddle::operators::ActivationOpGrad, \
+                    ::paddle::framework::SingleOpInplaceInToOut)
 
 #define REGISTER_ACTIVATION_OP(OP_NAME, KERNEL_TYPE)                    \
   REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp,     \

paddle/fluid/operators/batch_norm_op.cc

@@ -589,8 +589,10 @@ class BatchNormGradMaker : public framework::SingleGradOpDescMaker {
     op->SetInput("SavedVariance", Output("SavedVariance"));
 
     // used when setting use_global_stats True during training
-    op->SetInput("Mean", Output("MeanOut"));
-    op->SetInput("Variance", Output("VarianceOut"));
+    if (boost::get<bool>(GetAttr("use_global_stats"))) {
+      op->SetInput("Mean", Output("MeanOut"));
+      op->SetInput("Variance", Output("VarianceOut"));
+    }
 
     op->SetAttrMap(Attrs());
 
@@ -602,13 +604,48 @@ class BatchNormGradMaker : public framework::SingleGradOpDescMaker {
   }
 };
 
+class BatchNormInplaceInToOut : public framework::InplaceInToOut {
+ public:
+  using InplaceInToOut::InplaceInToOut;
+
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const framework::OpDesc &op_desc,
+      framework::BlockDesc *block) const override {
+    std::unordered_map<std::string, std::string> inplace_in_to_out = {
+        {"Mean", "MeanOut"}, {"Variance", "VarianceOut"}, {"X", "Y"},
+    };
+    return inplace_in_to_out;
+  }
+};
+
+class BatchNormGradInplaceInToOut : public framework::InplaceInToOut {
+ public:
+  using InplaceInToOut::InplaceInToOut;
+
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const framework::OpDesc &op_desc,
+      framework::BlockDesc *block) const override {
+    std::unordered_map<std::string, std::string> inplace_in_to_out = {
+        // Scale, Bias, SavedMean, SavedVariance shape is [batch_size, C]
+        {framework::GradVarName("Y"), framework::GradVarName("X")},
+        {"SavedMean", framework::GradVarName("Scale")},
+        {"SavedVariance", framework::GradVarName("Bias")},
+    };
+    return inplace_in_to_out;
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker,
-                  ops::BatchNormOpInferVarType, ops::BatchNormGradMaker);
-REGISTER_OPERATOR(batch_norm_grad, ops::BatchNormGradOp);
+                  ops::BatchNormOpInferVarType, ops::BatchNormGradMaker,
+                  ops::BatchNormInplaceInToOut);
+REGISTER_OPERATOR(batch_norm_grad, ops::BatchNormGradOp,
+                  ops::BatchNormGradInplaceInToOut);
 
 REGISTER_OP_CPU_KERNEL(
     batch_norm, ops::BatchNormKernel<paddle::platform::CPUDeviceContext, float>,

paddle/fluid/operators/detection/CMakeLists.txt

@@ -31,6 +31,7 @@ detection_library(polygon_box_transform_op SRCS polygon_box_transform_op.cc
 polygon_box_transform_op.cu)
 detection_library(rpn_target_assign_op SRCS rpn_target_assign_op.cc)
 detection_library(generate_proposal_labels_op SRCS generate_proposal_labels_op.cc)
+detection_library(box_clip_op SRCS box_clip_op.cc box_clip_op.cu)
 detection_library(yolov3_loss_op SRCS yolov3_loss_op.cc)
 
 if(WITH_GPU)

paddle/fluid/operators/detection/bbox_util.h

@@ -99,5 +99,29 @@ void BboxOverlaps(const framework::Tensor& r_boxes,
   }
 }
 
+template <class T>
+void ClipTiledBoxes(const platform::DeviceContext& ctx,
+                    const framework::Tensor& im_info,
+                    const framework::Tensor& input_boxes,
+                    framework::Tensor* out) {
+  T* out_data = out->mutable_data<T>(ctx.GetPlace());
+  const T* im_info_data = im_info.data<T>();
+  const T* input_boxes_data = input_boxes.data<T>();
+  T zero(0);
+  T im_w = round(im_info_data[1] / im_info_data[2]);
+  T im_h = round(im_info_data[0] / im_info_data[2]);
+  for (int64_t i = 0; i < input_boxes.numel(); ++i) {
+    if (i % 4 == 0) {
+      out_data[i] = std::max(std::min(input_boxes_data[i], im_w - 1), zero);
+    } else if (i % 4 == 1) {
+      out_data[i] = std::max(std::min(input_boxes_data[i], im_h - 1), zero);
+    } else if (i % 4 == 2) {
+      out_data[i] = std::max(std::min(input_boxes_data[i], im_w - 1), zero);
+    } else {
+      out_data[i] = std::max(std::min(input_boxes_data[i], im_h - 1), zero);
+    }
+  }
+}
+
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detection/box_clip_op.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+    http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/detection/box_clip_op.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+class BoxClipOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"),
+                   "Input(Input) of BoxClipOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("ImInfo"),
+                   "Input(ImInfo) of BoxClipOp should not be null.");
+
+    auto input_box_dims = ctx->GetInputDim("Input");
+    auto im_info_dims = ctx->GetInputDim("ImInfo");
+
+    if (ctx->IsRuntime()) {
+      auto input_box_size = input_box_dims.size();
+      PADDLE_ENFORCE_EQ(input_box_dims[input_box_size - 1], 4,
+                        "The last dimension of Input must be 4");
+      PADDLE_ENFORCE_EQ(im_info_dims.size(), 2,
+                        "The rank of Input(Input) in BoxClipOp must be 2");
+      PADDLE_ENFORCE_EQ(im_info_dims[1], 3,
+                        "The last dimension of ImInfo must be 3");
+    }
+    ctx->ShareDim("Input", /*->*/ "Output");
+    ctx->ShareLoD("Input", /*->*/ "Output");
+  }
+};
+
+class BoxClipOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("Input",
+             "(LoDTensor) "
+             "Input is a LoDTensor with shape [..., 4] holds 4 points"
+             "in last dimension in format [xmin, ymin, xmax, ymax]");
+    AddInput("ImInfo",
+             "(Tensor) Information for image reshape is in shape (N, 3), "
+             "in format (height, width, im_scale)");
+    AddOutput("Output",
+              "(LoDTensor) "
+              "Output is a LoDTensor with the same shape as Input"
+              "and it is the result after clip");
+    AddComment(R"DOC(
+This operator clips input boxes to original input images.
+
+For each input box, The formula is given as follows:
+
+       $$xmin = \max(\min(xmin, im_w - 1), 0)$$
+       $$ymin = \max(\min(ymin, im_h - 1), 0)$$     
+       $$xmax = \max(\min(xmax, im_w - 1), 0)$$
+       $$ymax = \max(\min(ymax, im_h - 1), 0)$$
+
+where im_w and im_h are computed from ImInfo, the formula is given as follows:
+
+       $$im_w = \round(width / im_scale)$$
+       $$im_h = \round(height / im_scale)$$ 
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(box_clip, ops::BoxClipOp, ops::BoxClipOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    box_clip, ops::BoxClipKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::BoxClipKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/detection/box_clip_op.cu

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <algorithm>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/detection/box_clip_op.h"
+#include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
+#include "paddle/fluid/platform/hostdevice.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTenso = framework::LoDTensor;
+
+static constexpr int ImInfoSize = 3;
+
+template <typename T, int BlockSize>
+static __global__ void GPUBoxClip(const T *input, const size_t *lod,
+                                  const size_t width, const T *im_info,
+                                  T *output) {
+  T im_w = round(im_info[blockIdx.x * ImInfoSize + 1] /
+                 im_info[blockIdx.x * ImInfoSize + 2]);
+  T im_h = round(im_info[blockIdx.x * ImInfoSize] /
+                 im_info[blockIdx.x * ImInfoSize + 2]);
+  for (int i = threadIdx.x; i < (lod[blockIdx.x + 1] - lod[blockIdx.x]) * width;
+       i += BlockSize) {
+    int idx = lod[blockIdx.x] * width + i;
+    T im_size = (idx % 2 == 0) ? im_w : im_h;
+    output[idx] = max(min(input[idx], im_size - 1), T(0.));
+  }
+}
+
+template <typename DeviceContext, typename T>
+class GPUBoxClipKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(context.GetPlace()),
+                   "This kernel only runs on GPU device.");
+    auto *input = context.Input<LoDTensor>("Input");
+    auto *im_info = context.Input<Tensor>("ImInfo");
+    auto *output = context.Output<LoDTensor>("Output");
+    const int64_t num = input->dims()[0];
+    const int64_t bbox_width = input->numel() / num;
+    auto lod = input->lod();
+    framework::LoD abs_offset_lod = framework::ToAbsOffset(lod);
+    auto &dev_ctx = context.template device_context<DeviceContext>();
+    auto stream = dev_ctx.stream();
+    const size_t batch_size = lod.back().size() - 1;
+    T *output_data = output->mutable_data<T>(dev_ctx.GetPlace());
+    GPUBoxClip<T, 512><<<batch_size, 512, 0, stream>>>(
+        input->data<T>(), abs_offset_lod[0].CUDAMutableData(dev_ctx.GetPlace()),
+        bbox_width, im_info->data<T>(), output_data);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    box_clip, ops::GPUBoxClipKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GPUBoxClipKernel<paddle::platform::CUDADeviceContext, double>);

paddle/fluid/operators/detection/box_clip_op.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+    http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <string>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/detection/bbox_util.h"
+#include "paddle/fluid/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+
+template <typename DeviceContext, typename T>
+class BoxClipKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* input_box = context.Input<LoDTensor>("Input");
+    auto* im_info = context.Input<LoDTensor>("ImInfo");
+    auto* output_box = context.Output<LoDTensor>("Output");
+    auto& dev_ctx =
+        context.template device_context<platform::CPUDeviceContext>();
+    output_box->mutable_data<T>(context.GetPlace());
+    if (input_box->lod().size()) {
+      PADDLE_ENFORCE_EQ(input_box->lod().size(), 1UL,
+                        "Only support 1 level of LoD.");
+    }
+    auto box_lod = input_box->lod().back();
+    int64_t n = static_cast<int64_t>(box_lod.size() - 1);
+    for (int i = 0; i < n; ++i) {
+      Tensor im_info_slice = im_info->Slice(i, i + 1);
+      Tensor box_slice = input_box->Slice(box_lod[i], box_lod[i + 1]);
+      Tensor output_slice = output_box->Slice(box_lod[i], box_lod[i + 1]);
+      ClipTiledBoxes<T>(dev_ctx, im_info_slice, box_slice, &output_slice);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/detection/box_coder_op.cc

@@ -38,20 +38,12 @@ class BoxCoderOp : public framework::OperatorWithKernel {
                         "The shape of PriorBox is [N, 4]");
       if (ctx->HasInput("PriorBoxVar")) {
         auto prior_box_var_dims = ctx->GetInputDim("PriorBoxVar");
-        PADDLE_ENFORCE(
-            prior_box_var_dims.size() == 1 || prior_box_var_dims.size() == 2,
-            "Input(PriorBoxVar) of BoxCoderOp should be 1 or 2.");
-        if (prior_box_var_dims.size() == 1) {
-          PADDLE_ENFORCE_EQ(
-              prior_box_var_dims[0], 4,
-              "The 1st dimension of Input(PriorBoxVar) should be 4"
-              "when the rank is 1.");
-        } else {
-          PADDLE_ENFORCE_EQ(
-              prior_box_dims, prior_box_var_dims,
-              "The dimension of Input(PriorBoxVar) should be equal to"
-              "the dimension of Input(PriorBox when the rank is 2.)");
-        }
+        PADDLE_ENFORCE(prior_box_var_dims.size() == 2,
+                       "Input(PriorBoxVar) of BoxCoderOp should be 2.");
+        PADDLE_ENFORCE_EQ(
+            prior_box_dims, prior_box_var_dims,
+            "The dimension of Input(PriorBoxVar) should be equal to"
+            "the dimension of Input(PriorBox) when the rank is 2.");
       }
     }
 

paddle/fluid/operators/detection/box_coder_op.cu

@@ -56,10 +56,7 @@ __global__ void EncodeCenterSizeKernel(
     output[idx * len + 2] = log(fabs(target_box_width / prior_box_width));
     output[idx * len + 3] = log(fabs(target_box_height / prior_box_height));
     if (prior_box_var_data) {
-      int prior_var_offset = 0;
-      if (prior_box_var_size == 2) {
-        prior_var_offset = col_idx * len;
-      }
+      int prior_var_offset = col_idx * len;
       output[idx * len] /= prior_box_var_data[prior_var_offset];
       output[idx * len + 1] /= prior_box_var_data[prior_var_offset + 1];
       output[idx * len + 2] /= prior_box_var_data[prior_var_offset + 2];
@@ -99,10 +96,7 @@ __global__ void DecodeCenterSizeKernel(
     T box_var_x = T(1), box_var_y = T(1);
     T box_var_w = T(1), box_var_h = T(1);
     if (prior_box_var_data) {
-      int prior_var_offset = 0;
-      if (prior_box_var_size == 2) {
-        prior_var_offset = axis == 0 ? col_idx * len : row_idx * len;
-      }
+      int prior_var_offset = axis == 0 ? col_idx * len : row_idx * len;
       box_var_x = prior_box_var_data[prior_var_offset];
       box_var_y = prior_box_var_data[prior_var_offset + 1];
       box_var_w = prior_box_var_data[prior_var_offset + 2];

paddle/fluid/operators/detection/box_coder_op.h

@@ -79,10 +79,7 @@ class BoxCoderKernel : public framework::OpKernel<T> {
         output[offset + 3] =
             std::log(std::fabs(target_box_height / prior_box_height));
         if (prior_box_var) {
-          int prior_var_offset = 0;
-          if (prior_box_var->dims().size() == 2) {
-            prior_var_offset = j * len;
-          }
+          int prior_var_offset = j * len;
           output[offset] /= prior_box_var_data[prior_var_offset];
           output[offset + 1] /= prior_box_var_data[prior_var_offset + 1];
           output[offset + 2] /= prior_box_var_data[prior_var_offset + 2];
@@ -95,11 +92,12 @@ class BoxCoderKernel : public framework::OpKernel<T> {
       }
     }
   }
+  template <int axis, int var_size>
   void DecodeCenterSize(const framework::Tensor* target_box,
                         const framework::Tensor* prior_box,
                         const framework::Tensor* prior_box_var,
-                        const bool normalized, const int axis,
-                        const std::vector<float> variance, T* output) const {
+                        const bool normalized, std::vector<float> variance,
+                        T* output) const {
     int64_t row = target_box->dims()[0];
     int64_t col = target_box->dims()[1];
     int64_t len = target_box->dims()[2];
@@ -107,19 +105,17 @@ class BoxCoderKernel : public framework::OpKernel<T> {
     auto* target_box_data = target_box->data<T>();
     auto* prior_box_data = prior_box->data<T>();
     const T* prior_box_var_data = nullptr;
-    if (prior_box_var) prior_box_var_data = prior_box_var->data<T>();
+    if (var_size == 2) prior_box_var_data = prior_box_var->data<T>();
     int prior_box_offset = 0;
+    T var_data[4] = {1., 1., 1., 1.};
+    T* var_ptr = var_data;
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for collapse(2)
 #endif
     for (int64_t i = 0; i < row; ++i) {
       for (int64_t j = 0; j < col; ++j) {
         size_t offset = i * col * len + j * len;
-        if (axis == 0) {
-          prior_box_offset = j * len;
-        } else if (axis == 1) {
-          prior_box_offset = i * len;
-        }
+        prior_box_offset = axis == 0 ? j * len : i * len;
         T prior_box_width = prior_box_data[prior_box_offset + 2] -
                             prior_box_data[prior_box_offset] +
                             (normalized == false);
@@ -133,26 +129,18 @@ class BoxCoderKernel : public framework::OpKernel<T> {
 
         T target_box_center_x = 0, target_box_center_y = 0;
         T target_box_width = 0, target_box_height = 0;
-        T box_var_x = T(1), box_var_y = T(1);
-        T box_var_w = T(1), box_var_h = T(1);
-        if (prior_box_var) {
-          int prior_var_offset = 0;
-          if (prior_box_var->dims().size() == 2) {
-            if (axis == 0)
-              prior_var_offset = j * len;
-            else if (axis == 1)
-              prior_var_offset = i * len;
-          }
-          box_var_x = prior_box_var_data[prior_var_offset];
-          box_var_y = prior_box_var_data[prior_var_offset + 1];
-          box_var_w = prior_box_var_data[prior_var_offset + 2];
-          box_var_h = prior_box_var_data[prior_var_offset + 3];
-        } else if (!(variance.empty())) {
-          box_var_x = static_cast<T>(variance[0]);
-          box_var_y = static_cast<T>(variance[1]);
-          box_var_w = static_cast<T>(variance[2]);
-          box_var_h = static_cast<T>(variance[3]);
+        int prior_var_offset = axis == 0 ? j * len : i * len;
+        if (var_size == 2) {
+          std::memcpy(var_ptr, prior_box_var_data + prior_var_offset,
+                      4 * sizeof(T));
+        } else if (var_size == 1) {
+          var_ptr = reinterpret_cast<T*>(variance.data());
         }
+        T box_var_x = *var_ptr;
+        T box_var_y = *(var_ptr + 1);
+        T box_var_w = *(var_ptr + 2);
+        T box_var_h = *(var_ptr + 3);
+
         target_box_center_x =
             box_var_x * target_box_data[offset] * prior_box_width +
             prior_box_center_x;
@@ -211,8 +199,31 @@ class BoxCoderKernel : public framework::OpKernel<T> {
       EncodeCenterSize(target_box, prior_box, prior_box_var, normalized,
                        variance, output);
     } else if (code_type == BoxCodeType::kDecodeCenterSize) {
-      DecodeCenterSize(target_box, prior_box, prior_box_var, normalized, axis,
-                       variance, output);
+      if (prior_box_var) {
+        if (axis == 0) {
+          DecodeCenterSize<0, 2>(target_box, prior_box, prior_box_var,
+                                 normalized, variance, output);
+        } else {
+          DecodeCenterSize<1, 2>(target_box, prior_box, prior_box_var,
+                                 normalized, variance, output);
+        }
+      } else if (!(variance.empty())) {
+        if (axis == 0) {
+          DecodeCenterSize<0, 1>(target_box, prior_box, prior_box_var,
+                                 normalized, variance, output);
+        } else {
+          DecodeCenterSize<1, 1>(target_box, prior_box, prior_box_var,
+                                 normalized, variance, output);
+        }
+      } else {
+        if (axis == 0) {
+          DecodeCenterSize<0, 0>(target_box, prior_box, prior_box_var,
+                                 normalized, variance, output);
+        } else {
+          DecodeCenterSize<1, 0>(target_box, prior_box, prior_box_var,
+                                 normalized, variance, output);
+        }
+      }
     }
   }
 };

paddle/fluid/operators/detection/density_prior_box_op.h

@@ -52,6 +52,10 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
       step_height = step_h;
     }
     int num_priors = 0;
+
+#ifdef PADDLE_WITH_MKLML
+#pragma omp parallel for reduction(+ : num_priors)
+#endif
     for (size_t i = 0; i < densities.size(); ++i) {
       num_priors += (fixed_ratios.size()) * (pow(densities[i], 2));
     }
@@ -64,6 +68,17 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
     auto e_boxes = framework::EigenTensor<T, 4>::From(*boxes).setConstant(0.0);
     int step_average = static_cast<int>((step_width + step_height) * 0.5);
 
+    std::vector<float> sqrt_fixed_ratios;
+#ifdef PADDLE_WITH_MKLML
+#pragma omp parallel for
+#endif
+    for (int i = 0; i < fixed_ratios.size(); i++) {
+      sqrt_fixed_ratios.push_back(sqrt(fixed_ratios[i]));
+    }
+
+#ifdef PADDLE_WITH_MKLML
+#pragma omp parallel for collapse(2)
+#endif
     for (int h = 0; h < feature_height; ++h) {
       for (int w = 0; w < feature_width; ++w) {
         T center_x = (w + offset) * step_width;
@@ -73,34 +88,25 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
         for (size_t s = 0; s < fixed_sizes.size(); ++s) {
           auto fixed_size = fixed_sizes[s];
           int density = densities[s];
+          int shift = step_average / density;
           // Generate density prior boxes with fixed ratios.
           for (size_t r = 0; r < fixed_ratios.size(); ++r) {
-            float ar = fixed_ratios[r];
-            int shift = step_average / density;
-            float box_width_ratio = fixed_size * sqrt(ar);
-            float box_height_ratio = fixed_size / sqrt(ar);
+            float box_width_ratio = fixed_size * sqrt_fixed_ratios[r];
+            float box_height_ratio = fixed_size / sqrt_fixed_ratios[r];
+            float density_center_x = center_x - step_average / 2. + shift / 2.;
+            float density_center_y = center_y - step_average / 2. + shift / 2.;
             for (int di = 0; di < density; ++di) {
               for (int dj = 0; dj < density; ++dj) {
-                float center_x_temp =
-                    center_x - step_average / 2. + shift / 2. + dj * shift;
-                float center_y_temp =
-                    center_y - step_average / 2. + shift / 2. + di * shift;
-                e_boxes(h, w, idx, 0) =
-                    (center_x_temp - box_width_ratio / 2.) / img_width >= 0
-                        ? (center_x_temp - box_width_ratio / 2.) / img_width
-                        : 0;
-                e_boxes(h, w, idx, 1) =
-                    (center_y_temp - box_height_ratio / 2.) / img_height >= 0
-                        ? (center_y_temp - box_height_ratio / 2.) / img_height
-                        : 0;
-                e_boxes(h, w, idx, 2) =
-                    (center_x_temp + box_width_ratio / 2.) / img_width <= 1
-                        ? (center_x_temp + box_width_ratio / 2.) / img_width
-                        : 1;
-                e_boxes(h, w, idx, 3) =
-                    (center_y_temp + box_height_ratio / 2.) / img_height <= 1
-                        ? (center_y_temp + box_height_ratio / 2.) / img_height
-                        : 1;
+                float center_x_temp = density_center_x + dj * shift;
+                float center_y_temp = density_center_y + di * shift;
+                e_boxes(h, w, idx, 0) = std::max(
+                    (center_x_temp - box_width_ratio / 2.) / img_width, 0.);
+                e_boxes(h, w, idx, 1) = std::max(
+                    (center_y_temp - box_height_ratio / 2.) / img_height, 0.);
+                e_boxes(h, w, idx, 2) = std::min(
+                    (center_x_temp + box_width_ratio / 2.) / img_width, 1.);
+                e_boxes(h, w, idx, 3) = std::min(
+                    (center_y_temp + box_height_ratio / 2.) / img_height, 1.);
                 idx++;
               }
             }
@@ -131,8 +137,14 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
     vars->Resize({box_num, static_cast<int>(variances.size())});
 
     auto e_vars = framework::EigenMatrix<T, Eigen::RowMajor>::From(*vars);
-
-    e_vars = var_et.broadcast(Eigen::DSizes<int, 2>(box_num, 1));
+#ifdef PADDLE_WITH_MKLML
+#pragma omp parallel for collapse(2)
+#endif
+    for (int i = 0; i < box_num; ++i) {
+      for (int j = 0; j < variances.size(); ++j) {
+        e_vars(i, j) = variances[j];
+      }
+    }
 
     vars->Resize(var_dim);
     boxes->Resize(box_dim);

paddle/fluid/operators/elementwise/elementwise_add_op.cc

@@ -18,6 +18,7 @@ namespace ops = paddle::operators;
 REGISTER_ELEMWISE_GRAD_MAKER(elementwise_add, Add);
 REGISTER_ELEMWISE_EXPLICIT_OP(elementwise_add, "Add", "Out = X + Y", "Out",
                               "X");
+
 REGISTER_OP_CPU_KERNEL(
     elementwise_add,
     ops::ElementwiseAddKernel<paddle::platform::CPUDeviceContext, float>,

paddle/fluid/operators/elementwise/elementwise_op.h

@@ -250,6 +250,20 @@ class ElemwiseGradKernel : public framework::OpKernel<T> {
   }
 };
 
+class ElementwiseOpInplace : public framework::InplaceInToOut {
+ public:
+  using framework::InplaceInToOut::InplaceInToOut;
+
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const framework::OpDesc &op_desc,
+      framework::BlockDesc *block) const override {
+    return std::unordered_map<std::string, std::string>{
+        {"X", "Out"},
+    };
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -299,6 +313,7 @@ class ElemwiseGradKernel : public framework::OpKernel<T> {
   REGISTER_OPERATOR(op_type, ::paddle::operators::ElementwiseOp,       \
                     __ElemwiseOp##op_type##Maker__,                    \
                     ::paddle::operators::ElementwiseOpInferVarType,    \
-                    op_type##GradMaker);                               \
+                    op_type##GradMaker,                                \
+                    ::paddle::operators::ElementwiseOpInplace);        \
   REGISTER_OPERATOR(op_type##_grad,                                    \
                     ::paddle::operators::ElementwiseOpExplicitGrad)

paddle/fluid/operators/flatten_op.cc

@@ -267,6 +267,35 @@ class Flatten2GradOp : public framework::OperatorBase {
   }
 };
 
+class FlattenOpInplaceInToOut : public framework::InplaceInToOut {
+ public:
+  using InplaceInToOut::InplaceInToOut;
+
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const framework::OpDesc &op_desc,
+      framework::BlockDesc *block) const override {
+    std::unordered_map<std::string, std::string> inplace_in_to_out = {
+        {"X", "Out"},
+    };
+    return inplace_in_to_out;
+  }
+};
+
+class FlattenGradInplaceinToOut : public framework::InplaceInToOut {
+  using InplaceInToOut::InplaceInToOut;
+
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const framework::OpDesc &op_desc,
+      framework::BlockDesc *block) const override {
+    std::unordered_map<std::string, std::string> inplace_in_to_out = {
+        {framework::GradVarName("Out"), framework::GradVarName("X")},
+    };
+    return inplace_in_to_out;
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -275,10 +304,13 @@ USE_OP(reshape);
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(flatten, ops::FlattenOp, ops::FlattenOpMaker,
                   ops::FlattenOpInferShape,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
-REGISTER_OPERATOR(flatten_grad, ops::FlattenGradOp, ops::FlattenGradInferShape);
+                  paddle::framework::DefaultGradOpDescMaker<true>,
+                  ops::FlattenOpInplaceInToOut);
+REGISTER_OPERATOR(flatten_grad, ops::FlattenGradOp, ops::FlattenGradInferShape,
+                  ops::FlattenGradInplaceinToOut);
 
 REGISTER_OPERATOR(flatten2, ops::Flatten2Op, ops::Flatten2OpMaker,
-                  ops::Flatten2OpInferShape, ops::Flatten2GradOpMaker);
+                  ops::Flatten2OpInferShape, ops::Flatten2GradOpMaker,
+                  ops::FlattenOpInplaceInToOut);
 REGISTER_OPERATOR(flatten2_grad, ops::Flatten2GradOp,
-                  ops::Flatten2GradInferShape);
+                  ops::Flatten2GradInferShape, ops::FlattenGradInplaceinToOut);

paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc

@@ -79,17 +79,17 @@ void FusionRepeatedFCReluOpMaker::Make() {
 }
 
 template <typename T>
-static void fc_relu(const T* x, const T* w, const T* b, T* y, int m, int n,
-                    int k) {
+static void fc_relu(const T* x, const T* w, const T* b, T* y,
+                    const jit::matmul_attr_t& attr) {
   auto matmul =
-      jit::Get<jit::kMatMul, jit::MatMulTuples<T>, platform::CPUPlace>(k);
+      jit::Get<jit::kMatMul, jit::MatMulTuples<T>, platform::CPUPlace>(attr);
   auto addbias_relu =
-      jit::Get<jit::kVAddRelu, jit::XYZNTuples<T>, platform::CPUPlace>(n);
-  matmul(x, w, y, m, n, k);
+      jit::Get<jit::kVAddRelu, jit::XYZNTuples<T>, platform::CPUPlace>(attr.n);
+  matmul(x, w, y, &attr);
   T* dst = y;
-  for (int i = 0; i < m; ++i) {
-    addbias_relu(b, dst, dst, n);
-    dst += n;
+  for (int i = 0; i < attr.m; ++i) {
+    addbias_relu(b, dst, dst, attr.n);
+    dst += attr.n;
   }
 }
 
@@ -107,32 +107,33 @@ class FusionRepeatedFCReluKernel : public framework::OpKernel<T> {
 
     auto i_dims = in->dims();
     auto w_dims = weights[0]->dims();
-    int m = i_dims[0];
-    int n = w_dims[1];
-    int k = w_dims[0];
-    relus[0]->Resize({m, n});
+    jit::matmul_attr_t attr;
+    attr.m = i_dims[0];
+    attr.n = w_dims[1];
+    attr.k = w_dims[0];
+    relus[0]->Resize({attr.m, attr.n});
     fc_relu(in->data<T>(), weights[0]->data<T>(), biases[0]->data<T>(),
-            relus[0]->mutable_data<T>(place), m, n, k);
+            relus[0]->mutable_data<T>(place), attr);
 
     for (int i = 1; i < weight_sz - 1; ++i) {
       auto i_dims = relus[i - 1]->dims();
       auto w_dims = weights[i]->dims();
-      int m = i_dims[0];
-      int n = w_dims[1];
-      int k = w_dims[0];
-      relus[i]->Resize({m, n});
+      attr.m = i_dims[0];
+      attr.n = w_dims[1];
+      attr.k = w_dims[0];
+      relus[i]->Resize({attr.m, attr.n});
       fc_relu(relus[i - 1]->data<T>(), weights[i]->data<T>(),
-              biases[i]->data<T>(), relus[i]->mutable_data<T>(place), m, n, k);
+              biases[i]->data<T>(), relus[i]->mutable_data<T>(place), attr);
     }
 
     auto i_dims_last = relus[weight_sz - 2]->dims();
     auto w_dims_last = weights[weight_sz - 1]->dims();
-    m = i_dims_last[0];
-    n = w_dims_last[1];
-    k = w_dims_last[0];
+    attr.m = i_dims_last[0];
+    attr.n = w_dims_last[1];
+    attr.k = w_dims_last[0];
     fc_relu(relus[weight_sz - 2]->data<T>(), weights[weight_sz - 1]->data<T>(),
-            biases[weight_sz - 1]->data<T>(), out->mutable_data<T>(place), m, n,
-            k);
+            biases[weight_sz - 1]->data<T>(), out->mutable_data<T>(place),
+            attr);
   }
 };
 

paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc

@@ -87,15 +87,18 @@ class FusionSquaredMatSubKernel : public framework::OpKernel<T> {
 
     auto x_dims = x->dims();
     auto y_dims = y->dims();
-    int m = x_dims[0];
-    int k = x_dims[1];
-    int n = y_dims[1];
-    int o_numel = m * n;
+    jit::matmul_attr_t attr;
+    attr.m = x_dims[0];
+    attr.k = x_dims[1];
+    attr.n = y_dims[1];
+    int o_numel = attr.m * attr.n;
 
     auto vsquare_x =
-        jit::Get<jit::kVSquare, jit::XYNTuples<T>, platform::CPUPlace>(m * k);
+        jit::Get<jit::kVSquare, jit::XYNTuples<T>, platform::CPUPlace>(attr.m *
+                                                                       attr.k);
     auto vsquare_y =
-        jit::Get<jit::kVSquare, jit::XYNTuples<T>, platform::CPUPlace>(k * n);
+        jit::Get<jit::kVSquare, jit::XYNTuples<T>, platform::CPUPlace>(attr.k *
+                                                                       attr.n);
     auto vsquare_xy =
         jit::Get<jit::kVSquare, jit::XYNTuples<T>, platform::CPUPlace>(o_numel);
     auto vsub =
@@ -103,7 +106,7 @@ class FusionSquaredMatSubKernel : public framework::OpKernel<T> {
     auto vscal =
         jit::Get<jit::kVScal, jit::AXYNTuples<T>, platform::CPUPlace>(o_numel);
     auto matmul =
-        jit::Get<jit::kMatMul, jit::MatMulTuples<T>, platform::CPUPlace>(k);
+        jit::Get<jit::kMatMul, jit::MatMulTuples<T>, platform::CPUPlace>(attr);
 
     const T* x_data = x->data<T>();
     const T* y_data = y->data<T>();
@@ -112,12 +115,12 @@ class FusionSquaredMatSubKernel : public framework::OpKernel<T> {
     T* squared_xy_data = squared_xy->mutable_data<T>(place);
     T* o_data = out->mutable_data<T>(place);
 
-    matmul(x_data, y_data, squared_xy_data, m, n, k);
+    matmul(x_data, y_data, squared_xy_data, &attr);
     vsquare_xy(squared_xy_data, squared_xy_data, o_numel);
 
-    vsquare_x(x_data, squared_x_data, m * k);
-    vsquare_y(y_data, squared_y_data, k * n);
-    matmul(squared_x_data, squared_y_data, o_data, m, n, k);
+    vsquare_x(x_data, squared_x_data, attr.m * attr.k);
+    vsquare_y(y_data, squared_y_data, attr.k * attr.n);
+    matmul(squared_x_data, squared_y_data, o_data, &attr);
 
     vsub(squared_xy_data, o_data, o_data, o_numel);
     vscal(&scalar, o_data, o_data, o_numel);

paddle/fluid/operators/jit/benchmark.cc

@@ -93,6 +93,7 @@ std::vector<int> TestSizes() {
 template <typename KernelTuples, typename... Args>
 struct BenchFunc {
   // return this function avg time
+  // TODO(TJ): clear cache every time
   double operator()(const typename KernelTuples::func_type tgt, Args... args) {
     for (int i = 0; i < FLAGS_burning; ++i) {
       tgt(args...);
@@ -172,6 +173,9 @@ void BenchXYZNKernel() {
     RandomVec<T>(d, y_data);
     BenchAllImpls<KT, jit::XYZNTuples<T>, PlaceType>(d, x.data<T>(),
                                                      y.data<T>(), z_data, d);
+    // test inplace
+    BenchAllImpls<KT, jit::XYZNTuples<T>, PlaceType>(d, x.data<T>(), z_data,
+                                                     z_data, d);
   }
 }
 
@@ -311,8 +315,9 @@ void BenchMatMulKernel() {
         const T* a_data = a.data<T>();
         const T* b_data = b.data<T>();
         T* c_data = c.mutable_data<T>(PlaceType());
-        BenchAllImpls<KT, jit::MatMulTuples<T>, PlaceType>(k, a_data, b_data,
-                                                           c_data, m, n, k);
+        const jit::matmul_attr_t attr{m, n, k};
+        BenchAllImpls<KT, jit::MatMulTuples<T>, PlaceType>(attr, a_data, b_data,
+                                                           c_data, &attr);
       }
     }
   }

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

@@ -9,6 +9,7 @@ function(USE_JITKERNEL_GEN TARGET)
 endfunction()
 
 # use gen jitcode kernel by name
+USE_JITKERNEL_GEN(kMatMul)
 USE_JITKERNEL_GEN(kVMul)
 USE_JITKERNEL_GEN(kVAdd)
 USE_JITKERNEL_GEN(kVSub)

paddle/fluid/operators/jit/gen/blas.cc

@@ -155,7 +155,7 @@ class NCHW16CMulNCCreator : public JitCodeCreator<int> {
   class name##Creator : public JitCodeCreator<int> {                         \
    public:                                                                   \
     bool UseMe(const int& attr) const override {                             \
-      return platform::MayIUse(platform::avx);                               \
+      return platform::MayIUse(platform::avx) && attr <= 1024;               \
     }                                                                        \
     size_t CodeSize(const int& d) const override {                           \
       return 96 + d / YMM_FLOAT_BLOCK * 4 * 8;                               \

paddle/fluid/operators/jit/gen/blas.h

@@ -61,6 +61,7 @@ class VXXJitCode : public JitCode {
       base += "_Vec";
     }
     base += (with_relu_ ? "_Relu" : "");
+    base += "_D" + std::to_string(num_);
     return base.c_str();
   }
   void genCode() override;

paddle/fluid/operators/jit/gen/matmul.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License. */
+
+#include "paddle/fluid/operators/jit/gen/matmul.h"
+#include <stddef.h>  // offsetof
+#include <vector>
+
+#include "paddle/fluid/operators/jit/registry.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+void MatMulJitCode::genCode() {
+  preCode();
+  int block, rest;
+  const auto groups = packed_groups(n_, k_, &block, &rest);
+  PADDLE_ENFORCE_GT(groups.front(), 0);
+
+  const int block_len = sizeof(float) * block;
+  const int x_reg_idx = (block == ZMM_FLOAT_BLOCK ? 32 : 16) - 1;
+  const int w_reg_idx = x_reg_idx - 1;
+  // from packed mov(reg_ptr_wgt, ptr[param_attr + offsetof(matmul_attr_t,
+  // packed_weight)]);
+  mov(reg_ptr_wgt, param_y);
+  size_t z_offset = 0;
+  size_t wgt_offset = 0;
+  for (size_t g = 0; g < groups.size(); ++g) {
+    size_t x_offset = 0;
+    for (int k = 0; k < k_; ++k) {
+      vbroadcastss(zmm_t(x_reg_idx), ptr[param_x + x_offset]);
+      // clean
+      if (k == 0) {
+        for (int i = 0; i < groups[g]; ++i) {
+          vxorps(zmm_t(i), zmm_t(i), zmm_t(i));
+        }
+      }
+      for (int i = 0; i < groups[g]; ++i) {
+        vmovups(zmm_t(w_reg_idx), ptr[reg_ptr_wgt + wgt_offset]);
+        vfmadd231ps(zmm_t(i), zmm_t(w_reg_idx), zmm_t(x_reg_idx));
+        wgt_offset += block_len;
+      }
+      // last one, save
+      if (k == k_ - 1) {
+        for (int i = 0; i < groups[g]; ++i) {
+          // only rest save should be careful
+          if (rest != 0 && g == groups.size() - 1 && i == groups[g] - 1) {
+            break;
+          }
+          vmovups(ptr[param_z + z_offset + i * block_len], zmm_t(i));
+        }
+      }
+      x_offset += sizeof(float);
+    }
+    z_offset += block_len * groups[g];
+  }
+
+  if (rest != 0) {
+    // below should refine with mask
+    int reg_idx = groups.back() - 1;
+    z_offset = (n_ - rest) * sizeof(float);
+    int inner_block = 8;
+    while (rest > 0) {
+      if (rest >= 8) {
+        inner_block = 8;
+        vmovups(ptr[param_z + z_offset], ymm_t(reg_idx));
+        // shift zmm of inner_block, change reg_idx if update
+      } else if (rest >= 4) {
+        inner_block = 4;
+        vmovups(ptr[param_z + z_offset], xmm_t(reg_idx));
+      } else if (rest >= 2) {
+        inner_block = 2;
+        vmovq(ptr[param_z + z_offset], xmm_t(reg_idx));
+      } else {
+        inner_block = 1;
+        vmovss(ptr[param_z + z_offset], xmm_t(reg_idx));
+      }
+      z_offset += inner_block * sizeof(float);
+      rest -= inner_block;
+    }
+  }
+
+  postCode();
+}
+
+class MatMulCreator : public JitCodeCreator<matmul_attr_t> {
+ public:
+  bool UseMe(const matmul_attr_t& attr) const override {
+    return attr.m == 1 && platform::MayIUse(platform::avx512f) &&
+           attr.n % ZMM_FLOAT_BLOCK == 0 && attr.k < 512;
+  }
+  size_t CodeSize(const matmul_attr_t& attr) const override {
+    int block = YMM_FLOAT_BLOCK;
+    if (platform::MayIUse(platform::avx512f)) {
+      block = ZMM_FLOAT_BLOCK;
+    }
+    return 96 + 4 * attr.k * (attr.n / block + 1) * 8;
+  }
+  std::unique_ptr<GenBase> CreateJitCode(
+      const matmul_attr_t& attr) const override {
+    PADDLE_ENFORCE_GT(attr.m, 0);
+    PADDLE_ENFORCE_GT(attr.n, 0);
+    PADDLE_ENFORCE_GT(attr.k, 0);
+    return make_unique<MatMulJitCode>(attr, CodeSize(attr));
+  }
+};
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle
+
+namespace gen = paddle::operators::jit::gen;
+
+REGISTER_JITKERNEL_GEN(kMatMul, gen::MatMulCreator);

paddle/fluid/operators/jit/gen/matmul.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License. */
+
+#pragma once
+
+#include <stdlib.h>  // for malloc and free
+#include <string>
+#include <vector>
+#include "glog/logging.h"
+#include "paddle/fluid/operators/jit/gen/jitcode.h"
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+class MatMulJitCode : public JitCode {
+ public:
+  explicit MatMulJitCode(const matmul_attr_t& attr,
+                         size_t code_size = 256 * 1024,
+                         void* code_ptr = nullptr)
+      : JitCode(code_size, code_ptr), m_(attr.m), n_(attr.n), k_(attr.k) {
+    PADDLE_ENFORCE_EQ(m_, 1, "Only support m==1 yet");
+    this->genCode();
+  }
+
+  virtual const char* name() const {
+    std::string base = "MatMulJitCode";
+    base = base + "_M" + std::to_string(m_) + "_N" + std::to_string(n_) + "_K" +
+           std::to_string(k_);
+    return base.c_str();
+  }
+  void genCode() override;
+
+ private:
+  int m_, n_, k_;
+
+  reg64_t param_x{abi_param1};
+  reg64_t param_y{abi_param2};
+  reg64_t param_z{abi_param3};
+  reg64_t param_attr{abi_param4};
+  reg64_t reg_tmp{rax};
+
+  reg64_t reg_ptr_wgt{r10};
+};
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/gen_base.cc

@@ -16,6 +16,8 @@
 #include <fstream>
 #include <iostream>
 #include <sstream>
+#include <vector>
+#include "paddle/fluid/platform/cpu_info.h"
 
 DEFINE_bool(dump_jitcode, false, "Whether to dump the jitcode to file");
 
@@ -38,6 +40,35 @@ void GenBase::dumpCode(const unsigned char* code) const {
   }
 }
 
+std::vector<int> packed_groups(int n, int k, int* block_out, int* rest_out) {
+  int block;
+  int max_num_regs;
+  if (platform::MayIUse(platform::avx512f)) {
+    block = ZMM_FLOAT_BLOCK;
+    max_num_regs = 32;
+  } else {
+    block = YMM_FLOAT_BLOCK;
+    max_num_regs = 16;
+  }
+  // one for x, one for y, others for z
+  const int max_used_regs_for_n = max_num_regs - 2;
+  const int aligned_n = n % block == 0 ? n : (n / block + 1) * block;
+  const int num_block = aligned_n / block;
+  const int num_groups = num_block / max_used_regs_for_n;
+  std::vector<int> groups(num_groups, max_used_regs_for_n);
+  int rest_num_regs = num_block % max_used_regs_for_n;
+  if (rest_num_regs != 0) {
+    groups.push_back(rest_num_regs);
+  }
+  if (block_out) {
+    *block_out = block;
+  }
+  if (rest_out) {
+    *rest_out = n % block;
+  }
+  return groups;
+}
+
 }  // namespace jit
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/jit/gen_base.h

@@ -16,6 +16,7 @@
 
 #include <gflags/gflags.h>
 #include <memory>  // for unique_ptr
+#include <vector>
 #include "paddle/fluid/operators/jit/kernel_base.h"
 
 DECLARE_bool(dump_jitcode);
@@ -67,6 +68,11 @@ class JitCodeCreator : public GenCreator {
   virtual std::unique_ptr<GenBase> CreateJitCode(const Attr& attr) const = 0;
 };
 
+// unify the method of packed groups
+// output the packed groups which used in weights, the block size and rest size
+std::vector<int> packed_groups(int n, int k, int* block = nullptr,
+                               int* rest = nullptr);
+
 }  // namespace jit
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/jit/helper.cc

@@ -14,6 +14,8 @@
 
 #include "paddle/fluid/operators/jit/helper.h"
 #include <algorithm>  // tolower
+#include <numeric>
+#include <string>
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {
@@ -91,6 +93,41 @@ KernelType to_kerneltype(const std::string& act) {
   return kNone;
 }
 
+template <>
+void pack_weights<float>(const float* src, float* dst, int n, int k) {
+  int block, rest;
+  const auto groups = packed_groups(n, k, &block, &rest);
+  std::for_each(groups.begin(), groups.end(), [&](int i) {
+    PADDLE_ENFORCE_GT(i, 0, "each element of groups should be larger than 0.");
+  });
+  int sum = std::accumulate(groups.begin(), groups.end(), 0);
+  std::memset(dst, 0, k * sum * block * sizeof(float));
+  PADDLE_ENFORCE_GE(sum * block, n,
+                    "The packed n should be equal to or larger than n");
+
+  const int block_len = sizeof(float) * block;
+  int n_offset = 0;
+
+  for (size_t g = 0; g < groups.size(); ++g) {
+    const float* from = src + n_offset;
+    for (int j = 0; j < k; ++j) {
+      size_t copy_sz = groups[g] * block_len;
+      if (g == groups.size() - 1 && rest != 0) {
+        copy_sz = (groups[g] - 1) * block_len + rest * sizeof(float);
+      }
+      std::memcpy(dst, from + j * n, copy_sz);
+      dst += groups[g] * block;
+    }
+    n_offset += groups[g] * block;
+  }
+}
+
+template <typename T>
+typename std::enable_if<!std::is_same<T, float>::value>::type pack_weights(
+    const T* src, T* dst, int n, int k) {
+  PADDLE_THROW("Only support pack with float type.");
+}
+
 }  // namespace jit
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/jit/helper.h

@@ -118,26 +118,33 @@ typename KernelTuples::func_type Get(
   return GetRefer<KT, KernelTuples>();
 }
 
-template <KernelType KT, typename KernelTuples>
-class KernelFuncsCache {
+template <KernelType KT, typename KernelTuples, typename PlaceType>
+class KernelFuncs {
  public:
-  KernelFuncsCache() = default;
-  static KernelFuncsCache& Instance() {
-    static thread_local KernelFuncsCache<KT, KernelTuples> g_func_cache;
+  KernelFuncs() = default;
+  static KernelFuncs& Cache() {
+    static thread_local KernelFuncs<KT, KernelTuples, PlaceType> g_func_cache;
     return g_func_cache;
   }
 
   bool Has(int key) const { return funcs_.find(key) != funcs_.end(); }
 
-  typename KernelTuples::func_type At(int key) { return funcs_.at(key); }
-
   void Insert(int key, typename KernelTuples::func_type func) {
     funcs_.emplace(key, func);
   }
 
+  typename KernelTuples::func_type At(int key) {
+    if (Has(key)) {
+      return funcs_.at(key);
+    }
+    auto func = Get<KT, KernelTuples, PlaceType>(key);
+    Insert(key, func);
+    return func;
+  }
+
  private:
   std::unordered_map<int, typename KernelTuples::func_type> funcs_;
-  DISABLE_COPY_AND_ASSIGN(KernelFuncsCache);
+  DISABLE_COPY_AND_ASSIGN(KernelFuncs);
 };
 
 const char* to_string(KernelType kt);
@@ -152,17 +159,28 @@ inline std::ostream& operator<<(std::ostream& os, const lstm_attr_t& attr) {
      << (attr.use_peephole ? "True" : "False") << "]";
   return os;
 }
+
 inline std::ostream& operator<<(std::ostream& os, const gru_attr_t& attr) {
   os << "dim_size[" << attr.d << "],act_gate[" << to_string(attr.act_gate)
      << "],act_cand[" << to_string(attr.act_cand) << "]";
   return os;
 }
+
 inline std::ostream& operator<<(std::ostream& os, const seq_pool_attr_t& attr) {
   os << "height_size[" << attr.h << "],width_size[" << attr.w << "],pool_type["
      << to_string(attr.type) << "]";
   return os;
 }
 
+inline std::ostream& operator<<(std::ostream& os, const matmul_attr_t& attr) {
+  os << "M[" << attr.m << "],N[" << attr.n << "],K[" << attr.k << "]";
+  return os;
+}
+
+// expose the method to pack matmul weight
+template <typename T>
+void pack_weights(const T* src, T* dst, int n, int k);
+
 }  // namespace jit
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/jit/kernel_base.h

@@ -145,11 +145,19 @@ struct SeqPoolTuples {
   typedef void (*func_type)(const T*, T*, const seq_pool_attr_t*);
 };
 
+typedef struct matmul_attr_s {
+  int m, n, k;
+  void* packed_weight{nullptr};
+  matmul_attr_s() = default;
+  explicit matmul_attr_s(int m_, int n_, int k_, void* packed_weight_ = nullptr)
+      : m(m_), n(n_), k(k_), packed_weight(packed_weight_) {}
+} matmul_attr_t;
+
 template <typename T>
 struct MatMulTuples {
   typedef T data_type;
-  typedef int attr_type;
-  typedef void (*func_type)(const T*, const T*, T*, int, int, int);
+  typedef matmul_attr_t attr_type;
+  typedef void (*func_type)(const T*, const T*, T*, const matmul_attr_t*);
 };
 
 template <typename T>

paddle/fluid/operators/jit/kernel_key.cc

@@ -49,6 +49,13 @@ size_t JitCodeKey<seq_pool_attr_t>(const seq_pool_attr_t& attr) {
   return (key << pool_type_shift) + static_cast<int>(attr.type);
 }
 
+template <>
+size_t JitCodeKey<matmul_attr_t>(const matmul_attr_t& attr) {
+  size_t key = attr.m;
+  constexpr int shift = 21;
+  return (key << shift * 2) + ((static_cast<size_t>(attr.n)) << shift) + attr.k;
+}
+
 }  // namespace jit
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/jit/more/mix/mix.cc

@@ -49,49 +49,16 @@ void VTanh(const T* x, T* y, int n) {
 }
 
 void Softmax(const T* x, T* y, int n, int bs) {
-  typename XRNTuples<T>::func_type compute_hmax{nullptr};
-  typename XRNTuples<T>::func_type compute_hsum{nullptr};
-  typename AXYNTuples<T>::func_type compute_vscal{nullptr};
-  typename AXYNTuples<T>::func_type compute_vaddbias{nullptr};
-  typename XYNTuples<T>::func_type compute_vexp{nullptr};
-
-  if (!KernelFuncsCache<kHMax, XRNTuples<T>>::Instance().Has(n)) {
-    compute_hmax = Get<kHMax, XRNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kHMax, XRNTuples<T>>::Instance().Insert(n, compute_hmax);
-  } else {
-    compute_hmax = KernelFuncsCache<kHMax, XRNTuples<T>>::Instance().At(n);
-  }
-
-  if (!KernelFuncsCache<kHSum, XRNTuples<T>>::Instance().Has(n)) {
-    compute_hsum = Get<kHSum, XRNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kHSum, XRNTuples<T>>::Instance().Insert(n, compute_hsum);
-  } else {
-    compute_hsum = KernelFuncsCache<kHSum, XRNTuples<T>>::Instance().At(n);
-  }
-
-  if (!KernelFuncsCache<kVScal, AXYNTuples<T>>::Instance().Has(n)) {
-    compute_vscal = Get<kVScal, AXYNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kVScal, AXYNTuples<T>>::Instance().Insert(n,
-                                                               compute_vscal);
-  } else {
-    compute_vscal = KernelFuncsCache<kVScal, AXYNTuples<T>>::Instance().At(n);
-  }
-
-  if (!KernelFuncsCache<kVAddBias, AXYNTuples<T>>::Instance().Has(n)) {
-    compute_vaddbias = Get<kVAddBias, AXYNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kVAddBias, AXYNTuples<T>>::Instance().Insert(
-        n, compute_vaddbias);
-  } else {
-    compute_vaddbias =
-        KernelFuncsCache<kVAddBias, AXYNTuples<T>>::Instance().At(n);
-  }
-
-  if (!KernelFuncsCache<kVExp, XYNTuples<T>>::Instance().Has(n)) {
-    compute_vexp = Get<KernelType::kVExp, XYNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kVExp, XYNTuples<T>>::Instance().Insert(n, compute_vexp);
-  } else {
-    compute_vexp = KernelFuncsCache<kVExp, XYNTuples<T>>::Instance().At(n);
-  }
+  auto compute_hmax =
+      KernelFuncs<kHMax, XRNTuples<T>, platform::CPUPlace>::Cache().At(n);
+  auto compute_hsum =
+      KernelFuncs<kHSum, XRNTuples<T>, platform::CPUPlace>::Cache().At(n);
+  auto compute_vscal =
+      KernelFuncs<kVScal, AXYNTuples<T>, platform::CPUPlace>::Cache().At(n);
+  auto compute_vaddbias =
+      KernelFuncs<kVAddBias, AXYNTuples<T>, platform::CPUPlace>::Cache().At(n);
+  auto compute_vexp =
+      KernelFuncs<kVExp, XYNTuples<T>, platform::CPUPlace>::Cache().At(n);
 
   for (int i = 0; i < bs; ++i) {
     T scalar;

paddle/fluid/operators/jit/more/mkl/mkl.cc

@@ -25,17 +25,19 @@ namespace more {
 namespace mkl {
 
 template <>
-void MatMul<float>(const float* a, const float* b, float* c, int m, int n,
-                   int k) {
-  platform::dynload::cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, m,
-                                 n, k, 1.f, a, k, b, n, 0.f, c, n);
+void MatMul<float>(const float* a, const float* b, float* c,
+                   const matmul_attr_t* attr) {
+  platform::dynload::cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans,
+                                 attr->m, attr->n, attr->k, 1.f, a, attr->k, b,
+                                 attr->n, 0.f, c, attr->n);
 }
 
 template <>
-void MatMul<double>(const double* a, const double* b, double* c, int m, int n,
-                    int k) {
-  platform::dynload::cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, m,
-                                 n, k, 1.0, a, k, b, n, 0.0, c, n);
+void MatMul<double>(const double* a, const double* b, double* c,
+                    const matmul_attr_t* attr) {
+  platform::dynload::cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans,
+                                 attr->m, attr->n, attr->k, 1.0, a, attr->k, b,
+                                 attr->n, 0.0, c, attr->n);
 }
 
 template <>
@@ -127,11 +129,6 @@ void ASum<double>(const double* x, double* res, int n) {
 }
 
 // TODO(TJ): tuning me carefully on AVX, AVX2 and AVX512
-template <>
-bool MatMulKernel<float>::UseMe(const int& d) const {
-  return platform::MayIUse(platform::avx);
-}
-
 template <>
 bool VMulKernel<float>::UseMe(const int& d) const {
   return platform::MayIUse(platform::avx512f) && d > 512;
@@ -139,7 +136,7 @@ bool VMulKernel<float>::UseMe(const int& d) const {
 
 template <>
 bool VAddKernel<float>::UseMe(const int& d) const {
-  return platform::MayIUse(platform::avx512f) && d > 512;
+  return platform::MayIUse(platform::avx) && d > 512;
 }
 
 template <>
@@ -177,6 +174,16 @@ bool SeqPoolKernel<double>::UseMe(const seq_pool_attr_t& attr) const {
   return true;
 }
 
+template <>
+bool MatMulKernel<float>::UseMe(const matmul_attr_t& attr) const {
+  return platform::MayIUse(platform::avx);
+}
+
+template <>
+bool MatMulKernel<double>::UseMe(const matmul_attr_t& attr) const {
+  return true;
+}
+
 template <>
 bool SoftmaxKernel<float>::UseMe(const int& d) const {
   // tuned on avx2
@@ -189,7 +196,6 @@ bool SoftmaxKernel<float>::UseMe(const int& d) const {
     return true;                                         \
   }
 
-AWALYS_USE_ME_WITH_DOUBLE(MatMul);
 AWALYS_USE_ME_WITH_DOUBLE(VMul);
 AWALYS_USE_ME_WITH_DOUBLE(VAdd);
 AWALYS_USE_ME_WITH_DOUBLE(VScal);

paddle/fluid/operators/jit/more/mkl/mkl.h

@@ -26,7 +26,7 @@ namespace more {
 namespace mkl {
 
 template <typename T>
-void MatMul(const T* a, const T* b, T* c, int m, int n, int k);
+void MatMul(const T* a, const T* b, T* c, const matmul_attr_t* attr);
 
 template <typename T>
 void VMul(const T* x, const T* y, T* z, int n);

paddle/fluid/operators/jit/refer/refer.h

@@ -363,17 +363,19 @@ void SeqPool(const T* x, T* y, const seq_pool_attr_t* attr) {
 
 // A(M,K) * B(K,N) = C(M,N)
 template <typename T>
-void MatMul(const T* A, const T* B, T* C, int M, int N, int K) {
+void MatMul(const T* A, const T* B, T* C, const matmul_attr_t* attr) {
+  int M = attr->m;
+  int N = attr->n;
+  int K = attr->k;
   for (int m = 0; m < M; ++m) {
     const T* pa = A + m * K;
     T* pc = C + m * N;
     for (int n = 0; n < N; ++n) {
       const T* pb = B + n;
-      T sum = static_cast<T>(0);
-      for (int k = 0; k < K; ++k) {
-        sum += (pa[k] * pb[k * N]);
+      pc[n] = pa[0] * pb[0];
+      for (int k = 1; k < K; ++k) {
+        pc[n] += pa[k] * pb[k * N];
       }
-      *(pc + n) = sum;
     }
   }
 }

paddle/fluid/operators/jit/test.cc

@@ -22,7 +22,7 @@
 #include "paddle/fluid/platform/cpu_info.h"
 #include "paddle/fluid/platform/place.h"
 
-static double acc = 1e-5;
+DEFINE_double(acc, 1e-5, "Test accuracy threshold.");
 
 template <typename T>
 void RandomVec(const int n, T* a, const T lower = static_cast<T>(-20.f),
@@ -39,7 +39,7 @@ template <typename T>
 void ExpectEQ(const T* target, const T* refer, int n) {
   if (std::is_floating_point<T>::value) {
     for (int i = 0; i < n; ++i) {
-      EXPECT_NEAR(target[i], refer[i], acc);
+      EXPECT_NEAR(target[i], refer[i], FLAGS_acc);
     }
   } else {
     for (int i = 0; i < n; ++i) {
@@ -272,21 +272,23 @@ struct TestFuncWithRefer<jit::SeqPoolTuples<T>, std::vector<T>, std::vector<T>,
 
 template <typename T>
 struct TestFuncWithRefer<jit::MatMulTuples<T>, std::vector<T>, std::vector<T>,
-                         std::vector<T>, int, int, int> {
+                         std::vector<T>,
+                         typename jit::MatMulTuples<T>::attr_type> {
   void operator()(const typename jit::MatMulTuples<T>::func_type tgt,
                   const std::vector<T>& a, const std::vector<T>& b,
-                  const std::vector<T>& cref, int m, int n, int k) {
+                  const std::vector<T>& cref,
+                  const typename jit::MatMulTuples<T>::attr_type& attr) {
     EXPECT_TRUE(tgt != nullptr);
-    EXPECT_EQ(a.size(), static_cast<size_t>(m * k));
-    EXPECT_EQ(b.size(), static_cast<size_t>(k * n));
-    EXPECT_EQ(cref.size(), static_cast<size_t>(m * n));
+    EXPECT_EQ(a.size(), static_cast<size_t>(attr.m * attr.k));
+    EXPECT_EQ(b.size(), static_cast<size_t>(attr.k * attr.n));
+    EXPECT_EQ(cref.size(), static_cast<size_t>(attr.m * attr.n));
     std::vector<T> c(cref.size());
     const T* a_data = a.data();
     const T* b_data = b.data();
     const T* cref_data = cref.data();
     T* c_data = c.data();
-    tgt(a_data, b_data, c_data, m, n, k);
-    ExpectEQ<T>(c_data, cref_data, m * n);
+    tgt(a_data, b_data, c_data, &attr);
+    ExpectEQ<T>(c_data, cref_data, attr.m * attr.n);
   }
 };
 
@@ -383,8 +385,8 @@ void TestAXYNKernel() {
 template <jit::KernelType KT, typename T, typename PlaceType>
 void TestXRNKernel() {
   VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
-  auto last_acc = acc;
-  acc = 1e-4;
+  auto last_acc = FLAGS_acc;
+  FLAGS_acc = 1e-4;
   for (int d : TestSizes()) {
     auto ref = jit::GetRefer<KT, jit::XRNTuples<T>>();
     EXPECT_TRUE(ref != nullptr);
@@ -395,7 +397,7 @@ void TestXRNKernel() {
     TestAllImpls<KT, jit::XRNTuples<T>, PlaceType, std::vector<T>, T>(d, x,
                                                                       ref_res);
   }
-  acc = last_acc;
+  FLAGS_acc = last_acc;
 }
 
 template <jit::KernelType KT, typename T, typename PlaceType>
@@ -535,9 +537,10 @@ void TestSeqPoolKernel() {
 template <jit::KernelType KT, typename T, typename PlaceType>
 void TestMatMulKernel() {
   VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
-  auto last_acc = acc;
-  // TODO(intel): this should be acc issue of MKL
-  acc = 1e-3;
+  auto last_acc = FLAGS_acc;
+  // TODO(intel): fix MKL acc issue
+  // https://github.com/PaddlePaddle/Paddle/issues/15447
+  FLAGS_acc = 1e-3;
   for (int m : {1, 2, 3, 4}) {
     for (int n : {1, 2, 3, 4}) {
       for (int k : TestSizes()) {
@@ -549,13 +552,14 @@ void TestMatMulKernel() {
         const T* a_data = a.data();
         const T* b_data = b.data();
         T* c_data = c.data();
-        ref(a_data, b_data, c_data, m, n, k);
+        const jit::matmul_attr_t attr{m, n, k};
+        ref(a_data, b_data, c_data, &attr);
         TestAllImpls<KT, jit::MatMulTuples<T>, PlaceType, std::vector<T>,
-                     std::vector<T>, std::vector<T>>(k, a, b, c, m, n, k);
+                     std::vector<T>, std::vector<T>>(attr, a, b, c, attr);
       }
     }
   }
-  acc = last_acc;
+  FLAGS_acc = last_acc;
 }
 
 template <jit::KernelType KT, typename T, typename PlaceType>

paddle/fluid/operators/math/CMakeLists.txt

@@ -37,7 +37,7 @@ math_library(concat_and_split)
 math_library(context_project DEPS im2col math_function)
 math_library(cross_entropy)
 math_library(cos_sim_functor)
-math_library(depthwise_conv)
+math_library(depthwise_conv DEPS cub)
 math_library(im2col)
 math_library(sample_prob)
 math_library(sampler)

paddle/fluid/operators/math/fc_compute.h

@@ -30,15 +30,17 @@ inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
     return;
   }
   if (relu) {
-    auto compute =
-        jit::Get<jit::kVAddRelu, jit::XYZNTuples<T>, platform::CPUPlace>(N);
+    auto compute = jit::KernelFuncs<jit::kVAddRelu, jit::XYZNTuples<T>,
+                                    platform::CPUPlace>::Cache()
+                       .At(N);
     for (int i = 0; i < M; i++) {
       T* dst = Y + i * N;
       compute(B, dst, dst, N);
     }
   } else {
-    auto compute =
-        jit::Get<jit::kVAdd, jit::XYZNTuples<T>, platform::CPUPlace>(N);
+    auto compute = jit::KernelFuncs<jit::kVAdd, jit::XYZNTuples<T>,
+                                    platform::CPUPlace>::Cache()
+                       .At(N);
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for
 #endif

paddle/fluid/operators/math/softmax_impl.h

@@ -82,8 +82,9 @@ class SoftmaxFunctor<DeviceContext, float, true, enable_if_CPU<DeviceContext>> {
     const int kClassDim = 1;
     // 2D data. Batch x C
     auto compute_softmax =
-        jit::Get<jit::kSoftmax, jit::SoftmaxTuples<float>, platform::CPUPlace>(
-            in_dims[kClassDim]);
+        jit::KernelFuncs<jit::kSoftmax, jit::SoftmaxTuples<float>,
+                         platform::CPUPlace>::Cache()
+            .At(in_dims[kClassDim]);
     compute_softmax(in_data, out_data, in_dims[kClassDim], in_dims[kBatchDim]);
   }
 };

paddle/fluid/operators/ngraph/ngraph_bridge.cc

@@ -31,6 +31,9 @@ std::map<std::string,
                             std::shared_ptr<std::unordered_map<
                                 std::string, std::shared_ptr<ngraph::Node>>>)>>
     NgraphBridge::NG_NODE_MAP = {
+        {"accuracy", NG_OPS::BuildAccuracyNode},
+        {"conv2d", NG_OPS::BuildConv2dNode},
+        {"conv2d_grad", NG_OPS::BuildConv2dGradNode},
         {"elementwise_add", NG_OPS::BuildElementwiseAddNode},
         {"elementwise_add_grad", NG_OPS::BuildElementwiseAddGradNode},
         {"fill_constant", NG_OPS::BuildFillConstantNode},

paddle/fluid/operators/ngraph/ngraph_ops.h

@@ -21,7 +21,9 @@ limitations under the License. */
 
 #pragma once
 
-#include "ops/binary_unnary_op.h"
+#include "ops/accuracy_op.h"
+#include "ops/binary_unary_op.h"
+#include "ops/conv2d_op.h"
 #include "ops/elementwise_add_op.h"
 #include "ops/fill_constant_op.h"
 #include "ops/mean_op.h"

paddle/fluid/operators/ngraph/ops/accuracy_op.h

+/*Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <string>
+#include <vector>
+#include "ngraph/ngraph.hpp"
+#include "paddle/fluid/platform/ngraph_helper.h"
+
+namespace paddle {
+namespace operators {
+namespace ngraphs {
+
+void BuildAccuracyNode(
+    const std::shared_ptr<framework::OperatorBase>& op,
+    std::shared_ptr<
+        std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
+        ngb_node_map) {
+  auto indices = platform::GetInputNode(op, "Indices", ngb_node_map);
+  auto label = platform::GetInputNode(op, "Label", ngb_node_map);
+  auto inference = platform::GetInputNode(op, "Out", ngb_node_map);
+  auto inference_shape = inference->get_shape();
+  size_t num_samples = inference_shape.at(0);
+  size_t k = inference_shape.at(1);
+
+  std::shared_ptr<ngraph::Node> label_k = label;
+  if (k > 1) {
+    auto label_1d = std::make_shared<ngraph::op::Reshape>(
+        label, ngraph::AxisVector{0, 1}, ngraph::Shape{num_samples});
+    label_k = std::make_shared<ngraph::op::Broadcast>(label_1d, inference_shape,
+                                                      ngraph::AxisSet{1});
+  }
+
+  auto node_equal = std::make_shared<ngraph::op::Equal>(indices, label_k);
+  auto node_eq_int =
+      std::make_shared<ngraph::op::Convert>(node_equal, ngraph::element::i64);
+  auto num_correct_0d =
+      std::make_shared<ngraph::op::Sum>(node_eq_int, ngraph::AxisSet{0, 1});
+  std::shared_ptr<ngraph::Node> num_correct =
+      platform::NgReshaper(num_correct_0d, ngraph::Shape{1});
+  std::shared_ptr<ngraph::Node> n_samples = ngraph::op::Constant::create(
+      ngraph::element::i64, ngraph::Shape{1}, {num_samples});
+  std::shared_ptr<ngraph::Node> accuracy = std::make_shared<ngraph::op::Divide>(
+      std::make_shared<ngraph::op::Convert>(num_correct, ngraph::element::f32),
+      std::make_shared<ngraph::op::Convert>(n_samples, ngraph::element::f32));
+
+  platform::SetOutputNode(op, "Accuracy", accuracy, ngb_node_map);
+  platform::SetOutputNode(op, "Correct", num_correct, ngb_node_map);
+  platform::SetOutputNode(op, "Total", n_samples, ngb_node_map);
+}
+}  // namespace ngraphs
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/ngraph/ops/conv2d_op.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <string>
+#include <vector>
+#include "ngraph/ngraph.hpp"
+#include "paddle/fluid/platform/ngraph_helper.h"
+
+namespace paddle {
+namespace operators {
+namespace ngraphs {
+
+std::shared_ptr<ngraph::Node> GroupedConvolution(
+    const std::shared_ptr<ngraph::Node>& data_batch,
+    const std::shared_ptr<ngraph::Node>& filters, const ngraph::Strides strides,
+    const ngraph::Strides dilations, const ngraph::CoordinateDiff& paddings,
+    size_t groups) {
+  auto& data_shape = data_batch->get_shape();
+  auto& filter_shape = filters->get_shape();
+  ngraph::NodeVector ng_slices;
+
+  for (size_t i = 0; i < groups; ++i) {
+    size_t channel_step = filter_shape.at(1);
+    const std::vector<size_t> lower_bound{0, i * channel_step, 0, 0};
+    const std::vector<size_t> upper_bound{data_shape.at(0),
+                                          (i + 1) * channel_step,
+                                          data_shape.at(2), data_shape.at(3)};
+    auto data_slice = std::make_shared<ngraph::op::Slice>(
+        data_batch, lower_bound, upper_bound);
+
+    size_t filter_step = filter_shape.at(0) / groups;
+    const std::vector<size_t> filter_lower_bound{i * filter_step, 0, 0, 0};
+    const std::vector<size_t> filter_upper_bound{
+        (i + 1) * filter_step, filter_shape.at(1), filter_shape.at(2),
+        filter_shape.at(3)};
+    auto filter_slice = std::make_shared<ngraph::op::Slice>(
+        filters, filter_lower_bound, filter_upper_bound);
+    auto ng_conv = std::make_shared<ngraph::op::Convolution>(
+        data_slice, filter_slice, strides, dilations, paddings, paddings);
+    ng_slices.push_back(ng_conv);
+  }
+
+  size_t concat_axis = 1;
+  return std::make_shared<ngraph::op::Concat>(ng_slices, concat_axis);
+}
+
+std::shared_ptr<ngraph::Node> GroupedGradConvolutionFilter(
+    const std::shared_ptr<ngraph::Node>& data_batch,
+    const std::shared_ptr<ngraph::Node>& filters,
+    const std::shared_ptr<ngraph::Node>& doutput, const ngraph::Strides strides,
+    const ngraph::Strides dilations, const ngraph::CoordinateDiff& paddings,
+    size_t groups) {
+  auto& data_shape = data_batch->get_shape();
+  auto& filter_shape = filters->get_shape();
+  auto& out_shape = doutput->get_shape();
+  ngraph::NodeVector ng_slices;
+
+  for (size_t i = 0; i < groups; ++i) {
+    size_t channel_step = filter_shape.at(1);
+    const std::vector<size_t> lower_bound{0, i * channel_step, 0, 0};
+    const std::vector<size_t> upper_bound{data_shape.at(0),
+                                          (i + 1) * channel_step,
+                                          data_shape.at(2), data_shape.at(3)};
+    auto data_slice = std::make_shared<ngraph::op::Slice>(
+        data_batch, lower_bound, upper_bound);
+
+    size_t filter_step = data_shape.at(0);
+
+    const std::vector<size_t> filter_lower_bound{i * filter_step, 0, 0, 0};
+    const std::vector<size_t> filter_upper_bound{
+        (i + 1) * filter_step, filter_shape.at(1), filter_shape.at(2),
+        filter_shape.at(3)};
+    auto filter_slice = std::make_shared<ngraph::op::Slice>(
+        filters, filter_lower_bound, filter_upper_bound);
+
+    const std::vector<size_t> olower_bound{0, i * filter_step, 0, 0};
+    const std::vector<size_t> oupper_bound{out_shape.at(0),
+                                           (i + 1) * filter_step,
+                                           out_shape.at(2), out_shape.at(3)};
+    auto out_slice = std::make_shared<ngraph::op::Slice>(doutput, olower_bound,
+                                                         oupper_bound);
+
+    auto ng_conv = std::make_shared<ngraph::op::ConvolutionBackpropFilters>(
+        data_slice, filter_slice->get_shape(), out_slice, strides, dilations,
+        paddings, paddings, ngraph::Strides{1, 1});
+
+    ng_slices.push_back(ng_conv);
+  }
+
+  size_t concat_axis = 0;
+  return std::make_shared<ngraph::op::Concat>(ng_slices, concat_axis);
+}
+
+std::shared_ptr<ngraph::Node> GroupedGradConvolutionData(
+    const std::shared_ptr<ngraph::Node>& data_batch,
+    const std::shared_ptr<ngraph::Node>& filters,
+    const std::shared_ptr<ngraph::Node>& doutput, const ngraph::Strides strides,
+    const ngraph::Strides dilations, const ngraph::CoordinateDiff& paddings,
+    size_t groups) {
+  auto& data_shape = data_batch->get_shape();
+  auto& filter_shape = filters->get_shape();
+  auto& out_shape = doutput->get_shape();
+  ngraph::NodeVector ng_slices;
+
+  for (size_t i = 0; i < groups; ++i) {
+    size_t channel_step = filter_shape.at(1);
+    const std::vector<size_t> lower_bound{0, i * channel_step, 0, 0};
+    const std::vector<size_t> upper_bound{data_shape.at(0),
+                                          (i + 1) * channel_step,
+                                          data_shape.at(2), data_shape.at(3)};
+    auto data_slice = std::make_shared<ngraph::op::Slice>(
+        data_batch, lower_bound, upper_bound);
+
+    size_t filter_step = data_shape.at(0);
+
+    const std::vector<size_t> filter_lower_bound{i * filter_step, 0, 0, 0};
+    const std::vector<size_t> filter_upper_bound{
+        (i + 1) * filter_step, filter_shape.at(1), filter_shape.at(2),
+        filter_shape.at(3)};
+    auto filter_slice = std::make_shared<ngraph::op::Slice>(
+        filters, filter_lower_bound, filter_upper_bound);
+
+    const std::vector<size_t> olower_bound{0, i * filter_step, 0, 0};
+    const std::vector<size_t> oupper_bound{out_shape.at(0),
+                                           (i + 1) * filter_step,
+                                           out_shape.at(2), out_shape.at(3)};
+    auto out_slice = std::make_shared<ngraph::op::Slice>(doutput, olower_bound,
+                                                         oupper_bound);
+
+    auto ng_conv = std::make_shared<ngraph::op::ConvolutionBackpropData>(
+        data_slice->get_shape(), filter_slice, out_slice, strides, dilations,
+        paddings, paddings, ngraph::Strides{1, 1});
+    ng_slices.push_back(ng_conv);
+  }
+
+  size_t concat_axis = 1;
+  return std::make_shared<ngraph::op::Concat>(ng_slices, concat_axis);
+}
+
+void BuildConv2dNode(
+    const std::shared_ptr<paddle::framework::OperatorBase>& op,
+    std::shared_ptr<
+        std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
+        ngb_node_map) {
+  auto op_attrs = paddle::framework::AttrReader(op->Attrs());
+  auto filters = paddle::platform::GetInputNode(op, "Filter", ngb_node_map);
+  auto input = paddle::platform::GetInputNode(op, "Input", ngb_node_map);
+
+  std::vector<int> strides = op_attrs.Get<std::vector<int>>("strides");
+  std::vector<int> paddings = op_attrs.Get<std::vector<int>>("paddings");
+  std::vector<int> dilations = op_attrs.Get<std::vector<int>>("dilations");
+
+  const ngraph::Strides ng_strides{static_cast<size_t>(strides.at(0)),
+                                   static_cast<size_t>(strides.at(1))};
+  const ngraph::Strides ng_dilations{static_cast<size_t>(dilations.at(0)),
+                                     static_cast<size_t>(dilations.at(1))};
+  const ngraph::CoordinateDiff ng_paddings{
+      static_cast<std::ptrdiff_t>(paddings.at(0)),
+      static_cast<std::ptrdiff_t>(paddings.at(1))};
+
+  int groups = static_cast<size_t>(op_attrs.Get<int>("groups"));
+  PADDLE_ENFORCE_GE(groups, 1, "conv groups needs be no less than 1");
+
+  std::shared_ptr<ngraph::Node> result;
+  if (groups == 1) {
+    result = std::make_shared<ngraph::op::Convolution>(
+        input, filters, ng_strides, ng_dilations, ng_paddings, ng_paddings);
+  } else {
+    result = GroupedConvolution(input, filters, ng_strides, ng_dilations,
+                                ng_paddings, groups);
+  }
+  paddle::platform::SetOutputNode(op, "Output", result, ngb_node_map);
+}
+
+void BuildConv2dGradNode(
+    const std::shared_ptr<paddle::framework::OperatorBase>& op,
+    std::shared_ptr<
+        std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
+        ngb_node_map) {
+  auto op_attrs = paddle::framework::AttrReader(op->Attrs());
+  auto filter = paddle::platform::GetInputNode(op, "Filter", ngb_node_map);
+  auto input = paddle::platform::GetInputNode(op, "Input", ngb_node_map);
+  auto doutput =
+      paddle::platform::GetInputNode(op, "Output@GRAD", ngb_node_map);
+
+  int groups = op_attrs.Get<int>("groups");
+  std::vector<int> strides = op_attrs.Get<std::vector<int>>("strides");
+  std::vector<int> paddings = op_attrs.Get<std::vector<int>>("paddings");
+  std::vector<int> dilations = op_attrs.Get<std::vector<int>>("dilations");
+
+  const ngraph::Strides ng_strides{static_cast<size_t>(strides.at(0)),
+                                   static_cast<size_t>(strides.at(1))};
+  const ngraph::Strides ng_dilations{static_cast<size_t>(dilations.at(0)),
+                                     static_cast<size_t>(dilations.at(1))};
+  const ngraph::CoordinateDiff ng_paddings{
+      static_cast<std::ptrdiff_t>(paddings.at(0)),
+      static_cast<std::ptrdiff_t>(paddings.at(1))};
+
+  std::shared_ptr<ngraph::Node> dfilter;
+  std::shared_ptr<ngraph::Node> dinput;
+  if (groups == 1) {
+    dfilter = std::make_shared<ngraph::op::ConvolutionBackpropFilters>(
+        input, filter->get_shape(), doutput, ng_strides, ng_dilations,
+        ng_paddings, ng_paddings, ngraph::Strides{1, 1});
+
+    dinput = std::make_shared<ngraph::op::ConvolutionBackpropData>(
+        input->get_shape(), filter, doutput, ng_strides, ng_dilations,
+        ng_paddings, ng_paddings, ngraph::Strides{1, 1});
+
+  } else {
+    dfilter = GroupedGradConvolutionFilter(input, filter, doutput, ng_strides,
+                                           ng_dilations, ng_paddings, groups);
+    dinput = GroupedGradConvolutionData(input, filter, doutput, ng_strides,
+                                        ng_dilations, ng_paddings, groups);
+  }
+
+  paddle::platform::SetOutputNode(op, "Filter@GRAD", dfilter, ngb_node_map);
+  paddle::platform::SetOutputNode(op, "Input@GRAD", dinput, ngb_node_map);
+}
+}  // namespace ngraphs
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/ngraph/ops/top_k_op.h

@@ -36,11 +36,6 @@ void BuildTopKNode(
       std::make_shared<ngraph::op::GetOutputElement>(top_k, 0);
   std::shared_ptr<ngraph::Node> out =
       std::make_shared<ngraph::op::GetOutputElement>(top_k, 1);
-  auto dummy_out = paddle::platform::GetOutputNode(op, "Out", ngb_node_map);
-  if (dummy_out && dummy_out->get_element_type() != out->get_element_type()) {
-    out = std::make_shared<ngraph::op::Convert>(out,
-                                                dummy_out->get_element_type());
-  }
   paddle::platform::SetOutputNode(op, "Indices", indices, ngb_node_map);
   paddle::platform::SetOutputNode(op, "Out", out, ngb_node_map);
 }

paddle/fluid/operators/norm_op.h

@@ -99,10 +99,10 @@ class NormGradKernel : public framework::OpKernel<T> {
     auto dx_e = framework::EigenVector<T>::Flatten(*out_dx);
 
     Eigen::DSizes<int, 3> shape(pre, n, post);
-    Eigen::DSizes<int, 2> norm_shape(pre, post);
+    Eigen::DSizes<int, 3> rshape(pre, 1, post);
     auto x = x_e.reshape(shape);
     auto dy = dy_e.reshape(shape);
-    auto norm = norm_e.reshape(norm_shape);
+    auto norm = norm_e.reshape(rshape);
     auto dx = dx_e.reshape(shape);
 
     framework::Tensor rsum;
@@ -111,7 +111,6 @@ class NormGradKernel : public framework::OpKernel<T> {
 
     Eigen::DSizes<int, 1> rdim(1);
     Eigen::DSizes<int, 3> bcast(1, n, 1);
-    Eigen::DSizes<int, 3> rshape(pre, 1, post);
 
     // dx = ( dy/sqrt(sum(x*x)) ) * [1 - x*sum(x) / (sum(x*x) + e)]
     //    = [dy - dy * x * sum(x) / (sum(x*x) + e)] / sqrt(sum(x*x))

paddle/fluid/operators/pool_op.cc

@@ -259,7 +259,7 @@ Example:
        W_{out} = \\frac{(W_{in} - ksize[1] + 2 * paddings[1] + strides[1] - 1)}{strides[1]} + 1
        $$
 
-  For exclusive = true:
+  For exclusive = false:
        $$
        hstart = i * strides[0] - paddings[0]
        hend = hstart + ksize[0]
@@ -267,7 +267,7 @@ Example:
        wend = wstart + ksize[1]
        Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{ksize[0] * ksize[1]}
        $$
-  For exclusive = false:
+  For exclusive = true:
        $$
        hstart = max(0, i * strides[0] - paddings[0])
        hend = min(H, hstart + ksize[0])
@@ -403,7 +403,7 @@ Example:
        H_{out} = \frac{(H_{in} - ksize[1] + 2 * paddings[1] + strides[1] -1)}{strides[1]} + 1 \\
        W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2] + strides[2] -1)}{strides[2]} + 1
   $$
-  For exclusive = true:
+  For exclusive = false:
   $$
   dstart = i * strides[0] - paddings[0]
   dend = dstart + ksize[0]
@@ -413,7 +413,7 @@ Example:
   wend = wstart + ksize[2]
   Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{ksize[0] * ksize[1] * ksize[2]}
   $$
-  For exclusive = false:
+  For exclusive = true:
   $$
   dstart = max(0, i * strides[0] - paddings[0])
   dend = min(D, dstart + ksize[0])

paddle/fluid/operators/reader/buffered_reader.cc

@@ -14,6 +14,7 @@
 
 #include "paddle/fluid/operators/reader/buffered_reader.h"
 #include <vector>
+#include "paddle/fluid/framework/data_type.h"
 
 namespace paddle {
 namespace operators {
@@ -24,6 +25,13 @@ BufferedReader::~BufferedReader() {
     position_.front().wait();
     position_.pop();
   }
+#ifdef PADDLE_WITH_CUDA
+  if (platform::is_gpu_place(place_)) {
+    platform::SetDeviceId(boost::get<platform::CUDAPlace>(place_).device);
+    PADDLE_ENFORCE(cudaStreamDestroy(stream));
+    for (auto &event : events) PADDLE_ENFORCE(cudaEventDestroy(event));
+  }
+#endif
 }
 
 BufferedReader::BufferedReader(
@@ -33,6 +41,19 @@ BufferedReader::BufferedReader(
       thread_pool_(1),
       place_(place),
       buffer_size_(buffer_size) {
+#ifdef PADDLE_WITH_CUDA
+  if (platform::is_gpu_place(place_)) {
+    platform::SetDeviceId(boost::get<platform::CUDAPlace>(place_).device);
+    compute_stream =
+        ((platform::CUDADeviceContext *)(platform::DeviceContextPool::Instance()
+                                             .Get(place_)))
+            ->stream();
+    events.resize(buffer_size);
+    for (auto &event : events)
+      PADDLE_ENFORCE(cudaEventCreateWithFlags(&event, cudaEventDisableTiming));
+    PADDLE_ENFORCE(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
+  }
+#endif
   cpu_buffer_.resize(buffer_size);
   gpu_buffer_.resize(buffer_size);
   ReadTillBufferFullAsync();
@@ -46,6 +67,12 @@ void BufferedReader::ReadTillBufferFullAsync() {
 }
 
 void BufferedReader::ReadAsync(size_t i) {
+#ifdef PADDLE_WITH_CUDA
+  if (platform::is_gpu_place(place_)) {
+    platform::SetDeviceId(boost::get<platform::CUDAPlace>(place_).device);
+    PADDLE_ENFORCE(cudaEventRecord(events[i], compute_stream));
+  }
+#endif
   position_.emplace(thread_pool_.enqueue([this, i]() -> size_t {
     TensorVec &cpu = cpu_buffer_[i];
     reader_->ReadNext(&cpu);
@@ -54,14 +81,41 @@ void BufferedReader::ReadAsync(size_t i) {
       return -1UL;
     }
 
+#ifdef PADDLE_WITH_CUDA
+    // NOTE(liangdun): using async copy instead of TensorCopySync
+    // TensorCopySync would block other stream
     if (platform::is_gpu_place(place_)) {
+      platform::SetDeviceId(boost::get<platform::CUDAPlace>(place_).device);
+      PADDLE_ENFORCE(cudaStreamWaitEvent(stream, events[i], 0));
       TensorVec &gpu = gpu_buffer_[i];
       gpu.resize(cpu.size());
       for (size_t i = 0; i < cpu.size(); ++i) {
-        framework::TensorCopySync(cpu[i], place_, &gpu[i]);
+        gpu[i].Resize(cpu[i].dims());
+        gpu[i].set_layout(cpu[i].layout());
+        auto cpu_place = cpu[i].place();
+        auto cpu_ptr = cpu[i].data<void>();
+        auto gpu_ptr = gpu[i].mutable_data(place_, cpu[i].type());
+        auto size =
+            cpu[i].numel() * paddle::framework::SizeOfType(cpu[i].type());
+        if (platform::is_cuda_pinned_place(cpu_place))
+          memory::Copy(boost::get<platform::CUDAPlace>(place_), gpu_ptr,
+                       boost::get<platform::CUDAPinnedPlace>(cpu_place),
+                       cpu_ptr, size, stream);
+        else if ((platform::is_gpu_place(cpu_place)))
+          memory::Copy(boost::get<platform::CUDAPlace>(place_), gpu_ptr,
+                       boost::get<platform::CUDAPlace>(cpu_place), cpu_ptr,
+                       size, stream);
+        else
+          // if cpu place is not pinned, async copy is slower than sync copy,
+          // so we use sync copy instead.
+          memory::Copy(boost::get<platform::CUDAPlace>(place_), gpu_ptr,
+                       boost::get<platform::CPUPlace>(cpu_place), cpu_ptr, size,
+                       0);
         gpu[i].set_lod(cpu[i].lod());
       }
+      PADDLE_ENFORCE(cudaStreamSynchronize(stream));
     }
+#endif
     return i;
   }));
 }

paddle/fluid/operators/reader/buffered_reader.h

@@ -19,6 +19,9 @@
 #include <vector>
 #include "ThreadPool.h"
 #include "paddle/fluid/framework/reader.h"
+#ifdef PADDLE_WITH_CUDA
+#include "paddle/fluid/platform/gpu_info.h"
+#endif
 
 namespace paddle {
 namespace operators {
@@ -59,6 +62,11 @@ class BufferedReader : public framework::DecoratedReader {
   std::vector<TensorVec> cpu_buffer_;
   std::vector<TensorVec> gpu_buffer_;
   size_t prev_pos_{-1UL};
+#ifdef PADDLE_WITH_CUDA
+  cudaStream_t stream;
+  cudaStream_t compute_stream;
+  std::vector<cudaEvent_t> events;
+#endif
 };
 
 }  // namespace reader

paddle/fluid/operators/reader/ctr_reader.cc

@@ -213,7 +213,7 @@ void ReadSvmData(const DataDesc& data_desc, std::shared_ptr<Reader> reader,
       framework::LoD lod{lod_data};
       lod_tensor.set_lod(lod);
       int64_t* tensor_data = lod_tensor.mutable_data<int64_t>(
-          framework::make_ddim({1, static_cast<int64_t>(batch_feasign.size())}),
+          framework::make_ddim({static_cast<int64_t>(batch_feasign.size()), 1}),
           platform::CPUPlace());
       memcpy(tensor_data, batch_feasign.data(),
              batch_feasign.size() * sizeof(int64_t));
@@ -223,7 +223,7 @@ void ReadSvmData(const DataDesc& data_desc, std::shared_ptr<Reader> reader,
     // insert label tensor
     framework::LoDTensor label_tensor;
     auto* label_tensor_data = label_tensor.mutable_data<int64_t>(
-        framework::make_ddim({1, static_cast<int64_t>(batch_label.size())}),
+        framework::make_ddim({static_cast<int64_t>(batch_label.size()), 1}),
         platform::CPUPlace());
     memcpy(label_tensor_data, batch_label.data(),
            batch_label.size() * sizeof(int64_t));

paddle/fluid/operators/reader/ctr_reader_test.cc

@@ -123,7 +123,7 @@ TEST(CTR_READER, read_data) {
   std::vector<std::tuple<LoD, std::vector<int64_t>>> data_slot_6003{b1, b2, b3,
                                                                     b4};
 
-  std::vector<DDim> label_dims = {{1, 3}, {1, 3}, {1, 3}, {1, 1}};
+  std::vector<DDim> label_dims = {{3, 1}, {3, 1}, {3, 1}, {1, 1}};
 
   LoDTensorBlockingQueueHolder queue_holder;
   int capacity = 64;

paddle/fluid/operators/reduce_ops/CMakeLists.txt

 include(operators)
-register_operators()
+if(WITH_GPU)
+    register_operators(DEPS cub)
+else()
+    register_operators()
+endif()
 
 if(WITH_GPU)
     file(GLOB OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.part.cu")

paddle/fluid/operators/reshape_op.cc

@@ -327,14 +327,45 @@ class Reshape2GradOp : public framework::OperatorWithKernel {
   }
 };
 
+class ReshapeOpInplaceInToOut : public framework::InplaceInToOut {
+ public:
+  using InplaceInToOut::InplaceInToOut;
+
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const framework::OpDesc &op_desc,
+      framework::BlockDesc *block) const override {
+    std::unordered_map<std::string, std::string> inplace_in_to_out = {
+        {"X", "Out"},
+    };
+    return inplace_in_to_out;
+  }
+};
+
+class ReshapeGradInplaceInToOut : public framework::InplaceInToOut {
+  using InplaceInToOut::InplaceInToOut;
+
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const framework::OpDesc &op_desc,
+      framework::BlockDesc *block) const override {
+    std::unordered_map<std::string, std::string> inplace_in_to_out = {
+        {framework::GradVarName("Out"), framework::GradVarName("X")},
+    };
+    return inplace_in_to_out;
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 
 REGISTER_OPERATOR(reshape, ops::ReshapeOp, ops::ReshapeOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
-REGISTER_OPERATOR(reshape_grad, ops::ReshapeGradOp);
+                  paddle::framework::DefaultGradOpDescMaker<true>,
+                  ops::ReshapeOpInplaceInToOut);
+REGISTER_OPERATOR(reshape_grad, ops::ReshapeGradOp,
+                  ops::ReshapeGradInplaceInToOut);
 REGISTER_OP_CPU_KERNEL_FUNCTOR(reshape, float, ops::ReshapeKernel, double,
                                ops::ReshapeKernel, int, ops::ReshapeKernel,
                                int64_t, ops::ReshapeKernel);
@@ -344,8 +375,9 @@ REGISTER_OP_CPU_KERNEL_FUNCTOR(reshape_grad, float, ops::ReshapeGradKernel,
                                ops::ReshapeGradKernel);
 
 REGISTER_OPERATOR(reshape2, ops::Reshape2Op, ops::Reshape2OpMaker,
-                  ops::Reshape2GradMaker);
-REGISTER_OPERATOR(reshape2_grad, ops::Reshape2GradOp);
+                  ops::Reshape2GradMaker, ops::ReshapeOpInplaceInToOut);
+REGISTER_OPERATOR(reshape2_grad, ops::Reshape2GradOp,
+                  ops::ReshapeGradInplaceInToOut);
 REGISTER_OP_CPU_KERNEL_FUNCTOR(reshape2, float, ops::ReshapeKernel, double,
                                ops::ReshapeKernel, int, ops::ReshapeKernel,
                                int64_t, ops::ReshapeKernel);

paddle/fluid/operators/scale_op.cc

@@ -100,13 +100,14 @@ class ScaleGradMaker : public framework::SingleGradOpDescMaker {
   }
 };
 
+using ScaleOpInplace = framework::SingleOpInplaceInToOut;
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker, ops::ScaleGradMaker,
-                  ops::ScaleOpVarTypeInference);
+                  ops::ScaleOpVarTypeInference, ops::ScaleOpInplace);
 REGISTER_OP_CPU_KERNEL(
     scale, ops::ScaleKernel<paddle::platform::CPUDeviceContext, float>,
     ops::ScaleKernel<paddle::platform::CPUDeviceContext, double>,

paddle/fluid/operators/softmax_op.cc

@@ -198,6 +198,21 @@ class SoftmaxOpGradMaker : public framework::SingleGradOpDescMaker {
     return std::unique_ptr<framework::OpDesc>(op);
   }
 };
+
+class SoftmaxInplaceInToOut : public framework::InplaceInToOut {
+ public:
+  using framework::InplaceInToOut::InplaceInToOut;
+
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const framework::OpDesc& op_desc,
+      framework::BlockDesc* block) const override {
+    return std::unordered_map<std::string, std::string>{
+        {"X", "Out"},
+    };
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 

paddle/fluid/platform/CMakeLists.txt

-proto_library(profiler_proto SRCS profiler.proto DEPS framework_proto)
+proto_library(profiler_proto SRCS profiler.proto DEPS framework_proto simple_threadpool)
 py_proto_compile(profiler_py_proto SRCS profiler.proto)
 
 add_custom_target(profiler_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
@@ -36,7 +36,7 @@ cc_test(cpu_info_test SRCS cpu_info_test.cc DEPS cpu_info)
 
 nv_library(gpu_info SRCS gpu_info.cc DEPS gflags glog enforce)
 
-cc_library(place SRCS place.cc DEPS enforce boost)
+cc_library(place SRCS place.cc DEPS enforce boost lib_any)
 cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
 
 add_subdirectory(dynload)

paddle/fluid/platform/cuda_device_function.h

@@ -53,10 +53,12 @@ inline static int RoundToPowerOfTwo(int dim) {
     __VA_ARGS__;                           \
   } break
 
-#define CUDA_LAUNCH_KERNEL_HELPER(...)         \
-  CUDA_LAUNCH_KERNEL_BASE(256, ##__VA_ARGS__); \
-  CUDA_LAUNCH_KERNEL_BASE(128, ##__VA_ARGS__); \
-  CUDA_LAUNCH_KERNEL_BASE(64, ##__VA_ARGS__);  \
+#define CUDA_LAUNCH_KERNEL_HELPER(...)          \
+  CUDA_LAUNCH_KERNEL_BASE(1024, ##__VA_ARGS__); \
+  CUDA_LAUNCH_KERNEL_BASE(512, ##__VA_ARGS__);  \
+  CUDA_LAUNCH_KERNEL_BASE(256, ##__VA_ARGS__);  \
+  CUDA_LAUNCH_KERNEL_BASE(128, ##__VA_ARGS__);  \
+  CUDA_LAUNCH_KERNEL_BASE(64, ##__VA_ARGS__);   \
   CUDA_LAUNCH_KERNEL_BASE(32, ##__VA_ARGS__);
 
 template <typename T>

paddle/fluid/platform/ngraph_helper.h

@@ -43,13 +43,14 @@ std::shared_ptr<ngraph::Node> NgReshaper(std::shared_ptr<ngraph::Node> input,
 
 std::shared_ptr<ngraph::Node> GetNode(
     const std::shared_ptr<paddle::framework::OperatorBase>& op,
-    const std::string prm, const paddle::framework::VariableNameMap& var_map,
+    const std::string name, const paddle::framework::VariableNameMap& var_map,
     std::shared_ptr<
         std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
         ngb_node_map) {
-  auto& var_names = var_map.at(prm);
+  auto& var_names = var_map.at(name);
   PADDLE_ENFORCE_EQ(var_names.size(), 1,
-                    "op %s prm %s expects one associated var", op->Type(), prm);
+                    "op %s name %s expects one associated var", op->Type(),
+                    name);
   if (ngb_node_map->find(var_names[0]) != ngb_node_map->end()) {
     return (*ngb_node_map)[var_names[0]];
   } else {
@@ -59,43 +60,53 @@ std::shared_ptr<ngraph::Node> GetNode(
 
 std::shared_ptr<ngraph::Node> GetInputNode(
     const std::shared_ptr<paddle::framework::OperatorBase>& op,
-    const std::string prm,
+    const std::string name,
     std::shared_ptr<
         std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
         ngb_node_map) {
-  return GetNode(op, prm, op->Inputs(), ngb_node_map);
+  return GetNode(op, name, op->Inputs(), ngb_node_map);
 }
 
 std::shared_ptr<ngraph::Node> GetOutputNode(
     const std::shared_ptr<paddle::framework::OperatorBase>& op,
-    const std::string prm,
+    const std::string name,
     std::shared_ptr<
         std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
         ngb_node_map) {
-  return GetNode(op, prm, op->Outputs(), ngb_node_map);
+  return GetNode(op, name, op->Outputs(), ngb_node_map);
 }
 
 void SetOutputNode(
     const std::shared_ptr<paddle::framework::OperatorBase>& op,
-    const std::string prm, std::shared_ptr<ngraph::Node> node,
+    const std::string name, std::shared_ptr<ngraph::Node> node,
     std::shared_ptr<
         std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
         ngb_node_map) {
-  auto& var_names = op->Outputs().at(prm);
+  auto& var_names = op->Outputs().at(name);
   if (var_names.size() == 1) {
+    /*  */
+    auto dummy_out = GetOutputNode(op, name, ngb_node_map);
+    if (dummy_out && dummy_out->get_shape() != node->get_shape()) {
+      node = NgReshaper(node, dummy_out->get_shape());
+    }
+    if (dummy_out &&
+        dummy_out->get_element_type() != node->get_element_type()) {
+      node = std::make_shared<ngraph::op::Convert>(
+          node, dummy_out->get_element_type());
+    }
     (*ngb_node_map)[var_names[0]] = node;
   } else if (var_names.size() == 0) {
     (*ngb_node_map)[""] = node;
   } else {
-    PADDLE_THROW("prm %s has more than 1 var_names.", prm);
+    PADDLE_THROW("name %s has more than 1 var_names.", name);
   }
 }
 
 bool HasOutput(const std::shared_ptr<paddle::framework::OperatorBase>& op,
-               const std::string prm) {
+               const std::string name) {
   auto& outputs = op->Outputs();
-  if (outputs.find(prm) == outputs.end()) return false;
-  return outputs.at(prm).size() > 0;
+  if (outputs.find(name) == outputs.end()) return false;
+  return outputs.at(name).size() > 0;
 }
 
 inline void GetMidDims(const ngraph::Shape& x_shape,

paddle/fluid/pybind/CMakeLists.txt

@@ -26,5 +26,5 @@ if(WITH_PYTHON)
   get_property (os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
   target_link_libraries(paddle_pybind ${os_dependency_modules})
 
-  cc_test(tensor_py_test SRCS tensor_py_test.cc DEPS python)
+  cc_test(tensor_py_test SRCS tensor_py_test.cc DEPS python pybind)
 endif(WITH_PYTHON)

paddle/fluid/pybind/pybind.cc

@@ -37,6 +37,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/version.h"
 #include "paddle/fluid/imperative/layer.h"
 #include "paddle/fluid/memory/allocation/allocator_strategy.h"
+#include "paddle/fluid/memory/allocation/legacy_allocator.h"
 #include "paddle/fluid/operators/activation_op.h"
 #include "paddle/fluid/operators/py_func_op.h"
 #include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h"
@@ -127,6 +128,13 @@ PYBIND11_MODULE(core, m) {
   m.add_object("_cleanup",
                py::capsule([]() { ScopePool::Instance().Clear(); }));
 
+  m.def("get_mem_usage", [](int device) {
+    return memory::allocation::GPUMemMonitor.GetMemUsage(device);
+  });
+
+  m.def("print_mem_usage",
+        []() { return memory::allocation::GPUMemMonitor.PrintMemUsage(); });
+
   py::class_<imperative::VarBase>(m, "VarBase", R"DOC()DOC")
       // .def(py::init<>())
       .def(py::init<bool>(), py::arg("stop_gradient") = false)
@@ -1088,6 +1096,10 @@ All parameter, weight, gradient are variables in Paddle.
           "memory_early_delete",
           [](const BuildStrategy &self) { return self.memory_early_delete_; },
           [](BuildStrategy &self, bool b) { self.memory_early_delete_ = b; })
+      .def_property(
+          "enable_inplace",
+          [](const BuildStrategy &self) { return self.enable_inplace_; },
+          [](BuildStrategy &self, bool b) { self.enable_inplace_ = b; })
       .def("_finalize_strategy_and_create_passes",
            [](BuildStrategy &self) -> std::shared_ptr<ir::PassBuilder> {
              return self.CreatePassesFromStrategy(true);

paddle/scripts/fast_install.sh

+#!/bin/bash
+
+path='http://paddlepaddle.org/download?url='
+#release_version=`curl -s https://pypi.org/project/paddlepaddle/|grep -E "/project/paddlepaddle/"|grep "release"|awk -F '/' '{print $(NF-1)}'|head -1`
+release_version=1.2.0
+python_list=(
+"27"
+"35"
+"36"
+"37"
+)
+
+
+function use_cpu(){
+   while true
+    do
+     read -p "是否安装CPU版本的PaddlePaddle？(y/n)" cpu_option
+     cpu_option=`echo $cpu_option | tr 'A-Z' 'a-z'`
+     if [[ "$cpu_option" == "" || "$cpu_option" == "n" ]];then
+        echo "退出安装中..."
+        exit
+     else
+        GPU='cpu'
+        echo "将为您安装CPU版本的PaddlePaddle"
+        break
+     fi
+    done
+}
+
+function checkLinuxCUDNN(){
+   echo
+   read -n1 -p "请按回车键进行下一步..."
+   echo
+   while true
+   do
+       version_file='/usr/local/cuda/include/cudnn.h'
+       if [ -f "$version_file" ];then
+          CUDNN=`cat $version_file | grep CUDNN_MAJOR |awk 'NR==1{print $NF}'`
+       fi
+       if [ "$CUDNN" == "" ];then
+           version_file=`sudo find /usr -name "cudnn.h"|head -1`
+           if [ "$version_file" != "" ];then
+               CUDNN=`cat ${version_file} | grep CUDNN_MAJOR -A 2|awk 'NR==1{print $NF}'`
+           else
+               echo "检测结果：未在常规路径下找到cuda/include/cudnn.h文件"
+               while true
+               do
+                  read -p "请核实cudnn.h位置，并在此输入路径（请注意，路径需要输入到“cudnn.h”这一级）:" cudnn_version
+                  echo
+                  if [ "$cudnn_version" == "" ] || [ ! -f "$cudnn_version" ];then
+                        read -p "仍未找到cuDNN，输入y将安装CPU版本的PaddlePaddle，输入n可重新录入cuDNN路径，请输入（y/n）" cpu_option
+                        echo
+                        cpu_option=`echo $cpu_option | tr 'A-Z' 'a-z'`
+                        if [ "$cpu_option" == "y" -o "$cpu_option" == "" ];then
+                            GPU='cpu'
+                            break
+                        else
+                            echo "请重新输入"
+                            echo
+                        fi
+                  else
+                     CUDNN=`cat $cudnn_version | grep CUDNN_MAJOR |awk 'NR==1{print $NF}'`
+                     echo "检测结果：找到cudnn.h"
+                     break
+                  fi
+                 done
+             if [ "$GPU" == "cpu" ];then
+                break
+             fi
+           fi
+       fi
+       if [ "$CUDA" == "9" -a "$CUDNN" != "7" ];then
+           echo
+           echo "目前CUDA9下仅支持cuDNN7，暂不支持您机器上的CUDNN${CUDNN}。您可以访问NVIDIA官网下载适合版本的CUDNN，请ctrl+c退出安装进程。按回车键将为您安装CPU版本的PaddlePaddle"
+           echo
+          use_cpu()
+          if [ "$GPU"=="cpu" ];then
+             break
+          fi
+       fi
+
+       if [ "$CUDNN" == 5 ] || [ "$CUDNN" == 7 ];then
+          echo
+          echo "您的CUDNN版本是: CUDNN$CUDNN"
+          break
+       else
+          echo
+          read -n1 -p "目前支持的CUDNN版本为5和7,暂不支持您机器上的CUDNN${CUDNN}，将为您安装CPU版本的PaddlePaddle,请按回车键开始安装"
+          echo
+          use_cpu
+          if [ "$GPU"=="cpu" ];then
+             break
+          fi
+       fi
+   done
+}
+
+function checkLinuxCUDA(){
+   while true
+   do
+       CUDA=`echo ${CUDA_VERSION}|awk -F "[ .]" '{print $1}'`
+       if [ "$CUDA" == "" ];then
+         if [ -f "/usr/local/cuda/version.txt" ];then
+           CUDA=`cat /usr/local/cuda/version.txt | grep 'CUDA Version'|awk -F '[ .]' '{print $3}'`
+           tmp_cuda=$CUDA
+         fi
+         if [ -f "/usr/local/cuda8/version.txt" ];then
+           CUDA=`cat /usr/local/cuda8/version.txt | grep 'CUDA Version'|awk -F '[ .]' '{print $3}'`
+           tmp_cuda8=$CUDA
+         fi
+         if [ -f "/usr/local/cuda9/version.txt" ];then
+           CUDA=`cat /usr/local/cuda9/version.txt | grep 'CUDA Version'|awk -F '[ .]' '{print $3}'`
+           tmp_cuda9=$CUDA
+         fi
+       fi
+
+       if [ "$tmp_cuda" != "" ];then
+         echo "检测结果：找到CUDA $tmp_cuda"
+       fi
+       if [ "$tmp_cudai8" != "" ];then
+         echo "检测结果：找到CUDA $tmp_cuda8"
+       fi
+       if [ "$tmp_cuda9" != "" ];then
+         echo "检测结果：找到CUDA $tmp_cuda9"
+       fi
+
+       if [ "$CUDA" == "" ];then
+            echo "检测结果：没有在常规路径下找到cuda/version.txt文件"
+            while true
+            do
+                read -p "请输入cuda/version.txt的路径:" cuda_version
+                if [ "$cuda_version" == "" || ! -f "$cuda_version" ];then
+                    read -p "仍未找到CUDA，输入y将安装CPU版本的PaddlePaddle，输入n可重新录入CUDA路径，请输入（y/n）" cpu_option
+                    cpu_option=`echo $cpu_option | tr 'A-Z' 'a-z'`
+                    if [ "$cpu_option" == "y" || "$cpu_option" == "" ];then
+                        GPU='cpu'
+                        break
+                    else
+                        echo "重新输入..."
+                    fi
+                else
+                    CUDA=`cat $cuda_version | grep 'CUDA Version'|awk -F '[ .]' '{print $3}'`
+                    if [ "$CUDA" == "" ];then
+                        echo "未能在version.txt中找到CUDA相关信息"
+                    else
+                        break
+                    fi
+                fi
+            done
+            if [ "$GPU" == "cpu" ];then
+                break
+            fi
+       fi
+
+       if [ "$CUDA" == "8" ] || [ "$CUDA" == "9" ];then
+          echo "您的CUDA版本是${CUDA}"
+          break
+       else
+          echo "目前支持CUDA8/9，暂不支持您的CUDA${CUDA}，将为您安装CPU版本的PaddlePaddle"
+          echo
+          use_cpu
+       fi
+
+       if [ "$GPU" == "cpu" ];then
+          break
+       fi
+   done
+}
+
+function checkLinuxMathLibrary(){
+  while true
+    do
+      if [ "$AVX" ==  "" ];then
+        echo "正在检测您环境中是否存在AVX指令集..."
+        echo
+        echo "检测结果：您电脑上没有AVX指令集，目前针对无AVX指令集的环境，我们仅提供支持mkl数学库的PaddlePaddle，将为您安装此版本的PaddlePaddle"
+        math='mkl'
+        break
+      elif [ "$GPU" == "gpu" ];then
+        math='mkl'
+        echo "检测到您的机器上配备GPU，推荐您使用mkl数学库"
+        break
+      else
+        read -p "请输入您希望使用的数学库：
+            1：openblas 一个高性能多核 BLAS 库
+            2：mkl（推荐） 英特尔数学核心函数库
+            => 请输入数字1或2。如输入其他字符或直接回车，将会默认选择【 2. mkl 】 。请在这里输入并回车：" math
+          if [ "$math" == "" ];then
+            math="mkl"
+            echo "您选择了数字【2】"
+            break
+          fi
+          if [ "$math" == "1" ];then
+            math=openblas
+            echo "您选择了数字【1】"
+            break
+          elif [ "$math" == "2" ];then
+            math=mkl
+            echo "您选择了数字【2】"
+            break
+          fi
+          echo "输入错误，请再次输入"
+      fi
+    done
+}
+
+function checkLinuxPaddleVersion(){
+  read -n1 -p "请按回车键继续..."
+  while true
+    do
+      read -p "
+               1. 开发版：对应Github上develop分支，如您需要开发、或希望使用PaddlePaddle最新功能，请选用此版本
+               2. 稳定版（推荐）：如您无特殊开发需求，建议使用此版本，目前最新的版本号为 ${release_version}
+                => 请输入数字1或2。如输入其他字符或直接回车，将会默认选择【 2. 稳定版 】 。请在这里输入并回车：" paddle_version
+        if [ "$paddle_version" == "" ];then
+          paddle_version="release-${release_version}"
+          echo "您选择了数字【2】，为您安装release-${release_version}"
+          break
+        fi
+        if [ "$paddle_version" == "1" ];then
+          echo "您选择了数字【1】，将为您安装开发版"
+          break
+        elif [ "$paddle_version" == "2" ];then
+          echo "您选择了数字【2】，为您安装release-${release_version}"
+          break
+        fi
+        echo "输入错误，请再次输入"
+    done
+}
+
+function checkLinuxPip(){
+  while true
+    do
+       echo "请输入您要使用的pip目录（您可以另起终端，并使用which pip来查看）："
+       read -p "" pip_path
+       if [ "$pip_path" == "" -o ! -f "$pip_path" ];then
+         echo "检测结果：pip不存在,请重新输入"
+         continue
+       fi
+       python_version=`$pip_path --version|awk -F "[ |)]" '{print $6}'|sed 's#\.##g'`
+       if [ "$python_version" == "27" ];then
+         uncode=`python -c "import pip._internal;print(pip._internal.pep425tags.get_supported())"|grep "cp27mu"`
+         if [[ "$uncode" == "" ]];then
+            uncode=
+         else
+            uncode=u
+         fi
+       fi
+       if [ "$python_version" == "" ];then
+         echo "检测结果：pip不存在,请重新输入"
+       else
+         version_list=`echo "${python_list[@]}" | grep "$python_version" `
+         if [ "$version_list" != "" ];then
+           echo "检测结果：找到python${python_version}版本"
+           break
+         else
+           echo "检测结果：找不到可用的 pip, 我们只支持Python27/35/36/37及其对应的pip, 请重新输入， 或使用ctrl + c退出 "
+         fi
+       fi
+    done
+}
+
+function checkLinuxAVX(){
+  while true
+  do
+    if [[ "$AVX" != "" ]];then
+      AVX="avx"
+      break
+    else
+      if [ "$CUDA" == "8" -a "$CUDNN" == "7" ] || [ "$GPU" == "cpu" ];then
+        AVX="noavx"
+        break
+      else
+        echo "Step 6. 检测是否有avx"
+        echo
+        echo "检测结果：未能找到avx，我们仅提供CPU版本或配置为CUDA8 cuDNN7的GPU版本的安装包"
+        break
+      fi
+    fi
+  done
+}
+
+function PipLinuxInstall(){
+  wheel_cpu_release="http://paddle-wheel.bj.bcebos.com/${release_version}-${GPU}-${AVX}-${math}/paddlepaddle-${release_version}-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl"
+  wheel_gpu_release="http://paddle-wheel.bj.bcebos.com/${release_version}-gpu-cuda${CUDA}-cudnn${CUDNN}-${AVX}-${math}/paddlepaddle_gpu-${release_version}.post${CUDA}${CUDNN}-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl"
+  wheel_gpu_release_noavx="http://paddle-wheel.bj.bcebos.com/${release_version}-gpu-cuda${CUDA}-cudnn${CUDNN}-${AVX}-${math}/paddlepaddle_gpu-${release_version}-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl"
+  wheel_cpu_develop="http://paddle-wheel.bj.bcebos.com/latest-cpu-${AVX}-${math}/paddlepaddle-latest-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl"
+  wheel_gpu_develop="http://paddle-wheel.bj.bcebos.com/latest-gpu-cuda${CUDA}-cudnn${CUDNN}-${AVX}-${math}/paddlepaddle_gpu-latest-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl"
+
+
+  if [[ "$paddle_version" == "2" ]];then
+    if [[ "$GPU" == "gpu" ]];then
+        if [[ ${AVX} == "avx" ]];then
+          rm -rf `echo $wheel_gpu_release|awk -F '/' '{print $NF}'`
+          wget -q $wheel_gpu_release
+          if [ "$?" == "0" ];then
+            $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_release
+          else
+            echo "paddlepaddle whl包下载失败"
+            exit 1
+          fi
+        else
+          rm -rf `echo $wheel_gpu_release_novax|awk -F '/' '{print $NF}'`
+          wget -q $wheel_gpu_release_novax
+          if [ "$?" == "0" ];then
+            $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_release_noavx
+          else
+            echo "paddlepaddle whl包下载失败"
+            exit 1
+          fi
+        fi
+    else
+        rm -rf `echo $wheel_cpu_release|awk -F '/' '{print $NF}'`
+        wget -q $wheel_cpu_release
+        if [ "$?" == "0" ];then
+          $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_cpu_release
+        else
+          echo "paddlepaddle whl包下载失败"
+          exit 1
+        fi
+    fi
+  else
+    if [[ "$GPU" == "gpu" ]];then
+        rm -rf `echo $wheel_gpu_develop|awk -F '/' '{print $NF}'`
+        wget -q $wheel_gpu_develop
+        if [ "$?" == "0" ];then
+          $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_develop
+        else
+          echo "paddlepaddle whl包下载失败"
+          exit 1
+        fi
+    else
+        rm -rf `echo $wheel_cpu_develop|awk -F '/' '{print $NF}'`
+        wget -q $wheel_cpu_develop
+        if [ "$?" == "0" ];then
+          $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_cpu_develop
+        else
+          echo "paddlepaddle whl包下载失败"
+          exit 1
+        fi
+    fi
+  fi
+}
+
+
+function checkLinuxGPU(){
+  read -n1 -p "即将检测您的机器是否含GPU，请按回车键继续..."
+  echo
+  AVX=`cat /proc/cpuinfo |grep avx|tail -1|grep avx`
+  which nvidia-smi >/dev/null 2>&1
+  if [ "$?" != "0" ];then
+    GPU='cpu'
+    echo "未在机器上找到GPU，或PaddlePaddle暂不支持此型号的GPU"
+  else
+    GPU='gpu'
+    echo "已在您的机器上找到GPU，即将确认CUDA和CUDNN版本..."
+    echo
+  fi
+  if [ "$GPU" == 'gpu' ];then
+    checkLinuxCUDA
+    checkLinuxCUDNN
+  fi
+}
+
+function linux(){
+gpu_list=(
+"GeForce 410M"
+"GeForce 610M"
+"GeForce 705M"
+"GeForce 710M"
+"GeForce 800M"
+"GeForce 820M"
+"GeForce 830M"
+"GeForce 840M"
+"GeForce 910M"
+"GeForce 920M"
+"GeForce 930M"
+"GeForce 940M"
+"GeForce GT 415M"
+"GeForce GT 420M"
+"GeForce GT 430"
+"GeForce GT 435M"
+"GeForce GT 440"
+"GeForce GT 445M"
+"GeForce GT 520"
+"GeForce GT 520M"
+"GeForce GT 520MX"
+"GeForce GT 525M"
+"GeForce GT 540M"
+"GeForce GT 550M"
+"GeForce GT 555M"
+"GeForce GT 610"
+"GeForce GT 620"
+"GeForce GT 620M"
+"GeForce GT 625M"
+"GeForce GT 630"
+"GeForce GT 630M"
+"GeForce GT 635M"
+"GeForce GT 640"
+"GeForce GT 640 (GDDR5)"
+"GeForce GT 640M"
+"GeForce GT 640M LE"
+"GeForce GT 645M"
+"GeForce GT 650M"
+"GeForce GT 705"
+"GeForce GT 720"
+"GeForce GT 720M"
+"GeForce GT 730"
+"GeForce GT 730M"
+"GeForce GT 735M"
+"GeForce GT 740"
+"GeForce GT 740M"
+"GeForce GT 745M"
+"GeForce GT 750M"
+"GeForce GTS 450"
+"GeForce GTX 1050"
+"GeForce GTX 1060"
+"GeForce GTX 1070"
+"GeForce GTX 1080"
+"GeForce GTX 1080 Ti"
+"GeForce GTX 460"
+"GeForce GTX 460M"
+"GeForce GTX 465"
+"GeForce GTX 470"
+"GeForce GTX 470M"
+"GeForce GTX 480"
+"GeForce GTX 480M"
+"GeForce GTX 485M"
+"GeForce GTX 550 Ti"
+"GeForce GTX 560M"
+"GeForce GTX 560 Ti"
+"GeForce GTX 570"
+"GeForce GTX 570M"
+"GeForce GTX 580"
+"GeForce GTX 580M"
+"GeForce GTX 590"
+"GeForce GTX 650"
+"GeForce GTX 650 Ti"
+"GeForce GTX 650 Ti BOOST"
+"GeForce GTX 660"
+"GeForce GTX 660M"
+"GeForce GTX 660 Ti"
+"GeForce GTX 670"
+"GeForce GTX 670M"
+"GeForce GTX 670MX"
+"GeForce GTX 675M"
+"GeForce GTX 675MX"
+"GeForce GTX 680"
+"GeForce GTX 680M"
+"GeForce GTX 680MX"
+"GeForce GTX 690"
+"GeForce GTX 750"
+"GeForce GTX 750 Ti"
+"GeForce GTX 760"
+"GeForce GTX 760M"
+"GeForce GTX 765M"
+"GeForce GTX 770"
+"GeForce GTX 770M"
+"GeForce GTX 780"
+"GeForce GTX 780M"
+"GeForce GTX 780 Ti"
+"GeForce GTX 850M"
+"GeForce GTX 860M"
+"GeForce GTX 870M"
+"GeForce GTX 880M"
+"GeForce GTX 950"
+"GeForce GTX 950M"
+"GeForce GTX 960"
+"GeForce GTX 960M"
+"GeForce GTX 965M"
+"GeForce GTX 970"
+"GeForce GTX 970M"
+"GeForce GTX 980"
+"GeForce GTX 980M"
+"GeForce GTX 980 Ti"
+"GeForce GTX TITAN"
+"GeForce GTX TITAN Black"
+"GeForce GTX TITAN X"
+"GeForce GTX TITAN Z"
+"Jetson TK1"
+"Jetson TX1"
+"Jetson TX2"
+"Mobile Products"
+"NVIDIA NVS 310"
+"NVIDIA NVS 315"
+"NVIDIA NVS 510"
+"NVIDIA NVS 810"
+"NVIDIA TITAN V"
+"NVIDIA TITAN X"
+"NVIDIA TITAN Xp"
+"NVS 4200M"
+"NVS 5200M"
+"NVS 5400M"
+"Quadro 410"
+"Quadro GP100"
+"Quadro K1100M"
+"Quadro K1200"
+"Quadro K2000"
+"Quadro K2000D"
+"Quadro K2100M"
+"Quadro K2200"
+"Quadro K2200M"
+"Quadro K3100M"
+"Quadro K4000"
+"Quadro K4100M"
+"Quadro K420"
+"Quadro K4200"
+"Quadro K4200M"
+"Quadro K5000"
+"Quadro K500M"
+"Quadro K5100M"
+"Quadro K510M"
+"Quadro K5200"
+"Quadro K5200M"
+"Quadro K600"
+"Quadro K6000"
+"Quadro K6000M"
+"Quadro K610M"
+"Quadro K620"
+"Quadro K620M"
+"Quadro M1000M"
+"Quadro M1200"
+"Quadro M2000"
+"Quadro M2000M"
+"Quadro M2200"
+"Quadro M3000M"
+"Quadro M4000"
+"Quadro M4000M"
+"Quadro M5000"
+"Quadro M5000M"
+"Quadro M500M"
+"Quadro M520"
+"Quadro M5500M"
+"Quadro M6000"
+"Quadro M6000 24GB"
+"Quadro M600M"
+"Quadro M620"
+"Quadro Mobile Products"
+"Quadro P1000"
+"Quadro P2000"
+"Quadro P3000"
+"Quadro P400"
+"Quadro P4000"
+"Quadro P5000"
+"Quadro P600"
+"Quadro P6000"
+"Quadro Plex 7000"
+"Tegra K1"
+"Tegra X1"
+"Tesla C2050/C2070"
+"Tesla C2075"
+"Tesla Data Center Products"
+"Tesla K10"
+"Tesla K20"
+"Tesla K40"
+"Tesla K80"
+"Tesla M40"
+"Tesla M60"
+"Tesla P100"
+"Tesla P4"
+"Tesla P40"
+"Tesla V100")
+
+  echo "Step 2. 检测GPU型号和CUDA/cuDNN版本"
+  echo
+  checkLinuxGPU
+  echo
+  echo "Step 3. 检测数学库"
+  echo
+  checkLinuxMathLibrary
+  echo
+  echo "Step 4. 选择要安装的PaddlePaddle版本"
+  echo
+  checkLinuxPaddleVersion
+  echo
+  echo "Step 5. 检测pip版本"
+  echo
+  checkLinuxPip
+  echo
+  checkLinuxAVX
+  echo "*********************2. 开始安装*****************************"
+  PipLinuxInstall
+}
+
+function checkMacPython2(){
+    while true
+       do
+          read -p "
+                => 未能在常规路径下找到Python2，请使用ctrl+c命令退出安装程序，并使用brew或pypi.org下载安装Python2（注意Python版本不能低于2.7.15）
+                如希望自定义Python路径，请输入路径：" python_root
+          echo
+          python_version=`$python_root --version 2>&1 1>&1`
+          if [ $? == "0" ];then
+            :
+          else
+            python_version=""
+          fi
+          check_python=`echo $python_version | grep "Python 2"`
+          if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ]  ;then
+               python_version=""
+          elif [ -n "$check_python" ];then
+              while true
+                do
+                  read -p "
+                => 在您的环境中找到 $python_version, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车: " use_python
+                  echo
+                  use_python=`echo $use_python | tr 'A-Z' 'a-z'`
+                  if [ "$use_python" == "y" ]||[ "$use_python" == "" ];then
+                       use_python="y"
+                       break
+                  elif [ "$use_python" == "n" ];then
+                       python_root=""
+                       break
+                  else
+                      echo "输入错误，请重新输入(y/n)"
+                  fi
+                done
+              if [ "$use_python" == "y" ];then
+                break
+              fi
+            else
+              echo "您输入Python的不是Python2"
+              python_version=""
+            fi
+       done
+}
+
+function checkMacPython3(){
+    while true
+       do
+          read -p "
+                => 未能在常规路径下找到Python3，请使用ctrl+c命令退出安装程序，并使用brew或pypi.org下载Python3
+                如希望自定义Python路径，请输入路径：" python_root
+          python_version=`$python_root --version  2>&1 1>&1`
+          if [ $? == "0" ];then
+              :
+          else
+              python_version=""
+          fi
+          check_python=`echo $python_version | grep "Python 3"`
+          if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ] ;then
+               python_version=""
+          elif [ -n "$check_python" ] ;then
+              while true
+                do
+                  read -p "
+                => 在您的环境中找到 $python_version, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车: " use_python
+                  echo
+                  use_python=`echo $use_python | tr 'A-Z' 'a-z'`
+                  if [ "$use_python" == "y" ]||[ "$use_python" == "" ];then
+                       use_python="y"
+                       break
+                  elif [ "$use_python" == "n" ];then
+                        python_root=""
+                        break
+                  else
+                      echo "输入错误，请重新输入(y/n)"
+                  fi
+                done
+              if [ "$use_python" == "y" ];then
+                    break
+              fi
+            else
+              echo "您输入Python的不是Python3"
+              python_version=""
+            fi
+       done
+}
+
+function checkMacPaddleVersion(){
+  while true
+    do
+      read -n1 -p "Step 2. 选择PaddlePaddle的版本，请按回车键继续..."
+      echo
+      read -p "
+               1. 开发版：对应Github上develop分支，如您需要开发、或希望使用PaddlePaddle最新功能，请选用此版本
+               2. 稳定版（推荐）：如您无特殊开发需求，建议使用此版本，目前最新的版本号为 ${release_version}
+
+               => 请输入数字1或2。如输入其他字符或直接回车，将会默认选择【 2. 稳定版 】 。请在这里输入并回车：" paddle_version
+      if [ "$paddle_version" == "1" ]||[ "$paddle_version" == "2" ];then
+          echo
+          echo "您选择了数字【"$paddle_version" 】"
+          echo
+          break
+      else
+          paddle_version="2"
+          echo
+          echo "您选择了数字【2】"
+          echo
+          break
+      fi
+    done
+}
+
+function checkMacPythonVersion(){
+  while true
+    do
+       read -n1 -p "Step 3. 选择Python版本，请按回车键继续..."
+       read -p "
+               2. 使用python 2.x
+               3. 使用python 3.x
+
+                => 请输入数字2或3。如输入其他字符或直接回车，将会默认使用【Python 2 】。请在这里输入并回车：" python_V
+                echo
+       if [ "$python_V" == "" ];then
+            python_V="2"
+       fi
+       echo "您选择了数字【"$python_V"】，正在寻找符合您要求的Python版本，请按回车键继续..."
+       echo
+       if [ "$python_V" == "2" ];then
+           python_root=`which python2.7`
+           if [ "$python_root" == "" ];then
+                python_root=`which python`
+           fi
+           python_version=`$python_root --version 2>&1 1>&1`
+           if [ $? == "0" ];then
+               :
+           else
+               python_version=""
+           fi
+           if [ "$python_root" == "" ]||[ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ]||[ "$python_root" == "/usr/bin/python2.7" -a "$python_version" == "Python 2.7.10" ];then
+               checkMacPython2
+           fi
+           while true
+             do
+               read -p "
+                => 在您的环境中找到 $python_version, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车：" use_python
+               echo
+               use_python=`echo $use_python | tr 'A-Z' 'a-z'`
+               if [ "$use_python" == "y" ]||[ "$use_python" == "" ];then
+                    break
+               elif [ "$use_python" == "n" ];then
+                    python_root=""
+                    checkMacPython2
+                    break
+               else
+                    echo "输入错误，请重新输入(y/n)"
+               fi
+            done
+
+       elif [ "$python_V" == "3" ];then
+           python_root=`which python3`
+           python_version=`$python_root --version 2>&1 1>&1`
+           if [ $? == "0" ];then
+               :
+           else
+               python_version=""
+           fi
+           if [ "$python_root" == "" ]||[ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ];then
+               checkMacPython3
+           fi
+           while true
+             do
+               read -p "
+                => 在您的环境中找到 $python_version, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车：" use_python
+               echo
+               use_python=`echo $use_python | tr 'A-Z' 'a-z'`
+               if [ "$use_python" == "y" ]||[ "$use_python" == "" ];then
+                   break
+               elif [ "$use_python" == "n" ];then
+                    checkMacPython3
+                    break
+               else
+                    echo "输入错误，请重新输入(y/n)"
+               fi
+           done
+       else
+           :
+       fi
+
+
+       if [ "$python_V" == "2" ]||[ "$python_V" == "3" ];then
+           python_brief_version=`$python_root -m pip -V |awk -F "[ |)]" '{print $6}'|sed 's#\.##g'`
+           if [[ $python_brief_version == "27" ]];then
+              uncode=`python -c "import pip._internal;print(pip._internal.pep425tags.get_supported())"|grep "cp27"`
+              if [[ $uncode == "" ]];then
+                 uncode="mu"
+              else
+                 uncode="m"
+              fi
+           fi
+           version_list=`echo "${python_list[@]}" | grep "$python_brief_version" `
+           if [ "$version_list" != "" ];then
+              break
+            else
+              echo "未找到可用的pip或pip3。PaddlePaddle目前支持：Python2.7/3.5/3.6/3.7及其对应的pip, 请重新输入，或使用ctrl + c退出"
+           fi
+        else
+            echo "输入错误，请重新输入"
+        fi
+  done
+}
+
+function checkMacAVX(){
+    read -n1 -p "Step 4. 检测您的Mac是否支持AVX指令集，请按回车键继续..."
+    echo
+    if [[ $AVX != "" ]];then
+        AVX="avx"
+        echo "检测结果：支持"
+    else
+        read -n1 -p "检测结果：不支持。非常抱歉，PaddlePaddle在Mac系统暂不提供no_avx类型的安装包，您可以选择在Linux系统中安装no_avx版的PaddlePaddle, 请按回车键退出..."
+        exit
+    fi
+    echo
+}
+
+function checkMacGPU(){
+    read -n1 -p "Step 5. 选择CPU/GPU版本，请按回车键继续..."
+    echo
+    if [[ $GPU != "" ]];then
+        echo "MacOS环境下，暂未提供GPU版本的PaddlePaddle安装包，将为您安装CPU版本的PaddlePaddle"
+    else
+        echo "MacOS环境下，暂未提供GPU版本的PaddlePaddle安装包，将为您安装CPU版本的PaddlePaddle"
+        GPU=cpu
+    fi
+    echo
+}
+
+function macos() {
+  path='http://paddlepaddle.org/download?url='
+  AVX=`sysctl -a | grep cpu | grep AVX1.0 | tail -1 | grep AVX`
+
+  while true
+      do
+        checkMacPaddleVersion
+        checkMacPythonVersion
+        checkMacAVX
+        checkMacGPU
+
+
+        echo "*********************2. 开始安装*****************************"
+        echo
+        read -n1 -p "即将为您下载并安装PaddlePaddle，请按回车键继续..."
+        echo
+        if [[ $paddle_version == "2" ]];then
+            $python_root -m pip install paddlepaddle
+            if [ $? == "0" ];then
+               echo "安装成功，可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器"
+               break
+            else
+               rm  $whl_cpu_release
+               echo "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用"
+               echo""
+               echo "=========================================================================================="
+               echo""
+               exit 1
+            fi
+        else
+            if [ -f $whl_cpu_develop ];then
+                $python_root -m pip install $whl_cpu_develop
+                if [ $? == "0" ];then
+                   rm -rf $whl_cpu_develop
+                   echo "安装成功！小提示：可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器"
+                   break
+                else
+                   echo "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用"
+                   echo""
+                   echo "=========================================================================================="
+                   echo""
+                   exit 1
+                fi
+            else
+                wget ${path}$whl_cpu_develop -O $whl_cpu_develop
+                if [ $? == "0" ];then
+                    $python_root -m pip install $whl_cpu_develop
+                    if [ $? == "0" ];then
+                       rm  $wheel_cpu_develop
+                       echo "安装成功，可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器"
+                       break
+                    else
+                       rm  $whl_cpu_release
+                       echo "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用"
+                       echo""
+                       echo "=========================================================================================="
+                       echo""
+                       exit 1
+                    fi
+                else
+                      rm  $whl_cpu_develop
+                      echo "未能正常安装PaddlePaddle，请检查您的网络 或者确认您是否安装有 wget，或者ctrl + c退出后反馈至https://github.com/PaddlePaddle/Paddle/issues"
+                      echo""
+                      echo "=========================================================================================="
+                      echo""
+                      exit 1
+                fi
+            fi
+        fi
+  done
+}
+
+function main() {
+  echo "*********************************"
+  echo "欢迎使用PaddlePaddle快速安装脚本"
+  echo "*********************************"
+  echo
+  echo "如果您在安装过程中遇到任何问题，请在https://github.com/PaddlePaddle/Paddle/issues反馈，我们的工作人员将会帮您答疑解惑"
+  echo
+  echo "本安装包将帮助您在Linux或Mac系统下安装PaddlePaddle，包括 1）安装前的准备和 2）开始安装 两部分"
+  echo
+  read -n1 -p "请按回车键进行下一步..."
+  echo
+  echo
+  echo "*********************1. 安装前的准备*****************************"
+  echo
+  echo "Step 1. 正在检测您的操作系统信息..."
+  echo
+  SYSTEM=`uname -s`
+  if [ "$SYSTEM" == "Darwin" ];then
+  	echo "您的系统为：MAC OSX"
+    echo
+  	macos
+  else
+ 	echo "您的系统为：Linux"
+  echo
+	  OS=`cat /etc/issue|awk 'NR==1 {print $1}'`
+	  if [ $OS == "\S" ] || [ "$OS" == "CentOS" ] || [ $OS == "Ubuntu" ];then
+	    linux
+	  else
+	    echo "您的系统不在本安装包的支持范围，如您需要在windows环境下安装PaddlePaddle，请您参考PaddlePaddle官网的windows安装文档"
+	  fi
+  fi
+}
+main

python/CMakeLists.txt

@@ -54,7 +54,7 @@ ELSE(WIN32)
 		DEPENDS copy_paddle_pybind ${FLUID_CORE} framework_py_proto profiler_py_proto ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER})
 ENDIF()
 
-set(paddle_python_deps ${PADDLE_PYTHON_BUILD_DIR}/.timestamp ${MKL_DEPENDS})
+set(paddle_python_deps ${PADDLE_PYTHON_BUILD_DIR}/.timestamp ${MKL_DEPENDS} ${external_project_dependencies})
 add_custom_target(paddle_python ALL DEPENDS ${paddle_python_deps})
 
 set(PADDLE_PYTHON_PACKAGE_DIR ${CMAKE_CURRENT_BINARY_DIR}/dist/)

python/paddle/fluid/__init__.py

@@ -158,7 +158,8 @@ def __bootstrap__():
             'enable_cublas_tensor_op_math', 'conv_workspace_size_limit',
             'cudnn_exhaustive_search', 'memory_optimize_debug', 'selected_gpus',
             'sync_nccl_allreduce', 'limit_of_tmp_allocation',
-            'times_excess_than_required_tmp_allocation'
+            'times_excess_than_required_tmp_allocation',
+            'enable_inplace_whitelist'
         ]
 
     core.init_gflags([sys.argv[0]] +

python/paddle/fluid/compiler.py

@@ -174,6 +174,11 @@ class CompiledProgram(object):
                 self._exec_strategy.num_threads = cpu_num * 2
 
         trainers_endpoints = self._program._trainers_endpoints
+
+        # FIXME(dzhwinter): enable_inplace should be after memory_optimize
+        # if turn on python memory optimize, turn off the inplace_pass.
+        self._build_strategy.enable_inplace = False if self._program._is_mem_optimized else True
+
         if self._build_strategy.num_trainers > 1 and trainers_endpoints:
             assert self._build_strategy.num_trainers == len(
                 trainers_endpoints), "num_trainers == len(end_points)"

python/paddle/fluid/framework.py

@@ -1725,6 +1725,19 @@ class Program(object):
         self._trainers_endpoints = []
         # the distributed lookup table names
         self._distributed_lookup_table = None
+        # @deprecated(the python memory optimize transpiler is deprecated)
+        # whether the program is optimized by memory_optimize_transpiler
+        self.__is_mem_optimized = False
+
+    @property
+    def _is_mem_optimized(self):
+        # if the program is optimized, operator input/outputs
+        # maybe same, which conflict with save_inference_model.
+        return self.__is_mem_optimized
+
+    @_is_mem_optimized.setter
+    def _is_mem_optimized(self, target):
+        self.__is_mem_optimized = target
 
     @property
     def op_role(self):
@@ -1744,7 +1757,7 @@ class Program(object):
         return self._current_role
 
     @op_role.setter
-    def set_op_role(self, role):
+    def op_role(self, role):
         self._current_role = role
 
     @property

python/paddle/fluid/io.py

@@ -16,14 +16,16 @@ from __future__ import print_function
 
 import os
 import errno
+import warnings
 import time
 import shutil
 import six
 from functools import reduce
 
+from paddle.fluid import layers
 from paddle.fluid.executor import Executor
 from paddle.fluid.evaluator import Evaluator
-from paddle.fluid.framework import Program, Parameter, default_main_program, default_startup_program, Variable
+from paddle.fluid.framework import Program, Parameter, default_main_program, default_startup_program, Variable, program_guard
 from . import core
 
 __all__ = [
@@ -930,6 +932,24 @@ def save_inference_model(dirname,
 
     if main_program is None:
         main_program = default_main_program()
+        if main_program._is_mem_optimized:
+            warnings.warn(
+                "save_inference_model must put before you call memory_optimize. \
+                                            the memory_optimize will modify the original program, \
+                                            is not suitable for saving inference model \
+                                            we save the original program as inference model.",
+                RuntimeWarning)
+
+    # fix the bug that the activation op's output as target will be pruned.
+    # will affect the inference performance.
+    # TODO(Superjomn) add an IR pass to remove 1-scale op.
+    with program_guard(main_program):
+        uniq_target_vars = []
+        for var in target_vars:
+            if isinstance(var, Variable):
+                var1 = layers.scale(var, 1.)
+            uniq_target_vars.append(var1)
+        target_vars = uniq_target_vars
 
     # when a pserver and a trainer running on the same machine, mkdir may conflict
     try:

python/paddle/fluid/layers/detection.py

@@ -49,6 +49,7 @@ __all__ = [
     'box_coder',
     'polygon_box_transform',
     'yolov3_loss',
+    'box_clip',
     'multiclass_nms',
 ]
 
@@ -396,10 +397,10 @@ def box_coder(prior_box,
                              input is image feature map, they are close to 
                              the origin of the coordinate system. [xmax, ymax]
                              is the right bottom coordinate of the anchor box.       
-        prior_box_var(Variable|list): prior_box_var supports two types of input. 
-                              One is variable with shape [M, 4] holds M group.
-                              The other one is list consist of 4 elements 
-                              shared by all boxes. 
+        prior_box_var(Variable|list|None): prior_box_var supports two types 
+                              of input. One is variable with shape [M, 4] 
+                              holds M group. The other one is list consist of 
+                              4 elements shared by all boxes. 
         target_box(Variable): This input can be a 2-D LoDTensor with shape 
                               [N, 4] when code_type is 'encode_center_size'. 
                               This input also can be a 3-D Tensor with shape 
@@ -2055,6 +2056,54 @@ def generate_proposals(scores,
     return rpn_rois, rpn_roi_probs
 
 
+def box_clip(input, im_info, name=None):
+    """
+    Clip the box into the size given by im_info
+    For each input box, The formula is given as follows:
+        
+    .. code-block:: text
+
+        xmin = max(min(xmin, im_w - 1), 0)
+        ymin = max(min(ymin, im_h - 1), 0) 
+        xmax = max(min(xmax, im_w - 1), 0)
+        ymax = max(min(ymax, im_h - 1), 0)
+    
+    where im_w and im_h are computed from im_info:
+ 
+    .. code-block:: text
+
+        im_h = round(height / scale)
+        im_w = round(weight / scale)
+
+    Args:
+        input(variable): The input box, the last dimension is 4.
+        im_info(variable): The information of image with shape [N, 3] with 
+                            layout (height, width, scale). height and width
+                            is the input size and scale is the ratio of input
+                            size and original size.
+        name (str): The name of this layer. It is optional.
+    
+    Returns:
+        Variable: The cliped tensor variable.
+        
+    Examples:
+        .. code-block:: python
+        
+            boxes = fluid.layers.data(
+                name='data', shape=[8, 4], dtype='float32', lod_level=1)
+            im_info = fluid.layers.data(name='im_info', shape=[3])
+            out = fluid.layers.box_clip(
+                input=boxes, im_info=im_info, inplace=True)
+    """
+
+    helper = LayerHelper("box_clip", **locals())
+    output = helper.create_variable_for_type_inference(dtype=input.dtype)
+    inputs = {"Input": input, "ImInfo": im_info}
+    helper.append_op(type="box_clip", inputs=inputs, outputs={"Output": output})
+
+    return output
+
+
 def multiclass_nms(bboxes,
                    scores,
                    score_threshold,
@@ -2132,9 +2181,11 @@ def multiclass_nms(bboxes,
              (After version 1.3, when no boxes detected, the lod is changed 
              from {0} to {1}) 
 
+
     Examples:
         .. code-block:: python
 
+
             boxes = fluid.layers.data(name='bboxes', shape=[81, 4],
                                       dtype='float32', lod_level=1)
             scores = fluid.layers.data(name='scores', shape=[81],

python/paddle/fluid/parallel_executor.py

@@ -146,6 +146,9 @@ class ParallelExecutor(object):
         # step4: get main_program, scope, local_scopes
         main = main_program if main_program \
             else framework.default_main_program()
+        # FIXME(dzhwinter): enable_inplace should be after memory_optimize
+        # if turn on python memory optimize, turn off the inplace_pass.
+        build_strategy.enable_inplace = False if main._is_mem_optimized else True
         scope = scope if scope is not None else executor.global_scope()
 
         if share_vars_from and not isinstance(share_vars_from,

python/paddle/fluid/tests/test_detection.py

@@ -482,6 +482,17 @@ class TestYoloDetection(unittest.TestCase):
             self.assertIsNotNone(loss)
 
 
+class TestBoxClip(unittest.TestCase):
+    def test_box_clip(self):
+        program = Program()
+        with program_guard(program):
+            input_box = layers.data(
+                name='input_box', shape=[7, 4], dtype='float32', lod_level=1)
+            im_info = layers.data(name='im_info', shape=[3], dtype='float32')
+            out = layers.box_clip(input_box, im_info)
+            self.assertIsNotNone(out)
+
+
 class TestMulticlassNMS(unittest.TestCase):
     def test_multiclass_nms(self):
         program = Program()

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -110,6 +110,10 @@ py_test_modules(test_parallel_executor_transformer MODULES test_parallel_executo
 if(NOT APPLE)
     py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL)
 endif()
+if(CMAKE_BUILD_TYPE STREQUAL "Debug")
+    # change the timeout from 600 to 900, because in debug mode, this test need more time.
+    set_tests_properties(test_image_classification_resnet PROPERTIES TIMEOUT 900)
+endif()
 
 if (WITH_NGRAPH)
     add_subdirectory(ngraph)

python/paddle/fluid/tests/unittests/ngraph/test_accuracy_ngraph_op.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+import paddle.fluid.core as core
+from paddle.fluid.tests.unittests.op_test import OpTest
+from paddle.fluid.tests.unittests.test_accuracy_op import TestAccuracyOp
+
+
+class TestNGRAPHAccuracyOp(TestAccuracyOp):
+    def setUp(self):
+        super(TestNGRAPHAccuracyOp, self).setUp()
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/ngraph/test_conv2d_ngraph_op.py

+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+from paddle.fluid.tests.unittests.test_conv2d_op import *
+
+
+class TestNGRAPH(TestConv2dOp):
+    def init_kernel_type(self):
+        super(TestNGRAPH, self).init_kernel_type()
+
+
+class TestNGRAPHWithPad(TestWithPad):
+    def init_kernel_type(self):
+        super(TestNGRAPHWithPad, self).init_kernel_type()
+
+
+class TestNGRAPHWithStride(TestWithStride):
+    def init_kernel_type(self):
+        super(TestNGRAPHWithStride, self).init_kernel_type()
+
+
+class TestNGRAPHWithGroup(TestWithGroup):
+    def init_kernel_type(self):
+        super(TestNGRAPHWithGroup, self).init_kernel_type()
+
+
+class TestNGRAPHWith1x1(TestWith1x1):
+    def init_kernel_type(self):
+        super(TestNGRAPHWith1x1, self).init_kernel_type()
+
+
+class TestNGRAPHWithInput1x1Filter1x1(TestWithInput1x1Filter1x1):
+    def init_kernel_type(self):
+        super(TestNGRAPHWithInput1x1Filter1x1, self).init_kernel_type()
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -40,7 +40,8 @@ class TestParallelExecutorBase(unittest.TestCase):
                                   seed=None,
                                   use_parallel_executor=True,
                                   use_reduce=False,
-                                  use_ir_memory_optimize=False,
+                                  use_ir_memory_optimize=True,
+                                  enable_inplace=True,
                                   fuse_elewise_add_act_ops=False,
                                   fuse_relu_depthwise_conv=False,
                                   optimizer=fluid.optimizer.Adam,
@@ -60,63 +61,65 @@ class TestParallelExecutorBase(unittest.TestCase):
                 main.random_seed = seed
 
             loss = method(use_feed=feed_dict is not None)
-
             if optimizer:
                 optimizer().minimize(loss)
 
             if memory_opt:
                 fluid.memory_optimize(main)
 
-            place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-            exe = fluid.Executor(place)
-            exe.run(startup)
-            exec_strategy = fluid.ExecutionStrategy()
-            exec_strategy.allow_op_delay = allow_op_delay
-            if use_fast_executor:
-                exec_strategy.use_experimental_executor = True
-            build_strategy = fluid.BuildStrategy()
-            build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce \
-                if use_reduce else fluid.BuildStrategy.ReduceStrategy.AllReduce
-            build_strategy.fuse_elewise_add_act_ops = fuse_elewise_add_act_ops
-            build_strategy.fuse_relu_depthwise_conv = fuse_relu_depthwise_conv
-            build_strategy.memory_optimize = use_ir_memory_optimize
-            build_strategy.enable_sequential_execution = enable_sequential_execution
-            if use_cuda and core.is_compiled_with_cuda():
-                build_strategy.remove_unnecessary_lock = True
-            if use_parallel_executor:
-                binary = compiler.CompiledProgram(main).with_data_parallel(
-                    loss_name=loss.name,
-                    build_strategy=build_strategy,
-                    exec_strategy=exec_strategy)
-            else:
-                binary = compiler.CompiledProgram(main)
-
-            if batch_size is not None:
-                batch_size *= fluid.core.get_cuda_device_count(
-                ) if use_cuda else int(
-                    os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
-            begin = time.time()
-            first_loss, = run_executor(
-                exe=exe, binary=binary, feed=feed_dict, fetch_list=[loss.name])
-
-            for i in range(iter):
-                run_executor(
-                    exe=exe, binary=binary, feed=feed_dict, fetch_list=[])
-
-            last_loss, = run_executor(
-                exe=exe, binary=binary, feed=feed_dict, fetch_list=[loss.name])
-            end = time.time()
-
-            if batch_size is not None:
-                print("%.4f Instance per second" % (
-                    (batch_size * iter + 2) / (end - begin)))
-
-            avg_last_loss_val = np.array(last_loss).mean()
-            avg_first_loss_val = np.array(first_loss).mean()
-            if math.isnan(float(avg_last_loss_val)) or math.isnan(
-                    float(avg_first_loss_val)):
-                sys.exit("got NaN loss, training failed.")
-
-            print(first_loss, last_loss)
-            # self.assertGreater(first_loss[0], last_loss[0])
-            return first_loss, last_loss
+        place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
+        exe = fluid.Executor(place)
+        exe.run(startup)
+        exec_strategy = fluid.ExecutionStrategy()
+        exec_strategy.allow_op_delay = allow_op_delay
+        if use_fast_executor:
+            exec_strategy.use_experimental_executor = True
+        build_strategy = fluid.BuildStrategy()
+        build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce \
+            if use_reduce else fluid.BuildStrategy.ReduceStrategy.AllReduce
+        build_strategy.fuse_elewise_add_act_ops = fuse_elewise_add_act_ops
+        build_strategy.fuse_relu_depthwise_conv = fuse_relu_depthwise_conv
+        build_strategy.memory_optimize = use_ir_memory_optimize
+        # python memory optimization is conflict with inplace pass.
+        # Use ir graph memory optimization after inplace pass is the correct way.
+        build_strategy.enable_inplace = False if memory_opt else enable_inplace
+        build_strategy.enable_sequential_execution = enable_sequential_execution
+
+        if use_cuda and core.is_compiled_with_cuda():
+            build_strategy.remove_unnecessary_lock = True
+        if use_parallel_executor:
+            binary = compiler.CompiledProgram(main).with_data_parallel(
+                loss_name=loss.name,
+                build_strategy=build_strategy,
+                exec_strategy=exec_strategy)
+        else:
+            binary = compiler.CompiledProgram(main)
+
+        if batch_size is not None:
+            batch_size *= fluid.core.get_cuda_device_count(
+            ) if use_cuda else int(
+                os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+        begin = time.time()
+        first_loss, = run_executor(
+            exe=exe, binary=binary, feed=feed_dict, fetch_list=[loss.name])
+
+        for i in range(iter):
+            run_executor(exe=exe, binary=binary, feed=feed_dict, fetch_list=[])
+
+        last_loss, = run_executor(
+            exe=exe, binary=binary, feed=feed_dict, fetch_list=[loss.name])
+        end = time.time()
+
+        if batch_size is not None:
+            print("%.4f Instance per second" % (
+                (batch_size * iter + 2) / (end - begin)))
+
+        avg_last_loss_val = np.array(last_loss).mean()
+        avg_first_loss_val = np.array(first_loss).mean()
+        if math.isnan(float(avg_last_loss_val)) or math.isnan(
+                float(avg_first_loss_val)):
+            sys.exit("got NaN loss, training failed.")
+
+        print(first_loss, last_loss)
+        # self.assertGreater(first_loss[0], last_loss[0])
+        return first_loss, last_loss

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

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+import sys
+import math
+from op_test import OpTest
+import copy
+
+
+def box_clip(input_box, im_info, output_box):
+    im_w = round(im_info[1] / im_info[2])
+    im_h = round(im_info[0] / im_info[2])
+    output_box[:, :, 0] = np.maximum(
+        np.minimum(input_box[:, :, 0], im_w - 1), 0)
+    output_box[:, :, 1] = np.maximum(
+        np.minimum(input_box[:, :, 1], im_h - 1), 0)
+    output_box[:, :, 2] = np.maximum(
+        np.minimum(input_box[:, :, 2], im_w - 1), 0)
+    output_box[:, :, 3] = np.maximum(
+        np.minimum(input_box[:, :, 3], im_h - 1), 0)
+
+
+def batch_box_clip(input_boxes, im_info, lod):
+    n = input_boxes.shape[0]
+    m = input_boxes.shape[1]
+    output_boxes = np.zeros((n, m, 4), dtype=np.float32)
+    cur_offset = 0
+    for i in range(len(lod)):
+        box_clip(input_boxes[cur_offset:(cur_offset + lod[i]), :, :],
+                 im_info[i, :],
+                 output_boxes[cur_offset:(cur_offset + lod[i]), :, :])
+        cur_offset += lod[i]
+    return output_boxes
+
+
+class TestBoxClipOp(OpTest):
+    def test_check_output(self):
+        self.check_output()
+
+    def setUp(self):
+        self.op_type = "box_clip"
+        lod = [[1, 2, 3]]
+        input_boxes = np.random.random((6, 10, 4)) * 5
+        im_info = np.array([[5, 8, 1.], [6, 6, 1.], [7, 5, 1.]])
+        output_boxes = batch_box_clip(input_boxes, im_info, lod[0])
+
+        self.inputs = {
+            'Input': (input_boxes.astype('float32'), lod),
+            'ImInfo': im_info.astype('float32'),
+        }
+        self.outputs = {'Output': output_boxes}
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -34,7 +34,9 @@ def box_decoder(t_box, p_box, pb_v, output_box, norm, axis=0):
     pb_y = pb_y.reshape(shape)
 
     if pb_v.ndim == 2:
-        pb_v = pb_v.reshape(1, pb_v.shape[0], pb_v.shape[1])
+        var_shape = (1, pb_v.shape[0], pb_v.shape[1]) if axis == 0 else (
+            pb_v.shape[0], 1, pb_v.shape[1])
+        pb_v = pb_v.reshape(var_shape)
     if pb_v.ndim == 1:
         tb_x = pb_v[0] * t_box[:, :, 0] * pb_w + pb_x
         tb_y = pb_v[1] * t_box[:, :, 1] * pb_h + pb_y
@@ -125,33 +127,6 @@ class TestBoxCoderOp(OpTest):
         self.outputs = {'OutputBox': output_box}
 
 
-class TestBoxCoderOpWithOneRankVar(OpTest):
-    def test_check_output(self):
-        self.check_output()
-
-    def setUp(self):
-        self.op_type = "box_coder"
-        lod = [[1, 1, 1, 1, 1]]
-        prior_box = np.random.random((81, 4)).astype('float32')
-        prior_box_var = np.random.random((4)).astype('float32')
-        target_box = np.random.random((20, 81, 4)).astype('float32')
-        code_type = "DecodeCenterSize"
-        box_normalized = False
-        output_box = batch_box_coder(prior_box, prior_box_var, target_box,
-                                     lod[0], code_type, box_normalized)
-
-        self.inputs = {
-            'PriorBox': prior_box,
-            'PriorBoxVar': prior_box_var,
-            'TargetBox': target_box,
-        }
-        self.attrs = {
-            'code_type': 'decode_center_size',
-            'box_normalized': False
-        }
-        self.outputs = {'OutputBox': output_box}
-
-
 class TestBoxCoderOpWithoutBoxVar(OpTest):
     def test_check_output(self):
         self.check_output()
@@ -210,7 +185,7 @@ class TestBoxCoderOpWithAxis(OpTest):
         self.op_type = "box_coder"
         lod = [[1, 1, 1, 1, 1]]
         prior_box = np.random.random((30, 4)).astype('float32')
-        prior_box_var = np.random.random((4)).astype('float32')
+        prior_box_var = np.random.random((30, 4)).astype('float32')
         target_box = np.random.random((30, 81, 4)).astype('float32')
         code_type = "DecodeCenterSize"
         box_normalized = False

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

@@ -16,12 +16,10 @@ import os
 import unittest
 os.environ['FLAGS_eager_delete_tensor_gb'] = "0.0"
 
-from test_parallel_executor_transformer import TestTransformer
-
-
-class EagerDeletionTestTransformer(TestTransformer):
-    pass
+os.environ[
+    'RECORDIO_FILENAME'] = '/tmp/eager_deletion_transformer.wmt16.recordio'
 
+from test_parallel_executor_transformer import TestTransformer
 
 if __name__ == '__main__':
     unittest.main()

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

@@ -25,6 +25,7 @@ import paddle.fluid.layers as layers
 import paddle.fluid.optimizer as optimizer
 from paddle.fluid.framework import Program, program_guard
 from paddle.fluid.io import save_inference_model, load_inference_model
+from paddle.fluid.transpiler import memory_optimize
 
 
 class TestBook(unittest.TestCase):
@@ -82,9 +83,36 @@ class TestBook(unittest.TestCase):
 
         self.assertEqual(feed_var_names, ["x", "y"])
         self.assertEqual(len(fetch_vars), 1)
-        self.assertEqual(str(fetch_vars[0]), str(avg_cost))
+        print("fetch %s" % str(fetch_vars[0]))
+        self.assertTrue("scale" in str(fetch_vars[0]))
         self.assertEqual(expected, actual)
 
 
+class TestSaveInferenceModel(unittest.TestCase):
+    def test_save_inference_model(self):
+        MODEL_DIR = "./tmp/inference_model2"
+        init_program = Program()
+        program = Program()
+
+        # fake program without feed/fetch
+        with program_guard(program, init_program):
+            x = layers.data(name='x', shape=[2], dtype='float32')
+            y = layers.data(name='y', shape=[1], dtype='float32')
+
+            y_predict = layers.fc(input=x, size=1, act=None)
+
+            cost = layers.square_error_cost(input=y_predict, label=y)
+            avg_cost = layers.mean(cost)
+
+        place = core.CPUPlace()
+        exe = executor.Executor(place)
+        exe.run(init_program, feed={}, fetch_list=[])
+
+        memory_optimize(program, print_log=True)
+        self.assertEqual(program._is_mem_optimized, True)
+        # will print warning message
+        save_inference_model(MODEL_DIR, ["x", "y"], [avg_cost], exe, program)
+
+
 if __name__ == '__main__':
     unittest.main()

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

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import os
+import unittest
+import numpy as np
+import paddle.fluid.core as core
+import paddle.fluid as fluid
+from parallel_executor_test_base import TestParallelExecutorBase
+
+
+def fc_with_batchnorm(use_feed):
+    img = fluid.layers.data(name='image', shape=[784], dtype='float32')
+    label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+
+    hidden = img
+    for _ in range(3):
+        hidden = fluid.layers.fc(
+            hidden,
+            size=200,
+            act='tanh',
+            bias_attr=fluid.ParamAttr(
+                initializer=fluid.initializer.Constant(value=1.0)))
+
+        hidden = fluid.layers.batch_norm(input=hidden)
+    prediction = fluid.layers.fc(hidden, size=10, act='softmax')
+    loss = fluid.layers.cross_entropy(input=prediction, label=label)
+    loss = fluid.layers.mean(loss)
+    return loss
+
+
+class TestIrInplace(TestParallelExecutorBase):
+    @classmethod
+    def setUpClass(cls):
+        os.environ['CPU_NUM'] = str(4)
+
+    def _fc_with_batchnorm(self,
+                           ir_memory_optimize,
+                           enable_inplace,
+                           memory_opt=False):
+
+        if not core.is_compiled_with_cuda():
+            return
+        np.random.seed(5)
+        img = np.random.random(size=[32, 784]).astype(np.float32)
+        label = np.ones(shape=[32, 1], dtype='int64')
+        self.check_network_convergence(
+            fc_with_batchnorm,
+            feed_dict={"image": img,
+                       "label": label},
+            use_cuda=True,
+            memory_opt=memory_opt,
+            use_ir_memory_optimize=ir_memory_optimize,
+            enable_inplace=enable_inplace)
+
+    def test_fc_with_batchnorm(self, delta=1e-3):
+        loss00 = self._fc_with_batchnorm(False, False)
+        loss10 = self._fc_with_batchnorm(True, False)
+        loss01 = self._fc_with_batchnorm(False, True)
+        loss11 = self._fc_with_batchnorm(True, True)
+        self.assertAlmostEqual(loss00, loss10, delta=delta)
+        self.assertAlmostEqual(loss00, loss01, delta=delta)
+        self.assertAlmostEqual(loss00, loss11, delta=delta)

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

@@ -200,7 +200,7 @@ class TestResnet(TestParallelExecutorBase):
                                       model,
                                       use_cuda,
                                       iter=20,
-                                      delta2=1e-6):
+                                      delta2=1e-5):
         if use_cuda and not core.is_compiled_with_cuda():
             return
 
@@ -228,7 +228,7 @@ class TestResnet(TestParallelExecutorBase):
             optimizer=optimizer)
 
         for loss in zip(all_reduce_first_loss, reduce_first_loss):
-            self.assertAlmostEquals(loss[0], loss[1], delta=1e-6)
+            self.assertAlmostEquals(loss[0], loss[1], delta=1e-5)
         for loss in zip(all_reduce_last_loss, reduce_last_loss):
             self.assertAlmostEquals(loss[0], loss[1], delta=delta2)
 
@@ -258,17 +258,17 @@ class TestResnet(TestParallelExecutorBase):
             enable_sequential_execution=True)
 
         for loss in zip(all_reduce_first_loss, all_reduce_first_loss_seq):
-            self.assertAlmostEquals(loss[0], loss[1], delta=1e-6)
+            self.assertAlmostEquals(loss[0], loss[1], delta=1e-5)
         for loss in zip(all_reduce_last_loss, all_reduce_last_loss_seq):
             self.assertAlmostEquals(loss[0], loss[1], delta=delta2)
 
         for loss in zip(reduce_first_loss, reduce_first_loss_seq):
-            self.assertAlmostEquals(loss[0], loss[1], delta=1e-6)
+            self.assertAlmostEquals(loss[0], loss[1], delta=1e-5)
         for loss in zip(reduce_last_loss, reduce_last_loss_seq):
             self.assertAlmostEquals(loss[0], loss[1], delta=delta2)
 
         for loss in zip(all_reduce_first_loss_seq, reduce_first_loss_seq):
-            self.assertAlmostEquals(loss[0], loss[1], delta=1e-6)
+            self.assertAlmostEquals(loss[0], loss[1], delta=1e-5)
         for loss in zip(all_reduce_last_loss_seq, reduce_last_loss_seq):
             self.assertAlmostEquals(loss[0], loss[1], delta=delta2)
 
@@ -277,7 +277,7 @@ class TestResnet(TestParallelExecutorBase):
                                   use_cuda=True,
                                   use_reduce=False,
                                   iter=20,
-                                  delta2=1e-6):
+                                  delta2=1e-5):
         if use_cuda and not core.is_compiled_with_cuda():
             return
 
@@ -308,7 +308,7 @@ class TestResnet(TestParallelExecutorBase):
             optimizer=optimizer)
 
         self.assertAlmostEquals(
-            np.mean(parallel_first_loss), single_first_loss[0], delta=1e-6)
+            np.mean(parallel_first_loss), single_first_loss[0], delta=1e-5)
         self.assertAlmostEquals(
             np.mean(parallel_last_loss), single_last_loss[0], delta=delta2)
 

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

@@ -24,7 +24,7 @@ import paddle.fluid.core as core
 import paddle.dataset.wmt16 as wmt16
 import os
 
-WMT16_RECORDIO_FILE = "/tmp/wmt16.recordio"
+WMT16_RECORDIO_FILE = os.environ.get('RECORDIO_FILENAME', '/tmp/wmt16.recordio')
 
 
 class ModelHyperParams(object):

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

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import os
+os.environ['FLAGS_benchmark'] = 'True'
+
+import numpy
+import paddle.fluid.core as core
+from paddle.fluid.executor import Executor
+from paddle.fluid.layers import mul, data
+
+
+class TestPeakMemoryMonitoring(unittest.TestCase):
+    def test_mul(self):
+
+        a = data(name='a', shape=[784], dtype='float32')
+        b = data(
+            name='b',
+            shape=[784, 100],
+            dtype='float32',
+            append_batch_size=False)
+        out = mul(x=a, y=b)
+
+        if core.is_compiled_with_cuda():
+            place = core.CUDAPlace(0)
+
+            a_np = numpy.random.random((100, 784)).astype('float32')
+            b_np = numpy.random.random((784, 100)).astype('float32')
+            self.assertEqual(0, core.get_mem_usage(0))
+            exe = Executor(place)
+            outs = exe.run(feed={'a': a_np, 'b': b_np}, fetch_list=[out])
+            out = outs[0]
+            #disable this assert since ctest will ignore the os.environ setting 
+            #self.assertGreater(core.get_mem_usage(0), 0)
+
+            raised = False
+            try:
+                core.print_mem_usage()
+            except:
+                raised = True
+            self.assertFalse(raised, 'Exception raised')
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -17,6 +17,7 @@ from __future__ import print_function
 from functools import partial
 import numpy as np
 
+import os
 import paddle.fluid as fluid
 import paddle.fluid.layers as layers
 from paddle.fluid.layers.io import open_recordio_file
@@ -408,7 +409,7 @@ def transformer(
         trg_pad_idx,
         pos_pad_idx, ):
     file_obj = open_recordio_file(
-        filename='/tmp/wmt16.recordio',
+        filename=os.environ.get('RECORDIO_FILENAME', '/tmp/wmt16.recordio'),
         shapes=[
             [batch_size * max_length, 1],
             [batch_size * max_length, 1],

python/paddle/fluid/transpiler/memory_optimization_transpiler.py

@@ -540,6 +540,7 @@ def memory_optimize(input_program,
     if skip_opt_set is not None:
         skip_opt_set = set(map(to_name_str, skip_opt_set))
     cfgs = _get_cfgs(input_program)
+    input_program._is_mem_optimized = True
     for cfg in cfgs:
         cfg.memory_optimize(skip_opt_set=skip_opt_set, level=level)
 
@@ -559,5 +560,6 @@ def release_memory(input_program, skip_opt_set=None):
         None
     """
     cfgs = _get_cfgs(input_program)
+    input_program._is_mem_optimized = True
     for cfg in cfgs:
         cfg.release_memory(skip_opt_set=skip_opt_set)