test=develop

Author: sneaxiy

.gitignore

@@ -25,5 +25,6 @@ third_party/
 bazel-*
 third_party/
 
+build_*
 # clion workspace.
 cmake-build-*

CMakeLists.txt

@@ -72,6 +72,7 @@ option(WITH_INFERENCE    "Compile fluid inference library"              ON)
 option(WITH_INFERENCE_API_TEST   "Test fluid inference high-level api interface"  OFF)
 option(WITH_SYSTEM_BLAS   "Use system blas library"           OFF)
 option(PY_VERSION       "Compile PaddlePaddle with python3 support"     ${PY_VERSION})
+option(WITH_FAST_MATH   "Make use of fast math library, might affect the precision to some extent" ON)
 
 # PY_VERSION
 if(NOT PY_VERSION)

Dockerfile

@@ -24,6 +24,7 @@ COPY ./paddle/scripts/docker/root/ /root/
 
 RUN apt-get update && \
     apt-get install -y --allow-downgrades patchelf \
+    python3 python3-dev python3-pip \
     git python-pip python-dev python-opencv openssh-server bison \
     libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 \
     wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
@@ -70,24 +71,33 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8
 # specify sphinx version as 1.5.6 and remove -U option for [pip install -U
 # sphinx-rtd-theme] since -U option will cause sphinx being updated to newest
 # version(1.7.1 for now), which causes building documentation failed.
-RUN easy_install -U pip && \
+RUN pip3 install -U wheel && \
+    pip3 install -U docopt PyYAML sphinx==1.5.6 && \
+    pip3 install sphinx-rtd-theme==0.1.9 recommonmark && \
+    easy_install -U pip && \
     pip install -U wheel && \
     pip install -U docopt PyYAML sphinx==1.5.6 && \
     pip install sphinx-rtd-theme==0.1.9 recommonmark
 
-RUN pip install pre-commit 'ipython==5.3.0' && \
+RUN pip3 install pre-commit 'ipython==5.3.0' && \
+    pip3 install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
+    pip3 install opencv-python && \
+    pip install pre-commit 'ipython==5.3.0' && \
     pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
     pip install opencv-python
 
 #For docstring checker
+RUN pip3 install pylint pytest astroid isort
 RUN pip install pylint pytest astroid isort LinkChecker
 
 COPY ./python/requirements.txt /root/
+RUN pip3 install -r /root/requirements.txt
 RUN pip install -r /root/requirements.txt
 
 # To fix https://github.com/PaddlePaddle/Paddle/issues/1954, we use
 # the solution in https://urllib3.readthedocs.io/en/latest/user-guide.html#ssl-py2
 RUN apt-get install -y libssl-dev libffi-dev
+RUN pip3 install certifi urllib3[secure]
 RUN pip install certifi urllib3[secure]
 
 

cmake/cblas.cmake

@@ -40,7 +40,7 @@ set(OPENBLAS_LIB_SEARCH_PATHS
         /usr/local/opt/openblas/lib)
 
 find_path(OPENBLAS_INC_DIR NAMES cblas.h
-  PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS})
+  PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS} NO_DEFAULT_PATH)
 find_path(OPENBLAS_LAPACKE_INC_DIR NAMES lapacke.h
   PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS})
 find_library(OPENBLAS_LIB NAMES openblas

cmake/cuda.cmake

@@ -175,7 +175,10 @@ list(APPEND CUDA_NVCC_FLAGS "-std=c++11")
 list(APPEND CUDA_NVCC_FLAGS "-Xcompiler -fPIC")
 endif(NOT WIN32)
 
-list(APPEND CUDA_NVCC_FLAGS "--use_fast_math")
+if(WITH_FAST_MATH)
+  # Make use of fast math library. https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html
+  list(APPEND CUDA_NVCC_FLAGS "--use_fast_math")
+endif()
 # in cuda9, suppress cuda warning on eigen 
 list(APPEND CUDA_NVCC_FLAGS "-w")
 # Set :expt-relaxed-constexpr to suppress Eigen warnings

cmake/external/eigen.cmake

@@ -3,6 +3,14 @@ INCLUDE(ExternalProject)
 SET(EIGEN_SOURCE_DIR ${THIRD_PARTY_PATH}/eigen3)
 SET(EIGEN_INCLUDE_DIR ${EIGEN_SOURCE_DIR}/src/extern_eigen3)
 INCLUDE_DIRECTORIES(${EIGEN_INCLUDE_DIR})
+if(NOT WITH_FAST_MATH)
+  # EIGEN_FAST_MATH: https://eigen.tuxfamily.org/dox/TopicPreprocessorDirectives.html
+  # enables some optimizations which might affect the accuracy of the result. 
+  # This currently enables the SSE vectorization of sin() and cos(), 
+  # and speedups sqrt() for single precision.
+  # Defined to 1 by default. Define it to 0 to disable.
+  add_definitions(-DEIGEN_FAST_MATH=0)
+endif()
 
 if(WITH_AMD_GPU)
     ExternalProject_Add(

cmake/external/openblas.cmake

@@ -27,7 +27,7 @@ IF(NOT ${CBLAS_FOUND})
 
     SET(CBLAS_SOURCES_DIR ${THIRD_PARTY_PATH}/openblas)
     SET(CBLAS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/openblas)
-    SET(CBLAS_INCLUDE_DIR "${CBLAS_INSTALL_DIR}/include" CACHE PATH "openblas include directory." FORCE)
+    SET(CBLAS_INC_DIR "${CBLAS_INSTALL_DIR}/include" CACHE PATH "openblas include directory." FORCE)
 
     SET(CBLAS_LIBRARIES
         "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}"
@@ -96,7 +96,7 @@ IF(NOT ${CBLAS_FOUND})
     ENDIF(NOT WIN32)
     SET(CBLAS_PROVIDER openblas)
     IF(WITH_C_API)
-        INSTALL(DIRECTORY ${CBLAS_INCLUDE_DIR} DESTINATION third_party/openblas)
+        INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
         # Because libopenblas.a is a symbolic link of another library, thus need to
         # install the whole directory.
         IF(ANDROID)
@@ -117,8 +117,8 @@ IF(NOT ${CBLAS_FOUND})
 ENDIF(NOT ${CBLAS_FOUND})
 
 MESSAGE(STATUS "BLAS library: ${CBLAS_LIBRARIES}")
-MESSAGE(STATUS "BLAS Include: ${CBLAS_INCLUDE_DIR}")
-INCLUDE_DIRECTORIES(${CBLAS_INCLUDE_DIR})
+MESSAGE(STATUS "BLAS Include: ${CBLAS_INC_DIR}")
+INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
 
 # FIXME(gangliao): generate cblas target to track all high performance
 # linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)

cmake/flags.cmake

@@ -157,6 +157,8 @@ if (APPLE)
         # On Mac OS X build fat binaries with x86_64 architectures by default.
         set (CMAKE_OSX_ARCHITECTURES "x86_64" CACHE STRING "Build architectures for OSX" FORCE)
     endif()
+    # On Mac OS X register class specifier is deprecated and will cause warning error on latest clang 10.0
+    set (COMMON_FLAGS -Wno-deprecated-register)
 endif(APPLE)
 
 if(LINUX)

paddle/fluid/API.spec

@@ -198,6 +198,9 @@ paddle.fluid.layers.argsort ArgSpec(args=['input', 'axis', 'name'], varargs=None
 paddle.fluid.layers.ones ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=None, keywords=None, defaults=(False,))
 paddle.fluid.layers.zeros ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=None, keywords=None, defaults=(False,))
 paddle.fluid.layers.reverse ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.layers.has_inf ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.layers.has_nan ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.layers.isfinite ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.While.__init__ ArgSpec(args=['self', 'cond', 'is_test', 'name'], varargs=None, keywords=None, defaults=(False, None))
 paddle.fluid.layers.While.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.Switch.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,))

paddle/fluid/framework/CMakeLists.txt

@@ -1,3 +1,4 @@
+
 # windows treat symbolic file as a real file, which is different with unix
 # We create a hidden file and compile it instead of origin source file.
 function(windows_symbolic TARGET)
@@ -9,11 +10,23 @@ function(windows_symbolic TARGET)
   if (NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cc OR NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cu)
       message(FATAL " ${src}.cc and ${src}.cu must exsits, and ${src}.cu must be symbolic file.")
   endif()
-  add_custom_command(OUTPUT .${src}.cu 
+
+  # only copy the xx.cu to .xx.cu when the content are modified
+  set(copy_flag 1)
+  if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/.${src}.cu)
+  file(READ ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cc SOURCE_STR)
+  file(READ ${CMAKE_CURRENT_SOURCE_DIR}/.${src}.cu TARGET_STR)
+  if (SOURCE_STR STREQUAL TARGET_STR)
+    set(copy_flag 0)
+  endif()
+  endif()
+  if (copy_flag)
+  add_custom_command(OUTPUT .${src}.cu
           COMMAND ${CMAKE_COMMAND} -E remove ${CMAKE_CURRENT_SOURCE_DIR}/.${src}.cu
           COMMAND ${CMAKE_COMMAND} -E copy "${CMAKE_CURRENT_SOURCE_DIR}/${src}.cc" "${CMAKE_CURRENT_SOURCE_DIR}/.${src}.cu"
           COMMENT "create hidden file of ${src}.cu")
-  add_custom_target(${TARGET} ALL DEPENDS .${src}.cu)  
+  endif(copy_flag)
+  add_custom_target(${TARGET} ALL DEPENDS .${src}.cu)
   endforeach()
 endfunction()
 
@@ -81,6 +94,8 @@ nv_test(data_device_transform_test SRCS data_device_transform_test.cu
 
 if(WITH_GPU)
   if (WIN32)
+    # windows treat symbolic file as a real file, which is different with unix
+    # We create a hidden file and compile it instead of origin source file.
       windows_symbolic(hidden_file SRCS data_type_transform.cu)
       nv_library(data_type_transform SRCS .data_type_transform.cu DEPS tensor)
       add_dependencies(data_type_transform hidden_file)
@@ -149,7 +164,7 @@ if(WITH_DISTRIBUTE)
   set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
 else()
   cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass)
-  cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass elementwise_add_op)
+  cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
 endif()
 
 if (NOT WIN32)