Merge remote-tracking branch 'ups/develop' into fea/jitkernel_peephole

test=develop

Author: tensor-tang

.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)

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/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/data_type.h

@@ -17,7 +17,6 @@ limitations under the License. */
 #include <typeindex>
 #include "paddle/fluid/framework/framework.pb.h"
 #include "paddle/fluid/platform/enforce.h"
-
 #include "paddle/fluid/platform/float16.h"
 
 namespace paddle {

paddle/fluid/framework/ir/CMakeLists.txt

@@ -38,6 +38,7 @@ pass_library(fc_lstm_fuse_pass inference)
 pass_library(embedding_fc_lstm_fuse_pass inference)
 pass_library(fc_gru_fuse_pass inference)
 pass_library(seq_concat_fc_fuse_pass inference)
+pass_library(conv_bn_fuse_pass inference)
 
 cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector )