merge develop

Author: dzhwinter

CMakeLists.txt

@@ -36,8 +36,7 @@ include(simd)
 ################################ Configurations #######################################
 option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_FOUND})
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
-option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    ${AVX_FOUND})
-option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      ${AVX_FOUND})
+option(WITH_MKL         "Compile PaddlePaddle with MKL support."        ${AVX_FOUND})
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        ON)
 option(WITH_SWIG_PY     "Compile PaddlePaddle with inference api"       ON)
@@ -82,10 +81,8 @@ if(ANDROID OR IOS)
         "Disable PYTHON when cross-compiling for Android and iOS" FORCE)
     set(WITH_RDMA OFF CACHE STRING
         "Disable RDMA when cross-compiling for Android and iOS" FORCE)
-    set(WITH_MKLDNN OFF CACHE STRING
-        "Disable MKLDNN when cross-compiling for Android and iOS" FORCE)
-    set(WITH_MKLML OFF CACHE STRING
-        "Disable MKLML package when cross-compiling for Android and iOS" FORCE)
+    set(WITH_MKL OFF CACHE STRING
+        "Disable MKL when cross-compiling for Android and iOS" FORCE)
 
     # Compile PaddlePaddle mobile inference library
     if (NOT WITH_C_API)
@@ -111,6 +108,14 @@ else()
     set(THIRD_PARTY_BUILD_TYPE Release)
 endif()
 
+set(WITH_MKLML ${WITH_MKL})
+if (WITH_MKL AND AVX2_FOUND)
+    set(WITH_MKLDNN ON)
+else()
+    message(STATUS "Do not have AVX2 intrinsics and disabled MKL-DNN")
+    set(WITH_MKLDNN OFF)
+endif()
+
 ########################################################################################
 
 include(external/mklml)     # download mklml package
@@ -158,14 +163,15 @@ set(EXTERNAL_LIBS
 )
 
 if(WITH_GPU)
-    list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
-    if(NOT WITH_DSO)
-        list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY})
-    endif(NOT WITH_DSO)
+  include(cuda)
 endif(WITH_GPU)
 
+if(WITH_MKLML)
+    list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB})
+endif()
+
 if(WITH_MKLDNN)
-    list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB} ${MKLDNN_IOMP_LIB})
+    list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB})
 endif()
 
 if(USE_NNPACK)

benchmark/IntelOptimizedPaddle.md

@@ -12,11 +12,11 @@ Machine:
 
 System: CentOS release 6.3 (Final), Docker 1.12.1.
 
-PaddlePaddle: paddlepaddle/paddle:latest (TODO: will rerun after 0.11.0)
-
-- MKL-DNN tag v0.10
-- MKLML 2018.0.20170720
+PaddlePaddle: paddlepaddle/paddle:latest (for MKLML and MKL-DNN), paddlepaddle/paddle:latest-openblas (for OpenBLAS)
+- MKL-DNN tag v0.11
+- MKLML 2018.0.1.20171007
 - OpenBLAS v0.2.20
+(TODO: will rerun after 0.11.0)
 
 On each machine, we will test and compare the performance of training on single node using MKL-DNN / MKLML / OpenBLAS respectively.
 
@@ -31,17 +31,37 @@ Input image size - 3 * 224 * 224, Time: images/second
 
 | BatchSize    | 64    | 128  | 256     |
 |--------------|-------| -----| --------|
-| OpenBLAS     | 7.82  | 8.62  | 10.34  | 
-| MKLML        | 11.02 | 12.86 | 15.33  |
-| MKL-DNN      | 27.69 | 28.8 | 29.27  |
+| OpenBLAS     | 7.80  | 9.00  | 10.80  | 
+| MKLML        | 12.12 | 13.70 | 16.18  |
+| MKL-DNN      | 28.46 | 29.83 | 30.44  |
+
+
+chart on batch size 128
+TBD
+
+ - ResNet-50
+
+| BatchSize    | 64    | 128   | 256    |
+|--------------|-------| ------| -------|
+| OpenBLAS     | 25.22 | 25.68 | 27.12  | 
+| MKLML        | 32.52 | 31.89 | 33.12  |
+| MKL-DNN      | 81.69 | 82.35 | 84.08  |
 
 
 chart on batch size 128
 TBD
 
- - ResNet
  - GoogLeNet
 
+| BatchSize    | 64    | 128   | 256    |
+|--------------|-------| ------| -------|
+| OpenBLAS     | 89.52 | 96.97 | 108.25 | 
+| MKLML        | 128.46| 137.89| 158.63 |
+| MKL-DNN      | 250.46| 264.83| 269.50 |
+
+chart on batch size 128
+TBD
+
 ### Laptop
 TBD
 ### Desktop

benchmark/paddle/image/googlenet.py

@@ -5,6 +5,7 @@
 width = 224
 num_class = 1000
 batch_size = get_config_arg('batch_size', int, 128)
+use_gpu = get_config_arg('use_gpu', bool, True)
 
 args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
 define_py_data_sources2(
@@ -16,6 +17,8 @@
     learning_method=MomentumOptimizer(0.9),
     regularization=L2Regularization(0.0005 * batch_size))
 
+conv_projection = conv_projection if use_gpu else img_conv_layer
+
 def inception2(name, input, channels, \
     filter1,
     filter3R, filter3,
@@ -138,7 +141,7 @@ def inception(name, input, channels, \
     cat = concat_layer(
         name=name,
         input=[cov1, cov3, cov5, covprj],
-        bias_attr=True,
+        bias_attr=True if use_gpu else False,
         act=ReluActivation())
     return cat
 

benchmark/paddle/image/run_mkldnn.sh

@@ -1,9 +1,7 @@
 set -e
 
 function train() {
-  unset OMP_NUM_THREADS MKL_NUM_THREADS
-  export OMP_DYNAMIC="FALSE"
-  export KMP_AFFINITY="granularity=fine,compact,0,0"
+  unset OMP_NUM_THREADS MKL_NUM_THREADS OMP_DYNAMIC KMP_AFFINITY
   topology=$1
   layer_num=$2
   bs=$3
@@ -14,8 +12,6 @@ function train() {
   elif [ $4 == "False" ]; then
     thread=`nproc`
     # each trainer_count use only 1 core to avoid conflict
-    export OMP_NUM_THREADS=1
-    export MKL_NUM_THREADS=1
     log="logs/${topology}-${layer_num}-${thread}mklml-${bs}.log"
   else
     echo "Wrong input $3, use True or False."
@@ -44,6 +40,7 @@ fi
 for use_mkldnn in True False; do
   for batchsize in 64 128 256; do
     train vgg 19 $batchsize $use_mkldnn
-    train resnet 50  $batchsize $use_mkldnn
+    train resnet 50 $batchsize $use_mkldnn
+    train googlenet v1 $batchsize $use_mkldnn
   done
 done

cmake/configure.cmake

@@ -76,27 +76,14 @@ else()
     include_directories(${CUDA_TOOLKIT_INCLUDE})
 endif(NOT WITH_GPU)
 
-if(WITH_MKLDNN)
-    add_definitions(-DPADDLE_USE_MKLDNN)
-    if (WITH_MKLML AND MKLDNN_IOMP_DIR)
-        message(STATUS "Enable Intel OpenMP at ${MKLDNN_IOMP_DIR}")
-        set(OPENMP_FLAGS "-fopenmp")
-        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} ${OPENMP_FLAGS}")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OPENMP_FLAGS}")
-    else()
-        find_package(OpenMP)
-        if(OPENMP_FOUND)
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
-        else()
-            message(WARNING "Can not find OpenMP."
-                 "Some performance features in MKLDNN may not be available")
-        endif()
-    endif()
-
-endif(WITH_MKLDNN)
+if (WITH_MKLML AND MKLML_IOMP_LIB)
+    message(STATUS "Enable Intel OpenMP with ${MKLML_IOMP_LIB}")
+    set(OPENMP_FLAGS "-fopenmp")
+    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} ${OPENMP_FLAGS}")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OPENMP_FLAGS}")
+endif()
 
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${SIMD_FLAG}")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SIMD_FLAG}")

cmake/cross_compiling/ios.cmake

@@ -76,11 +76,9 @@ set(IOS_PLATFORM ${IOS_PLATFORM} CACHE STRING "Type of iOS Platform")
 # Set the architecture for iOS
 if(NOT DEFINED IOS_ARCH)
   if(IOS_PLATFORM STREQUAL "OS")
-    # FIXME(liuyiqun): support "armv7;armv7s;arm64" future
-    set(IOS_ARCH "arm64")
+    set(IOS_ARCH "armv7;armv7s;arm64")
   elseif(IOS_PLATFORM STREQUAL "SIMULATOR")
-    # FIXME(liuyiqun): support "i386;x86_64" future
-    set(IOS_ARCH "x86_64")
+    set(IOS_ARCH "i386;x86_64")
   endif()
 endif()
 set(CMAKE_OSX_ARCHITECTURES ${IOS_ARCH} CACHE string  "Build architecture for iOS")
@@ -248,7 +246,7 @@ set(IOS_COMPILER_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} ${XCODE_IOS_BITCODE_
 
 # Hidden visibilty is required for cxx on iOS 
 set(CMAKE_C_FLAGS "${IOS_COMPILER_FLAGS} ${CMAKE_C_FLAGS}" CACHE STRING "C flags")
-set(CMAKE_CXX_FLAGS "${IOS_COMPILER_FLAGS} -fvisibility-inlines-hidden ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags")
+set(CMAKE_CXX_FLAGS "${IOS_COMPILER_FLAGS} -fvisibility=hidden -fvisibility-inlines-hidden ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags")
 
 set(IOS_LINK_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} -Wl,-search_paths_first")