Created by: mapingshuo

###错误描述

基于C++训练库的demo trainer在MKLDNN下报segmentation fault: http://ci.paddlepaddle.org/viewLog.html?tab=buildLog&logTab=tree&filter=debug&expand=all&buildId=171521&_focus=6744

本地编译时选择 WITH_MKL=ON, WITH_MKLDNN=ON是出错，WITH_MKL=ON, WITH_MKLDNN=OFF时正常运行, 应该是MKLDNN的问题。

错误排查

错误出现在运行这一行时：https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/train/demo/demo_trainer.cc#L83

经过排查，是在tensor.cc里面，holder_.reset()时出的错：

https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/tensor.cc#L55

如何复现这一错误？

cmake .. -DWITH_GPU=OFF -DWITH_DISTRIBUTE=OFF -DON_INFER=ON

make -j40 fluid_lib_dist
make -j40 inference_lib_dist

然后，在Paddle/paddle/fluid/train/demo文件夹下：

set -x

PADDLE_ROOT=your_Paddle/
TURN_ON_MKL=ON # use MKL or Openblas

# download models
function download() {
    wget -q http://paddle-tar.bj.bcebos.com/train_demo/LR/main_program
    wget -q http://paddle-tar.bj.bcebos.com/train_demo/LR/startup_program
}

download

# build demo trainer
fluid_install_dir=${PADDLE_ROOT}/build/fluid_install_dir

mkdir -p build
cd build
rm -rf *
cmake .. -DPADDLE_LIB=$fluid_install_dir \
         -DWITH_MKLDNN=ON \
         -DWITH_MKL=$TURN_ON_MKL
make

cd ..

# run demo trainer
build/demo_trainer