嵌入式深度學習之mxnet交叉編譯Arm Linux
阿新 • • 發佈：2019-02-11
參考文件
本文件涉及到的目標硬體為全志H8(8核Cortex-A7，Armv7架構)，但是對其他Arm晶片也有一定的借鑑意義，只需要更換交叉編譯鏈即可。
開發環境介紹
- 主機作業系統：Ubuntu14.04 64位
- 目標平臺：CQA83t 全志H8
- 交叉工具鏈：arm-linux-gnueabihf，gcc4.9.2
- mxnet版本：20180124-master
注意：mxnet的最低GCC版本為4.8.0
設定交叉編譯鏈
# 在/etc/bash.bashrc的最後增加如下指令
# Allwinner H8 CQA83t cross compiler
export ARCH=arm
export 下載依賴庫
編譯OpenBLAS
編譯
cd OpenBLAS-0.2.20
# 編譯單獨的靜態庫版本
make TARGET=ARMV7 HOSTCC=gcc CROSS=1 CROSS_SUFFIX=arm-linux-gnueabihf- BINARY=32 NOFORTRAN=1 libs -j8
# 編譯動態庫版本
make TARGET=ARMV7 HOSTCC=gcc CROSS=1 CROSS_SUFFIX=arm-linux-gnueabihf- BINARY=32 NOFORTRAN=1 libs shared
# 另附上ARMV8晶片的編譯命令以供參考
make TARGET=ARMV8 HOSTCC=gcc CROSS=1 CROSS_SUFFIX=aarch64-linux-gnu- BINARY=64 NOFORTRAN=1 libs -j8
make TARGET=ARMV8 HOSTCC=gcc CROSS=1 CROSS_SUFFIX=aarch64-linux-gnu- BINARY=64 NOFORTRAN=1 libs shared -j8安裝
# PREFIX使用您自己的自定義安裝目錄
make PREFIX=/usr/local/H8/openBLAS-0.2.20 install編譯mxnet
修改CmakeLists.txt
修改
# ---[ LAPack
if(USE_LAPACK)
add_definitions(-DMXNET_USE_LAPACK=1)
list(APPEND mxnet_LINKER_LIBS lapack)
else(USE_LAPACK)
Workaround for Windows until using new Jenkinsfile.
if(BLAS STREQUAL "Open" OR BLAS STREQUAL "open")
add_definitions(-DMXNET_USE_LAPACK=1)
endif()
endif()
為
# ---[ LAPack
if(USE_LAPACK)
add_definitions(-DMXNET_USE_LAPACK=1)
list(APPEND mxnet_LINKER_LIBS lapack)
else(USE_LAPACK)
# Workaround for Windows until using new Jenkinsfile.
# if(BLAS STREQUAL "Open" OR BLAS STREQUAL "open")
# add_definitions(-DMXNET_USE_LAPACK=1)
# endif()
endif()
# 這裡遮蔽add_definitions(-DMXNET_USE_LAPACK=1)是因為這裡的新增是Windows專用的,針對Linux需要遮蔽修改tests/CmakeLists.txt
修改
if(NOT MSVC)
set(UNITTEST_STATIC_LINK ON)
endif()
為
if(NOT MSVC)
set(UNITTEST_STATIC_LINK OFF)
endif()
修改
if(UNITTEST_STATIC_LINK)
target_link_libraries(${PROJECT_NAME}_unit_tests
${GTEST_LIBRARY}
${BEGIN_WHOLE_ARCHIVE} mxnet ${END_WHOLE_ARCHIVE}
dmlc
${mxnet_LINKER_LIBS}
${pslite_LINKER_LIBS}
)
else()
target_link_libraries(${PROJECT_NAME}_unit_tests
${GTEST_LIBRARY}
dmlc
${nnvm_LINKER_LIBS}
${mxnet_LINKER_LIBS}
mxnet
${pslite_LINKER_LIBS}
)
endif()
為
if(UNITTEST_STATIC_LINK)
target_link_libraries(${PROJECT_NAME}_unit_tests
${GTEST_LIBRARY}
${BEGIN_WHOLE_ARCHIVE} mxnet ${END_WHOLE_ARCHIVE}
dmlc
${mxnet_LINKER_LIBS}
${pslite_LINKER_LIBS}
)
else()
target_link_libraries(${PROJECT_NAME}_unit_tests
${GTEST_LIBRARY}
dmlc
${nnvm_LINKER_LIBS}
${mxnet_LINKER_LIBS}
mxnet
${pslite_LINKER_LIBS}
rt
)
endif()新建cmake Build目錄
cd incubator-mxnet
mkdir _install開啟cmake-gui進行交叉編譯鏈配置
注意：mxnet要求cmake版本需要高於3.0.2
選擇Specify options for cross-compiling，進入交叉編譯鏈設定介面
# 設定交叉編譯相關引數
# 本文使用的是NVIDIA提供的交叉工具鏈，存放位置為/opt/toolchain/gcc-linaro-arm-linux-gnueabihf-4.9-2014.09_linux/
# (更改為你正在使用的目錄)
# 設定Operating System 為 arm-linux
# 設定C編譯器為 /opt/toolchain/gcc-linaro-arm-linux-gnueabihf-4.9-2014.09_linux/bin/arm-linux-gnueabihf-gcc
# (更改為你正在使用的gcc)
# 設定C++編譯器為 /opt/toolchain/gcc-linaro-arm-linux-gnueabihf-4.9-2014.09_linux/bin/arm-linux-gnueabihf-g++
# 設定System Root 為 /opt/toolchain/gcc-linaro-arm-linux-gnueabihf-4.9-2014.09_linux/arm-linux-gnueabihf/sysroot點選Configure生成配置選項，並且開啟Advancd選項
此時Cmake會顯示CUDA錯誤，暫時先不用管，後面修改設定後錯誤會消失
修改設定選項：
# 去掉勾選為FLASE 勾選為TRUE
USE_CUDA # FALSE
USE_CUDNN # FALSE
USE_GPERFTOOLS # FALSE
USE_JEMALLOC # FALSE
USE_LAPACK # FALSE
USE_OPENCV # FALSE
USE_OPENMP # TRUE 多核ARM建議開啟openmp點選Configure
再次修改設定選項：
# 指定安裝型別和目錄
CMAKE_BUILD_TYPE Release
CMAKE_INSTALL_PREFIX /usr/local/H8/mxnet-20180122
CMAKE_C_FLAGS -mfpu=neon -march=armv7-a
CMAKE_CXX_FLAGS -mfpu=neon -march=armv7-a
BLAS Open
# 指定OpenBLAS標頭檔案和庫檔案路徑(填寫之前交叉編譯OpenBLAS的安裝目錄)
OpenBLAS_INCLUDE_DIR /usr/local/H8/openBLAS-0.2.20/include
OpenBLAS_LIBRARY /usr/local/H8/openBLAS-0.2.20/lib/libopenblas.so點選Configure
點選Generate生成Makefile
進入_install目錄&編譯
cd _install
make -j8
make install檢視libmxnet.so的依賴庫
arm-linux-gnueabihf-readelf -d /usr/local/H8/mxnet-20180122/lib/libmxnet.so
Dynamic section at offset 0x13fde68 contains 33 entries:
Tag Type Name/Value
0x00000001 (NEEDED) Shared library: [libopenblas.so.0]
0x00000001 (NEEDED) Shared library: [3rdparty/openmp/runtime/src/libomp.so]
0x00000001 (NEEDED) Shared library: [libpthread.so.0]
0x00000001 (NEEDED) Shared library: [libdl.so.2]
0x00000001 (NEEDED) Shared library: [libstdc++.so.6]
0x00000001 (NEEDED) Shared library: [libm.so.6]
0x00000001 (NEEDED) Shared library: [libgomp.so.1]
0x00000001 (NEEDED) Shared library: [libgcc_s.so.1]
0x00000001 (NEEDED) Shared library: [libc.so.6]
0x00000001 (NEEDED) Shared library: [ld-linux-armhf.so.3]從上面的依賴庫可以看到3rdparty/openmp/runtime/src/libomp.so這個相對路徑，這樣並不利於最終的部署測試
可以嘗試改動Cmake生成的Makefile來解決
修改_install/CMakeFiles/mxnet.dir/link.txt
將3rdparty/openmp/runtime/src/libomp.so
修改為-L/usr/local/H8/mxnet-20180122/lib/ -lomp
# /usr/local/H8/mxnet-20180122/lib/ 即是之前的安裝目錄重新進行一次編譯&安裝
make -j8
make install