libtorch cross compile on aarch64-linux-gnu-gcc include torchvision

Refer to:
https://github.com/pytorch/pytorch#get-the-pytorch-source
https://github.com/pytorch/pytorch/blob/master/docs/libtorch.rst
https://discuss.pytorch.org/t/compile-libtorch-c-api-from-source/81624/2
https://qiita.com/syoyo/items/360bcf15b9ee88eeeacd
https://github.com/ljk53/pytorch-android-cpp-demo
https://github.com/dpilger26/NumCpp/tree/master
https://github.com/t-kuha/mpsoc-library/wiki/How-to-build-libtorch-&-torchvision          
https://github.com/t-kuha/mpsoc-library/wiki/How-to-build-Caffe-%26-libtorch-dependencies 
https://fatalfeel.blogspot.com/2013/12/build-nrf9160dk-projects-with-zephyr-os.html      
https://fatalfeel.blogspot.com/2019/12/ai-with-cuda-install-step-and-yolov3-in.html

#install gcc compiler
sudo apt install build-essential

#install python3 version >= 3.7 need here are a lot of works.
https://fatalfeel.blogspot.com/2019/12/ai-with-cuda-install-step-and-yolov3-in.html

#install cmake 3.20.5
wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | gpg --dearmor - | sudo tee /etc/apt/trusted.gpg.d/kitware.gpg >/dev/null
sudo apt-add-repository 'deb https://apt.kitware.com/ubuntu/ xenial main'
sudo apt update
apt-key adv --keyserver keyserver.ubuntu.com --recv-keys DE19EB17684BA42D
sudo apt install cmake cmake-qt-gui cmake-curses-gui

sudo apt install libjpeg8-dev libpng12-dev libgtk2.0-dev libxext-dev libopenblas-dev liblapack-dev
sudo apt install libavcodec-dev libavutil-dev libavformat-dev libswscale-dev libavdevice-dev
sudo apt install libv4l-dev zlib1g-dev

sudo apt install ninja-build

===================================================
==================Ubuntu16.04 x86_64=================
===================================================
///////////////////////opencv
CPU_NUM=$(grep -c processor /proc/cpuinfo)
echo "CPU number = "$CPU_NUM
export PATH=$PATH:/usr/local/cuda/bin

1.
https://github.com/opencv/opencv_contrib/tags
dwonload
opencv_contrib-4.6.0.zip
unzip ./opencv_contrib-4.6.0.zip

https://opencv.org/releases
download opencv-4.6.0.zip
unzip ./opencv-4.6.0.zip

~~if no avx instruction cpu
-DCPU_DISPATCH="SSE4_1;SSE4_2;FP16" \
~~if have avx2 instruction cpu
-DCPU_DISPATCH="SSE4_1;SSE4_2;AVX;FP16;AVX2" \
~~if have avx-512 instruction cpu
-DCPU_DISPATCH="SSE4_1;SSE4_2;AVX;FP16;AVX2;AVX512_SKX" \

~~~~~~~~opencv debug
cd opencv-4.6.0
mkdir build_debug
cd build_debug

cmake .. -G"Unix Makefiles" \
-DCMAKE_INSTALL_PREFIX=/opt/opencv/debug \
-DCMAKE_BUILD_TYPE=Debug \
-DCMAKE_C_FLAGS_DEBUG="-O0 -g" \
-DCMAKE_CXX_FLAGS_DEBUG="-O0 -g" \
-DCUDA_NVCC_FLAGS="-O0 -Wno-deprecated-gpu-targets" \
-DBUILD_PERF_TESTS=OFF \
-DBUILD_TESTS=OFF \
-DBUILD_JPEG=ON \
-DBUILD_PNG=ON \
-DBUILD_ZLIB=ON \
-DWITH_LIBV4L=ON \
-DBUILD_opencv_python3=OFF \
-DBUILD_opencv_python_tests=OFF \
-DBUILD_EXAMPLES=ON \
-DWITH_GSTREAMER=OFF \
-DWITH_TESSERACT=OFF \
-DOPENCV_EXTRA_MODULES_PATH=../../opencv_contrib-4.6.0/modules \
-DCPU_DISPATCH="SSE4_1;SSE4_2;FP16" \
-DWITH_CUDA=OFF

cmake --build . --config Debug --target install -- -j$CPU_NUM VERBOSE=1

~~~~~~~~opencv release
cd opencv-4.6.0
mkdir build_release
cd build_release

cmake .. -G"Unix Makefiles" \
-DCMAKE_INSTALL_PREFIX=/opt/opencv \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_C_FLAGS_RELEASE="-O3 -g" \
-DCMAKE_CXX_FLAGS_RELEASE="-O3 -g" \
-DCUDA_NVCC_FLAGS="-O3 -Wno-deprecated-gpu-targets" \
-DBUILD_PERF_TESTS=OFF \
-DBUILD_TESTS=OFF \
-DBUILD_JPEG=ON \
-DBUILD_PNG=ON \
-DBUILD_ZLIB=ON \
-DWITH_LIBV4L=ON \
-DBUILD_opencv_python3=OFF \
-DBUILD_opencv_python_tests=OFF \
-DBUILD_EXAMPLES=ON \
-DWITH_GSTREAMER=OFF \
-DWITH_TESSERACT=OFF \
-DOPENCV_EXTRA_MODULES_PATH=../../opencv_contrib-4.6.0/modules \
-DCPU_DISPATCH="SSE4_1;SSE4_2;FP16" \
-DWITH_CUDA=ON

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

///////////////////////libtorch
CPU_NUM=$(grep -c processor /proc/cpuinfo)
echo "CPU number = "$CPU_NUM
export PATH=$PATH:/usr/local/cuda/bin

1.
git clone --recursive https://github.com/pytorch/pytorch
cd pytorch
git checkout v1.7.1
git submodule sync
git submodule update --init --recursive

export LIBTORCH_ROOT=$(pwd)

2. 
gedit ./CMakeLists.txt
~change
###string(APPEND CMAKE_CXX_FLAGS " -O2 -fPIC")
~to
string(APPEND CMAKE_CXX_FLAGS " -fPIC")

3.
gedit ./cmake/TorchConfig.cmake.in
~comment
###if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
###  set(TORCH_CXX_FLAGS "-D_GLIBCXX_USE_CXX11_ABI=@GLIBCXX_USE_CXX11_ABI@")
###endif()

4.
gedit ./caffe2/utils/math_gpu.cu
~add header host_vector.h
#include <thrust/host_vector.h> //---> add here
#include <thrust/device_vector.h>
#include <thrust/functional.h>

5.
~~if no avx instruction cpu
-DC_AVX_FOUND=OFF \
-DC_AVX2_FOUND=OFF \
-DCXX_AVX_FOUND=OFF \
-DCXX_AVX2_FOUND=OFF \
-DDISABLE_AVX2=ON \
-DDISABLE_AVX512F=ON \
-DC_AVX_FLAGS="" \
-DC_AVX2_FLAGS="" \
-DCXX_AVX_FLAGS="" \
-DCXX_AVX2_FLAGS="" \

~~if have avx-512 instruction cpu and gcc version >= 5.0
~debug version
-DCMAKE_C_FLAGS_DEBUG="-O0 -g -mavx512f" \
-DCMAKE_CXX_FLAGS_DEBUG="-O0 -g -mavx512f" \
~release version
-DCMAKE_C_FLAGS_RELEASE="-O3 -g -mavx512f" \
-DCMAKE_CXX_FLAGS_RELEASE="-O3 -g -mavx512f" \

~~if set -DUSE_FBGEMM=ON
FBGEMM offers optimized on-CPU performance for reduced precision calculations used to accelerate deep learning models. It’s delivered greater than 2x performance gains.

~~~~~~~~libtorch debug
mkdir build_debug
cd build_debug

cmake .. -G"Unix Makefiles" \
-DCMAKE_INSTALL_PREFIX=/opt/libtorch/x64/debug \
-DCMAKE_BUILD_TYPE=Debug \
-DCMAKE_C_FLAGS_DEBUG="-O0 -g" \
-DCMAKE_CXX_FLAGS_DEBUG="-O0 -g" \
-DPYTHON_EXECUTABLE=$(which python3) \
-DBUILDING_WITH_TORCH_LIBS=ON \
-DUSE_OPENCV=ON \
-DOpenCV_DIR=/opt/opencv/debug/lib/cmake/opencv4 \
-DBUILD_EXAMPLES=ON \
-DC_AVX_FOUND=OFF \
-DC_AVX2_FOUND=OFF \
-DCXX_AVX_FOUND=OFF \
-DCXX_AVX2_FOUND=OFF \
-DDISABLE_AVX2=ON \
-DDISABLE_AVX512F=ON \
-DC_AVX_FLAGS="" \
-DC_AVX2_FLAGS="" \
-DCXX_AVX_FLAGS="" \
-DCXX_AVX2_FLAGS="" \
-DUSE_FBGEMM=OFF \
-DBUILD_PYTHON=OFF \
-DBUILD_ONNX_PYTHON=OFF \
-DUSE_CUDA=OFF

cmake --build . --config Debug --target install -- -j$CPU_NUM VERBOSE=1

~~~~~~~~libtorch release
mkdir build_release
cd build_release

cmake .. -G"Unix Makefiles" \
-DCMAKE_INSTALL_PREFIX=/opt/libtorch/x64 \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_C_FLAGS_RELEASE="-O3 -g" \
-DCMAKE_CXX_FLAGS_RELEASE="-O3 -g" \
-DPYTHON_EXECUTABLE=$(which python3) \
-DBUILDING_WITH_TORCH_LIBS=ON \
-DUSE_OPENCV=ON \
-DOpenCV_DIR=/opt/opencv/lib/cmake/opencv4 \
-DBUILD_EXAMPLES=ON \
-DC_AVX_FOUND=OFF \
-DC_AVX2_FOUND=OFF \
-DCXX_AVX_FOUND=OFF \
-DCXX_AVX2_FOUND=OFF \
-DDISABLE_AVX2=ON \
-DDISABLE_AVX512F=ON \
-DC_AVX_FLAGS="" \
-DC_AVX2_FLAGS="" \
-DCXX_AVX_FLAGS="" \
-DCXX_AVX2_FLAGS="" \
-DUSE_FBGEMM=OFF \
-DBUILD_PYTHON=OFF \
-DBUILD_ONNX_PYTHON=OFF \
-DUSE_CUDA=ON

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

///////////////////////torchvision
~~if you have not set the environment variables yet, set them now~~
CPU_NUM=$(grep -c processor /proc/cpuinfo)
echo "CPU number = "$CPU_NUM
export PATH=$PATH:/usr/local/cuda/bin

1.
https://github.com/pytorch/vision/releases
download vision-0.8.2.tar.gz
tar -xvf ./vision-0.8.2.tar.gz
cd vision-0.8.2

~modify these files add definition USE_PYTHON for cmake, header, source
./CMakeLists.txt
./cmake/TorchVisionConfig.cmake.in
./torchvision/csrc/cpu/image/image.cpp
./torchvision/csrc/cpu/video/Video.h
./torchvision/csrc/cpu/video_reader/VideoReader.cpp
./torchvision/csrc/cpu/vision_cpu.h
./torchvision/csrc/cuda/vision_cuda.h
./torchvision/csrc/empty_tensor_op.h
./torchvision/csrc/vision.cpp
patch: https://www.mediafire.com/file/wcxce2v6vdzvtvw/vision-0.8.2_patch.tar.gz

2.
~~~~~~~~~~~~~~vision debug
mkdir build_debug
cd build_debug

cmake .. -G"Unix Makefiles" \
-DCMAKE_INSTALL_PREFIX=/opt/torchvision/x64/debug \
-DCMAKE_BUILD_TYPE=Debug \
-DCMAKE_C_FLAGS_DEBUG="-O0 -g" \
-DCMAKE_CXX_FLAGS_DEBUG="-O0 -g" \
-DTorch_DIR=/opt/libtorch/x64/debug/share/cmake/Torch \
-DUSE_PYTHON=OFF \
-DWITH_CUDA=OFF

~~~TorchConfig.cmake.in modified already, so we can skip sed -i
###sed -i 's/-D_GLIBCXX_USE_CXX11_ABI=//g' ./CMakeFiles/torchvision.dir/flags.make

cmake --build . --config Debug --target install -- -j$CPU_NUM VERBOSE=1

~~~~~~~~~~~~~~vision release
mkdir build_release
cd build_release

cmake .. -G"Unix Makefiles" \
-DCMAKE_INSTALL_PREFIX=/opt/torchvision/x64 \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_C_FLAGS_RELEASE="-O3 -g" \
-DCMAKE_CXX_FLAGS_RELEASE="-O3 -g" \
-DTorch_DIR=/opt/libtorch/x64/share/cmake/Torch \
-DUSE_PYTHON=OFF \
-DWITH_CUDA=ON

~~~TorchConfig.cmake.in modified already, so we can skip sed -i
###sed -i 's/-D_GLIBCXX_USE_CXX11_ABI=//g' ./CMakeFiles/torchvision.dir/flags.make

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

==========================================
==================Android==================
==========================================
https://developer.android.com/ndk/downloads/older_releases
download and extract Android NDK 19c to /opt

#python link to python3
ln -sf /usr/bin/python3.7 /usr/bin/python

pip3 install pyyaml

cd pytorch
#release version
gedit ./scripts/build_android.sh
-DCMAKE_BUILD_TYPE=Release
#change to
-DCMAKE_BUILD_TYPE=RelWithDebInfo

#debug version
gedit ./scripts/build_android.sh
-DCMAKE_BUILD_TYPE=Release
#change to
-DCMAKE_BUILD_TYPE=Debug

export ANDROID_NDK=/opt/android-ndk
export ANDROID_ABI=arm64-v8a
export ANDROID_NATIVE_API_LEVEL=28
export ANDROID_DEBUG_SYMBOLS=1
export VERBOSE=1
./scripts/build_android.sh

==================================================================
=========================aarch64-linux-gnu-gcc========================
==================================================================
//////////////////Method A: download prebuilt libs
1.
CPU_NUM=$(grep -c processor /proc/cpuinfo)
echo "CPU number = "$CPU_NUM
export PATH=$PATH:/opt/ivot/aarch64-ca53-linux-gnueabihf-8.4/bin
export LD_LIBRARY_PATH=/opt/ivot/aarch64-ca53-linux-gnueabihf-8.4/lib

2.
gedit ./toolchain.make
set( CMAKE_SYSTEM_NAME Linux )
set( CMAKE_SYSTEM_PROCESSOR aarch64 )
set( CMAKE_C_COMPILER /opt/ivot/aarch64-ca53-linux-gnueabihf-8.4/bin/aarch64-ca53-linux-gnu-gcc )
set( CMAKE_CXX_COMPILER /opt/ivot/aarch64-ca53-linux-gnueabihf-8.4/bin/aarch64-ca53-linux-gnu-g++ )

~~~copy toolchain.make to each dependency and libtorch source directories~~~

3.
git clone --recursive https://github.com/pytorch/pytorch
cd pytorch
git checkout v1.7.1
git submodule sync
git submodule update --init --recursive

export LIBTORCH_ROOT=$(pwd)

4.
gedit ./CMakeLists.txt
~change
###string(APPEND CMAKE_CXX_FLAGS " -O2 -fPIC")
~to
string(APPEND CMAKE_CXX_FLAGS " -fPIC")

5.
gedit ./cmake/TorchConfig.cmake.in
~comment
###if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
###  set(TORCH_CXX_FLAGS "-D_GLIBCXX_USE_CXX11_ABI=@GLIBCXX_USE_CXX11_ABI@")
###endif()

6.
./scripts/build_host_protoc.sh

7.
cd ./third_party/sleef
mkdir _host
cd _host

cmake .. -G"Unix Makefiles" -DCMAKE_INSTALL_PREFIX=_install -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTS=OFF

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

8. (if use manual make libs, skip this)
cd ../..
git clone -b 2019.2 https://github.com/t-kuha/mpsoc-library.git
export DEPEND_DIR=$(pwd)/mpsoc-library

9. (option can skip, if cmake error)
when
CMake Error at cmake/Codegen.cmake:163 (message):

cd $LIBTORCH_ROOT
gedit ./cmake/Codegen.cmake
set(GEN_COMMAND
  "${PYTHON_EXECUTABLE}" ${CMAKE_CURRENT_LIST_DIR}/../aten/src/ATen/gen.py
#to
find_package (Python COMPONENTS Interpreter)
set(GEN_COMMAND
  "${Python_EXECUTABLE}" ${CMAKE_CURRENT_LIST_DIR}/../aten/src/ATen/gen.py

10.
cd $LIBTORCH_ROOT

mkdir _mpsoc
cd _mpsoc

cmake ${LIBTORCH_ROOT} -G"Unix Makefiles" \
-DCMAKE_INSTALL_PREFIX=/opt/libtorch/arm64 \
-DCMAKE_BUILD_TYPE=Release \
-DCAFFE2_CUSTOM_PROTOC_EXECUTABLE=${LIBTORCH_ROOT}/build_host_protoc/bin/protoc \
-DCMAKE_TOOLCHAIN_FILE=toolchain.make \
-DCMAKE_PREFIX_PATH=${DEPEND_DIR}/dl-framework/caffe-dependency \
-DPYTHON_EXECUTABLE=$(which python3) \
-C${DEPEND_DIR}/dl-framework/libtorch/TryRunResults.cmake \
-DBUILDING_WITH_TORCH_LIBS=ON \
-DBUILD_BINARY=ON \
-DBUILD_CAFFE2_MOBILE=ON \
-DBUILD_CAFFE2_OPS=ON \
-DBUILD_CUSTOM_PROTOBUF=ON \
-DBUILD_DFT=OFF \
-DBUILD_DOCS=OFF \
-DBUILD_GMOCK=ON \
-DBUILD_GNUABI_LIBS=OFF \
-DBUILD_LIBM=ON \
-DBUILD_ONNX_PYTHON=OFF \
-DBUILD_PYTHON=OFF \
-DBUILD_QUAD=OFF \
-DBUILD_SHARED_LIBS=ON \
-DBUILD_TEST=ON \
-DBUILD_TESTS=ON \
-DUSE_ASAN=OFF \
-DUSE_CUDA=OFF \
-DUSE_DISTRIBUTED=OFF \
-DUSE_FBGEMM=OFF \
-DUSE_FFMPEG=ON \
-DUSE_GFLAGS=ON \
-DUSE_GLOG=ON \
-DUSE_GLOO=OFF \
-DUSE_LEVELDB=ON \
-DUSE_LITE_PROTO=OFF \
-DUSE_LMDB=ON \
-DUSE_METAL=OFF \
-DUSE_MKLDNN=OFF \
-DUSE_MPI=OFF \
-DUSE_NATIVE_ARCH=OFF \
-DUSE_NNAPI=OFF \
-DUSE_NNPACK=ON \
-DUSE_NUMA=OFF \
-DUSE_NUMPY=ON \
-DUSE_OBSERVERS=OFF \
-DUSE_OPENCL=OFF \
-DUSE_OPENCV=ON \
-DUSE_OPENMP=OFF \
-DUSE_PROF=OFF \
-DUSE_PYTORCH_QNNPACK=OFF \
-DUSE_QNNPACK=ON \
-DUSE_REDIS=OFF \
-DUSE_ROCKSDB=ON \
-DUSE_ROCM=OFF \
-DUSE_SNPE=OFF \
-DUSE_SYSTEM_EIGEN_INSTALL=OFF \
-DUSE_TBB=OFF \
-DUSE_TENSORRT=OFF \
-DUSE_ZMQ=ON \
-DUSE_ZSTD=OFF \
-DHAVE_STD_REGEX=0 \
-DHAVE_POSIX_REGEX=0 \
-DHAVE_STEADY_CLOCK=0 \
-DATEN_THREADING=NATIVE \
-DBLAS=OpenBLAS \
-DCMAKE_CROSSCOMPILING=ON \
-DNATIVE_BUILD_DIR=${LIBTORCH_ROOT}/third_party/sleef/_host/ \
-DCMAKE_CXX_FLAGS="-L${DEPEND_DIR}/dl-framework/caffe-dependency/lib -llmdb -lleveldb -lsnappy -lopencv_core -lopencv_highgui -lopencv_imgcodecs -lopencv_imgproc -lopencv_video -lopencv_videoio -lzmq -lrocksdb"

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

~~~note~~~
search python3: -DPYTHON_EXECUTABLE=$(which python3)
arm gcc have no -fopenmp: -DATEN_THREADING=NATIVE

//////////////////Method B: build manual libs
1.
CPU_NUM=$(grep -c processor /proc/cpuinfo)
echo "CPU number = "$CPU_NUM
export PATH=$PATH:/opt/ivot/aarch64-ca53-linux-gnueabihf-8.4/bin
export LD_LIBRARY_PATH=/opt/ivot/aarch64-ca53-linux-gnueabihf-8.4/lib

2.
gedit ./toolchain.make
set( CMAKE_SYSTEM_NAME Linux )
set( CMAKE_SYSTEM_PROCESSOR aarch64 )
set( CMAKE_C_COMPILER /opt/ivot/aarch64-ca53-linux-gnueabihf-8.4/bin/aarch64-ca53-linux-gnu-gcc )
set( CMAKE_CXX_COMPILER /opt/ivot/aarch64-ca53-linux-gnueabihf-8.4/bin/aarch64-ca53-linux-gnu-g++ )

~~~copy toolchain.make to each dependency and libtorch source directories~~~

3.
wget --continue https://raw.githubusercontent.com/t-kuha/mpsoc-library/2019.2/dl-framework/libtorch/TryRunResults.cmake
or
download https://www.mediafire.com/file/6ospje4cqqrswvq/TryRunResults.cmake
mkdir -p /opt/torch_deps/arm64
cp ./TryRunResults.cmake /opt/torch_deps/arm64

4.
https://www.boost.org/users/history/version_1_68_0.html
download bboost_1_68_0.tar.gz
tar -xvf ./boost_1_68_0.tar.gz
cd boost_1_68_0
./bootstrap.sh
gedit ./project-config.jam
using gcc ;
~change to
using gcc : arm64 : aarch64-ca53-linux-gnu-g++ ;

using python : 2.7 : /usr ;
~change to
using python : 3.7 : /usr ;

./bjam install toolset=gcc-arm64 variant=release link=shared threading=multi runtime-link=shared --prefix=/opt/torch_deps/arm64 --with-thread --with-system --with-filesystem -j$CPU_NUM

5.
https://github.com/gflags/gflags/releases
download gflags-2.2.2.tar.gz
tar -xvf ./gflags-2.2.2.tar.gz
cd gflags-2.2.2
mkdir _mpsoc
cd _mpsoc

cmake .. -G"Unix Makefiles" -DCMAKE_TOOLCHAIN_FILE=toolchain.make -DCMAKE_INSTALL_PREFIX=/opt/torch_deps/arm64 -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

6.
https://github.com/google/glog/releases
download glog-0.4.0.tar.gz
tar -xvf ./glog-0.4.0.tar.gz
cd glog-0.4.0
mkdir _mpsoc
cd _mpsoc

cmake .. -G"Unix Makefiles" -DCMAKE_TOOLCHAIN_FILE=toolchain.make -DCMAKE_INSTALL_PREFIX=/opt/torch_deps/arm64 -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -DBUILD_TESTING=OFF

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

7.
https://github.com/google/leveldb/releases
download leveldb-1.22.tar.gz
tar -xvf ./leveldb-1.22.tar.gz
cd leveldb-1.22
mkdir _mpsoc
cd _mpsoc

cmake .. -G"Unix Makefiles" -DCMAKE_TOOLCHAIN_FILE=toolchain.make -DCMAKE_INSTALL_PREFIX=/opt/torch_deps/arm64 -DCMAKE_BUILD_TYPE=Release -DLEVELDB_BUILD_TESTS=OFF -DLEVELDB_BUILD_BENCHMARKS=OFF -DBUILD_SHARED_LIBS=ON

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

8.
https://git.openldap.org/openldap/openldap/-/tags
download openldap-LMDB_0.9.24.tar.gz
tar -xvf ./openldap-LMDB_0.9.24.tar.gz
cd openldap-LMDB_0.9.24/libraries/liblmdb

gedit ./Makefile
CC      = gcc
AR      = ar
OPT     = -O2 -g
prefix  = /usr/local
#change to
CC      = aarch64-ca53-linux-gnu-gcc
AR      = aarch64-ca53-linux-gnu-ar
OPT     = -O3
prefix  = /opt/torch_deps/arm64
#save

make -j$CPU_NUM install

9.
https://github.com/xianyi/OpenBLAS/releases?page=2
download OpenBLAS-0.3.7.tar.gz
tar -xvf ./OpenBLAS-0.3.7.tar.gz
cd OpenBLAS-0.3.7
mkdir _mpsoc
cd _mpsoc
cmake .. -G"Unix Makefiles" \
-DCMAKE_TOOLCHAIN_FILE=toolchain.make \
-DCMAKE_INSTALL_PREFIX=/opt/torch_deps/arm64 \
-DCMAKE_BUILD_TYPE=Release \
-DNOFORTRAN=1 \
-DBUILD_SHARED_LIBS=1

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

10.
Need arm64 opencv only. To avoid FindCUDA.cmake point to x64 opencv in libtorch:
Could NOT find CUDA (missing: CUDA_TOOLKIT_ROOT_DIR CUDA_NVCC_EXECUTABLE
Please remove all the built targets of x64 opencv that in /opt or any other directory

https://opencv.org/releases/page/3/
download opencv-3.4.8.zip
unzip ./opencv-3.4.8.zip
cd opencv-3.4.8
mkdir _mpsoc
cd _mpsoc

cmake .. -G"Unix Makefiles" \
-DCMAKE_TOOLCHAIN_FILE=toolchain.make \
-DCMAKE_INSTALL_PREFIX=/opt/torch_deps/arm64 \
-DCMAKE_BUILD_TYPE=Release \
-DBUILD_PERF_TESTS=OFF \
-DBUILD_TESTS=OFF \
-DBUILD_ZLIB=ON \
-DWITH_CUDA=OFF \
-DZLIB_INCLUDE_DIR=../3rdparty/zlib

~~~~enable these modules, disable unuse BUILD_opencv_xxxx~~~
cmake-gui ./CMakeCache.txt
BUILD_opencv_core
BUILD_opencv_video
BUILD_opencv_videoio
BUILD_opencv_highgui
BUILD_opencv_imgproc
BUILD_opencv_imgcodecs
#press Configure then Generate

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

11.
https://github.com/protocolbuffers/protobuf/releases?page=6
download protobuf-3.10.1.tar.gz
tar -xvf ./protobuf-3.10.1.tar.gz

***copy toolchain.make to ./protobuf-3.10.1/cmake***

cd protobuf-3.10.1
mkdir _host
cd _host

cmake ../cmake/ -G"Unix Makefiles" -DCMAKE_INSTALL_PREFIX=/opt/torch_deps/x64/host -DCMAKE_BUILD_TYPE=Release -Dprotobuf_BUILD_TESTS=OFF

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

export PATH=${PATH}:/opt/torch_deps/x64/host/bin

cd ..
mkdir _mpsoc
cd _mpsoc

cmake ../cmake/ -G"Unix Makefiles" -DCMAKE_TOOLCHAIN_FILE=toolchain.make -DCMAKE_INSTALL_PREFIX=/opt/torch_deps/arm64 -DCMAKE_BUILD_TYPE=Release -Dprotobuf_BUILD_TESTS=OFF -Dprotobuf_BUILD_SHARED_LIBS=ON

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

12.
https://github.com/google/snappy/releases
download snappy-1.1.7.tar.gz
tar -xvf ./snappy-1.1.7.tar.gz
cd snappy-1.1.7
mkdir _mpsoc
cd _mpsoc

cmake .. -G"Unix Makefiles" -DCMAKE_TOOLCHAIN_FILE=toolchain.make -DCMAKE_INSTALL_PREFIX=/opt/torch_deps/arm64 -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -DSNAPPY_BUILD_TESTS=OFF

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

13.
https://github.com/facebook/rocksdb/releases?page=4
download rocksdb-6.4.6.tar.gz
tar -xvf ./rocksdb-6.4.6.tar.gz
cd rocksdb-6.4.6
mkdir _mpsoc
cd _mpsoc

cmake .. -G"Ninja" -DCMAKE_TOOLCHAIN_FILE=toolchain.make -DCMAKE_INSTALL_PREFIX=/opt/torch_deps/arm64 -DCMAKE_BUILD_TYPE=Release -DWITH_TESTS=OFF -DWITH_TOOLS=OFF -DWITH_GFLAGS=OFF -DPORTABLE=ON

cmake --build . --config Release --target install -- -j$CPU_NUM
~~~can not use VERBOSE=1~~~

14.
https://github.com/zeromq/libzmq/releases
download libzmq-4.3.4.tar.gz
tar -xvf ./libzmq-4.3.4.tar.gz
cd libzmq-4.3.4
mkdir _mpsoc
cd _mpsoc

cmake .. -G"Unix Makefiles" -DCMAKE_TOOLCHAIN_FILE=toolchain.make -DCMAKE_INSTALL_PREFIX=/opt/torch_deps/arm64 -DCMAKE_BUILD_TYPE=Release -DBUILD_STATIC=OFF -DBUILD_TESTS=OFF -DWITH_DOCS=OFF -DWITH_PERF_TOOL=OFF

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

15.
first do step 1.2.3.4.5.6.7. of [Method A: download prebuilt libs]

CPU_NUM=$(grep -c processor /proc/cpuinfo)
echo "CPU number = "$CPU_NUM
export PATH=$PATH:/opt/ivot/aarch64-ca53-linux-gnueabihf-8.4/bin
export LD_LIBRARY_PATH=/opt/ivot/aarch64-ca53-linux-gnueabihf-8.4/lib

cd pytorch
export LIBTORCH_ROOT=$(pwd)

gedit ./CMakeLists.txt
~change
###string(APPEND CMAKE_CXX_FLAGS " -O2 -fPIC")
~to
string(APPEND CMAKE_CXX_FLAGS " -fPIC")

gedit ./cmake/TorchConfig.cmake.in
~comment
###if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
###  set(TORCH_CXX_FLAGS "-D_GLIBCXX_USE_CXX11_ABI=@GLIBCXX_USE_CXX11_ABI@")
###endif()

./scripts/build_host_protoc.sh

cd ./third_party/sleef
mkdir _host
cd _host

cmake .. -G"Unix Makefiles" -DCMAKE_INSTALL_PREFIX=_install -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTS=OFF

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

cd $LIBTORCH_ROOT
mkdir _mpsoc
cd _mpsoc

cmake ${LIBTORCH_ROOT} -G"Unix Makefiles" \
-DCMAKE_INSTALL_PREFIX=/opt/libtorch/arm64 \
-DCMAKE_BUILD_TYPE=Release \
-DCAFFE2_CUSTOM_PROTOC_EXECUTABLE=${LIBTORCH_ROOT}/build_host_protoc/bin/protoc \
-DCMAKE_TOOLCHAIN_FILE=toolchain.make \
-DCMAKE_PREFIX_PATH=/opt/torch_deps/arm64 \
-DPYTHON_EXECUTABLE=$(which python3) \
-C/opt/torch_deps/arm64/TryRunResults.cmake \
-DBUILDING_WITH_TORCH_LIBS=ON \
-DBUILD_BINARY=ON \
-DBUILD_CAFFE2_MOBILE=ON \
-DBUILD_CAFFE2_OPS=ON \
-DBUILD_CUSTOM_PROTOBUF=ON \
-DBUILD_DFT=OFF \
-DBUILD_DOCS=OFF \
-DBUILD_GMOCK=ON \
-DBUILD_GNUABI_LIBS=OFF \
-DBUILD_LIBM=ON \
-DBUILD_ONNX_PYTHON=OFF \
-DBUILD_PYTHON=OFF \
-DBUILD_QUAD=OFF \
-DBUILD_SHARED_LIBS=ON \
-DBUILD_TEST=ON \
-DBUILD_TESTS=ON \
-DUSE_ASAN=OFF \
-DUSE_CUDA=OFF \
-DUSE_DISTRIBUTED=OFF \
-DUSE_FBGEMM=OFF \
-DUSE_FFMPEG=ON \
-DUSE_GFLAGS=ON \
-DUSE_GLOG=ON \
-DUSE_GLOO=OFF \
-DUSE_LEVELDB=ON \
-DUSE_LITE_PROTO=OFF \
-DUSE_LMDB=ON \
-DUSE_METAL=OFF \
-DUSE_MKLDNN=OFF \
-DUSE_MPI=OFF \
-DUSE_NATIVE_ARCH=OFF \
-DUSE_NNAPI=OFF \
-DUSE_NNPACK=ON \
-DUSE_NUMA=OFF \
-DUSE_NUMPY=ON \
-DUSE_OBSERVERS=OFF \
-DUSE_OPENCL=OFF \
-DUSE_OPENCV=ON \
-DUSE_OPENMP=OFF \
-DUSE_PROF=OFF \
-DUSE_PYTORCH_QNNPACK=OFF \
-DUSE_QNNPACK=ON \
-DUSE_REDIS=OFF \
-DUSE_ROCKSDB=ON \
-DUSE_ROCM=OFF \
-DUSE_SNPE=OFF \
-DUSE_SYSTEM_EIGEN_INSTALL=OFF \
-DUSE_TBB=OFF \
-DUSE_TENSORRT=OFF \
-DUSE_ZMQ=ON \
-DUSE_ZSTD=OFF \
-DHAVE_STD_REGEX=0 \
-DHAVE_POSIX_REGEX=0 \
-DHAVE_STEADY_CLOCK=0 \
-DATEN_THREADING=NATIVE \
-DBLAS=OpenBLAS \
-DCMAKE_CROSSCOMPILING=ON \
-DNATIVE_BUILD_DIR=${LIBTORCH_ROOT}/third_party/sleef/_host/ \
-DCMAKE_CXX_FLAGS="-L/opt/torch_deps/arm64/lib -llmdb -lleveldb -lsnappy -lopencv_core -lopencv_highgui -lopencv_imgcodecs -lopencv_imgproc -lopencv_video -lopencv_videoio -lzmq -lrocksdb"

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

~~~note~~~
search python3: -DPYTHON_EXECUTABLE=$(which python3)
arm gcc have no -fopenmp: -DATEN_THREADING=NATIVE

16.
https://github.com/pytorch/vision/releases
download vision-0.8.2.tar.gz
tar -xvf ./vision-0.8.2.tar.gz
cd vision-0.8.2

~modify these files add definition USE_PYTHON for cmake, header, source
./CMakeLists.txt
./cmake/TorchVisionConfig.cmake.in
./torchvision/csrc/cpu/image/image.cpp
./torchvision/csrc/cpu/video/Video.h
./torchvision/csrc/cpu/video_reader/VideoReader.cpp
./torchvision/csrc/cpu/vision_cpu.h
./torchvision/csrc/cuda/vision_cuda.h
./torchvision/csrc/empty_tensor_op.h
./torchvision/csrc/vision.cpp
patch: https://www.mediafire.com/file/wcxce2v6vdzvtvw/vision-0.8.2_patch.tar.gz

mkdir _mpsoc
cd _mpsoc

cmake .. -G"Unix Makefiles" \
-DCMAKE_TOOLCHAIN_FILE=toolchain.make \
-DCMAKE_INSTALL_PREFIX=/opt/torchvision/arm64 \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_PREFIX_PATH=/opt/torch_deps/arm64 \
-DTorch_DIR=/opt/libtorch/arm64/share/cmake/Torch \
-DUSE_PYTHON=OFF \
-DWITH_CUDA=OFF \
-DPNG_LIBRARY=/opt/torch_deps/arm64/lib/libopencv_imgcodecs.so \
-DJPEG_LIBRARY=/opt/torch_deps/arm64/lib/libopencv_imgcodecs.so

~~~TorchConfig.cmake.in modified already, so we can skip sed -i
###sed -i 's/-D_GLIBCXX_USE_CXX11_ABI=//g' ./CMakeFiles/torchvision.dir/flags.make

cmake --build . --config Release --target install -- -j$CPU_NUM VERBOSE=1

sample code:
https://www.mediafire.com/file/72pdqg041n2iw85/torch_test_arm.tar.gz
https://www.mediafire.com/file/18lyaq8d4qhx0ne/torch_test_x64.tar.gz

demo: (both forward & backward support for aarch64-linux-gnu-gcc)
https://www.mediafire.com/view/1zgp44v08n7tpj6/libtorch_build_ok.png
https://www.mediafire.com/view/fdwap6ldb8zl6w0/torchvision.png
https://www.mediafire.com/view/cmqha2rgexv7l3p/libtorch_nettransform.png
https://www.mediafire.com/view/gl8rlbt9e911f7p/libtorch_debug.png

///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
Tesla told in New York Herald: I prefer to be remembered as the inventor who succeeded in abolishing war. That will be my highest pride.
http://www.teslacollection.com/tesla_articles/1898/new_york_herald/f_l_christman/tesla_declares_he_will_abolish_war (in middle section)

Albert Einstein: The release of atom power has changed everything except our way of thinking... the solution to this problem lies in the heart of mankind. If only I had known, I should have become a watchmaker.
https://atomictrauma.wordpress.com/the-scientists/albert-einstein

AI training will produce carbon footprint
https://arxiv.org/pdf/1906.02243.pdf