Dockerfile to create Jetson TX1 and TX2 compatible builds (apache#9837)

Builds are ARM64 with CUDA 8 support. Dependencies based on Jetpack 3.1.

Author: KellenSunderland

docker_multiarch/Dockerfile.build.jetson

@@ -0,0 +1,92 @@
+# -*- mode: dockerfile -*-
+# dockerfile to build libmxnet.so, and a python wheel for the Jetson TX1/TX2
+
+FROM nvidia/cuda:8.0-cudnn5-devel as cudabuilder
+
+FROM dockcross/linux-arm64
+
+ENV ARCH aarch64
+ENV NVCCFLAGS "-m64"
+ENV CUDA_ARCH "-gencode arch=compute_53,code=sm_53 -gencode arch=compute_62,code=sm_62"
+ENV BUILD_OPTS "USE_OPENCV=0 USE_BLAS=openblas USE_SSE=0 USE_CUDA=1 USE_CUDNN=1 ENABLE_CUDA_RTC=0 USE_NCCL=0 USE_CUDA_PATH=/usr/local/cuda/"
+ENV CC /usr/bin/aarch64-linux-gnu-gcc
+ENV CXX /usr/bin/aarch64-linux-gnu-g++
+ENV FC /usr/bin/aarch64-linux-gnu-gfortran-4.9
+ENV HOSTCC gcc
+
+WORKDIR /work
+
+# Build OpenBLAS
+ADD https://api.github.com/repos/xianyi/OpenBLAS/git/refs/heads/master /tmp/openblas_version.json
+RUN git clone https://github.com/xianyi/OpenBLAS.git && \
+    cd OpenBLAS && \
+    make -j$(nproc) TARGET=ARMV8 && \
+    make install && \
+    ln -s /opt/OpenBLAS/lib/libopenblas.so /usr/lib/libopenblas.so && \
+    ln -s /opt/OpenBLAS/lib/libopenblas.a /usr/lib/libopenblas.a && \
+    ln -s /opt/OpenBLAS/lib/libopenblas.a /usr/lib/liblapack.a
+
+ENV LD_LIBRARY_PATH $LD_LIBRARY_PATH:/opt/OpenBLAS/lib
+ENV CPLUS_INCLUDE_PATH /opt/OpenBLAS/include
+
+# Setup CUDA build env (including configuring and copying nvcc)
+COPY --from=cudabuilder /usr/local/cuda /usr/local/cuda
+ENV PATH $PATH:/usr/local/cuda/bin
+ENV TARGET_ARCH aarch64
+ENV TARGET_OS linux
+
+# Install ARM depedencies based on Jetpack 3.1
+RUN wget http://developer.download.nvidia.com/devzone/devcenter/mobile/jetpack_l4t/013/linux-x64/cuda-repo-l4t-8-0-local_8.0.84-1_arm64.deb && \
+    wget http://developer.download.nvidia.com/devzone/devcenter/mobile/jetpack_l4t/013/linux-x64/libcudnn6_6.0.21-1+cuda8.0_arm64.deb && \
+    dpkg -i cuda-repo-l4t-8-0-local_8.0.84-1_arm64.deb && \
+    dpkg -i libcudnn6_6.0.21-1+cuda8.0_arm64.deb && \
+    apt update -y && \
+    apt install cuda-cudart-cross-aarch64-8-0 cuda-cublas-cross-aarch64-8-0 \
+    cuda-nvml-cross-aarch64-8-0 cuda-nvrtc-cross-aarch64-8-0 cuda-cufft-cross-aarch64-8-0 \
+    cuda-curand-cross-aarch64-8-0 cuda-cusolver-cross-aarch64-8-0 cuda-cusparse-cross-aarch64-8-0 \
+    cuda-misc-headers-cross-aarch64-8-0 cuda-npp-cross-aarch64-8-0 libcudnn6 -y && \
+    cp /usr/local/cuda-8.0/targets/aarch64-linux/lib/*.so /usr/local/cuda/lib64/ && \
+    cp /usr/local/cuda-8.0/targets/aarch64-linux/lib/stubs/*.so /usr/local/cuda/lib64/stubs/ && \
+    cp -r /usr/local/cuda-8.0/targets/aarch64-linux/include/ /usr/local/cuda/include/ && \
+    rm cuda-repo-l4t-8-0-local_8.0.84-1_arm64.deb && rm libcudnn6_6.0.21-1+cuda8.0_arm64.deb
+
+# Build MXNet
+ADD mxnet mxnet
+
+WORKDIR /work/mxnet
+
+# Add ARM specific settings
+ADD arm.crosscompile.mk make/config.mk
+
+# Build and link
+RUN make -j$(nproc) $BUILD_OPTS
+
+# Create a binary wheel for easy installation.
+# When using tool.py output will be in the jetson folder.
+# Scp the .whl file to your target device, and install via
+# pip install
+WORKDIR /work/mxnet/python
+RUN python setup.py  bdist_wheel --universal
+
+# Copy build artifacts to output folder for tool.py script
+RUN mkdir -p /work/build & cp dist/*.whl /work/build && cp ../lib/* /work/build
+
+# Fix pathing issues in the wheel.  We need to move libmxnet.so from the data folder to the root
+# of the wheel, then repackage the wheel.
+# Create a temp dir to do the work.
+WORKDIR /work/build
+RUN apt-get install -y unzip && \
+    mkdir temp && \
+    cp *.whl temp
+
+# Extract the wheel, move the libmxnet.so file, repackage the wheel.
+WORKDIR /work/build/temp
+RUN unzip *.whl &&  \
+    rm *.whl && \
+    mv *.data/data/mxnet/libmxnet.so mxnet && \
+    zip -r ../temp.zip *
+
+# Replace the existing wheel with our fixed version.
+WORKDIR /work/build
+RUN rm -rf temp && \
+    for f in *.whl; do rm "$f" && mv temp.zip "$f"; done