Add android64 support and fix build.py

CMakeLists.txt

@@ -15,7 +15,7 @@ mxnet_option(USE_NCCL             "Use NVidia NCCL with CUDA" OFF)
 mxnet_option(USE_OPENCV           "Build with OpenCV support" ON)
 mxnet_option(USE_OPENMP           "Build with Openmp support" ON)
 mxnet_option(USE_CUDNN            "Build with cudnn support"  ON) # one could set CUDNN_ROOT for search path
-mxnet_option(USE_SSE              "Build with x86 SSE instruction support" ON)
+mxnet_option(USE_SSE              "Build with x86 SSE instruction support" ON IF NOT ARM)
 mxnet_option(USE_F16C             "Build with x86 F16C instruction support" ON) # autodetects support if ON
 mxnet_option(USE_LAPACK           "Build with lapack support" ON)
 mxnet_option(USE_MKL_IF_AVAILABLE "Use MKL if found" ON)
@@ -360,7 +360,7 @@ if(USE_OPENMP)
       set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_EXE_LINKER_FLAGS}")
     endif()
   endif()
-elseif(UNIX)
+elseif(UNIX AND NOT ANDROID)
   list(APPEND mxnet_LINKER_LIBS pthread)
 endif()
 

Jenkinsfile

@@ -458,6 +458,16 @@ try {
           }
         }
       }
+    },
+    'Android / ARM64':{
+      node('mxnetlinux-cpu') {
+        ws('workspace/android64') {
+          timeout(time: max_time, unit: 'MINUTES') {
+            init_git()
+            docker_run('android_arm64', 'build_android_arm64', false)
+          }
+        }
+      }
     }
   } // End of stage('Build')
 

ci/build.py

@@ -221,63 +221,69 @@ def script_name() -> str:
                         help="go in a shell inside the container",
                         action='store_true')
 
-    parser.add_argument("--docker-registry",
-                        help="Dockerhub registry name to retrieve cache from",
+    parser.add_argument("-d", "--docker-registry",
+                        help="Dockerhub registry name to retrieve cache from. Default is 'mxnetci'",
                         default='mxnetci',
                         type=str)
 
+    parser.add_argument("-c", "--cache", action="store_true",
+                        help="Enable docker registry cache")
+
     parser.add_argument("command",
                         help="command to run in the container",
                         nargs='*', action='append', type=str)
 
     args = parser.parse_args()
-    docker_registry = args.docker_registry
+    def use_cache():
+        return args.cache or 'JOB_NAME' in os.environ # we are in Jenkins
+
     command = list(chain(*args.command))
     docker_binary = get_docker_binary(args.nvidiadocker)
     shared_memory_size = args.shared_memory_size
 
-    print("into container: {}".format(args.into_container))
     if args.list:
         list_platforms()
     elif args.platform:
         platform = args.platform
-        tag = get_docker_tag(platform=platform, registry=docker_registry)
-        load_docker_cache(tag=tag, docker_registry=args.docker_registry)
-        build_docker(platform, docker_binary, registry=docker_registry)
+        tag = get_docker_tag(platform=platform, registry=args.docker_registry)
+        if use_cache():
+            load_docker_cache(tag=tag, docker_registry=args.docker_registry)
+        build_docker(platform, docker_binary, registry=args.docker_registry)
         if args.build_only:
             logging.warning("Container was just built. Exiting due to build-only.")
             return 0
 
         if command:
             container_run(platform=platform, docker_binary=docker_binary, shared_memory_size=shared_memory_size,
-                          command=command, docker_registry=docker_registry)
+                          command=command, docker_registry=args.docker_registry)
         elif args.print_docker_run:
             print(container_run(platform=platform, docker_binary=docker_binary, shared_memory_size=shared_memory_size,
-                                command=[], dry_run=True, docker_registry=docker_registry))
+                                command=[], dry_run=True, docker_registry=args.docker_registry))
         elif args.into_container:
             container_run(platform=platform, docker_binary=docker_binary, shared_memory_size=shared_memory_size,
-                          command=[], dry_run=False, into_container=True, docker_registry=docker_registry)
+                          command=[], dry_run=False, into_container=True, docker_registry=args.docker_registry)
         else:
             cmd = ["/work/mxnet/ci/docker/runtime_functions.sh", "build_{}".format(platform)]
             logging.info("No command specified, trying default build: %s", ' '.join(cmd))
             container_run(platform=platform, docker_binary=docker_binary, shared_memory_size=shared_memory_size,
-                          command=cmd, docker_registry=docker_registry)
+                          command=cmd, docker_registry=args.docker_registry)
 
     elif args.all:
         platforms = get_platforms()
         logging.info("Building for all architectures: {}".format(platforms))
         logging.info("Artifacts will be produced in the build/ directory.")
         for platform in platforms:
-            tag = get_docker_tag(platform=platform, registry=docker_registry)
-            load_docker_cache(tag=tag, docker_registry=args.docker_registry)
-            build_docker(platform, docker_binary)
+            tag = get_docker_tag(platform=platform, registry=args.docker_registry)
+            if use_cache():
+                load_docker_cache(tag=tag, docker_registry=args.docker_registry)
+            build_docker(platform, docker_binary, args.docker_registry)
             if args.build_only:
                 continue
             build_platform = "build_{}".format(platform)
             cmd = ["/work/mxnet/ci/docker/runtime_functions.sh", build_platform]
             shutil.rmtree(buildir(), ignore_errors=True)
             container_run(platform=platform, docker_binary=docker_binary, shared_memory_size=shared_memory_size,
-                          command=cmd, docker_registry=docker_registry)
+                          command=cmd, docker_registry=args.docker_registry)
             plat_buildir = os.path.join(get_mxnet_root(), build_platform)
             shutil.move(buildir(), plat_buildir)
             logging.info("Built files left in: %s", plat_buildir)

ci/docker/Dockerfile.build.android_arm64

@@ -51,6 +51,7 @@ LABEL org.label-schema.build-date=$BUILD_DATE \
       org.label-schema.schema-version="1.0"
 
 ENV ARCH aarch64
+ENV ANDROID_NDK_REVISION 15c
 
 ENV CC=${CROSS_ROOT}/bin/${CROSS_TRIPLE}-clang
 ENV CXX=${CROSS_ROOT}/bin/${CROSS_TRIPLE}-clang++

ci/docker/install/android_arm64_ndk.sh

@@ -22,7 +22,8 @@
 
 set -ex
 pushd .
-export ANDROID_NDK_REVISION=15c
+# This environment variable comes from the docker file
+echo "Downloading android SDK rev ${ANDROID_NDK_REVISION}"
 curl -O https://dl.google.com/android/repository/android-ndk-r${ANDROID_NDK_REVISION}-linux-x86_64.zip && \
 unzip ./android-ndk-r${ANDROID_NDK_REVISION}-linux-x86_64.zip && \
 cd android-ndk-r${ANDROID_NDK_REVISION} && \
@@ -32,4 +33,4 @@ cd android-ndk-r${ANDROID_NDK_REVISION} && \
     --api 21 \
     --install-dir=${CROSS_ROOT} && \
 
-popd
+popd

ci/docker/runtime_functions.sh

@@ -143,7 +143,7 @@ build_arm64() {
     cp dist/*.whl /work/build
 }
 
-build_android_arm64() {
+build_android_armv7() {
     set -ex
     cd /work/build
     cmake\
@@ -163,6 +163,30 @@ build_android_arm64() {
     cp dist/*.whl /work/build
 }
 
+build_android_arm64() {
+    set -ex
+    cd /work/build
+# There are other ways for CMake to cross compile android, like setting the following variables
+# below. But right not it doesn't work as expected, we need to find what's the best strategy to 
+# build with CMake in Android.
+#        -DCMAKE_ANDROID_NDK=${CROSS_ROOT} \
+#        -DCMAKE_SYSTEM_VERSION=${ANDROID_NDK_REVISION} \
+#        -DCMAKE_SYSTEM_NAME=Android \
+#
+    cmake\
+        -DANDROID=ON \
+        -DUSE_CUDA=OFF\
+        -DUSE_SSE=OFF\
+        -DUSE_LAPACK=OFF\
+        -DUSE_OPENCV=OFF\
+        -DUSE_OPENMP=OFF\
+        -DUSE_SIGNAL_HANDLER=ON\
+        -DCMAKE_BUILD_TYPE=RelWithDebInfo\
+        -DUSE_MKL_IF_AVAILABLE=OFF\
+        -G Ninja /work/mxnet
+    ninja -v
+}
+
 build_centos7_cpu() {
     set -ex
     cd /work/mxnet
@@ -270,6 +294,10 @@ build_ubuntu_cpu_mkldnn() {
         -j$(nproc)
 }
 
+build_ubuntu_gpu() {
+    build_ubuntu_gpu_cuda91_cudnn7
+}
+
 build_ubuntu_gpu_mkldnn() {
     set -ex
     make  \

src/operator/random/shuffle_op.cc

@@ -22,8 +22,9 @@
  * \file shuffle_op.cc
  * \brief Operator to shuffle elements of an NDArray
  */
-#if (__GNUC__ > 4 && !defined(__clang__major__)) || (__clang_major__ > 4 && __linux__)
-  #define USE_GNU_PARALLEL_SHUFFLE
+#if !defined (__ANDROID__) && ((__GNUC__ > 4 &&\
+    !defined(__clang__major__)) || (__clang_major__ > 4 && __linux__))
+        #define USE_GNU_PARALLEL_SHUFFLE
 #endif
 
 #include <mxnet/operator_util.h>