feat(cuda/aarch64): upgrade aarch64-cuda to 11.8

GitOrigin-RevId: 4b1ad73fbc5ec0a688309f58e7ccf4fe377fca39

CMakeLists.txt

@@ -1274,16 +1274,30 @@ if(MGE_ARCH STREQUAL "aarch64")
     set(MARCH "-march=armv8.2-a+fp16")
   endif()
 
-  if(MGE_WITH_CUDA)
+  if(${CMAKE_BUILD_TYPE} STREQUAL "Debug")
     message(
       WARNING
-        "aarch64 ld will add -mfix-cortex-a53-843419 and -mfix-cortex-a53-835769,\
-        when cuda enable and CMAKE with DEBUG build type,ld will take about 14min+,\
-        for save link time(14min->1min), you may open below flags if not deploy on\
-        arm a53 platform, or just build release type!")
+        "aarch64 ld will add -mfix-cortex-a53-843419 and -mfix-cortex-a53-835769 by default.\
+    when build with DEBUG build type,ld will take about 14min+, for save link time(14min->1min), \
+    you may open below flags if not deploy on arm a53 platform, or just build release type!"
+    )
     # set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mno-fix-cortex-a53-843419
     # -mno-fix-cortex-a53-835769")
   endif()
+
+  if(MGE_WITH_CUDA)
+    message(STATUS "check compiler version...")
+    if(CMAKE_COMPILER_IS_GNUCC AND CMAKE_CXX_COMPILER_VERSION VERSION_LESS 7.5)
+      message(FATAL_ERROR "gcc version must >= 7.5 when build with cuda")
+    endif()
+    # cuda libs build with -mno-fix-cortex-a53-843419 -mno-fix-cortex-a53-835769, if you are not deploy on arm a53 platform and want to save link time, you may open below flags even
+    message(
+      WARNING
+        "cuda libs build with -mno-fix-cortex-a53-843419 -mno-fix-cortex-a53-835769, so force disable it to avoid link error"
+    )
+    set(CMAKE_CXX_FLAGS
+        "${CMAKE_CXX_FLAGS} -mno-fix-cortex-a53-843419 -mno-fix-cortex-a53-835769")
+  endif()
 endif()
 
 if(MGE_ARCH STREQUAL "riscv64")

LICENSE

@@ -175,7 +175,7 @@
 
    END OF TERMS AND CONDITIONS
 
-   Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+   Copyright (c) 2014-2023 Megvii Inc. All rights reserved.
 
    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.

dnn/cuda-stub/src/dlopen_helper.h

@@ -115,10 +115,7 @@ static void* resolve_library_func(void* handle, const char* func) {
 	    cnt++;
 	    //! do not print all annoying msg at broken driver env, for example empty libcuda.so or libcuda.dll
 	    if (cnt < 3) {
-            LOGD("failed to load %s func: %s.(May caused by currently driver is too old, \
-                if you find cuda is not available(by import megengine as mge; mge.get_device_count(\"gpu\") \
-                    or find some inexplicable crash of the program, try upgrade driver from \
-                    https://developer.nvidia.com/cuda-downloads)", g_default_api_name, func);
+            LOGD("failed to load %s func: %s.(May caused by currently driver is too old, \nif you find cuda is not available(by import megengine as mge; mge.get_device_count(\"gpu\")\nor find some inexplicable crash of the program,\ntry upgrade driver from https://developer.nvidia.com/cuda-downloads)", g_default_api_name, func);
 	    }
     }
     return ret;

scripts/whl/manylinux2014/Dockerfile_aarch64

@@ -8,4 +8,3 @@ RUN ./setup_mirror.sh "$platform"
 
 ADD init_image.sh /tmp
 RUN /tmp/init_image.sh aarch64 && rm -f /tmp/init_image.sh
-

scripts/whl/manylinux2014/Dockerfile_aarch64_ubuntu2004

+FROM ubuntu:20.04
+
+ENV UID=1024 PATH=${PATH}:/usr/local/cuda/bin
+
+# update
+RUN apt-get update -y
+
+# build env
+RUN DEBIAN_FRONTEND=noninteractive \
+TZ=Etc/UTC apt-get install -y \
+pkg-config build-essential git-lfs \
+autoconf gdb git-core gnupg flex bison gperf \
+zip curl zlib1g-dev gcc \
+g++ rsync wget vim llvm clang
+
+# install cmake 3.15.2
+RUN cd /tmp ; wget https://cmake.org/files/v3.15/cmake-3.15.2.tar.gz;tar -xzvf cmake-3.15.2.tar.gz;cd cmake-3.15.2;./configure; make -j8; make install
+
+# install base python3
+RUN apt-get install -y python3
+RUN apt-get install -y python3-dev python3-pip python3-numpy python3-setuptools
+
+# install pyenv
+RUN curl https://pyenv.run | bash
+
+# install more env for build python3 from src
+RUN apt-get install -y checkinstall libncursesw5-dev \
+libssl-dev libsqlite3-dev tk-dev libgdbm-dev libc6-dev libbz2-dev libreadline-dev \
+libffi-dev liblzma-dev libncurses5-dev libdb-dev libexpat-dev libtinfo5 libtinfo-dev
+
+# install more env for build swig and ninja
+RUN apt-get install -y libpcre3 libpcre3-dev re2c
+
+# call other build env
+ADD init_ubuntu2004.sh /tmp
+RUN /tmp/init_ubuntu2004.sh && rm -f /tmp/init_ubuntu2004.sh
+
+# install lsb-release to distinguish is ubuntu or manylinux
+RUN apt-get install -y lsb-release
+
+# install gcc/g++ 8 as gcc/g++ 9 as issue: some code will take about 1h to run as on aarch64-ubuntu
+RUN apt-get install -y gcc-8 g++-8

scripts/whl/manylinux2014/build_image.sh

@@ -7,3 +7,10 @@ config_docker_file
 cd $(dirname $0)
 echo "docker_file is ${docker_file}"
 docker build -t env_manylinux2014:latest -f ${docker_file} .
+
+OS=$(uname -m)
+if [[ $OS = "aarch64" ]]; then
+    # as nivida trt-8.5.3.1 build from ubuntu20.04, which depends on new glibc, env_manylinux2014 do not satisfy, so need build aarch64-ubuntu20.04 docker
+    echo "as in aarch64, need build aarch64-ubuntu20.04 docker"
+    docker build -t ubuntu2004:latest -f Dockerfile_aarch64_ubuntu2004 .
+fi

scripts/whl/manylinux2014/build_wheel_common.sh

@@ -13,7 +13,9 @@ TensorRT_LIB_DIR="/opt/tensorrt/lib/"
 
 SDK_NAME="unknown"
 x86_64_support_version="cu101 cu111 cu112 cpu cu111_cudnn821_tensorRT825 cu114 cu118"
-aarch64_support_version="cu102_JetsonNano cu111 cpu"
+aarch64_support_version="cu102_JetsonNano cu111 cpu cu118"
+docker_tag="env_manylinux2014:latest"
+
 if [[ -z ${IN_CI} ]]
 then
     IN_CI="false"
@@ -236,6 +238,12 @@ elif [ $SDK_NAME == "cu118" ];then
     CUDNN_LIB_DIR="/opt/cudnn/lib/"
     CUDA_LIB_DIR="/usr/local/cuda/targets/x86_64-linux/lib/"
     TensorRT_LIB_DIR="/opt/tensorrt/lib/"
+    if [ ${machine} == "aarch64" ];then
+        CUDA_LIB_DIR="/usr/local/cuda/targets/sbsa-linux/lib/"
+        # aarch64-trt8 libs build from ubuntu2004, which depends on new glibc and libc++
+        # manylinux2014 do satify the glibc/libc++ requirement, so we need to use ubuntu2004 as base image
+        docker_tag="ubuntu2004:latest"
+    fi
 
     CUDA_COPY_LIB_LIST="\
         ${CUDA_LIB_DIR}/libnvrtc.so.11.2:\
@@ -262,6 +270,23 @@ elif [ $SDK_NAME == "cu118" ];then
         -gencode arch=compute_86,code=sm_86 \
         -gencode arch=compute_89,code=sm_89 \
         -gencode arch=compute_89,code=compute_89\" "
+    if [ ${machine} == "aarch64" ];then
+        if [ ${IN_CI} = "true" ]; then
+            # ci taishan is use SM_75 card
+            EXTRA_CMAKE_FLAG=" -DMGE_WITH_CUDNN_SHARED=ON -DMGE_WITH_CUBLAS_SHARED=ON -DMGE_CUDA_GENCODE=\"-gencode arch=compute_75,code=sm_75\" "
+        else
+            # support orin and remove 1080Ti
+            EXTRA_CMAKE_FLAG=" -DMGE_WITH_CUDNN_SHARED=ON -DMGE_WITH_CUBLAS_SHARED=ON \
+                -DMGE_CUDA_GENCODE=\" \
+                -gencode arch=compute_70,code=sm_70 \
+                -gencode arch=compute_75,code=sm_75 \
+                -gencode arch=compute_80,code=sm_80 \
+                -gencode arch=compute_86,code=sm_86 \
+                -gencode arch=compute_87,code=sm_87 \
+                -gencode arch=compute_89,code=sm_89 \
+                -gencode arch=compute_89,code=compute_89\" "
+        fi
+    fi
 
 
 elif [ $SDK_NAME == "cpu" ];then
@@ -407,6 +432,10 @@ if [ ${machine} == "aarch64" ];then
     # infact ubuntu gcc version: gcc (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0 is OK
     echo "force use gcc8 on aarch64 linux"
     BUILD_GCC8="ON"
+    if [ $SDK_NAME == "cu118" ];then
+        echo "cu118 with aarch64 will build at ubuntu2004 docker env, so do not use gcc8"
+        BUILD_GCC8="OFF"
+    fi
 fi
 
 if [ "$BUILD_GCC8" == "ON" ];then
@@ -443,4 +472,4 @@ docker run --rm ${docker_args} $TMPFS_ARGS \
     -v ${TENSORRT_ROOT_DIR}:/opt/tensorrt \
     -v ${BASEDIR}:/home/code \
     -v ${OUTPUTDIR}:/home/output:rw \
-    env_manylinux2014:latest /bin/bash -c "$run_cmd"
+    ${docker_tag} /bin/bash -c "$run_cmd"

scripts/whl/manylinux2014/do_build_common.sh

 #!/bin/bash -e
 
+IN_UBUNTU_DOCKER_ENV="OFF"
+if which lsb_release && lsb_release -a | grep "Ubuntu"; then
+    IN_UBUNTU_DOCKER_ENV="ON"
+    # some code will take about 1h to run as on aarch64-ubuntu
+    export CC=gcc-8
+    export CXX=g++-8
+fi
+
 function handle_strip() {
     echo "now handle strip $1"
     objcopy --only-keep-debug $1 $1.dbg
@@ -42,7 +50,9 @@ function patch_elf_depend_lib_mgb_mge() {
     echo "handle common depend lib for mgb or mge"
     LIBS_DIR=${BUILD_DIR}/staging/megengine/core/lib
     mkdir -p ${LIBS_DIR}
-    cp /usr/lib64/libatomic.so.1 ${LIBS_DIR}
+    if [ ${IN_UBUNTU_DOCKER_ENV} = "OFF" ]; then
+        cp /usr/lib64/libatomic.so.1 ${LIBS_DIR}
+    fi
 
     patchelf --remove-rpath ${BUILD_DIR}/staging/megengine/core/_imperative_rt.so
     patchelf --force-rpath --set-rpath '$ORIGIN/lib' ${BUILD_DIR}/staging/megengine/core/_imperative_rt.so
@@ -73,6 +83,10 @@ SRC_DIR=$(readlink -f "`dirname $0`/../../../")
 source ${SRC_DIR}/scripts/whl/utils/utils.sh
 
 SUPPORT_ALL_VERSION="36m 37m 38 39 310"
+if [ ${IN_UBUNTU_DOCKER_ENV} = "ON" ]; then
+    SUPPORT_ALL_VERSION="3.6.10 3.7.7 3.8.3 3.9.4 3.10.1"
+    echo "in ubuntu docker env, override support all python version: ${SUPPORT_ALL_VERSION}"
+fi
 ALL_PYTHON=${ALL_PYTHON}
 if [[ -z ${ALL_PYTHON} ]]
 then
@@ -92,6 +106,42 @@ if [ ${BUILD_WHL_CPU_ONLY} = "OFF" ]; then
     BUILD_DIR=${SRC_DIR}/build_dir/host/MGE_WITH_CUDA_ON/MGE_INFERENCE_ONLY_OFF/Release/build/
 fi
 
+function config_ubuntu_python_env() {
+    PYTHON_DIR=~/.pyenv/versions/$1/
+    PYTHON_BIN=~/.pyenv/versions/$1/bin
+    if [ ! -f "$PYTHON_BIN/python3" ]; then
+        echo "ERR: can not find $PYTHON_BIN , Invalid python package"
+        echo "now support list: ${FULL_PYTHON_VER}"
+        err_env
+    else
+        echo "put python3 to env..."
+        export PATH=${PYTHON_BIN}:$PATH
+        which python3
+    fi
+    echo ${ver}
+
+    if [ "$1" = "3.6.10" ]; then
+        PYTHON_INCLUDE_DIR=${PYTHON_DIR}include/python3.6m
+        PYTHON_LIBRARY=${PYTHON_DIR}/lib/libpython3.6m.so
+    elif [ "$1" = "3.7.7" ]; then
+        PYTHON_INCLUDE_DIR=${PYTHON_DIR}include/python3.7m
+        PYTHON_LIBRARY=${PYTHON_DIR}/lib/libpython3.7m.so
+    elif [[ "$1" = "3.8.3" || "$1" = "3.8.10" ]]; then
+        PYTHON_INCLUDE_DIR=${PYTHON_DIR}include/python3.8
+        PYTHON_LIBRARY=${PYTHON_DIR}/lib/libpython3.8.so
+    elif [ "$1" = "3.9.4" ]; then
+        PYTHON_INCLUDE_DIR=${PYTHON_DIR}include/python3.9
+        PYTHON_LIBRARY=${PYTHON_DIR}/lib/libpython3.9.so
+    elif [ "$1" = "3.10.1" ]; then
+        PYTHON_INCLUDE_DIR=${PYTHON_DIR}include/python3.10
+        PYTHON_LIBRARY=${PYTHON_DIR}/lib/libpython3.10.so
+    else
+        echo "ERR: DO NOT SUPPORT PYTHON VERSION"
+        echo "now support list: ${FULL_PYTHON_VER}"
+        exit -1
+    fi
+}
+
 # here we just treat cu file should not in the increment build file list
 INCREMENT_KEY_WORDS=".cu.o is dirty"
 IS_IN_FIRST_LOOP=TRUE
@@ -107,22 +157,42 @@ do
         rm -rf ${BUILD_DIR}
     fi
 
-    python_ver=`echo $ver | tr -d m`
-    MAJOR=${python_ver:0:1}
-    MINOR=${python_ver:1}
-    PYTHON_DIR=/opt/python/cp${python_ver}-cp${ver}
-
-    SUFFIX=
-    if [[ $MINOR -lt 8 ]];then
-        SUFFIX="m"
-    fi
-
+    # export common args
     export EXTRA_CMAKE_ARGS="${ORG_EXTRA_CMAKE_FLAG} -DCMAKE_BUILD_TYPE=RelWithDebInfo"
     export EXTRA_CMAKE_ARGS="${EXTRA_CMAKE_ARGS} -DMGE_WITH_CUSTOM_OP=ON"
+    export EXTRA_CMAKE_ARGS="${EXTRA_CMAKE_ARGS} -DMGE_WITH_ATLAS=ON"
+
+    if [ ${IN_UBUNTU_DOCKER_ENV} = "ON" ]; then
+        echo "into Ubuntu env"
+        # config env
+        config_ubuntu_python_env ${ver}
+        #check env
+        if [ ! -f "$PYTHON_LIBRARY" ]; then
+            echo "ERR: can not find $PYTHON_LIBRARY , Invalid python package"
+            err_env
+        fi
+        if [ ! -d "$PYTHON_INCLUDE_DIR" ]; then
+            echo "ERR: can not find $PYTHON_INCLUDE_DIR , Invalid python package"
+            err_env
+        fi
+        export EXTRA_CMAKE_ARGS="${EXTRA_CMAKE_ARGS} -DPYTHON_INCLUDE_DIR=${PYTHON_INCLUDE_DIR}"
+    else
+        echo "into manylinux env"
+        python_ver=`echo $ver | tr -d m`
+        MAJOR=${python_ver:0:1}
+        MINOR=${python_ver:1}
+        PYTHON_DIR=/opt/python/cp${python_ver}-cp${ver}
+
+        SUFFIX=
+        if [[ $MINOR -lt 8 ]];then
+            SUFFIX="m"
+        fi
+        export EXTRA_CMAKE_ARGS="${EXTRA_CMAKE_ARGS} -DPYTHON_INCLUDE_DIR=${PYTHON_DIR}/include/python${MAJOR}.${MINOR}${SUFFIX}"
+    fi
+    #append cmake args for config python
     export EXTRA_CMAKE_ARGS="${EXTRA_CMAKE_ARGS} -DPYTHON_EXECUTABLE=${PYTHON_DIR}/bin/python3"
+    # please do not config to real python path, it will cause import failed if user python3 do not have libpython3.xx.so[dynamic lib]
     export EXTRA_CMAKE_ARGS="${EXTRA_CMAKE_ARGS} -DPYTHON_LIBRARY=${PYTHON_DIR}/lib/"
-    export EXTRA_CMAKE_ARGS="${EXTRA_CMAKE_ARGS} -DPYTHON_INCLUDE_DIR=${PYTHON_DIR}/include/python${MAJOR}.${MINOR}${SUFFIX}"
-    export EXTRA_CMAKE_ARGS="${EXTRA_CMAKE_ARGS} -DMGE_WITH_ATLAS=ON"
 
     if [ -d "${BUILD_DIR}" ]; then
         # insure rm have args
@@ -174,8 +244,15 @@ do
     cd /home/output
     mkdir -p ${SRC_DIR}/scripts/whl/manylinux2014/output/wheelhouse/${SDK_NAME}
     cd ${BUILD_DIR}/staging/dist/
-    org_whl_name=`ls Meg*${ver}*.whl`
-    compat_whl_name=`echo ${org_whl_name} | sed 's/linux/manylinux2014/'`
+    org_whl_name="null"
+    if [ ${IN_UBUNTU_DOCKER_ENV} = "ON" ]; then
+        # ubuntu glibc higher than manylinux2014, so keep org name
+        org_whl_name=`ls Meg*.whl`
+        compat_whl_name=${org_whl_name}
+    else
+        org_whl_name=`ls Meg*${ver}*.whl`
+        compat_whl_name=`echo ${org_whl_name} | sed 's/linux/manylinux2014/'`
+    fi
     echo "org whl name: ${org_whl_name}"
     echo "comapt whl name: ${compat_whl_name}"
     mv ${org_whl_name} ${SRC_DIR}/scripts/whl/manylinux2014/output/wheelhouse/${SDK_NAME}/${compat_whl_name}

scripts/whl/manylinux2014/init_ubuntu2004.sh

+#!/bin/bash -e
+
+ALL_PYTHON="3.6.10 3.7.7 3.8.3 3.9.4 3.10.1"
+SWIG_URL='https://codeload.github.com/swig/swig/tar.gz/refs/tags/rel-3.0.12'
+
+NINJA_URL='https://codeload.github.com/ninja-build/ninja/tar.gz/refs/tags/v1.10.0'
+for ver in ${ALL_PYTHON}
+do
+    numpy_version="1.19.5"
+    if [ ${ver} = "3.10.1" ];then
+        numpy_version="1.21.6"
+    fi
+    echo "Install python ${ver}"
+    env PYTHON_CONFIGURE_OPTS="--enable-shared" ~/.pyenv/bin/pyenv install ${ver}
+    ~/.pyenv/versions/${ver}/bin/python3 -m pip install --upgrade pip
+    ~/.pyenv/versions/${ver}/bin/python3 -m pip install numpy==${numpy_version} setuptools==46.1.3 wheel
+done
+
+pushd /home >/dev/null
+echo "Install swig"
+curl -sSL ${SWIG_URL} | tar xz
+pushd swig-rel-3.0.12 >/dev/null
+./autogen.sh
+mkdir build
+pushd build >/dev/null
+../configure
+make -j$(nproc)
+make install
+popd >/dev/null
+popd >/dev/null
+rm -rf swig-3.0.12
+
+echo "Install ninja build"
+curl -sSL ${NINJA_URL} | tar xz
+pushd ninja-1.10.0 >/dev/null
+mkdir build
+pushd build >/dev/null
+cmake .. -DCMAKE_BUILD_TYPE=Release
+make -j$(nproc)
+cp ninja /usr/bin/
+popd >/dev/null
+popd >/dev/null
+rm -rf ninja-1.10.0
+popd >/dev/null
+
+pushd /tmp >/dev/null
+echo "Install patchelf"
+curl -sSL https://github.com/NixOS/patchelf/archive/0.12.tar.gz | tar xz
+pushd /tmp/patchelf-0.12 >/dev/null
+sed -i '331s/32/256/' ./src/patchelf.cc
+./bootstrap.sh && ./configure && make install-strip
+popd
+rm -rf /tmp/patchelf-0.12
+popd

scripts/whl/utils/utils.sh

@@ -78,18 +78,29 @@ function check_build_ninja_python_api() {
             INCLUDE_KEYWORD="${ver}\\\\include"
             PYTHON_API_INCLUDES="3.6.8\\\\include 3.7.7\\\\include 3.8.3\\\\include 3.9.4\\\\include 3.10.1\\\\include"
         elif [[ $OS =~ "Linux" ]]; then
-            ver=`echo $ver | tr -d m`
-            INCLUDE_KEYWORD="include/python3.${ver:1}" # like 39/310
-            info=`command -v termux-info || true`
-            if [[ "${info}" =~ "com.termux" ]]; then
-                echo "find termux-info at: ${info}"
+            if which lsb_release && lsb_release -a | grep "Ubuntu"; then
+                echo "into Ubuntu env"
                 is_punctuation=${ver:4:1}
                 INCLUDE_KEYWORD="include/python3.${ver:2:1}"
                 if [ ${is_punctuation} = "." ]; then
                     INCLUDE_KEYWORD="include/python3.${ver:2:2}"
                 fi
+                PYTHON_API_INCLUDES="include/python3.5 include/python3.6 include/python3.7 include/python3.8 include/python3.9 include/python3.10"
+            else
+                echo "into manylinux env"
+                ver=`echo $ver | tr -d m`
+                INCLUDE_KEYWORD="include/python3.${ver:1}" # like 39/310
+                info=`command -v termux-info || true`
+                if [[ "${info}" =~ "com.termux" ]]; then
+                    echo "find termux-info at: ${info}"
+                    is_punctuation=${ver:4:1}
+                    INCLUDE_KEYWORD="include/python3.${ver:2:1}"
+                    if [ ${is_punctuation} = "." ]; then
+                        INCLUDE_KEYWORD="include/python3.${ver:2:2}"
+                    fi
+                fi
+                PYTHON_API_INCLUDES="include/python3.5 include/python3.6 include/python3.7 include/python3.8 include/python3.9 include/python3.10"
             fi
-            PYTHON_API_INCLUDES="include/python3.5 include/python3.6 include/python3.7 include/python3.8 include/python3.9 include/python3.10"
         elif [[ $OS =~ "Darwin" ]]; then
             is_punctuation=${ver:4:1}
             INCLUDE_KEYWORD="include/python3.${ver:2:1}"