Create a new Ubunutu 22.04 (jammy) build for platform010 (#77591)

###  Motivation

In order to match the internal platform010 builds, we are creating a new config to run on PRs that uses compiler and package versions >= used in platform010. Here are the versions used in the new build

- Ubuntu 22.04 (Jammy-Jellyfish)
- Clang-12
- Python 3.8
- CUDA 11.6

### Summary of Changes

- As `nvidia/docker` images only support Cuda 11.7 with Ubuntu 22.04, we are starting with base Ubuntu 22.04 docker images, and then installing Cuda 11.6

- Fetching `install_cuda.sh` from `pytorch/builder` repo in order to install Cuda using `wget`

- `libssl-dev` has been upgraded to libssl3 in Ubuntu 22.04. Instead, we are using `include` and `lib` folders downloaded with `Openssl1.1.1`

- `Clang-12` requires `libomp-12-dev` to work with `OpenMP` which is added to the `install_base.sh` file.

- Minor fixes to handle compilation errors generated when using `clang-12`.
      - In `pow_test.cpp` adding a `static_cast` to input of `sqrt` method
      - In `vec512_qint.h`, explicitly defining `copy-assignment` operator as its implicit definition is deprecated due to
         user-declared `copy-constructor` in C++11

Pull Request resolved: https://github.com/pytorch/pytorch/pull/77591
Approved by: https://github.com/seemethere, https://github.com/atalman

.circleci/docker/build.sh

@@ -54,6 +54,8 @@ elif [[ "$image" == *-bionic* ]]; then
   UBUNTU_VERSION=18.04
 elif [[ "$image" == *-focal* ]]; then
   UBUNTU_VERSION=20.04
+elif [[ "$image" == *-jammy* ]]; then
+  UBUNTU_VERSION=22.04
 elif [[ "$image" == *ubuntu* ]]; then
   extract_version_from_image_name ubuntu UBUNTU_VERSION
 elif [[ "$image" == *centos* ]]; then
@@ -70,7 +72,8 @@ else
 fi
 
 DOCKERFILE="${OS}/Dockerfile"
-if [[ "$image" == *cuda* ]]; then
+# When using ubuntu - 22.04, start from Ubuntu docker image, instead of nvidia/cuda docker image.
+if [[ "$image" == *cuda* && "$UBUNTU_VERSION" != "22.04" ]]; then
   DOCKERFILE="${OS}-cuda/Dockerfile"
 elif [[ "$image" == *rocm* ]]; then
   DOCKERFILE="${OS}-rocm/Dockerfile"
@@ -249,6 +252,15 @@ case "$image" in
     VISION=yes
     KATEX=yes
     ;;
+  pytorch-linux-jammy-cuda11.6-cudnn8-py3.8-clang12)
+    ANACONDA_PYTHON_VERSION=3.8
+    CUDA_VERSION=11.6
+    CUDNN_VERSION=8
+    CLANG_VERSION=12
+    PROTOBUF=yes
+    DB=yes
+    VISION=yes
+    ;;
   *)
     # Catch-all for builds that are not hardcoded.
     PROTOBUF=yes

.circleci/docker/common/install_base.sh

@@ -15,11 +15,20 @@ install_ubuntu() {
   elif [[ "$UBUNTU_VERSION" == "20.04"* ]]; then
     cmake3="cmake=3.16*"
     maybe_libiomp_dev=""
+  elif [[ "$UBUNTU_VERSION" == "22.04"* ]]; then
+    cmake3="cmake=3.22*"
+    maybe_libiomp_dev=""
   else
     cmake3="cmake=3.5*"
     maybe_libiomp_dev="libiomp-dev"
   fi
 
+  if [[ "$CLANG_VERSION" == 12 ]]; then
+    libomp_dev="libomp-12-dev"
+  else
+    libomp_dev=""
+  fi
+
   # TODO: Remove this once nvidia package repos are back online
   # Comment out nvidia repositories to prevent them from getting apt-get updated, see https://github.com/pytorch/pytorch/issues/74968
   # shellcheck disable=SC2046
@@ -51,6 +60,7 @@ install_ubuntu() {
     libjpeg-dev \
     libasound2-dev \
     libsndfile-dev \
+    ${libomp_dev} \
     software-properties-common \
     wget \
     sudo \
@@ -60,6 +70,20 @@ install_ubuntu() {
   # see: https://github.com/pytorch/pytorch/issues/65931
   apt-get install -y libgnutls30
 
+  # cuda-toolkit does not work with gcc-11.2.0 which is default in Ubunutu 22.04
+  # see: https://github.com/NVlabs/instant-ngp/issues/119
+  if [[ "$UBUNTU_VERSION" == "22.04"* ]]; then
+    apt-get install -y g++-10
+    update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 30
+    update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 30
+    update-alternatives --install /usr/bin/gcov gcov /usr/bin/gcov-10 30
+
+    # https://www.spinics.net/lists/libreoffice/msg07549.html
+    sudo rm -rf /usr/lib/gcc/x86_64-linux-gnu/11
+    wget https://github.com/gcc-mirror/gcc/commit/2b2d97fc545635a0f6aa9c9ee3b017394bc494bf.patch -O noexecpt.patch
+    sudo patch  /usr/include/c++/10/bits/range_access.h noexecpt.patch
+  fi
+
   # Cleanup package manager
   apt-get autoclean && apt-get clean
   rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*

.circleci/docker/common/install_cache.sh

@@ -5,7 +5,9 @@ set -ex
 install_ubuntu() {
   echo "Preparing to build sccache from source"
   apt-get update
-  apt-get install -y cargo pkg-config libssl-dev
+  # libssl-dev will not work as it is upgraded to libssl3 in Ubuntu-22.04.
+  # Instead use lib and headers from OpenSSL1.1 installed in `install_openssl.sh``
+  apt-get install -y cargo
   echo "Checking out sccache repo"
   git clone https://github.com/pytorch/sccache
   cd sccache

.circleci/docker/common/install_openssl.sh

@@ -10,5 +10,7 @@ cd "${OPENSSL}"
 ./config --prefix=/opt/openssl -d '-Wl,--enable-new-dtags,-rpath,$(LIBRPATH)'
 # NOTE: openssl install errors out when built with the -j option
 make -j6; make install_sw
+# Link the ssl libraries to the /usr/lib folder.
+sudo ln -s /opt/openssl/lib/lib* /usr/lib
 cd ..
 rm -rf "${OPENSSL}"

.circleci/docker/ubuntu-cuda/Dockerfile

@@ -65,6 +65,7 @@ ENV INSTALLED_VISION ${VISION}
 ADD ./common/install_openssl.sh install_openssl.sh
 ENV OPENSSL_ROOT_DIR /opt/openssl
 RUN bash ./install_openssl.sh
+ENV OPENSSL_DIR /opt/openssl
 
 # (optional) Install non-default CMake version
 ARG CMAKE_VERSION

.circleci/docker/ubuntu/Dockerfile

@@ -6,6 +6,8 @@ ARG UBUNTU_VERSION
 
 ENV DEBIAN_FRONTEND noninteractive
 
+ARG CLANG_VERSION
+
 # Install common dependencies (so that this step can be cached separately)
 ARG EC2
 ADD ./common/install_base.sh install_base.sh
@@ -13,7 +15,6 @@ RUN bash ./install_base.sh && rm install_base.sh
 
 # Install clang
 ARG LLVMDEV
-ARG CLANG_VERSION
 ADD ./common/install_clang.sh install_clang.sh
 RUN bash ./install_clang.sh && rm install_clang.sh
 
@@ -50,6 +51,13 @@ RUN bash ./install_gcc.sh && rm install_gcc.sh
 ADD ./common/install_lcov.sh install_lcov.sh
 RUN  bash ./install_lcov.sh && rm install_lcov.sh
 
+# Install cuda and cudnn
+ARG CUDA_VERSION
+RUN wget -q https://raw.githubusercontent.com/pytorch/builder/main/common/install_cuda.sh -O install_cuda.sh
+RUN bash ./install_cuda.sh ${CUDA_VERSION} && rm install_cuda.sh
+ENV DESIRED_CUDA ${CUDA_VERSION}
+ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:$PATH
+
 # (optional) Install protobuf for ONNX
 ARG PROTOBUF
 ADD ./common/install_protobuf.sh install_protobuf.sh
@@ -111,6 +119,8 @@ RUN rm install_ninja.sh
 ADD ./common/install_openssl.sh install_openssl.sh
 RUN bash ./install_openssl.sh
 ENV OPENSSL_ROOT_DIR /opt/openssl
+ENV OPENSSL_DIR /opt/openssl
+RUN rm install_openssl.sh
 
 # Install ccache/sccache (do this last, so we get priority in PATH)
 ADD ./common/install_cache.sh install_cache.sh
@@ -122,12 +132,19 @@ ADD ./common/install_jni.sh install_jni.sh
 ADD ./java/jni.h jni.h
 RUN bash ./install_jni.sh && rm install_jni.sh
 
+# Install Open MPI for CUDA
+ADD ./common/install_openmpi.sh install_openmpi.sh
+RUN if [ -n "${CUDA_VERSION}" ]; then bash install_openmpi.sh; fi
+RUN rm install_openmpi.sh
+
 # Include BUILD_ENVIRONMENT environment variable in image
 ARG BUILD_ENVIRONMENT
 ENV BUILD_ENVIRONMENT ${BUILD_ENVIRONMENT}
 
-# Install LLVM dev version (Defined in the pytorch/builder github repository)
+# AWS specific CUDA build guidance
-COPY --from=pytorch/llvm:9.0.1 /opt/llvm /opt/llvm
+ENV TORCH_CUDA_ARCH_LIST Maxwell
+ENV TORCH_NVCC_FLAGS "-Xfatbin -compress-all"
+ENV CUDA_PATH /usr/local/cuda
 
 USER jenkins
 CMD ["bash"]

.github/workflows/docker-builds.yml

@@ -30,6 +30,7 @@ jobs:
           - docker-image-name: pytorch-linux-bionic-py3.7-clang9
           - docker-image-name: pytorch-linux-bionic-rocm5.0-py3.7
           - docker-image-name: pytorch-linux-bionic-rocm5.1-py3.7
+          - docker-image-name: pytorch-linux-jammy-cuda11.6-cudnn8-py3.8-clang12
           - docker-image-name: pytorch-linux-xenial-cuda10.2-cudnn7-py3-gcc7
           - docker-image-name: pytorch-linux-xenial-cuda11.3-cudnn8-py3-gcc7
           - docker-image-name: pytorch-linux-xenial-py3-clang5-android-ndk-r19c

.github/workflows/pull.yml

@@ -193,6 +193,13 @@ jobs:
       docker-image-name: pytorch-linux-xenial-py3-clang5-asan
       build-generates-artifacts: false
 
+  linux-jammy-cuda-11_6-cudnn8-py3_8-clang12-build:
+    name: linux-jammy-cuda11.6-cudnn8-py3.8-clang12
+    uses: ./.github/workflows/_linux-build.yml
+    with:
+      build-environment: linux-jammy-cuda11.6-cudnn8-py3.8-clang12
+      docker-image-name: pytorch-linux-jammy-cuda11.6-cudnn8-py3.8-clang12
+
   linux-xenial-py3-clang5-mobile-custom-build-static:
     name: linux-xenial-py3-clang5-mobile-custom-build-static
     uses: ./.github/workflows/_linux-build.yml

aten/src/ATen/cpu/vec/vec512/vec512_qint.h

@@ -430,6 +430,11 @@ struct Vectorized<c10::qint8> : public Vectorizedqi {
     // constructor for moving the enum
     Vectorized(const Vectorized<c10::qint8>& other) : Vectorizedqi(other.vals) { }
 
+    // This is added to avoid error: definition of implicit copy assignment operator
+    // for 'Vectorized<c10::qint8>' is deprecated because it has a user-declared
+    // copy constructor [-Werror,-Wdeprecated-copy]
+    Vectorized& operator=(const Vectorized<c10::qint8>&) = default;
+
     void store(void* ptr, int count = size()) const {
         if (count != size()) {
             memcpy(ptr, &vals, count * sizeof(value_type));
@@ -589,6 +594,11 @@ struct Vectorized<c10::quint8> : public Vectorizedqi {
 
     Vectorized(const Vectorized<c10::quint8>& other) : Vectorizedqi(other.vals) { }
 
+    // This is added to avoid error: definition of implicit copy assignment operator
+    // for 'Vectorized<c10::quint8>' is deprecated because it has a user-declared
+    // copy constructor [-Werror,-Wdeprecated-copy]
+    Vectorized& operator=(const Vectorized<c10::quint8>&) = default;
+
     void store(void* ptr, int count = size()) const {
         if (count != size()) {
             memcpy(ptr, &vals, count * sizeof(value_type));

aten/src/ATen/test/pow_test.cpp

@@ -35,16 +35,16 @@ const std::vector<int> ints {
   int_min,
   int_min + 1,
   int_min + 2,
-  static_cast<int>(-sqrt(int_max)),
+  static_cast<int>(-sqrt(static_cast<double>(int_max))),
   -3, -2, -1, 0, 1, 2, 3,
-  static_cast<int>(sqrt(int_max)),
+  static_cast<int>(sqrt(static_cast<double>(int_max))),
   int_max - 2,
   int_max - 1,
   int_max
 };
 const std::vector<int> non_neg_ints {
   0, 1, 2, 3,
-  static_cast<int>(sqrt(int_max)),
+  static_cast<int>(sqrt(static_cast<double>(int_max))),
   int_max - 2,
   int_max - 1,
   int_max
@@ -53,16 +53,16 @@ const std::vector<int64_t> longs {
   long_min,
   long_min + 1,
   long_min + 2,
-  static_cast<int64_t>(-sqrt(long_max)),
+  static_cast<int64_t>(-sqrt(static_cast<double>(long_max))),
   -3, -2, -1, 0, 1, 2, 3,
-  static_cast<int64_t>(sqrt(long_max)),
+  static_cast<int64_t>(sqrt(static_cast<double>(long_max))),
   long_max - 2,
   long_max - 1,
   long_max
 };
 const std::vector<int64_t> non_neg_longs {
   0, 1, 2, 3,
-  static_cast<int64_t>(sqrt(long_max)),
+  static_cast<int64_t>(sqrt(static_cast<double>(long_max))),
   long_max - 2,
   long_max - 1,
   long_max
@@ -128,7 +128,7 @@ void tensor_pow_scalar(const Vals vals, const Pows pows, const torch::ScalarType
 
   for (const auto pow : pows) {
     // NOLINTNEXTLINE(clang-diagnostic-implicit-const-int-float-conversion)
-    if ( dtype == kInt && pow > std::numeric_limits<int>::max()) {
+    if ( dtype == kInt && pow > static_cast<float>(std::numeric_limits<int>::max())) {
       // value cannot be converted to type int without overflow
       // NOLINTNEXTLINE(hicpp-avoid-goto,cppcoreguidelines-avoid-goto)
       EXPECT_THROW(tensor.pow(pow), std::runtime_error);