Update Ubuntu images used on CI to 20.04 (#19588)

* Update Ubuntu images used on CI to 20.04

This helps ensure MXNet to work well on recent Linux distributions (while ensuring it continues to work well on ancient distributions based on the CentOS7 CI pipeline)

* Preserve Ubuntu 18.04 images for TensorRT pipeline as NVidia failed to make TensorRT available for Ubuntu 20.04

* Temporarily disable NVML on CI

[2020-12-03T18:33:10.380Z] OSError: /work/mxnet/python/mxnet/../../build/libmxnet.so: undefined symbol: nvmlDeviceGetComputeRunningProcesses_v2

3rdparty/mshadow/mshadow/tensor.h

@@ -108,7 +108,10 @@ struct Shape {
    * \return the corresponding dimension size
    */
   MSHADOW_XINLINE const index_t &operator[](int idx) const {
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Warray-bounds"
     return shape_[idx];
+#pragma GCC diagnostic pop
   }
   /*!
    * \return whether two shape equals

CMakeLists.txt

@@ -77,7 +77,6 @@ option(USE_JEMALLOC "Build with Jemalloc support" OFF)
 option(USE_LIBJPEG_TURBO "Use libjpeg-turbo" OFF)
 option(USE_DIST_KVSTORE "Build with DIST_KVSTORE support" OFF)
 option(USE_PLUGINS_WARPCTC "Use WARPCTC Plugins" OFF)
-option(USE_CPP_PACKAGE "Build C++ Package" OFF)
 option(USE_MXNET_LIB_NAMING "Use MXNet library naming conventions." ON)
 option(USE_GPROF "Compile with gprof (profiling) flag" OFF)
 option(USE_VTUNE "Enable use of Intel Amplifier XE (VTune)" OFF) # one could set VTUNE_ROOT for search path

ci/docker/Dockerfile.build.ubuntu

@@ -21,8 +21,8 @@
 # See docker-compose.yml for supported BASE_IMAGE ARGs and targets.
 
 ####################################################################################################
-# The Dockerfile uses a dynamic BASE_IMAGE (for example ubuntu:18.04
-# nvidia/cuda:10.2-cudnn7-devel-ubuntu18.04 etc).
+# The Dockerfile uses a dynamic BASE_IMAGE (for example ubuntu:20.04
+# nvidia/cuda:11.1-cudnn8-devel-ubuntu20.04 etc).
 # On top of BASE_IMAGE we install all dependencies shared by all MXNet build
 # environments into a "base" target. At the end of this file, we can specialize
 # "base" for specific usecases. The target built by docker can be selected via
@@ -36,9 +36,7 @@ WORKDIR /work/deps
 RUN export DEBIAN_FRONTEND=noninteractive && \
     apt-get update && \
     apt-get install -y wget software-properties-common && \
-    wget -qO - http://apt.llvm.org/llvm-snapshot.gpg.key | apt-key add - && \
     wget -qO - wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS-2019.PUB | apt-key add - && \
-    apt-add-repository "deb http://apt.llvm.org/xenial/ llvm-toolchain-xenial-10 main" &&  \
     apt-add-repository "deb https://apt.repos.intel.com/mkl all main" &&  \
     apt-get update && \
     apt-get install -y \
@@ -47,6 +45,7 @@ RUN export DEBIAN_FRONTEND=noninteractive && \
         unzip \
         pandoc \
         ## Development tools
+        cmake \
         build-essential \
         ninja-build \
         git \
@@ -123,43 +122,33 @@ COPY runtime_functions.sh /work/
 # The target built by docker can be selected via "--target" option or docker-compose.yml
 ####################################################################################################
 FROM base as gpu
-# Install Thrust 1.9.8 to be shipped with Cuda 11.
-# Fixes https://github.com/thrust/thrust/issues/1072 for Clang 10
-# This file can be deleted when using Cuda 11 on CI
-RUN cd /usr/local && \
-    git clone https://github.com/thrust/thrust.git && \
-    cd thrust && \
-    git checkout 1.9.8
 
-# Install TensorRT
+# Install TensorRT and CuDNN
+# Use bash as it has better support for string comparisons in if clauses
+SHELL ["/bin/bash", "-c"]
 # We need to redeclare ARG due to
 # https://docs.docker.com/engine/reference/builder/#understand-how-arg-and-from-interact
 ARG BASE_IMAGE
-# due to issue https://gitlab.com/nvidia/container-images/cuda/-/issues/92, we
-#  get the cuda version from installed deb package if CUDA_VERSION is not set
-RUN if [ -z "$CUDA_VERSION" ]; then export CUDA_VERSION=$( \
-    dpkg --list cuda-libraries*| grep "cuda-libraries-dev-" | awk '{print $3}' | sed 's/-1$//'); \
-    fi && export SHORT_CUDA_VERSION=${CUDA_VERSION%.*} && \
+RUN export SHORT_CUDA_VERSION=${CUDA_VERSION%.*} && \
+    export OS_RELEASE="$(cat /etc/os-release)" && \
     apt-get update && \
-    if [ ${SHORT_CUDA_VERSION} = 10.0 ]; then \
-        TRT_VERSION="7.0.0-1+cuda10.0"; \
-        TRT_MAJOR_VERSION=7; \
-    elif [ ${SHORT_CUDA_VERSION} = 10.1 ]; then \
-        TRT_VERSION="6.0.1-1+cuda10.1"; \
-        TRT_MAJOR_VERSION=6; \
-    elif [ ${SHORT_CUDA_VERSION} = 10.2 ]; then \
-        TRT_VERSION="7.0.0-1+cuda10.2"; \
-        TRT_MAJOR_VERSION=7; \
-    elif [ ${SHORT_CUDA_VERSION} = 11.0 ]; then \
-        TRT_VERSION="7.2.0-1+cuda11.0"; \
-        TRT_MAJOR_VERSION=7; \
-    else \
-        echo "ERROR: Cuda ${SHORT_CUDA_VERSION} not yet supported in Dockerfile.build.ubuntu"; \
-        exit 1; \
+    if [[ ${OS_RELEASE} == *"Bionic"* ]]; then \
+        if [ ${SHORT_CUDA_VERSION} = 11.0 ]; then \
+            TRT_VERSION="7.2.0-1+cuda11.0"; \
+            TRT_MAJOR_VERSION=7; \
+        elif [ ${SHORT_CUDA_VERSION} = 11.1 ]; then \
+            TRT_VERSION="7.2.1-1+cuda11.1"; \
+            TRT_MAJOR_VERSION=7; \
+        else \
+            echo "ERROR: Cuda ${SHORT_CUDA_VERSION} not yet supported in Dockerfile.build.ubuntu"; \
+            exit 1; \
+        fi; \
+        apt-get install -y libnvinfer${TRT_MAJOR_VERSION}=${TRT_VERSION} \
+                           libnvinfer-dev=${TRT_VERSION} \
+                           libnvinfer-plugin${TRT_MAJOR_VERSION}=${TRT_VERSION} \
+                           libnvinfer-plugin-dev=${TRT_VERSION}; \
     fi && \
-    apt-get install -y libnvinfer${TRT_MAJOR_VERSION}=${TRT_VERSION} \
-                       libnvinfer-dev=${TRT_VERSION} \
-                       libnvinfer-plugin${TRT_MAJOR_VERSION}=${TRT_VERSION} \
-                       libnvinfer-plugin-dev=${TRT_VERSION} && \
+    apt-get install -y libcudnn8-dev && \
     rm -rf /var/lib/apt/lists/*
 
+ENV CUDNN_VERSION=8.0.5

ci/docker/docker-compose.yml

@@ -85,49 +85,29 @@ services:
       dockerfile: Dockerfile.build.ubuntu
       target: base
       args:
-        BASE_IMAGE: ubuntu:18.04
+        BASE_IMAGE: ubuntu:20.04
       cache_from:
         - ${DOCKER_CACHE_REGISTRY}/build.ubuntu_cpu:latest
-  ubuntu_gpu_cu101:
-    image: ${DOCKER_CACHE_REGISTRY}/build.ubuntu_gpu_cu101:latest
+  ubuntu_tensorrt_cu111:
+    image: ${DOCKER_CACHE_REGISTRY}/build.ubuntu_tensorrt_cu111:latest
     build:
       context: .
       dockerfile: Dockerfile.build.ubuntu
       target: gpu
       args:
-        BASE_IMAGE: nvidia/cuda:10.1-cudnn7-devel-ubuntu18.04
+        BASE_IMAGE: nvidia/cuda:11.1-devel-ubuntu18.04
       cache_from:
-        - ${DOCKER_CACHE_REGISTRY}/build.ubuntu_gpu_cu101:latest
-  ubuntu_gpu_cu102:
-    image: ${DOCKER_CACHE_REGISTRY}/build.ubuntu_gpu_cu102:latest
+        - ${DOCKER_CACHE_REGISTRY}/build.ubuntu_tensorrt_cu111:latest
+  ubuntu_gpu_cu111:
+    image: ${DOCKER_CACHE_REGISTRY}/build.ubuntu_gpu_cu111:latest
     build:
       context: .
       dockerfile: Dockerfile.build.ubuntu
       target: gpu
       args:
-        BASE_IMAGE: nvidia/cuda:10.2-cudnn7-devel-ubuntu18.04
+        BASE_IMAGE: nvidia/cuda:11.1-devel-ubuntu20.04
       cache_from:
-        - ${DOCKER_CACHE_REGISTRY}/build.ubuntu_gpu_cu102:latest
-  ubuntu_gpu_cu110:
-    image: ${DOCKER_CACHE_REGISTRY}/build.ubuntu_gpu_cu110:latest
-    build:
-      context: .
-      dockerfile: Dockerfile.build.ubuntu
-      target: gpu
-      args:
-        BASE_IMAGE: nvidia/cuda:11.0-cudnn8-devel-ubuntu18.04
-      cache_from:
-        - ${DOCKER_CACHE_REGISTRY}/build.ubuntu_gpu_cu110:latest
-  ubuntu_build_cuda:
-    image: ${DOCKER_CACHE_REGISTRY}/build.ubuntu_build_cuda:latest
-    build:
-      context: .
-      dockerfile: Dockerfile.build.ubuntu
-      target: gpu
-      args:
-        BASE_IMAGE: nvidia/cuda:10.1-cudnn7-devel-ubuntu18.04
-      cache_from:
-        - ${DOCKER_CACHE_REGISTRY}/build.ubuntu_build_cuda:latest
+        - ${DOCKER_CACHE_REGISTRY}/build.ubuntu_gpu_cu111:latest
   ###################################################################################################
   # Dockerfile.build.android based images used for testing cross-compilation for plain ARM
   ###################################################################################################