DALI and NGC¶
DALI is preinstalled in the NVIDIA GPU Cloud TensorFlow, PyTorch, and MXNet containers in versions 18.07 and later.
Installing prebuilt DALI packages¶
One or more of the following deep learning frameworks:
Execute the below command CUDA 9.0 based build:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/cuda/9.0 nvidia-dali
Starting DALI 0.8.0 for CUDA 10.0 based build use:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/cuda/10.0 nvidia-dali
nvidia-dali package doesn’t contain prebuilt versions of the DALI TensorFlow plugin, DALI TensorFlow plugin needs to be installed explicitly for the currently present version of TensorFlow:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/cuda/9.0 nvidia-dali-tf-plugin
Starting DALI 0.8.0 for CUDA 10.0 based build execute:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/cuda/10.0 nvidia-dali-tf-plugin
Due to a known issue with installing dependent packages), DALI needs to be installed before installing
nvidia-dali-tf-plugin (in a separate pip install call). The package
tensorflow-gpu must be installed before attempting to install
nvidia-dali-tf-plugin has a strict requirement with
nvidia-dali as its exact same version.
nvidia-dali-tf-plugin at its latest version will replace any older
nvidia-dali versions already installed with the latest.
To work with older versions of DALI, provide the version explicitly to the
pip install command.
OLDER_VERSION=0.6.1 pip install --extra-index-url https://developer.download.nvidia.com/compute/redist nvidia-dali-tf-plugin==$OLDER_VERSION
Nightly and weekly release channels¶
While binaries available to download from nightly and weekly builds include most recent changes available in the GitHub some functionalities may not work or provide inferior performance comparing to the official releases. Those builds are meant for the early adopters seeking for the most recent version available and being ready to boldly go where no man has gone before.
It is recommended to uninstall regular DALI and TensorFlow plugin before installing nvidia-dali-nightly or nvidia-dali-weekly as they are installed in the same path
To access most recent nightly builds please use flowing release channel:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/nightly/cuda/9.0 nvidia-dali-nightly pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/nightly/cuda/9.0 nvidia-dali-tf-plugin-nightly
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/nightly/cuda/10.0 nvidia-dali-nightly pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/nightly/cuda/10.0 nvidia-dali-tf-plugin-nightly
Also, there is a weekly release channel with more thorough testing (only CUDA10 builds are provided there):
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/weekly/cuda/10.0 nvidia-dali-weekly pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/weekly/cuda/10.0 nvidia-dali-tf-plugin-weekly
Compiling DALI from source (using Docker builder) - recommended¶
Following these steps, it is possible to recreate Python wheels in a similar fashion as we provide as an official prebuild binary.
Follow installation guide and manual at the link (version 17.05 or later is required).
Building Python wheel and (optionally) Docker image¶
Change directory (
cd) into Docker directory and run
./build.sh. If needed, set the following environment variables:
PYVER - Python version. Default is
CUDA_VERSION - CUDA toolkit version (9 for 9.0 or 10 for 10.0). Default is
NVIDIA_BUILD_ID - Custom ID of the build. Default is
CREATE_WHL - Create a standalone wheel. Default is
CREATE_RUNNER - Create Docker image with cuDNN, CUDA and DALI installed inside. It will create the
Docker_run_cudaimage, which needs to be run using
nvidia-dockerand DALI wheel in the
DALI_BUILD_FLAVOR - adds a suffix to DALI package name and put a note about it in the whl package description, i.e. nightly will result in the nvidia-dali-nightly
CMAKE_BUILD_TYPE - build type, available options: Debug, DevDebug, Release, RelWithDebInfo. Default is
BUILD_INHOST - ask docker to mount source code instead of copying it. Thank to that consecutive builds are resuing existing object files and are faster for the development. Uses $DALI_BUILD_DIR as a directory for build objects. Default is
REBUILD_BUILDERS - if builder docker images need to be rebuild or can be reused from the previous build. Default is
REBUILD_MANYLINUX - if manylinux base image need to be rebuild. Default is
DALI_BUILD_DIR - where DALI build should happen. It matters only bit the in-tree build where user may provide different path for every python/CUDA version. Default is
It is worth to mention that build.sh should accept the same set of environment variables as the project CMake.
The recommended command line is:
PYVER=X.Y CUDA_VERSION=Z ./build.sh
PYVER=3.6 CUDA_VERSION=10 ./build.sh
Will build CUDA 10 based DALI for Python 3.6 and place relevant Python wheel inside DALI_root/wheelhouse
Compiling DALI from source (bare metal)¶
GCC 4.9.2 or later
Boost 1.66 or later
CUDA 8.0 compatibility is provided unofficially.
This can be unofficially disabled. See below.
Version 2 or later
(Version 3 or later is required for TensorFlow TFRecord file format support).
CMake 3.11 or later
libjpeg-turbo 1.5.x or later
This can be unofficially disabled. See below.
FFmpeg 3.4.2 or later
We recommend using version 3.4.2 compiled following the instructions below.
OpenCV 3 or later
Supported version: 3.4
(Optional) liblmdb 0.9.x or later
TensorFlow installation is required to build the TensorFlow plugin for DALI.
Items marked “unofficial” are community contributions that are believed to work but not officially tested or maintained by NVIDIA.
This software uses the FFmpeg licensed code under the LGPLv2.1. Its source can be downloaded from here.
FFmpeg was compiled using the following command line:
./configure \ --prefix=/usr/local \ --disable-static \ --disable-all \ --disable-autodetect \ --disable-iconv \ --enable-shared \ --enable-avformat \ --enable-avcodec \ --enable-avfilter \ --enable-protocol=file \ --enable-demuxer=mov,matroska \ --enable-bsf=h264_mp4toannexb,hevc_mp4toannexb && \ make
Get the DALI source¶
git clone --recursive https://github.com/NVIDIA/dali cd dali
Make the build directory¶
mkdir build cd build
Building DALI without LMDB support:¶
cmake .. make -j"$(nproc)"
Building DALI with LMDB support:¶
cmake -DBUILD_LMDB=ON .. make -j"$(nproc)"
Building DALI using Clang (experimental):¶
This build is experimental. It is neither maintained nor tested. It is not guaranteed to work. We recommend using GCC for production builds.
cmake -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang .. make -j"$(nproc)"
Optional CMake build parameters:
BUILD_PYTHON- build Python bindings (default: ON)
BUILD_TEST- include building test suite (default: ON)
BUILD_BENCHMARK- include building benchmarks (default: ON)
BUILD_LMDB- build with support for LMDB (default: OFF)
BUILD_NVTX- build with NVTX profiling enabled (default: OFF)
BUILD_TENSORFLOW- build TensorFlow plugin (default: OFF)
BUILD_NVJPEG- build with
nvJPEGsupport (default: ON)
BUILD_NVOF- build with
NVIDIA OPTICAL FLOW SDKsupport (default: ON)
BUILD_NVDEC- build with
NVIDIA NVDECsupport (default: ON)
BUILD_NVML- build with
NVIDIA Management Library(
NVML) support (default: ON)
WERROR- treat all build warnings as errors (default: OFF)
DALI_BUILD_FLAVOR- Allow to specify custom name sufix (i.e. ‘nightly’) for nvidia-dali whl package
BUILD_JPEG_TURBO- build with
DALI release packages are built with the options listed above set to ON and NVTX turned OFF. Testing is done with the same configuration. We ensure that DALI compiles with all of those options turned OFF, but there may exist cross-dependencies between some of those features.
Following CMake parameters could be helpful in setting the right paths:
FFMPEG_ROOT_DIR - path to installed FFmpeg
NVJPEG_ROOT_DIR - where nvJPEG can be found (from CUDA 10.0 it is shipped with the CUDA toolkit so this option is not needed there)
libjpeg-turbo options can be obtained from libjpeg CMake docs page
protobuf options can be obtained from protobuf CMake docs page
Install Python bindings¶
pip install dali/python
Cross-compiling DALI C++ API for aarch64 Linux (Docker)¶
Support for aarch64 Linux platform is experimental. Some of the features are available only for x86-64 target and they are turned off in this build. There is no support for DALI Python library on aarch64 yet. Some Operators may not work as intended due to x86-64 specific implementations.
Build the aarch64 Linux Build Container¶
docker build -t dali_builder:aarch64-linux -f Dockerfile.build.aarch64-linux .
From the root of the DALI source tree
docker run -v $(pwd):/dali dali_builder:aarch64-linux
The relevant artifacts will be in