nvImageCodec

The nvImageCodec is an open-source library of accelerated codecs with unified interface. It is designed as a framework for extension modules which delivers codec plugins.

This nvImageCodec release includes the following key features:

• Unified API for decoding and encoding images
• Batch processing, with variable shape and heterogeneous formats images
• Codec prioritization with automatic fallback
• Builtin parsers for image format detection: jpeg, jpeg2000, tiff, bmp, png, pnm, webp
• Python bindings
• Zero-copy interfaces to CV-CUDA, PyTorch and CuPy
• End-end accelerated sample applications for common image transcoding

Currently there are following native codec extensions:

• nvjpeg_ext

  • Hardware jpeg decoder
  • CUDA jpeg decoder
  • CUDA lossless jpeg decoder
  • CUDA jpeg encoder

• nvjpeg2k_ext

  • CUDA jpeg 2000 decoder (including High Throughput Jpeg2000)
  • CUDA jpeg 2000 encoder

• nvbmp_ext (as an example extension module)

  • CPU bmp reader
  • CPU bmp writer

• nvpnm_ext (as an example extension module)

  • CPU pnm (ppm, pbm, pgm) writer

Additionally as a fallback there are following 3rd party codec extensions:

• libturbo-jpeg_ext

  • CPU jpeg decoder

• libtiff_ext

  • CPU tiff decoder

• opencv_ext

  • CPU jpeg decoder
  • CPU jpeg2k_decoder
  • CPU png decoder
  • CPU bmp decoder
  • CPU pnm decoder
  • CPU tiff decoder
  • CPU webp decoder

Pre-requisites

This section describes the recommended dependencies to use nvImageCodec.

• Linux distro:
  • x86_64
    • Debian 11, 12
    • Fedora 39
    • RHEL 8, 9
    • OpenSUSE 15
    • SLES 15
    • Ubuntu 20.04, 22.04
    • WSL2 Ubuntu 20.04
  • arm64-sbsa
    • RHEL 8, 9
    • SLES 15
    • Ubuntu 20.04, 22.04
• NVIDIA driver >= 520.56.06
• CUDA Toolkit > = 11.8
• Python >= 3.8

Install nvImageCodec library

You can download and install the appropriate built binary packages from the nvImageCodec Developer Page or install nvImageCodec Python from PyPI as it is described below.

nvImageCodec Python for CUDA 11.x

pip install nvidia-nvimgcodec-cu11

nvImageCodec Python for CUDA 12.x

pip install nvidia-nvimgcodec-cu12

Optional installation of nvJPEG library

If you do not have CUDA Toolkit installed, or you would like install nvJPEG library independently, you can use the instructions described below.

Install the nvidia-pyindex module

pip install nvidia-pyindex

Install nvJPEG for CUDA 11.x

pip install nvidia-nvjpeg-cu11

Install nvJPEG for CUDA 12.x

pip install nvidia-nvjpeg-cu12

Documentation

NVIDIA nvImageCodec Documentation

Build and install from Sources

Additional pre-requisites

• GCC >= 9.4
• cmake >= 3.18
• patchelf >= 0.17.2
• Dependencies for extensions. If you would not like to build particular extension you can skip it.
  • nvJPEG2000 >= 0.7.0
  • libjpeg-turbo >= 2.0.0
  • libtiff >= 4.5.0
  • opencv >= 4.7.0
• Python packages:
  • clang==14.0.1
  • wheel
  • setuptools
  • sphinx_rtd_theme
  • breathe
  • future
  • flake8
  • sphinx==4.5.0

Please see also Dockerfiles.

Build

git lfs clone https://github.com/NVIDIA/nvImageCodec.git
cd nvimagecodec
git submodule update --init --recursive --depth 1
mkdir build
cd build
export CUDACXX=nvcc
cmake .. -DCMAKE_BUILD_TYPE=Release
make

Build CVCUDA samples

To build CV-CUDA samples, additionally CV-CUDA has to be installed and CVCUDA_DIR and NVCV_TYPES_DIR need to point folders with *-config.cmake files. Apart of that, BUILD_CVCUDA_SAMPLES variable must be set to ON.

Build Python wheel

After succesfully built project, execute below commands.

cd build
make wheel

Packaging

From a successfully built project, installers can be generated using cpack:

cd build
cpack --config CPackConfig.cmake -DCMAKE_BUILD_TYPE=Release

This will generate in build directory *.zip or *tar.xz files

Installation from locally built packages

Tar file installation

tar -xvf nvimgcodec-0.2.0.0-cuda12-x86_64-linux-lib.tar.gz -C /opt/nvidia/

DEB File Installation

sudo apt-get install -y ./nvimgcodec-0.2.0.0-cuda12-x86_64-linux-lib.deb

Python WHL File Installation

pip install nvidia_nvimgcodec_cu12-0.2.0-py3-none-manylinux2014_x86_64.whl

Installation from sources

Linux

cd build
cmake --install . --config Release -prefix /opt/nvidia/nvimgcodec_<major_cuda_ver>

After execution there should be:

• all extension modules in /opt/nvidia/nvimgcodec_cuda<major_cuda_ver>/extensions (it is default directory for extension discovery)
• libnvimgcodec.so in /opt/nvidia/nvimgcodec_cuda<major_cuda_ver>/lib64

Add directory with libnvimgcodec.so to LD_LIBRARY_PATH

export LD_LIBRARY_PATH=/opt/nvidia/nvimgcodec_cuda<major_cuda_ver>/lib64:$LD_LIBRARY_PATH

Testing

Run CTest to execute L0 and L1 tests

cd build
cmake --install . --config Release --prefix bin
ctest -C Release

Run sample transcoder app tests

cd build
cmake --install . --config Release --prefix bin
cd bin/test

LD_LIBRARY_PATH=$PWD/../lib64 pytest -v test_transcoder.py

Run Python API tests

First install python wheel. You would also need to have installed all Python tests dependencies (see Dockerfiles).

pip install nvidia_nvimgcodec_cu12-0.2.0.x-py3-none-manylinux2014_x86_64.whl

Run tests

cd tests
pytest -v ./python

CMake package integration

To use nvimagecodec as a dependency in your CMake project, use:

list(APPEND CMAKE_PREFIX_PATH "/opt/nvidia/nvimgcodec_cuda<major_cuda_ver>/")  # or the prefix where the package was installed if custom

find_package(nvimgcodec CONFIG REQUIRED)
# Mostly for showing some of the variables defined
message(STATUS "nvimgcodec_FOUND=${nvimgcodec_FOUND}")
message(STATUS "nvimgcodec_INCLUDE_DIR=${nvimgcodec_INCLUDE_DIR}")
message(STATUS "nvimgcodec_LIB_DIR=${nvimgcodec_LIB_DIR}")
message(STATUS "nvimgcodec_BIN_DIR=${nvimgcodec_BIN_DIR}")
message(STATUS "nvimgcodec_LIB=${nvimgcodec_LIB}")
message(STATUS "nvimgcodec_EXTENSIONS_DIR=${nvimgcodec_EXTENSIONS_DIR}")
message(STATUS "nvimgcodec_VERSION=${nvimgcodec_VERSION}")

target_include_directories(<your-target> PUBLIC ${nvimgcodec_INCLUDE_DIR})
target_link_directories(<your-target> PUBLIC ${nvimgcodec_LIB_DIR})
target_link_libraries(<your-target> PUBLIC ${nvimgcodec_LIB})