NVIDIA nvJPEG2000#
The nvJPEG2000 library accelerates the decoding and encoding of JPEG2000 images on NVIDIA GPUs. The library is built on the CUDA platform and is supported on Pascal+ GPU architectures.
Note
Throughout this document, the terms “CPU” and “Host” are used synonymously. Similarly, the terms “GPU” and “Device” are synonymous.
Decoder#
The library utilizes both CPU and GPU for decoding. Tier 2 decode stage (first stage of decode; please refer to the JPEG2000 specification for details) is run on the CPU. All other stages of the decoding process are offloaded to the GPU.
The nvJPEG2000 decoder supports the following:
JPEG2000 Options:
Up to 16 bits per component for integer outputs
FP16 and FP32 outputs, for streams carrying NLT type 3 on every component (which requires the reversible 5-3 wavelet); FP16 requires BDnlt of 16 on every component, FP32 requires 16 or 32
No of components : up to 16384 (the CSiz limit); device memory scales linearly with the component count
Reversible(5-3) and irreversible(9-7) wavelet transforms
Multiple tiles per image
Up to 32 layers per image
Chroma sub-sampling types - 444, 422, and 420, on the first three components only; components at index 3 and above must not be sub-sampled
All progression orders
Tile/Resolution based decoding
Partial decoding of tiles and images
jp2 file format and jpeg2000 codestream
High throughput jpeg2000(All code-blocks have to be HT code-blocks with one HT set and no refinement)
Features:
Decode multiple images in pipeline mode. Refer to nvJPEG2000-Decoder-Pipelined sample for details.
Color conversion to RGB.
Both planar and interleaved outputs are supported.
Encoder#
The library utilizes both GPU and CPU to create JPEG2000 bitstreams. The encode APIs require that the input image is on device memory, and the compressed bitstream will be written to host memory.
The nvJPEG2000 encoder supports the following:
JPEG2000 options
Up to 16 bits per component for integer inputs
FP16 and FP32 inputs, through the NLT type 3 transform (requires high throughput jpeg2000 and the reversible 5-3 wavelet); BDnlt is 16 for FP16, and 16 or 32 for FP32
No of components : up to 16384 (the CSiz limit); device and pinned memory scale linearly with the component count
Reversible(5-3) and irreversible(9-7) wavelet transforms
Multiple tiles per image
Single layer per image
All progression orders
Chroma sub-sampling types - 444, 422, and 420, on the first three components only; components at index 3 and above must not be sub-sampled
jp2 file format and jpeg2000 codestream
High throughput jpeg2000
- Lossy encode using one of the following options
Quantization step
Q-Factor (JPEG like quality)
PSNR
Features
Both planar and interleaved inputs are supported.
Prerequisites#
CUDA Toolkit version 12.0 and above
CUDA Driver version r525 and above
Platforms Supported#
Linux versions:
Distribution |
OS Version |
Default GCC |
GLIBC |
|---|---|---|---|
x86_64 |
|||
RHEL 10 |
10 |
14.3.1 |
2.39 |
RHEL 9 |
9 |
11.5.0 |
2.34 |
RHEL 8 |
8 |
8.5.0 |
2.28 |
Rocky Linux 10 |
10 |
14.3.1 |
2.39 |
Rocky Linux 9 |
9 |
11.5.0 |
2.34 |
Rocky Linux 8 |
8 |
8.5.0 |
2.28 |
Amazon Linux 2023 |
2023 |
11.4.1 |
2.34 |
MSFT Azure Linux |
3 |
13.2.0 |
2.38-8 |
Ubuntu 26.04 LTS |
26.04 |
15.2.0 |
2.43 |
Ubuntu 24.04 LTS |
24.04 |
14.3.0 |
2.39 |
Ubuntu 22.04 LTS |
22.04 |
12.3.0 |
2.35 |
OpenSUSE Leap 16 |
16 |
15.1.1 |
2.40 |
OpenSUSE Leap 15 |
15 |
7.5.0 |
2.38 |
SUSE SLES 16 |
16 |
15.1.1 |
2.40 |
SUSE SLES 15 |
15 |
7.5.0 |
2.38 |
Debian 13 |
13 |
14.2.0 |
2.41 |
Debian 12 |
12 |
12.2.0 |
2.36 |
Fedora 44 |
44 |
15.2.1 |
2.43 |
KylinOS V11 |
V11 |
12.3.1 |
2.38 |
arm64 systems (sbsa) |
|||
RHEL 10 |
10 |
14.3.1 |
2.39 |
RHEL 9 |
9 |
11.5.0 |
2.34 |
RHEL 8 |
8 |
8.5.0 |
2.28 |
Amazon Linux 2023 |
2023 |
11.4.1 |
2.34 |
MSFT Azure Linux |
3 |
13.2.0 |
2.38-8 |
Ubuntu 26.04 LTS |
26.04 |
15.2.0 |
2.43 |
Ubuntu 24.04 LTS |
24.04 |
13.4.0 |
2.39 |
Ubuntu 22.04 LTS |
22.04 |
11.4.0 |
2.35 |
OpenSUSE Leap 16 |
16 |
15.1.1 |
2.40 |
SUSE SLES 16 |
16 |
15.1.1 |
2.40 |
SUSE SLES 15 |
15 |
7.5.0 |
2.38 |
Debian 13 |
13 |
14.2.0 |
2.41 |
Debian 12 |
12 |
12.2.0 |
2.36 |
KylinOS V11 |
V11 |
12.3.1 |
2.38 |
arm64 sbsa Jetson (dGPU + iGPU with OpenRM/nvgpu) |
|||
Ubuntu 24.04 LTS Rel39 (JP7.x) |
24.04 |
13.3.0 |
2.39 |
Windows Versions
Windows 10/11.
Tegra
arm64-sbsa builds are supported on Thor starting CUDA 13.0.
arm64-sbsa builds are supported on Orin starting CUDA 13.2.
aarch64-jetson binaries based on CUDA 12 are available for Orin.
Thread Safety#
Not all nvJPEG2000 types are thread safe. The following should be instantiated separately for each thread: nvjpeg2kStream_t, nvjpeg2kDecodeState_t, nvjpeg2kDecodeParams_t, nvjpeg2kEncodeState_t, and nvjpeg2kEncodeParams_t.
nvjpeg2kHandle_t and nvjpeg2kEncoder_t can be shared by multiple threads.
Custom allocators used to create a handle must be thread safe, because every thread that uses the handle, directly or through an encoder
created from it with nvjpeg2kEncoderCreate, calls those allocators.
Changing the configuration of a handle, for example with nvjpeg2kSetDeviceMemoryPadding or nvjpeg2kSetPinnedMemoryPadding,
is not synchronized with the library’s use of that configuration. The padding is read every time the library sizes a buffer allocated
through the handle, whether or not the buffer then grows. Both the decoder and the encoder do this:
Decoder:
nvjpeg2kDecodeStateCreate, andnvjpeg2kDecode,nvjpeg2kDecodeImageandnvjpeg2kDecodeTileon a decode state created from the handle.Encoder:
nvjpeg2kEncodeStateCreateandnvjpeg2kEncodeon an encoder created from the handle withnvjpeg2kEncoderCreate.
When multiple threads reuse a handle, change its configuration under a lock that those threads also hold around their nvJPEG2000 calls that use the handle, directly or through such an encoder or decode or encode state. Alternatively, configure the handle before sharing it between threads.