JPEG encoder and decoder library and console application for NVIDIA GPUs for high-performance image encoding and decoding.
This documents provides an introduction to the library and how to use it. You can also look to FAQ.md for additional information. To see latest changes you can display file NEWS.md.
- Martin Srom, CESNET z.s.p.o
- Jan Brothánek
- Petr Holub
- Martin Jirman
- Jiri Matela
- Martin Pulec
- Lukáš Ručka
- uses NVIDIA CUDA platform
- baseline Huffman 8-bit coding
- use of JFIF file format by default, Adobe and SPIFF is supported as well (used by encoder if JPEG internal color space is not representable by JFIF - eg. limited range YCbCr BT.709 or RGB)
- use of restart markers that allow fast parallel encoding/decoding
- Encoder by default creates non-interleaved stream, optionally it can produce an interleaved stream (all components in one scan) or/and subsampled stream.
- support for color transformations and coding RGB JPEG
- Decoder can decompress JPEG codestreams that can be generated by encoder. If scan contains restart flags, decoder can use parallelism for fast decoding.
- command-line tool with support for encoding/decoding raw images as well as PNM/PAM or Y4M
Encoding/Decoding of JPEG codestream is divided into following phases:
Encoding: Decoding
1) Input data loading 1) Input data loading
2) Preprocessing 2) Parsing codestream
3) Forward DCT 3) Huffman decoder
4) Huffman encoder 4) Inverse DCT
5) Formatting codestream 5) Postprocessing
and they are implemented on CPU or/and GPU as follows:
- CPU:
- Input data loading
- Parsing codestream
- Huffman encoder/decoder (when restart flags are disabled)
- Output data formatting
- GPU:
- Preprocessing/Postprocessing (color component parsing, color transformation RGB <-> YCbCr)
- Forward/Inverse DCT (discrete cosine transform)
- Huffman encoder/decoder (when restart flags are enabled)
Source 16K (DCI) image (8, 9) was cropped to 15360x8640+0+0 (1920x1080 multiplied by 8 in both dimensions) and for lower resolutions downscaled. Encoding was done with default values with input in RGB (quality 75, non-interleaved, rst 24-36, average from 99 measurements excluding first iteration) with following command:
gpujpegtool -v -e mediavision_frame_<res>.pnm mediavision_frame_<res>.jpg -n 100 [-q <Q>]
| GPU \ resolution | HD (2 Mpix) | 4K (8 Mpix) | 8K (33 Mpix) | 16K (132 Mpix) |
|---|---|---|---|---|
| GTX 3080 | 0.54 ms | 1.71 ms | 6.20 ms | 24.48 ms |
| GTX 1060M | 1.36 ms | 4.55 ms | 17.34 ms | (low mem) |
| GTX 580 | 2.38 ms | 8.68 ms | (low mem) | (low mem) |
Note: First iteration took 233 ms for 8K on GTX 3080 and scales proportionally with respect to resolution.
Further measurements were performed on GTX 3080 only:
| quality | 10 | 20 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 |
|---|---|---|---|---|---|---|---|---|---|---|
| duration HD (ms) | 0.48 | 0.49 | 0.50 | 0.51 | 0.51 | 0.53 | 0.54 | 0.57 | 0.60 | 0.82 |
| duration 4K (ms) | 1.61 | 1.65 | 1.66 | 1.67 | 1.69 | 1.68 | 1.70 | 1.72 | 1.79 | 2.44 |
| duration 8K (ms) | 6.02 | 6.04 | 6.09 | 6.14 | 6.12 | 6.17 | 6.21 | 6.24 | 6.47 | 8.56 |
| duration 8K (ms, w/o PCIe xfers) | 2.13 | 2.14 | 2.18 | 2.24 | 2.23 | 2.25 | 2.28 | 2.33 | 2.50 | 5.01 |
Decoded images were those encoded in previous section, averaging has been done similarly by taking 99 samples excluding the first one. Command used:
gpujpegtool -v mediavision_frame_<res>.jpg output.pnm -n 100
| GPU \ resolution | HD (2 Mpix) | 4K (8 Mpix) | 8K (33 Mpix) | 16K (132 Mpix) |
|---|---|---|---|---|
| GTX 3080 | 0.75 ms | 1.94 ms | 6.76 ms | 31.50 ms |
| GTX 1060M | 1.68 ms | 4.81 ms | 17.56 ms | (low mem) |
| GTX 580 | 2.61 ms | 7.96 ms | (low mem) | (low mem) |
Note: (low mem) above means that the card didn't have sufficient memory to encode or decode the picture.
Following measurements were performed on GTX 3080 only:
| quality | 10 | 20 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 |
|---|---|---|---|---|---|---|---|---|---|---|
| duration HD (ms) | 0.58 | 0.60 | 0.63 | 0.65 | 0.67 | 0.69 | 0.73 | 0.78 | 0.89 | 1.58 |
| duration 4K (ms) | 1.77 | 1.80 | 1.83 | 1.84 | 1.87 | 1.89 | 1.92 | 1.95 | 2.11 | 3.69 |
| duration 8K (ms) | 6.85 | 6.88 | 6.90 | 6.92 | 6.98 | 6.70 | 6.74 | 6.84 | 7.17 | 12.43 |
| duration 8K (ms, w/o PCIe xfers) | 2.14 | 2.18 | 2.21 | 2.24 | 2.27 | 2.29 | 2.34 | 2.42 | 2.71 | 7.27 |
Following tables summarizes encoding/decoding quality using NVIDIA GTX 580 for non-interleaved and non-subsampled stream with different quality settings (PSNR and encoded size values are averages of encoding several images, each of them multiple times):
| quality | PSNR 4K | size 4K | PSNR HD | size HD |
|---|---|---|---|---|
| 10 | 29.33 dB | 539.30 kB | 27.41 dB | 145.90 kB |
| 20 | 32.70 dB | 697.20 kB | 30.32 dB | 198.30 kB |
| 30 | 34.63 dB | 850.60 kB | 31.92 dB | 243.60 kB |
| 40 | 35.97 dB | 958.90 kB | 32.99 dB | 282.20 kB |
| 50 | 36.94 dB | 1073.30 kB | 33.82 dB | 319.10 kB |
| 60 | 37.96 dB | 1217.10 kB | 34.65 dB | 360.00 kB |
| 70 | 39.22 dB | 1399.20 kB | 35.71 dB | 422.10 kB |
| 80 | 40.67 dB | 1710.00 kB | 37.15 dB | 526.70 kB |
| 90 | 42.83 dB | 2441.40 kB | 39.84 dB | 768.40 kB |
| 100 | 47.09 dB | 7798.70 kB | 47.21 dB | 2499.60 kB |
| quality | PSNR 4K | size 4K | PSNR 4K | size 4K |
|---|---|---|---|---|
| 10 | 29.33 dB | 539.30 kB | 27.41 dB | 145.90 kB |
| 20 | 32.70 dB | 697.20 kB | 30.32 dB | 198.30 kB |
| 30 | 34.63 dB | 850.60 kB | 31.92 dB | 243.60 kB |
| 40 | 35.97 dB | 958.90 kB | 32.99 dB | 282.20 kB |
| 50 | 36.94 dB | 1073.30 kB | 33.82 dB | 319.10 kB |
| 60 | 37.96 dB | 1217.10 kB | 34.65 dB | 360.00 kB |
| 70 | 39.22 dB | 1399.20 kB | 35.71 dB | 422.10 kB |
| 80 | 40.67 dB | 1710.00 kB | 37.15 dB | 526.70 kB |
| 90 | 42.83 dB | 2441.40 kB | 39.84 dB | 768.40 kB |
| 100 | 47.09 dB | 7798.70 kB | 47.21 dB | 2499.60 kB |
To build console application check Requirements and go
to gpujpeg directory (where README.md and COPYING
files are placed) and run cmake command:
cmake -DCMAKE_BUILD_TYPE=Release -Bbuild .
cmake --build build
You can also use autotools to create a build recipe for the library and the application or a plain old Makefile.bkp. However, cmake is recommended.
To build libgpujpeg library check Compile.
To use library in your project you have to include library to your sources and linked shared library object to your executable:
#include <libgpujpeg/gpujpeg.h>
For simple library usage examples you look into subdirectory examples.
For encoding by libgpujpeg library you have to declare two structures and set proper values to them. The first is definition of encoding/decoding parameters, and the second is structure with parameters of input image:
struct gpujpeg_parameters param;
gpujpeg_set_default_parameters(¶m);
param.quality = 80;
// (default value is 75)
param.restart_interval = 16;
// (default value is 8)
param.interleaved = 1;
// (default value is 0)
struct gpujpeg_image_parameters param_image;
gpujpeg_image_set_default_parameters(¶m_image);
param_image.width = 1920;
param_image.height = 1080;
param_image.comp_count = 3; // output JPEG channel count (3 default)
param_image.color_space = GPUJPEG_RGB; // input colorspace (GPUJPEG_RGB
// default), can be also
// eg. GPUJPEG_YCBCR_JPEG
param_image.pixel_format = GPUJPEG_444_U8_P012;
// or eg. GPUJPEG_U8 for grayscale
// (default value is GPUJPEG_444_U8_P012)
If you want to use subsampling in JPEG format call following function, that will set default sampling factors (2x2 for Y, 1x1 for Cb and Cr):
// Use default sampling factors
gpujpeg_parameters_chroma_subsampling(¶m);
Or define sampling factors by hand:
// User custom sampling factors
param.sampling_factor[0].horizontal = 4;
param.sampling_factor[0].vertical = 4;
param.sampling_factor[1].horizontal = 1;
param.sampling_factor[1].vertical = 2;
param.sampling_factor[2].horizontal = 2;
param.sampling_factor[2].vertical = 1;
Next you have to initialize CUDA device by calling:
if ( gpujpeg_init_device(device_id, 0) )
return -1;
where first parameters is CUDA device (e.g. device_id = 0) id and second
parameter is flag if verbose output should be used (0 or GPUJPEG_VERBOSE).
Next step is to create encoder:
struct gpujpeg_encoder* encoder = gpujpeg_encoder_create(0);
if ( encoder == NULL )
return -1;
When creating encoder, library allocates all device buffers which will be needed for image encoding and when you encode concrete image, they are already allocated and encoder will used them for every image. Now we need raw image data that we can encode by encoder, for example we can load it from file:
int image_size = 0;
uint8_t* input_image = NULL;
if ( gpujpeg_image_load_from_file("input_image.rgb", &input_image,
&image_size) != 0 )
return -1;
Next step is to encode uncompressed image data to JPEG compressed data by encoder:
struct gpujpeg_encoder_input encoder_input;
gpujpeg_encoder_input_set_image(&encoder_input, input_image);
uint8_t* image_compressed = NULL;
int image_compressed_size = 0;
if ( gpujpeg_encoder_encode(encoder, &encoder_input, &image_compressed,
&image_compressed_size) != 0 )
return -1;
Compressed data are placed in internal encoder buffer so we have to save them somewhere else before we start encoding next image, for example we can save them to file:
if ( gpujpeg_image_save_to_file("output_image.jpg", image_compressed,
image_compressed_size, NULL) != 0 )
return -1;
Now we can load, encode and save next image or finish and move to clean up encoder. Finally we have to clean up so destroy loaded image and destroy the encoder.
gpujpeg_image_destroy(input_image);
gpujpeg_encoder_destroy(encoder);
For decoding we don't need to initialize two structures of parameters. We only have to initialize CUDA device if we haven't initialized it yet and create decoder:
if ( gpujpeg_init_device(device_id, 0) )
return -1;
struct gpujpeg_decoder* decoder = gpujpeg_decoder_create(0);
if ( decoder == NULL )
return -1;
Now we have two options. The first is to do nothing and decoder will postpone buffer allocations to decoding first image where it determines proper image size and all other parameters (recommended). The second option is to provide input image size and other parameters (reset interval, interleaving) and the decoder will allocate all buffers and it is fully ready when encoding even the first image:
// you can skip this code below and let the decoder initialize automatically
struct gpujpeg_parameters param;
gpujpeg_set_default_parameters(¶m);
param.restart_interval = 16;
param.interleaved = 1;
struct gpujpeg_image_parameters param_image;
gpujpeg_image_set_default_parameters(¶m_image);
param_image.width = 1920;
param_image.height = 1080;
param_image.pixel_format = GPUJPEG_444_U8_P012;
param_image.comp_count = 3;
// Pre initialize decoder before decoding
gpujpeg_decoder_init(decoder, ¶m, ¶m_image);
If you want to specify output image color space and/or subsampling factor,
you can use following two parameters. You can specify them though the
param structure befor passing it to gpujpeg_decoder_init. But if you
postpone this initialization process to the first image, you have no
other option than specify them in this way:
gpujpeg_decoder_set_output_format(decoder, GPUJPEG_RGB,
GPUJPEG_444_U8_P012);
// or eg. GPUJPEG_YCBCR_JPEG and GPUJPEG_422_U8_P1020
Next we have to load JPEG image data from file and decoded it to raw image data:
int image_size = 0;
uint8_t* image = NULL;
if ( gpujpeg_image_load_from_file("input_image.jpg", &image,
&image_size) != 0 )
return -1;
struct gpujpeg_decoder_output decoder_output;
gpujpeg_decoder_output_set_default(&decoder_output);
if ( gpujpeg_decoder_decode(decoder, image, image_size,
&decoder_output) != 0 )
return -1;
Now we can save decoded raw image data to file and perform cleanup:
if ( gpujpeg_image_save_to_file("output_image.rgb", decoder_output.data,
decoder_output.data_size, NULL) != 0 )
return -1;
gpujpeg_image_destroy(image);
gpujpeg_decoder_destroy(decoder);
The console application gpujpeg uses libgpujpeg library to demonstrate it's functions. To build console application check Compile.
To encode image from raw RGB image file to JPEG image file use following command:
gpujpegtool --encode --size=WIDTHxHEIGHT --quality=QUALITY \
INPUT_IMAGE.rgb OUTPUT_IMAGE.jpg
You must specify input image size by --size=WIDTHxHEIGHT parameter.
Optionally you can specify desired output quality by parameter
--quality=QUALITY which accepts values 0-100. Console application accepts
a few more parameters and you can list them by folling command:
gpujpegtool --help
To decode image from JPEG image file to raw RGB image file use following command:
gpujpegtool --decode OUTPUT_IMAGE.jpg INPUT_IMAGE.rgb
You can also encode and decode image to test the console application:
gpujpegtool --encode --decode --size=WIDTHxHEIGHT --quality=QUALITY \
INPUT_IMAGE.rgb OUTPUT_IMAGE.jpg
Decoder will create new decoded file OUTPUT_IMAGE.jpg.decoded.rgb and do
not overwrite your INPUT_IMAGE.rgb file.
Console application is able to load raw RGB image file data from *.rgb
files and raw YUV and YUV422 data from *.yuv files. For YUV422 you must
specify *.yuv file and use --sampling-factor=4:2:2 parameter.
All supported parameters for console application are following:
--help
Prints console application help
--size=1920x1080
Input image size in pixels, e.g. 1920x1080
--pixel-format=444-u8-p012
Input/output image pixel format ('u8', '444-u8-p012', '444-u8-p012z',
'444-u8-p0p1p2', '422-u8-p1020', '422-u8-p0p1p2' or '420-u8-p0p1p2')
--colorspace=rgb
Input image colorspace (supported are 'rgb', 'yuv' and 'ycbcr-jpeg',
where 'yuv' means YCbCr ITU-R BT.601), when *.yuv file is specified,
instead of default 'rgb', automatically the colorspace 'yuv' is used
--quality
Set output quality level 0-100 (default 75)
--restart=8
Set restart interval for encoder, number of MCUs between
restart markers
--subsampled
Produce chroma subsampled JPEG stream
--interleaved
Produce interleaved stream
--encode
Encode images
--decode
Decode images
--device=0
By using this parameter you can specify CUDA device id which will
be used for encoding/decoding.
Restart interval is important for parallel huffman encoding and decoding.
When --restart=N is used (default is 8), the coder can process each
N MCUs independently, and so he can code each N MCUs in parallel. When
--restart=0 is specified, restart interval is disabled and the coder
must use CPU version of huffman coder (because on GPU would run only one
thread, which is very slow).
The console application can encode/decode multiple images by following command:
gpujpegtool ARGUMENTS INPUT_IMAGE_1.rgb OUTPUT_IMAGE_1.jpg \
INPUT_IMAGE_2.rgb OUTPUT_IMAGE_2.jpg ...
To be able to build and run libgpujpeg library and gpujpeg console application you need:
- NVIDIA CUDA Toolkit
- C/C++ compiler + CMake
- NVIDIA drivers
- CUDA enabled NVIDIA GPU (cc >= 2.0; older may or may not work)
- optional OpenGL support:
- GLEW, OpenGL (usually present in Windows, may need headers installation in Linux)
- GLFW or GLX (Linux only) for context creation
- GLUT for OpenGL tests