JPEGGPU is an experimental JPEG decoder implemented in CUDA. It works by decoding many sequences of the encoded stream in parallel, and then synchronizing these decoded sequences. This process is based on the paper Accelerating JPEG Decompression on GPUs1.
- Implements DCT-based baseline JPEGs, see JPEG Support below for more details.
- Flexible API: no implicit synchronization between host and device in the decoding process, explicit device memory management allowing reuse, thread safety, C99 compatible, and OS independent.
- Simple library design: JPEG application segments are not interpreted, i.e. no attempt is made to interpret color space. No attempt is made to support non-standard JPEGs (no EOF marker, table index out of bounds, etc.).
Build with CMake, for example:
cmake -S . -B build
cmake --build buildexample/example_tool.c is built as jpeggpu_example. It demonstrates basic usage of jpeggpu and outputs some information about the file.
./build/jpeggpu_example in.jpg
marker Start of image
marker Define quantization table(s)
marker Define restart interval
restart_interval: 252
marker Baseline DCT
size_x: 4032, size_y: 3024, num_components: 3
c_id: 1, ssx: 2, ssy: 2, qi: 0
c_id: 2, ssx: 1, ssy: 1, qi: 1
c_id: 3, ssx: 1, ssy: 1, qi: 1
...
marker End of image
intra sync of 89 blocks of 256 subsequences
intra sync of 1 blocks of 131072 subsequences
gpu decode done
decoded image at: out.pngbenchmark/benchmark.cpp builds jpeggpu_benchmark that compares performance with nvJPEG.
Possible output with AMD Ryzen 5 2600 and NVIDIA GeForce RTX 2070, 12MP 4:2:0 image with restart intervals, (chunk_size = 32):
./build/jpeggpu_benchmark in.jpg
throughput (image/s) | avg latency (ms) | max latency (ms)
jpeggpu singlethread 176.18 5.68 6.72
nvJPEG singlethread 71.82 13.92 15.51
nvJPEG batch 25 22.31 44.82 1123.72
nvJPEG batch 50 22.69 44.07 2214.65
nvJPEG batch 75 22.89 43.68 3303.78
nvJPEG 4 threads 268.95 14.69 39.53
nvJPEG 5 threads 332.87 14.91 39.72
nvJPEG 6 threads 386.19 15.19 35.15
nvJPEG 7 threads 389.12 15.92 33.67
nvJPEG 8 threads 447.42 16.44 33.00
nvJPEG 9 threads 493.39 17.16 33.90
nvJPEG 10 threads 531.90 17.85 35.43
nvJPEG 11 threads 568.07 18.78 49.18
nvJPEG 12 threads 620.45 19.21 45.74
nvJPEG 13 threads 603.87 20.59 55.99
nvJPEG 14 threads 609.44 22.21 66.65test/test.cpp builds jpeggpu_test that compares output against nvJPEG. Helper script test.sh uses ImageMagick to convert an input image to a few different JPEG variations.
./build/jpeggpu_example test.jpg --write_out # writing out is optional
component 0 MSE: 0.23201 component 1 MSE: 0.198817 component 2 MSE: 0.199355
writing out to "test.jpg.nvjpeg.png" and "test.jpg.jpeggpu.png"
./build/test.sh test.jpg # can also optionally pass --write_out
creating tmp file test.jpg.1x1.jpg..
component 0 MSE: 0.202032 component 1 MSE: 0.155791 component 2 MSE: 0.155672
creating tmp file test.jpg.2x1.jpg..
...JPEGGPU implements the full baseline process (see Table 12), with the extension of allowing up to four Huffman tables of each type:
- DCT-based process
- 8-bit samples within each component
- Sequential
- Huffman coding: 4 AC and 4 DC tables
- 1, 2, 3, or 4 components
- Interleaved and non-interleaved scans
Compared to nvJPEG, JPEGGPU does not support progressive JPEGs but has no restrictions on chroma subsampling. One estimate suggests 30% of JPEGs used in websites are progressive and 10% of JPEG photographs are progressive3. The parallel decoding method used in JPEGGPU is fundamentally incompatible with progressive JPEGs, specifically because of the AC refinement scan.