AVIF Speed/Quality Test Condition
(a) originalPNG / 582.9kb
(b) compressedjpeg / 16.2kb
psnr=34.19, ssim=0.914, vmaf=81.4
(d) compressedpulsar avif / 7.1kb
psnr=35.49, ssim=0.927, vmaf=77.6
Codec
Command line
| Codec | Code base | Hash |
|---|---|---|
libjpeg-turbo | a JPEG codec to accelerate libjpeg by 2-6x | a01857cff6a2b7b3283ef1d0ecb6b01d06b597f8 |
avifenc | https://github.com/AOMediaCodec/libavif v0.11.1
| 6ab53189045e7a6fe0bd93d14977b2a4f8efa5e9 |
pulsar-avif | BLUEDOT's FPGA-based AVIF encoder | 5307f4605513110ad3c267fac6873e1deb913385 |
libjpeg-turbo./cjpeg -quality ${quality} -sample 2x2 -outfile ${output jpg} ${input ppm}avifenc with AOM./avifenc ${input png} --min ${qp} --max ${qp} -y 420 -d 8 -j ${thread} -r l -c aom ${output avif}avifenc with SVT./avifenc ${input png} --min ${qp} --max ${qp} -y 420 -d 8 -j ${thread} -r l -c svt ${output avif}pulsar-avif./avif_sp_client -i ${input yuv} -o ${output avif} --width ${width} --height ${height} --enable-avif 1 --avif-seq 0 --fps 1 --frames 1 --qp ${qp}
The test images are in the format of jpg. For the evaluation, we used decoded images of them. Therefore, the speed measured here does not include a time for decoding jpg and converting color space.
Test Images
https://drive.google.com/drive/folders/1VK12pBFnYJzJRD1k8uo38yammO-_XtkO
- Kodak - 24 images, 768x512; Kodak Lossless True Color Image Suite
- CLIC - 303 images, 2048x1320; from workshop and challenges on learned image compression(CLIC)
Quality Metrics and Coding Efficiency
BD-Rate for the following metrics:
- PSNR_Y
- PSNR_AVG: (PSNR_Y * 4 + PSNR_U + PSNR_V) / 6
- SSIM
- VIF
ResultsCoding EfficiencyInterpretation: For example, when avifenc(aom) has the same PSNR_Y score as libjpeg-turbo, avifienc(aom)'s file size is 56.21% smaller than libjpeg-turbo.The coding efficiency of pulsar-avif is not much different from the software AVIF encoders.
ExamplesCLIC, 2048x1320, john-price-5258Kodak, 512x768, kodim09
(a) originalpng / 3,604kb
(b) compressedjpeg / 132kb
psnr=40.7dB, ssim=0.964, vmaf=89.96
(c) compressedaom avif / 66kb
psnr=41.93, ssim=0.973, vmaf=88.82
(d) compressedpulsar avif / 67kb
psnr=42.01, ssim=0.972, vmaf=88.62
(a) originalpng / 582.9kb
(b) compressedjpeg / 16.2kb
psnr=34.19, ssim=0.914, vmaf=81.4
(c) compressedaom avif / 6.2kb
psnr=34.99, ssim=0.923, vmaf=75.7
(d) compressedpulsar avif / 7.1kb
psnr=35.49, ssim=0.927, vmaf=77.6
Encoding SpeedHardware Platforms
| Encoder | Microsoft Azure VM | Specs. |
|---|---|---|
pulsar-avif | AMD Alveo U250 1ea + Intel(R) Xeon(R) Platinum 8171 CPU @ 2.6GHz, 10 cores | |
avifenc, | D32s | Intel(R) Xeon(R) Platinum 8370C CPU @ 2.8GHz, 32 cores |
Kodak (24 images of 768x512)
Note. When we tested pulsar-avif with 100 times of encoding 24 images, the average frame rate was increased to 582 fps.
- Encoding 24 images simultaneously (24 processes)
- Each software encoding process is executed with 4 threads. (-j 4)
| time [ms] | fps | CPU utilization | |
|---|---|---|---|
pulsar-avif | 58.6 | 409.85 | 305% |
avifenc(aom) | 405 | 59.26 | 3200% |
avifenc(svt) | 1325 | 18.11 | 3200% |
CLIC (303 images of 2048x1320)
Note. When we tested pulsar-avif with 100 times of encoding 303 images, the average frame rate was increased to 93.5 fps.
Note. For the evaluation of the performance of pulsar-avif, we developed and used a daemon where a queue is embedded. The CPU utilization reported above is mostly taken by the daemon and client processes waiting for allocation of AVIF encoding hardware resources. For example, the 305% utilization to encode 24 Kodak images in pulsar-avif consists of 65% from daemon and 240% from 24 client processes (10% per process).
- avifienc(aom): encoding 32 images simultaneously with -j 8 to encode 303 images
- avifenc(svt): encoding 16 images simultaneously with -j 4 to encode 303 images
| time [ms] | fps | CPU utilization | |
|---|---|---|---|
pulsar-avif | 3267 | 92.74 | 570% |
avifenc(aom) | 16191 | 18.71 | 3200% |
avifenc(svt) | 65826 | 4.60 | 3200% |
Note. For the evaluation of the performance of pulsar-avif, we developed and used a daemon where a queue is embedded. The CPU utilization reported above is mostly taken by the daemon and client processes waiting for allocation of AVIF encoding hardware resources. For example, the 305% utilization to encode 24 Kodak images in pulsar-avif consists of 65% from daemon and 240% from 24 client processes (10% per process).
How we determined the number of threads for each processWe picked up threads# where CPU utilization was not increased anymore, even if more threads were specified in the command line for encoding a single image.