Arm and Tencen collaborate to accelerate image encoding and decoding on mobile using SME2 extension
One of the main functions that our mobile phones have today is the photograph. In addition to advanced systems that capture images with higher quality, it is also very useful to reduce the storage that these images occupy. Something that Arm is working on together with Tencent thanks to the technology included in the latest versions of the Arm architecture, we talk about SME2. A technology that takes advantage of neural networks to perform tasks such as image encoding and decoding.
Arm and Tencent accelerate image compression with SME2
Machine learning models are increasingly a fundamental part of the tasks we perform every day, such as image compression. Tencent’s neural imaging codec currently in development is capable of offering a Up to 30% higher compression efficiency compared to other HEIF codecs, maintaining comparable image quality. This technology uses neural networks to compress the image and use less information.
Tencent Media and Arm are collaborating to increase this performance using image neural coding with Arm Scalable Matrix Extension 2 (SME2). At the moment they have managed to reconstruct the image more quickly on mobile devices and others that use Arm architecture and are compatible with this extension.
SME2 reduces decoding time to 30 ms
As shown in an example video, this collaboration has allowed the decoding module to consume less time, remaining in 30mscompared to not using SME2, this time is extended to 61 ms. These times are achieved using 2 CPU threads, if we reduce it to a single thread, with SME2 we achieve 54 ms compared to 95 ms if we have SME2 unused.
This reduction is achieved thanks to the fact that most of the time operations are carried out with matrices. With SME2, multiple die operations can be performed simultaneously, taking advantage of the capabilities of this dedicated engine. Starting from a base of 99.2 ms, the acceleration with matrices has managed to reduce 39.7 ms, the numerical optimization 17.8 ms and the data flow optimization 11.7 ms less. This has left an optimized time of only 30 ms.
An idea that can reduce the amount of data needed for images, always maintaining a quality comparable to other methods used. By taking advantage of SME2 included with the latest Arm architecture, you can make images encode and decode more smoothly, faster and more efficiently.
