GPU based Real-time UHD Intra Decoding for AVS3

To encode ultra high definition (UHD) content such as 4K or 8K sequences, AVS3 introduced many complex coding tools, which increased the compression efficiency. However, UHD bitstreams, especially streams encoded with all intra configuration, still have quite high bit rates. To decoding such bitstreams in real-time, a large amount of computing resources are needed. In this paper, an efficient heterogeneous CPU+GPU framework is presented for AVS3 decoding under AI testing configuration. Through efficient asynchronous collaboration mechanism for CPU and GPU, the computation of different computing units and the data transfer between them are highly overlapped. Moreover, to increase GPU utilization, high parallel and low memory access latency schemes are carefully designed for each module in AVS3. The proposed framework archives 113 fps for 4K bitstreams decoding with the NVIDIA GeForce RTX 2080Ti GPU. When the bitrate is up to 300 Mbps, the decoding speed is still higher than 50 fps. For 8K bitstreams decoding, an average frame rate of 47 fps is obtained, which is 78 times faster than HPM 4.0.