Real-time layered video compression using SIMD computation

Abstract

We present the design and implementation of a high performance software layered video codec, designed for deployment in bandwidth heterogeneous networks. The codec facilitates layered spatial and SNR (signal-to-noise ratio) coding for bit-rate adaption to a wide range of receiver capabilities. The codec uses a wavelet subband decomposition for spatial layering and a discrete cosine transform combined with repeated quantization for SNR layering. Through the use of the Visual Instruction Set on SUN’s UltraSPARC platform we demonstrate how SIMD parallel image processing enables layered real-time software encoding and decoding. The codec partitions our 384 × 320 × 24-bit test video stream into 21 layers at a speed of 39 frames per second and reconstructed at 28 frames per second. The Visual Instruction Set accelerated encoder stages are about 3-4 times as fast as an optimized C version. We find that this speedup is well worth the extra implementation effort.