Key Algorithms Of H.264 Encoder On A Stream Architecture

With the appearance of stream computing and stream application, stream architecture is becoming one of the hotspots in microprocessor research. Stanford University in America has proposed new stream architecture, and has developed Imagine prototype stream processor. Imagine has achieved good performance with its relatively simple hardware in media processing and signal processing domains.H.264, as a new video coding standard, aims at high quality and high compression of video picture. And its performance has attracted academia and business circles. However, H.264 requires large and intensive computation. Thus, it is eager to find an efficient implementation for H.264 coding. Stream architecture is promising to make H.264 real-time encoder with its strong support for media processing.This paper introduces Imagine stream processor and H.264 coding standard and describes the features of kernel modules in H.264 framework. Then this paper explains our implementation of H.264 encoding on Imagine. Finally, this paper gives performance analysis of our H.264 encoder. The main work of this paper is present as follow:First, we improve on classical algorithm according to Imagine's 8-cluster SIMD structure, and map H.264 video encoder onto Imagine based on the idea of stream processing. We explain our implementation in terms of data stream organization, cluster distribution, inter-cluster communication and parallelism mode, and try to provide a reference for stream programmers.Second, we show that our H.264 encoder can achieve higher than lOOfps for up to Standard Definition (SD) image format. And we analyze the experimental results in detail, including bandwidth hierarchy, ALU utilization, kernel run-time, SIMD efficiency and so on, and find out the shortcoming in algorithm or implementation.Third, we propose a new SKMD mode in our MASA stream architecture. We map motion estimation algorithm in this mode, and achieve 3.6 times speedup compared with that in SIMD mode.Our experiment proves that H.264 coding is suitable for Imagine implementation. So Imagine can be used to implement different video coding standards such as MPEG-2 and H.264 with good performance and flexibility. Our work is also a little contribution for Imagine's application research.