dc.description.abstract | This thesis presents a software model that allows a parallel decomposition of the
MPEG-4 video encoder onto shared memory architectures, in order to reduce its
total video encoding time.
Since a video sequence consists of video objects each of which is likely to have
different encoding requirements, the model incorporates a scheduler which
(a) always selects the most appropriate video object for encoding and,
(b) employs a mechanism for dynamically allocating video objects allocation onto
the system processors, based on video object size information.
Further spatial video object parallelism is exploited by applying the single program
multiple data (SPMD) paradigm within the different modules of the MPEG-4
video encoder. Due to the fact that not all macroblocks have the same processing
requirements, the model also introduces a data partition scheme that generates tiles
with identical processing requirements. Since, macroblock data dependencies
preclude data parallelism at the shape encoder the model also introduces a new
mechanism that allows parallelism using a circular pipeline macroblock technique
The encoding time depends partly on an encoder’s computational complexity. This
thesis also addresses the problem of the motion estimation, as its complexity has a
significant impact on the encoder’s complexity. In particular, two fast motion
estimation algorithms have been developed for the model which reduce the
computational complexity significantly. The thesis includes experimental results on a four processor shared memory
platform, Origin200. | en_US |