Abstract
H.264/AVC [1] is an extremely popular international standard of digital vieo compression in recent years, which is developed by ITU-T VCEG (Video Coding Experts Group) and ISO/IEC MPEG (Moving Picture Experts Group ) . H.264 / AVC has higher computed complexity compared with previous standard, so we use the asymmetric dual-core SoC platform to implement H.264/AVC decoder. On the dual-core platform, there are three points that shoud be think about with a view to implement efficient H.264/AVC decoder 1. software partition: Which procedures should execute on MPU? Which procedures should execute on DSP? Average allocation needs ore synchronizing control. 2. Data movement: dual-core needs to do a lot of data exchange, for example the restructed data that DSP decoded need to be moved to external memory in order to display later. We adopt DMA or MPU to move the materials. 3. synchronization: MPU and DSP both have some procedures which shoulde be executed in the specific opportunity, so need to do synchronizing control. We adopt polling or interrupt. We consider this three points to implement three h.264 decoder programming models on dual-core SoC platform and utilize software pipeline to increase parallelism. The following is three programming models that we put forward 1. MPU decode full Entropy - polling programming model 2. MPU decode full Entropy - interrupt programming model 3. MPU decode partial Entropy - interrupt programming model In our experiment environment, we are running an embedded linux on MPU. MPU decode full Entropy - polling programming model is relatively unsuitable to use DMA, because the setup of DMA shoule execte in kernel mode. However, the procedure and buffer of MPU decode full Entropy - polling programming model are all in user space, so using DMA need to spend a lot of time copying data and doing system call. Consequently, in the fist programming model we don’t use DMA to move data. The experimental results show that the time that DSP spend on decoding is only 1.8s. Other time is spended on waiting data. Therfore, MPU decode full Entropy - Interrupt programming model can improve the efficiency of decoding, but need additional overhead of onterrup processing and data copy between user space and kernel space. In MPU decode partial Entropy - interrupt programming model, the code size of DSP instruction is too big. It leads to a lot of cache misses and makes decoding time increase.The number of cache misses can be reduced through the design of decoding flow.