An efficient dynamic scheduling scheme for H.264/AVC encoding on multi-core architecture

The popular wave front parallelization has been proposed to encode H.264/AVC video employing macro-block level parallelism. This approach, however, fails to achieve an optimum performance due to a significant overhead of barrier-based synchronization. All threads must wait for the slowest ones to complete encoding before starting a next processing wave. In this paper, we propose a dynamic scheduling for parallel encoding without employing the barrier synchronization. In this approach, current threads will 1) keep recursively encoding any new ready macro-blocks, and; 2) load-balance any other ready ones to a global shared queue, which is accessed by all threads. We further propose an adaptive dynamic scheduling to mitigate the access contention at the global queue, and considers cache locality with due to the underlying hierarchical core/cache topology. The adaptive dynamic scheduling employs multiple distributed queues, and schedule tasks. To achieve both locality and load balance, neighboring macro-blocks are preferably scheduled to nearby cores. We design, implement, and evaluate our scheme on a 32-core SGI server. Running real benchmarks, we observe that our scheme outperforms wave front parallelization by 200%, and achieves 65% to 70% of the maximum encoding speedup.