Parallel AMVP candidate list construction for HEVC

Advanced motion vector prediction (AMVP) is one of the most important inter prediction coding tools adopted in the state-of-the-art HEVC coding standard, which does great effect on the coding efficiency. However, the current AMVP design is highly sequential and thus restricts the throughput both on the encoder and the decoder sides. To facilitate the parallel processing and enlarge the throughput, a parallel AMVP candidate list (AMVPCL) construction solution is proposed. The proposed parallel scheme consists of a three level fine granularity solutions. The first level is a CU-based approach and it constructs AMVPCL of all PUs in the same CU in parallel. The second level is also at CU level but it generates a single set of AMVPCL for all PUs inside a CU. Specifically, we only apply this method to 8×8 CU to balance the parallelism degree and rate-distortion performance. The third level is a CU-group based approach, in which AMVPCL of all PUs in the same CU-group are constructed in parallel. Experimental results show the proposed algorithm can efficiently achieve parallel motion estimation with negligible 0.0%~1.3% BD-rate loss at different degree of parallelism.