An efficient parallel motion estimation algorithm and X264 parallelization in CUDA

Enhanced multicore–manycore interaction in high-performance video encoding

“…Such works clearly highlight that, when compared to the fully accelerated version of x264 or WebM VP8, the known collaborative GPU-based techniques do not exhibit significant compute performance improvements [21]. For instance, [20] reports a speed improvement of only 20% with respect to x264.…”

“…In particular, the open source X264 [8] and WebM [2] projects, implementing the widely used H.264 and VP8/VP9 schemes, respectively, can obtain an impressive compute performance. For instance, X264 can achieve up to a ×1000 speed-up compared to the JM H.264 standard reference encoder [20,23]. Such results have been obtained by thoroughly exploiting the optimized SSE/MMX/AVX assembly instruction set, which enables instruction-level parallelism (ILP) on register vectors up to 512-bit-wide.…”

“…According to the results reported in literature, GPU-based schemes can achieve significant compute performance improvements with respect to CPU-only software implementations. However, most of the algorithms proposed in literature are compared to the corresponding standard reference encoder, which is typically not optimized at all (such as the JM encoder for H.264), and often implements the still prohibitive (from the computational point of view) full search ME on CPU [20]. Conversely, to the best of the authors' knowledge, only few works compare the achieved compute performance to the one provided by the fully accelerated version (that is, a version exploiting efficiently all ILP features, like MMX/SSE/AVX) of x264 or WebM VP8 encoders.…”

“…Conversely, to the best of the authors' knowledge, only few works compare the achieved compute performance to the one provided by the fully accelerated version (that is, a version exploiting efficiently all ILP features, like MMX/SSE/AVX) of x264 or WebM VP8 encoders. In particular, in [20,23] GPU-accelerated H.264 encoders have been compared to x264, while [17,24] report some preliminary results about a GPU-accelerated VP8 version. Such works clearly highlight that, when compared to the fully accelerated version of x264 or WebM VP8, the known collaborative GPU-based techniques do not exhibit significant compute performance improvements [21].…”

Design of an application specific instruction set processor for a universal bitstream codec

“…There are also some approaches that accelerate codec execution by using multicore CPUs or both CPUs and GPUs [14]. They effectively run most codec functions, but not the EC which is known to be hard to parallelize.…”