On Combining Wavefront and Tile Parallelism with a Novel GPU-Friendly Fast Search

Abstract

As the necessity of supporting ever-increasing demands in video resolution leads to new video coding standards, the challenge of harnessing their computational overhead becomes important. Such overhead stems not only from the increased image data due to higher resolutions but also from the coding techniques per se that are introduced by each standard to improve compression. All modern standards in the field of video coding offer high compression efficiency, but this is achieved by increasing the computational complexity of the encoding part. Ultra-High-Definition (UHD) videos, bring new encoding implementation schemes that are being recommended for CPU and GPU parallelization. Therefore, several works are published to achieve better performance and reduce encoding complexity. Following this idea, we proposed and evaluated a hybrid encoding scheme that utilizes the constant growth of the CPU power with the massive GPU popularity in parallel. Taking advantage of the encoding schemes from the leading video coding standards, such as High-Efficiency Video Coding (HEVC) and Versatile Video Coding (VVC), which support parallel processing thru Wavefront or Tiling, in our work, we combined both of them at the same time as a whole, and in addition, we introduced a GPU-friendly fast search algorithm that is highly parallel and alternative to the default non-parallel TZ-Search. Through an experimental evaluation with common test sequences, the proposed GPU Fast Motion Estimation with our previous Wavefront per Tile Parallelism (WTP) was shown to provide valid trade-off between speedup and video coding efficiency, effectively combining the best of two worlds, i.e., WTP using CPUs and parallel Motion Estimation with GPUs.

Fraction Execution Resolver Using a Hybrid Multi-CPU/GPU Encoding Scheme

“…This modified encoder uses tiles as a primary parallelism scheme but internally uses WPP for each tile separately with multiple CPU cores working in a parallel pattern. Extending the original WTP encoder, the GPU's fast search algorithm has been demonstrated (4PhaseFS) [10] to speed up the motion estimation calculations, which has been proven to be one of the most intensive tasks in a video encoder. As in our previous works, to achieve the best performance from our experiments, the application thread pool was fed with a pre-selected number of threads that matched exactly with the physical CPU cores we had available for our setup.…”

“…In contrast, optimized software algorithms have been introduced to reduce the computational load [6][7][8] in variants of existing software video encoders without any additional costs. The proposed work in this paper is also a software algorithm that extends our proven hybrid encoder [9,10] to increase the speedup times by minimizing the calculations of the fraction motion estimation (FME) part by skipping them when specific criteria are met inside a fraction execution resolver (FER) function. Our encoder already has a GPU-friendly fast integer motion estimation algorithm [10], so this extension focuses only on the fraction optimization part.…”

“…The proposed work in this paper is also a software algorithm that extends our proven hybrid encoder [9,10] to increase the speedup times by minimizing the calculations of the fraction motion estimation (FME) part by skipping them when specific criteria are met inside a fraction execution resolver (FER) function. Our encoder already has a GPU-friendly fast integer motion estimation algorithm [10], so this extension focuses only on the fraction optimization part. As we target mostly the end users, we only exploit and capitalize common hardware setups such as video cards.…”

“…We extended our previous hybrid encoding model [9,10] to support the proposed fraction execution resolver (FER) algorithm, which has been proven to be fast, easy to implement, and cost effective, minimizing in this way the computational load and thus the encoding times even further. This encoding model effectively utilizes both CPUs and GPUs in a flexible tiling selection scheme to balance speed and quality.…”

“…So, each tile uses the standard WPP encoding pattern. GPU power is engaged [10] to improve the coding efficiency for each tile separately only for the integer motion estimation (IME) part and falls back to the CPU for the fraction refinement as a subsequent and mandatory job. The fraction part always follows the IME part, but it is very difficult to parallelize the default algorithm because of the dependencies it has.…”