x265 version 3.1 is out with attractive performance and quality improvements. The key new additions include integration of SVT-HEVC library into the x265 library, improved support for interlaced encoding, and better encoder control to ensure correct rendering of HDR content. Check out our release notes at https://x265.readthedocs.io/en/3.1/releasenotes.html for more details.

Happy compressing!

By  Praveen Kumar Karadugattu and Kalyan Goswami

In the x265 Encoder ver 3.0, the default parameters for “veryslow” and “slower” presets are changed. In this blog we will discuss this modification in detail.

The all new ‘veryslow’ Preset

The “veryslow” preset is one of the most preferred presets by the offline encoding vendors, where the video quality is paramount. Hence, the quality of encoded bitstream is the key feature for this preset than the encoding speed. Hence, this preset targets at improving the quality of the encoded video at a given bitrate while performing extensive computations. This preset can be enabled by specifying the CLI option “–preset veryslow” or “–preset 8” with x265. After performing a bunch of evaluations and finally come up with the following changes in the default settings of the four key parameters in x265 ver 3.0 under the “veryslow” preset.

The significance of these modified params is described below.

1. “–limit-refs” implies that x265 will limit references for a PU at a given CU-depth by using information from its higher depth and/or other PU modes at the same depth. Previously, the default value for this preset was set as “1”, which restrict analysis at the current depth, based on references used to code four sub-blocks at the next depth. For example, a 16×16 CU will only use the references used to code its four 8×8 CUs. In version 3.0 of x265, the restriction for limiting the references is relaxed.
2. “–limit-modes” restricts the mode analysis for each CU using cost metrics from the 4 sub-CUs. When multiple inter-modes, such as “–rect” and/or “–amp” are enabled, this feature uses motion cost heuristics from the 4 sub-CUs to bypass modes that are unlikely to be the best choice. It significantly improves the encoding performance, when “–rect” and/or “–amp” are enabled bit with a minimal loss of compression efficiency. In x265 version 3.0, this restriction on mode analysis is relaxed.
3. During motion estimation, the encoder merges the motion of neighbouring blocks to predict the motion of the current block. The CLI option “–max-merge” defines a maximum number of such neighbour candidate blocks (spatial and temporal) that the encoder may consider for merging motion predictions. If a merge candidate results in no residual, it is immediately coded as a “skip”. Otherwise, the merge candidates are tested when searching for the least cost inter-mode option.
4. “–limit-tu” enables an early exit from TU depth recursion, for inter coded blocks. Previously, the “–limit-tu” was 4, which means the 1st sub-TU depth is taken as the limiting depth for the other sub-TUs. In x265 version 3.0, this feature is disabled under the “veryslow” preset.

All the above-mentioned features have a direct impact on the visual quality (both subjective and objective) of the encoded bitstream, but with the penalty of slightly increased encoding time. We have performed a series of experiments to evaluate the speed vs quality trade-off with the new “veryslow” preset against the previous version.

The all-new ‘slower’ Preset

Since most of our OTT clients use “veryslow” preset, the burden of longer encoding times might hamper their application. In order to address this issue, we have redefined the new “slower” preset to match the quality of the previous “veryslow” preset. Hence, the “slower” preset of x265 version 3.0 should provide the same visual quality as the older versions’ “veryslow” preset while maintaining the same encoding speed (fps) as before.

Experimental Results

The sequences that we have chosen for our experiments are listed in Table 1. We have used 4 different resolutions with different source FPS (not to be confused with the encoding fps) and bit-depth. All the experiments are carried out on a Dual-Xeon Linux-based server with 56 cores and 112 threads in a a sequential order.

Table 1: Test Sequences used

In Table 2 the improvements in quality under the “slower” and “veryslow” presets in version 3.0 over version 2.9 are shown. We have computed the BD-PSNR and BD-SSIM metrics to represent the improvement in encoding efficiency in the x65 version 3.0 by making the version 2.9 as the anchor. The positive values of BD-PSNR and BD-SSIM imply the improvement of encoding efficiency (quality). The values in each row of this table indicate the average value of all the test samples under a particular resolution. From this table, it is quite clear that the new version of x265 has significant improvement in quality compared to the older one for these presets.

Table 2: Quality improvement with “slower” and “veryslow” presets in the x265 ver 3.0 over ver 2.9

We can clearly see that there is a significant improvement in the visual quality of the “veryslow” preset of x265 ver 3.0 over the “veryslow” preset of x265 ver 2.9.

As pointed out in the previous section, the improvement in encoding efficiency is expected to slow down the performance (encoding fps). Hence, it is important to gauge the effect of the newer presets on performance. Table 3 shows the performance results of x265 version 3.0 in comparison with those of version 2.9 under the “slower” and “veryslow” presets. These are the encoding FPS values averaged under each resolution.

Table 3: Performance decrement for “slower’ and ‘veryslow” presets in the x265 ver 3.0 over ver 2.9

From Table 3 it is evident that the new “veryslow” and “slower” presets are significantly slower than the older version. However, the performance drop with the new “slower” from the older “veryslow” (represented under the column “ΔFPS% (slower_3.0 vs veryslow_2.9)”) is quite negligible. Hence, for the customers who need maintain the same performance as the “veryslow” preset of x265 version 2.9, are recommended to use the new “slower” preset of x265 version 3.0.

Conclusion

The new “veryslow” preset of x265 ver 3.0 gives much higher compression efficiency compared to the old “veryslow” of ver 2.9, however, it comes with a significant drop in the performance (encoding fps). Hence, we have moved the old “veryslow” to the new “slower” in ver 3.0 which retains a similar compression efficiency as the old “veryslow” (of ver 2.9) without sacrificing much of the performance (encoding speed in fps).

P.S:

Both the versions of x265 can be downloaded from here.

The CLIs we ran to perform these tests are:

x265 –psnr –ssim –input <input.yuv> –input-res <width>x<height> –fps <fps> –bitrate <target bitrate in kbps> –preset slower –output “out_slower.hevc”

x265 –psnr –ssim –input <input.yuv> –input-res <width>x<height> –fps <fps> –bitrate <target bitrate in kbps> –preset veryslow –output “out_veryslow.hevc”

With changeset a41325fc854f, the x265 library can invoke the SVT-HEVC library for encoding through the —svt option. We have mapped presets and command-line options supported by the x265 application into the equivalent options of SVT-HEVC, and have added a few specific options that are available only when the SVT-HEVC library is invoked. This page in our documentation describes the steps to build, and invoke the SVT-HEVC library in more detail.

Our reason for this integration was to enable our users to evaluate additional relative trade-offs between performance and compression efficiency while working behind the familiar API of the x265 library. In the long term, we plan to leverage this integration to further improve x265’s ability to handle real-time and low turn-around scenarios in pure software; this is the space that SVT-HEVC was focused on. In parallel, we will continue to innovate on our flagship presets that are used in offline encoding where x265 dominates.  You can expect to see these changes in the coming releases of x265, increasing the reach of open-source for video compression!

We are happy to announce now have version 3.0 of x265. The main focus of this version is to improve the quality, especially for the ‘veryslow’ and ‘slower’ presets. Moreover, Dolbyvision is included in this version. The detail description of all the new features and releases is available in our release notes.

By Bhavna Hariharan

x265’s ability to leverage AI to accelerate encoding Adaptive Bitrate Streaming (ABR) has been presented in the paper titled “Adaptive Multi-Resolution Encoding Scheme for ABR Streaming” that is to appear in the proceedings of IEEE International Conference on Image Processing (ICIP), 2018. Adaptive streaming enables dynamic adaptation to changes in network conditions by encoding video content in multiple bitrates and resolutions. Multi-pass x265 encodes exploit the structural redundancy across multiple resolutions by sharing analysis data in the order of increasing resolution, thereby reducing the computational burden significantly.

Static and dynamic refinement features were conceptualized for the Adaptive streaming topology in x265. As the first step, the input video sequence is scaled down to the lowest resolution of the bitrate ladder and is encoded. Coding metadata such as CU quadtree structure, PU predictions, coding modes and motion vectors (MVs) are stored during this first pass encode at CTU level abstraction. The subsequent higher-resolution encodes invoke either the static or the dynamic refinement algorithm. The figure below depicts this architecture for a three pass system. Theoretically, this can be extended to N passes.

The static refinement algorithm may further be classified as intra-picture and inter-picture refinement wherein the information from the previous pass is used at different levels of granularity thereby giving the current encode a head start. The encoder intelligently refines the analysis data based on certain heuristics and later uses it in the Rate Distortion Optimization (RDO) process. By tweaking the degree of information that is being reused, the user can control the performance vs quality balance of the encode.

The dynamic refinement algorithm applies Bayesian classification on the information saved in the previous pass and identifies patterns between the complexity of the content and the refinement level being used. Based on these patterns, the dynamic refinement algorithm switches between the static refinement levels. This will allow the encoder to find the optimal balance between performance and quality while taking the content complexity into account, making it more efficient than the static refinement methods.

Compared to conventional single-resolution approaches, the computational complexity of higher resolution encodes is drastically reduced with the use of the AI-based adaptive multi-resolution technique. The adaptive multi-resolution scheme is able to achieve a whopping speedup of about 2.5X, with a negligible drop in the quality. The above table from the research paper compares the speed vs quality of static and dynamic refinement levels. Since the AI tools used in x265 are not based on CNN models, x265 is able to leverage the advances in CPU technology to achieve these improvements.