Top level

IPP Encoding Pass Appendix

1. Description of the Algorithm

The IPP encoding pass of SVT-AV1 encoder is a light encoding process that processes all of the input video pictures and collects statistics while consuming negligible time as compared to the time it takes to run the final encoding. The IPP pass makes use of a flat IPP… prediction structure where the input pictures are used as reference pictures. The IPP encoding pass can be used as a first pass in the multi-pass encoding mode as shown in Figure 1, or as a statistics generation tool based on a lookahead window when it is enabled in the single pass encoding mode as shown in Figure 2. In both encoding modes, when IPP pass is enabled, the input data from the source clip is analyzed and stored in a local buffer. The collected data from the IPP pass is then used to achieve the best encoding complexity-quality trade-offs. The collected data may be used to select the suitable prediction structure for the next encoding pass or may be used to improve the rate control behavior.

Figure 1. Multi-pass encoding mode with IPP pass.

Figure 2. Single pass encoding mode with look-a-head.

2. Implementation of the Algorithm

2.1. IPP Pass Inputs/Outputs

Input to open_loop_first_pass: Source streams.

Outputs of open_loop_first_pass: Statistics per frame.

Table 1 summarises the output statistic variables.

Table 1: Frame-level IPP pass output statistics.

2.2. IPP Pass API

Table 2 summarises the invoked functions when IPP pass is enabled. The process where each function is called is indicated as well as a brief description of each function.

Table 2: IPP pass main function calls.

Details of the Implementation

The IPP pass is enabled when a multi-pass encoding mode is used or when the lookahead is enabled. When the IPP pass is enabled, the following algorithmic steps are considered for each input picture:

1. Initialisation of all statistics structures and variables.
2. If a reference picture is not enabled yet, as in the example of the first picture in the input sequence, then an intra DC prediction takes place and the statistics are updated accordingly.
3. If at least one reference picture is available then, motion estimation takes place followed by an intra DC prediction and motion compensation. According to the generated prediction errors and the motion information, the prediction and coding statistics are updated for each block.
4. After processing all blocks within a picture, the generated statistics are accumulated, normalized and stored for further use.

Figures 3 and 4 summarises the data flow of the IPP pass.

Figure 3. IPP algorithm data flow.

Figure 4. IPP pass localisation in the encoder pipeline.

3. Optimization of the Algorithm

The IPP pass algorithm can be configured in different ways to achieve its best quality-complexity tradoffs depending on the application. The control parameters used to configure the IPP pass are described in Table 3.

Table 3: Control flags associated with the IPP pass.

Note that in the multi-pass encoding mode, the encode pipeline of the IPP pass has been modified to involve only the necessary encoder kernels, such motion estimation kernel, and all unnecessary kernels such as Rate Control, Mode Decision, Encode-decode, Deblocking, CDEF, Restoration, Entropy Coder… are bypassed.

Notes

The feature settings that are described in this document were compiled at v1.1.0 of the code and may not reflect the current status of the code. The description in this document represents an example showing how features would interact with the SVT architecture. For the most up-to-date settings, it's recommended to review the section of the code implementing this feature.