Appendix-IPP-Pass.md

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 is used as a first pass in the multi-pass encoding mode as shown in Figure 1. 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 improve the rate control behavior.

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

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.

Parameter	Description
intra_error	Intra prediction error
coded_error	Best of intra pred error and inter pred error using last frame as ref.
intra_skip_count	Count of blocks where the intra prediction error is very small
image_data_start_row	Start row

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.

Process	Function	Purpose
motion_estimation_kernel	setup_firstpass_data_seg	Initialize the IPP data
motion_estimation_kernel	first_pass_me	Source-based Motion estimation
motion_estimation_kernel	first_pass_frame_seg	Source-based Intra prediction and Motion compensation
motion_estimation_kernel	first_pass_frame_end	Normalize and accumulate statistics
motion_estimation_kernel	svt_av1_end_first_pass	Output statistics when multi-pass encoding mode is used

Details of the Implementation

The IPP pass is enabled when a multi-pass VBR encoding mode is used and the following algorithmic steps are performed 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 2 and 3 summarise the data flow of the IPP pass.

Figure 2. IPP algorithm data flow.

Figure 3. 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 tradeoffs 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.

Flag	Level (Sequence/Picture)	Description
skip_frame_first_pass	Sequence	0: OFF, 1: Skip the processing of some frames in the IPP pass
ds	Sequence	0: OFF, 1: Use down-sampled input picture in the IPP search algorithms
bypass_blk_step	Sequence	0: OFF, 1: Bypass every other row of blocks and every other column of blocks in the IPP search algorithm
dist_ds	Sequence	0: OFF, 1: Use down-sampled distortion in the IPP search algorithms
bypass_zz_check	Sequence	0: OFF, 1: Bypass the (0,0)_MV check against HME_MV before performing motion estimation
use8blk	Sequence	0: OFF, 1: Use 8x8 blocks instead of 16x16 blocks in the IPP search algorithms
reduce_me_search	Sequence	0: OFF, 1: Reduce HME and ME search area size

Note that 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 v1.8.0 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.