Appendix-DLF.md

Deblocking Loop Filter Appendix

1. Description of the algorithm

The deblocking loop filter is used to address blocking artifacts in reconstructed pictures. The filter was developed based on VP9 deblocking filter. The filter switches between multiple filters according to the minimum transform size across the edge and according to whether the areas around the edge are relatively flat. For a given block, vertical edges are filtered first, followed by the filtering of the horizontal edges. The main idea behind the filter can be summarized as follows:

• Pixels on either side of the boundary are not filtered if they show a clear difference in values, i.e. the edge is most likely a true edge and not a blockiness artifact.

• If the pixels on both sides of the boundary are to be filtered, the selection of the filter is performed according to the guidelines presented in Table 1:

Table 1. Summary of the loop filter decisions.

Min TX size across the edge	Flat areas around the edge	Filter (Luma)	Filter_Size (Luma)	Filter (Chroma)	Filter_Size (Chroma)
>TX_8x8	Yes	Filter14 (13-tap)	14	Filter6 (5-tap)	6
	No	Filter8 (7-tap)/ Filter4	4	Filter4	4
TX_8x8	Yes	Filter8 (7-tap)	8	Filter6 (5-tap)	6
	No	Filter4	4	Filter4	4
TX_4x4	-	Filter4	4	Filter4	4

A block diagram illustrating the flow of the different steps involved in filtering is given in Figure 1 below.

Figure 1. Steps in the loop filter decisions making process.

The steps shown in the block diagram above are now discussed in more detail.

Determine the loopfilter level and sharpness

Both the loop filter level and sharpness are frame-level parameters.

• Loop filter level: The loop filter level takes value in [0, 63] and can be set using different methods: (see the functions av1_pick_filter_level and search_filter_level)

• Set to 0 to disable filtering.

• Set as a function of AC quantization step size.

• Set through the evaluation of filtering using different level values and selecting the level that yields the lowest distortion in the filtered frame.

• Sharpness: The sharpness parameter takes value in [0, 7] and is an input parameter to the encoder (--sharpness).

Filter Strength Parameters: lvl, limit, blimit, thresh

• lvl: At the block level, the loopfilter level is referred to as lvl. The parameter lvl builds on the loopfilter level for the frame and includes refinements based on segmentation, coding mode and reference picture. The lvl parameter is computed as follows:

  lvl = (filter_level for the frame) + (segment delta) + (mode delta) + (reference delta)

where

• filter_level is the loopfilter level for the frame that has already been determined.
• Segment delta is specified on an 8x8 basis for each of the eight segments.
• Mode delta and reference delta are determined as follows:
  • Define scale = 1 << (((filter_level for the frame) + (segment delta)) >>5 )
  • mode delta = mode_deltas * scale, where mode_deltas is obtained from Table 2.
  • Reference delta = ref_deltas * scale, where ref_deltas is obtained from Table 3 below.

Table 2. Mode deltas for the loop filter level.

Intra modes	mode_deltas	Inter modes	mode_deltas
DC_PRED	0	NEARESTMV	1
V_PRED	0	NEARMV	1
H_PRED	0	GLOBALMV	0
D45_PRED	0	NEWMV	1
D135_PRED	0	NEAREST_NEARESTMV	1
D113_PRED	0	NEAR_NEARMV	1
D157_PRED	0	NEAREST_NEWMV	1
D203_PRED	0	NEW_NEARESTMV	1
D67_PRED	0	NEAR_NEWMV	1
SMOOTH_PRED	0	NEW_NEARMV	1
SMOOTH_V_PRED	0	GLOBAL_GLOBALMV	1
SMOOTH_H_PRED	0	NEW_NEWMV	1
PAETH_PRED	0

Table 3. Reference deltas for the loop filter level.

Reference Picture	Default ref_deltas
INTRA_FRAME	1
LAST_FRAME	0
LAST2_FRAME	0
LAST3_FRAME	0
BWDREF_FRAME	0
GOLDEN_FRAME	-1
ALTREF2_FRAME	-1
ALTREF_FRAME	-1

• shift: The shift parameter is computed as follows:

  • If sharpness > 4, shift = 2.
  • Otherwise, if sharpness > 0, shift = 1.
  • Otherwise, shift = 0.

• limit

  • If sharpness > 0, limit = Clip3( 1, 9 - sharpness, lvl >> shift ).
  • Otherwise, limit = Max( 1, lvl >> shift ).

• blimit = 2 * (lvl + 2) + limit.

• Thresh = lvl >> 4.

Identify edges to filter

The edges to filter should satisfy the following conditions:

• Transform unit edges AND

• (non-zero level on either side of the edge) AND ((Non-skip inter blocks on either side of the edge) OR block edge).

Filtering Process for luma

The decisions on the filtering operation is based on a number of masks used to evaluate the level of difference between the samples across the edge to be filtered. The filtering masks are outlined first, followed by a description of the filtering decision making process.

Filter Masks. The filter masks to be discussed are Filter_Mask, High Edge Variance Mask (Hev_Mask), Flat_Mask, Flat_Mask2

• Filter_Mask: Indicates whether the edge is a true edge or an artifact, and consequently whether the tested samples are to be filtered or not. (See the functions filter_mask3_chroma and filter_mask)

  Original idea:

  • Let n = 2 when filter_length = 6 and n = 3 when filter_length = 8 or 14.

    If

    • (ABS( – ) > limit, i=1,…,n; OR

    • (ABS( – ) > limit, i=1,…,n; OR

    • (ABS( - ) * 2 + ABS( - ) / 2 > blimit)

    then the edge is most likely a true edge. In that case, do not filter the tested samples and set Filter_Mask to zero. Otherwise, Filter_Mask is set to 1 and the tested samples are to be filtered.

• Hev_Mask: Used to identify edges with large change in pixel values on either side of the edge. (See the function hev_mask)

  If ABS(p1-p0) > thresh OR ABS(p1-p0) > thresh, then Hev_Mask = 1, else Hev_Mask = 0.

• Flat_Mask: Considered when Filter_Length >= 6. Indicates whether samples 0,…,n on each side of the boundary belong to relatively flat areas, where n = 2 for Filter_length = 6 and n=3 when Filter_length = 8 or 14

  Flat_Mask = 1 when the following conditions are true:

  • abs( - ) <= thresh, i=1,…,n; AND

  • abs( - ) <= thresh, i=1,…,n

  Otherwise, Flat_Mask = 0.

• Flat_Mask2: Considered when Filter_Length = 14. Indicates whether samples 4, 5 and 6 on each side of the boundary belong to relatively flat areas.

  Flat_Mask2 = 1 when the following conditions are true: ABS( - ) <= thresh, i=4,…,6; AND ABS( - ) <= thresh, i=4,…,6. Otherwise, Flat_Mask2 = 0.

Filtering decision making process

The steps involved in the filtering operation are as follows:

• Determine min_ts = The smaller of the two transform sizes on either side of the edge (e.g. min(TX_16x16, TX_8x8) = TX_8x8)

• Determine filter masks: hevMask, filterMask, flatMask and flatMask2.

• If filterMask == 0, no filtering takes place.

• Otherwise, if ((min_ts == TX_4x4) OR (flatMask == 0)), then use filter4.

• Otherwise, if ((min_ts == TX_8x8) OR (flatMask2 == 0)), then use filter8.

• Otherwise, use filter14.

The filtering decisions are outlined in the diagram shown in Figure 2 below.

Figure 2. Flow of the loop filter decision making process.

The different filters that could be considered in the filtering operation are outlined below. Figure 3 below indicates the positions of the samples across the horizontal edge to be filtered, with similar arrangement of the samples for the case of a vertical edge.

Figure 3. Sample positions across the horizontal edge to be filtered.

Filter4: Modifies up to two samples on each side of the boundary, depending on High Variance Edge Mask (Hev_Mask). Rough outline of the main idea:

• Hev_Mask = 1 only q0 and p0 are filtered.
  • Delta = (( – ) + 3(-))/8
  • q0 q0 – Delta; p0 p0 + Delta
• Hev_Mask = 0 q0, q1, p0 and p1 are filtered.
  • Delta = 3(q0-p0)/8
  • q0 q0 - Delta; p0 p0 + Delta
  • q1 q1 - Delta/2; p1 p1+Delta/2

Filter6

• 5-tap filter: [1, 2, 2, 2, 1]
• Applies to chroma planes only.
• Modifies two samples on each side of the edge. (See the function filter6)
  • p1 (p2 * 3 + p1 * 2 + p0 * 2 + q0 + 4)>>3;
  • p0 (p2 + p1 * 2 + p0 * 2 + q0 * 2 + q1 + 4)>>3;
  • q0 (p1 + p0 * 2 + q0 * 2 + q1 * 2 + q2 + 4)>>3;
  • q1 (p0 + q0 * 2 + q1 * 2 + q2 * 3 + 4)>>3;

Filter8

• 7-tap filter: [1, 1, 1, 2, 1, 1, 1]
• Applies to luma plane only.
• Modifies three samples on each side of the edge. (See the function filter8)
  • p2 (p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0 + 4)>>3;
  • p1 (p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1 + 4)>>3;
  • p0 (p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2 + 4)>>3;
  • q0 (p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3 + 4)>>3;
  • q1 (p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3 + 4)>>3;
  • q2 (p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3 + 4)>>3;

Filter14

• 13-tap filter: [1, 1, 1, 1, 1, 2, 2, 2, 1, 1, 1, 1, 1]
• Applies to luma plane only.
• Modifies six samples on each side of the edge. (See the function filter14)
  • p5 (p6 * 7 + p5 * 2 + p4 * 2 + p3 + p2 + p1 + p0 + q0 + 8)>>4,
  • p4 (p6 * 5 + p5 * 2 + p4 * 2 + p3 * 2 + p2 + p1 + p0 + q0 + q1 + 8)>>4
  • p3 (p6 * 4 + p5 + p4 * 2 + p3 * 2 + p2 * 2 + p1 + p0 + q0 + q1 + q2 + 8)>>4
  • p2 (p6 * 3 + p5 + p4 + p3 * 2 + p2 * 2 + p1 * 2 + p0 + q0 + q1 + q2 + q3 + 8)>>4
  • p1 (p6 * 2 + p5 + p4 + p3 + p2 * 2 + p1 * 2 + p0 * 2 + q0 + q1 + q2 + q3 + q4 + 8)>>4
  • p0 (p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 * 2 + q0 * 2 + q1 + q2 + q3 + q4 + q5 + 8)>>4
  • q0 (p5 + p4 + p3 + p2 + p1 + p0 * 2 + q0 * 2 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + 8)>>4
  • q1 (p4 + p3 + p2 + p1 + p0 + q0 * 2 + q1 * 2 + q2 * 2 + q3 + q4 + q5 + q6 * 2 + 8)>>4
  • q2 (p3 + p2 + p1 + p0 + q0 + q1 * 2 + q2 * 2 + q3 * 2 + q4 + q5 + q6 * 3 + 8)>>4
  • q3 (p2 + p1 + p0 + q0 + q1 + q2 * 2 + q3 * 2 + q4 * 2 + q5 + q6 * 4 +8)>>4
  • q4 (p1 + p0 + q0 + q1 + q2 + q3 * 2 + q4 * 2 + q5 * 2 + q6 * 5 + 8)>>4
  • q5 (p0 + q0 + q1 + q2 + q3 + q4 * 2 + q5 * 2 + q6 * 7 + 8)>>4

2. Implementation

Inputs: Block mode, transform size and reconstructed picture.

Outputs: Filtered frame, filter parameters.

Controlling tokens/flags:

Table 4. List of loop filter control flags.

Flag	Level	Description
--enable-dlf	Configuration	Deblocking loop filter control (0:ON (Default), 1: OFF)
DlfCtrls	Picture	Describes the Dlf control signal.
combine_vert_horz_lf	Picture	When set, it implies performing filtering of vertical edges in the current SB followed by filtering of horizontal edges in the preceding SB in the same SB row. When OFF, it implies performing filtering of vertical edges in the current SB followed by filtering of horizontal edges in the same SB.

Figure 4. High-level encoder process dataflow with DLF feature.

The high level dataflow of deblocking loop filter in SVT-AV1 is shown in Figure 4. The main steps involved in the implementation of the algorithm are outlined below, followed by more details on some of the important functions.

The loop filtering operation consists of the following three main steps:

• Initializing the loop filter parameters (limits and thresholds).

• Choosing the optimal loop filter levels

• Applying loop filtering to the frame

The steps above are performed only when DlfCtrls is enabled. Otherwise, loop filtering is not applied to the frame. The details of the three steps mentioned above are outlined in the following.

Step 1: Initialize the loop filter limits and thresholds (av1_loop_filter_init)

• Initialize lfi->lfthr[lvl].lim = block_inside_limit, for 0<=lvl<=63 (update_sharpness)

• Initialize lfi->lfthr[lvl].mblim = *2 * (lvl + 2) +* block_inside_limit, for 0<=lvl<=63 (update_sharpness)

• Initialize lfi->lfthr[lvl].hev_thr = (lvl >> 4), for 0<=lvl<=63 where block_inside_limit is given by

  • If sharpness_lvl > 0, block_inside_lim = Clip3(1, 9 – sharpness_lvl, lvl >> shift).
  • Otherwise, block_inside_limit = Max(1, lvl >> shift ).

  Moreover, lf->combine_vert_horz_lf = 1, implying that both vertical and horizontal filtering are to be considered.

Step 2: Choosing the optimal loop filter levels (av1_pick_filter_level)

In this step, a search is performed for the best loop filter level to work with. The loop filter levels are:

filter_level[0]: Loop filter level for luma vertical edge filtering.

filter_level[1]: Loop filter level for luma horizontal edge filtering.

filter_level_u: Loop filter level for Cb edge filtering.

filter_level_v: Loop filter level for Cr edge filtering.

• if method == LPF_PICK_MINIMAL_LPF

  • Set filter_level[0] = filter_level[1] = 0;

• else if method >= LPF_PICK_FROM_Q

  • generate filt_guess based on the quantization parameter, the encoder bit depth, the frame type and the picture plane. Update filter_level[0], filter_level[1], filter_level_u, filter_level_v;

• else {

  • Get the last frame filter levels filter_level[0], filter_level[1], filter_level_u, filter_level_v.

  • For each of the picture data planes, perform a binary search for the best filter level for the picture data plane (search_filter_level)

    • Set the filter level filt_mid to the frame level for the last frame, and the filter search step to 4 if (fil\_mid < 16), otherwise it is set to filt_mid/4.

    • Filter the frame with filter level set to filt_mid and evaluate the SSE for the filtered frame (try_filter_frame and then svt_av1_loop_filter_frame. See below for more details on the two functions), update the best SSE best_err and corresponding filter level filt_best.

    • Search for the best filter level, starting with filter level (filt_mid-filter_step) or (filt_mid+filter_step), depending on the search direction (try_filter_frame). Keep track of the best filtering SSE and filter level, as well as the search direction. At each iteration, the best filter level becomes the search starting point for the next iteration. If the best filter level in the current iteration is the same as in the previous iteration, halve the filter_step.

    • Return the best filter level and corresponding cost.

Step 3: Applying loop filtering to the frame based on the selected loop filter parameters (svt_av1_loop_filter_frame).

More details on (try_filter_frame)

(try_filter_frame) is just an intermediate function to prepare for (svt_av1_loop_filter_frame), mainly setting the filter levels, computing the filtering sse, and resetting the recon buffer. Returns the filtering SSE.

More details on (svt_av1_loop_filter_frame)

The function calls that start at svt_av1_loop_filter_frame are indicated in Figure 5 below according to the depth of the function call.

Figure 5. Function calls starting at eb_av1_loop_filter_frame.

The main steps involved in are outlines as follows.

1. (svt_av1_loop_filter_frame_init)

   • For the given plane, loop over all segments (i.e. segments as defined by the segmentation feature in AV1 specifications) in the picture
     • Loop over the filtering directions (vertical and horizontal)
       • Adjust the level calculations for each segment to account for the level deltas related to each segment, reference pictures, encoding modes (intra or inter).

2. Loop over all superblocks in the picture and filter each superblock (loop_filter_sb)

   • Perform combined filtering of both vertical edges in the current superblock and filtering of horizontal edges in the preceding superblock in the same superblock row OR Perform filtering of all vertical edges in the superblock followed by filtering all the horizontal edges in the same superblock. (av1_filter_block_plane_vert) and (av1_filter_block_plane_horz).
   • (av1_filter_block_plane_vert) [The description for (av1_filter_block_plane_horz) is similar, except that the filtering would be applied to horizontal edges].

     • Loop over rows of 4x4 blocks in the superblock

       • For each block that intersects the current row of 4x4 blocks (set_lpf_parameters)

         • Determine the transform size to work with for the vertical edges. (get_transform_size)

           • For luma plane, If inter block, then tx_size = tx_depth_to_tx_size[0][mbmi->block_mi.sb_type], else tx_size = tx_depth_to_tx_size[mbmi->tx_depth][mbmi->block_mi.sb_type]. Otherwise, tx_size is determined through the function call (av1_get_max_uv_txsize).
           • For luma plane, If inter block and no skip, then tx_size = tx_depth_to_tx_size[mbmi->tx_depth][mbmi->block_mi.sb_type]
           • tx_size is ultimately set to the width the transform block.

         • Determine the loop filter level to use (get_filter_level_delta_), which accounts for the loop filter level deltas associated with segmentation, reference pictures and encoding modes.

         • If not at the picture left or top boundaries, filtering is to be considered if the filter level for the current or the previous 4x4 blocks are non-zero, and [the current or the previous 4x4 blocks are inter non-skip blocks, or the edge is a CU edge]. Under these conditions, min_ts, the minimum of the transform sizes associated with the current and previous 4x4 blocks, is considered. The selection of the filter length depends on the data plane and min_ts, as indicated in the Table above.

     • Apply the selected filter to the four samples along the vertical edge.

3. Return the frame filtering sse for the loop filter level and the picture data plane being considered.

3. Optimization of the algorithm

The algorithmic optimization of the loop filter is performed by considering different loop filter search methods. If LPF_PICK_FROM_Q is chosen as the search method, the filter levels are determined using the picture qindex, however is LPF_PICK_FROM_FULL_IMAGE is selected, a binary search is performed to find the best filter levels. Table 5 shows the DLF control signals and their descriptions. The encoder mode and the picture being used as a reference are used to determine the search method.

Table 5. DLF control signals description.

Signal	Description
enabled	0/1: Enable/Disable DLF
sb_based_dlf	0: perform picture-based DLF with LPF_PICK_FROM_FULL_IMAGE search method 1: perform DLF per SB using LPF_PICK_FROM_Q method

4. Signaling

The loop filter parameters are signaled at the frame level and include the following parameters: filter_level[0], filter_level[1], filter_level_u, filter_level_v and sharpness_level, seen in Table 6.

Table 6. Frame level loop filter parameters signaled in the bitstream.

Parameters	Values
filter_level[0]	{0,…,63}
filter_level[1]	{0,…,63}
filter_level_u	{0,…,63}
filter_level_v	{0,…,63}
sharpness_level	{0,…,7}

Notes

The feature settings that are described in this document were compiled at v0.9.0 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.

References

[1] Zhijun Lei, Srinath Reddy, Victor Cherepanov, and Zhiping Deng, “GPGPU Implementation of VP9 Inloop Deblocking Filter and Improvements for AV1 CODEC,” International Conference on Image Processing, pp. 925-929, 2017.

[2] Jingning Han, Bohan Li, Debargha Mukherjee, Ching-Han Chiang, Adrian Grange, Cheng Chen, Hui Su, Sarah Parker, Sai Deng, Urvang Joshi, Yue Chen, Yunqing Wang, Paul Wilkins, Yaowu Xu, James Bankoski, “A Technical Overview of AV1,” Proceedings of the IEEE, vol. 109, no. 9, pp. 1435-1462, Sept. 2021.