[chengchen-google]

Enable fast two-pass (dry + wet) superblock partition search on intra-only/key frames in GOOD mode for speeds 1-6, controlled by part_sf.intra_two_pass_partition_search and resolution-/speed-adaptive part_sf.intra_two_pass_max_qindex (using 8-bit-equivalent qindex):

• Dry pass (SB_DRY_PASS): search luma partition tree down to BLOCK_16X16 (re-splitting unsplit blocks up to 32x32 in the wet pass for qindex_8b < 150 in RA / 210 in AI), cap TX partition sweep to NONE + SPLIT, and skip chroma SDP, IntraBC, FSC, secondary MRL lines, palette, and DIP.
• Wet pass (SB_WET_PASS): evaluate PARTITION_NONE on >= 32x32 blocks that the dry pass split (try_none_with_forced_split), reuse luma PARTITION_NONE for chroma when luma sub-blocks are >= bsize, and keep full intra mode search at wet-pass leaf blocks.

CTC Random Access performance (17f on A1, 33f on A2-A5), Anchor is 24ff468:

Speed    Set     BD-Rate (YUV)   Speedup   Ratio
  1      A1       0.10%           4.65%    46.6
  1      A2       0.01%           0.80%   111.2

Speed    Set     BD-Rate (YUV)   Speedup   Ratio
  2      A1       0.08%           3.00%    39.0
  2      A2       0.01%           0.25%    44.7

Speed    Set     BD-Rate (YUV)   Speedup   Ratio
  3      A1       0.06%           2.83%    46.7
  3      A2       0.00%           0.13%    48.9

Speed    Set     BD-Rate (YUV)   Speedup   Ratio
  4      A1       0.34%           8.17%    23.8
  4      A2       0.27%           6.01%    22.3

Speed    Set     BD-Rate (YUV)   Speedup   Ratio
  5      A1       0.71%          12.83%    18.1
  5      A2       0.27%           8.89%    32.9

Speed    Set     BD-Rate (YUV)   Speedup   Ratio
  6      A1       0.22%           4.56%    20.7
  6      A2       0.14%           1.92%    13.7

STATS_CHANGED

[yeqing-wu]

Hi @chengchen-google, thanks for supporting the two-pass partition into key/intra frame. I’ve also implemented this, but I haven’t achieved a good tradeoff, and I don’t have time to tune it. It’s great to see you implement this and speed up the encoder.

[yeqing-wu]

LGTM, thanks

[leolzhao]

This MR looks good to me. Please rebase the code and fix the conflicts.

[yeqing-wu]

@chengchen-google, there is an error because of documentation, please address it.

[urvangjoshi]

@chengchen-google : looks like there is a conflict again. Could you rebase one more time?

[chengchen-google]

@chengchen-google : looks like there is a conflict again. Could you rebase one more time?

Done.