Single-ref NEWMV DRL search speed features

[yeqing-wu]

Adds three single-ref NEWMV DRL search speed features, enabled at speed >= 1:

• mv_sf.skip_second_best_subpel: the subpel search normally refines both the best and the second-best full-pel candidate and keeps the lower-variance result; with this feature set, only the best full-pel candidate is refined.

• mv_sf.predict_repeated_newmv: when a previously searched ref_mv_idx has a reference MV within one full pel of the current one, reuse the cached NEWMV result and only recompute the MV rate, avoiding the full-pel and subpel search entirely.

• mv_sf.newmv_drl_search_limit: once ref_mv_idx reaches the limit, reuse the nearest already-searched result regardless of ref-MV distance instead of running a fresh search. Enabled with a cap of 2. Shares the same search-loop pass as the predict_repeated_newmv reuse.

STATS_CHANGED

Anchor: commit 165345b
Speed 1 (cpu-used=1): FG16 CTC (33 frames, class A1 and A2, RA)
Speed 4 (cpu-used=4): FG16 CTC (33 frames, class A1 and A2, RA)

1) Speed 1

  a) Overall performance (all three features combined)

  +------------+------+-------+-------+------+------+------+
  | Class      |    Y |    Cb |    Cr | wAvg | Enc% | Dec% |
  +------------+------+-------+-------+------+------+------+
  | A1         | 0.04 | -0.08 | -0.36 | 0.00 |   99 |  101 |
  | A2         | 0.04 | -0.04 | -0.18 | 0.03 |   98 |  100 |
  | Avg w/o B2 | 0.04 | -0.05 | -0.23 | 0.02 |   98 |  100 |
  +------------+------+-------+-------+------+------+------+

  b) Performance of skip_second_best_subpel only

  +------------+-------+-------+-------+-------+------+------+
  | Class      |     Y |    Cb |    Cr |  wAvg | Enc% | Dec% |
  +------------+-------+-------+-------+-------+------+------+
  | A1         | -0.10 | -0.32 |  0.09 | -0.11 | 99.5 |  101 |
  | A2         | -0.01 | -0.09 | -0.34 | -0.02 | 99.5 |  100 |
  | Avg w/o B2 | -0.04 | -0.16 | -0.21 | -0.05 | 99.5 |  100 |
  +------------+-------+-------+-------+-------+------+------+

  c) Performance of predict_repeated_newmv only

  +------------+-------+-------+-------+-------+------+------+
  | Class      |     Y |    Cb |    Cr |  wAvg | Enc% | Dec% |
  +------------+-------+-------+-------+-------+------+------+
  | A1         |  0.00 | -0.08 |  0.08 |  0.00 |  100 |  100 |
  | A2         | -0.03 | -0.04 | -0.04 | -0.03 |  100 |  100 |
  | Avg w/o B2 | -0.02 | -0.05 | -0.00 | -0.02 |  100 |  100 |
  +------------+-------+-------+-------+-------+------+------+

  d) Performance of newmv_drl_search_limit only

  +------------+------+-------+------+------+------+------+
  | Class      |    Y |    Cb |   Cr | wAvg | Enc% | Dec% |
  +------------+------+-------+------+------+------+------+
  | A1         | 0.05 | -0.17 | 0.09 | 0.03 |   99 |  101 |
  | A2         | 0.00 |  0.36 | 0.01 | 0.02 |   99 |  100 |
  | Avg w/o B2 | 0.01 |  0.20 | 0.03 | 0.02 |   99 |  100 |
  +------------+------+-------+------+------+------+------+

2) Speed 4

  Overall performance (all three features combined)

  +------------+------+-------+-------+------+------+------+
  | Class      |    Y |    Cb |    Cr | wAvg | Enc% | Dec% |
  +------------+------+-------+-------+------+------+------+
  | A1         | 0.01 | -0.13 |  0.07 | 0.00 |   96 |  101 |
  | A2         | 0.05 | -0.13 | -0.25 | 0.03 |   96 |  100 |
  | Avg w/o B2 | 0.04 | -0.13 | -0.15 | 0.02 |   96 |  100 |
  +------------+------+-------+-------+------+------+------+

[yeqing-wu]

Hello all,

Could you please review this PR?

Thanks
Yeqing

[leolzhao]

LGTM.