[YixinDu1]

Run the full-frame TIP work, that is the TIP frame build, its loop filter and the SSE of the TIP searches, on the tile encoder's worker threads. The band builds and band SSEs of the direct output rank search stay on the main thread. The work is split into row or column jobs that write disjoint output or integer partial sums, so the bitstream does not depend on the thread count. It uses the existing tile workers and stays serial until they exist.

Anchor: commit 224c347
CTC: AV2 CTC v8 (33 frames, RA)

Speed 4 (cpu-used=4)
  +--------------+-------+-------+-------+-------+-------+--------+
  | Class        |     Y |    Cb |    Cr |  wAvg |  Enc% |  Ratio |
  +--------------+-------+-------+-------+-------+-------+--------+
  | A1           |  0.00 |  0.00 |  0.00 |  0.00 |  98.2 |    N/A |
  | A2           |  0.00 |  0.00 |  0.00 |  0.00 | 100.0 |    N/A |
  | Avg w/o B2   |  0.00 |  0.00 |  0.00 |  0.00 |  99.9 |    N/A |
  +--------------+-------+-------+-------+-------+-------+--------+

Speed 5 (cpu-used=5)
  +--------------+-------+-------+-------+-------+-------+--------+
  | Class        |     Y |    Cb |    Cr |  wAvg |  Enc% |  Ratio |
  +--------------+-------+-------+-------+-------+-------+--------+
  | A1           |  0.00 |  0.00 |  0.00 |  0.00 |  99.0 |    N/A |
  | A2           |  0.00 |  0.00 |  0.00 |  0.00 |  99.6 |    N/A |
  | Avg w/o B2   |  0.00 |  0.00 |  0.00 |  0.00 |  99.8 |    N/A |
  +--------------+-------+-------+-------+-------+-------+--------+

Speed 6 (cpu-used=6)
  +--------------+-------+-------+-------+-------+-------+--------+
  | Class        |     Y |    Cb |    Cr |  wAvg |  Enc% |  Ratio |
  +--------------+-------+-------+-------+-------+-------+--------+
  | A1           |  0.00 |  0.00 |  0.00 |  0.00 |  98.8 |    N/A |
  | A2           |  0.00 |  0.00 |  0.00 |  0.00 |  99.2 |    N/A |
  | Avg w/o B2   |  0.00 |  0.00 |  0.00 |  0.00 |  99.7 |    N/A |
  +--------------+-------+-------+-------+-------+-------+--------+

Speed 6 (cpu-used=6), threads=8, row-mt=1, classes A1 and A2 only
  +--------------+-------+-------+-------+-------+-------+--------+
  | Class        |     Y |    Cb |    Cr |  wAvg |  Enc% |  Ratio |
  +--------------+-------+-------+-------+-------+-------+--------+
  | A1           |  0.00 |  0.00 |  0.00 |  0.00 |  96.9 |    N/A |
  | A2           |  0.00 |  0.00 |  0.00 |  0.00 |  98.3 |    N/A |
  +--------------+-------+-------+-------+-------+-------+--------+

[YixinDu1]

Hi @yunqingwang1, this is a similar patch to apply multi-threading to TIP. I also tested the tool off results of TIP and compare it against the patch savings:

At the CTC thread counts the patch recovers about 24% of that time on A1 (1.2 of 5.0) and about 42% on A2 (0.8 of 1.9).

[yeqing-wu]

@yunqingwang1 , could you please help review again for this PR to see if you have any other comments?