[chengchen-google]

The CCSO search evaluates up to 116 filter candidates per plane. For each candidate it classifies every luma sample of the frame, accumulates the per class error statistics, and for every trained filter copies the reconstructed plane, filters it and computes the per filter unit distortion.

These full frame passes run on the main thread while the tile worker threads wait.

This commit proposes to process these passes with multi-thread. It process the frame filter row by row, each writes only its own samples of the classification and reconstruction buffers. It splits the plane evenly across the encoder worker threads.

The new speed feature enable_ccso_mt is introduced. It is off with one thread and on for multi-threads.

Results on RA A1 (17 frames), A2 (33 frames) show a significant encoder speed up (~20%) for A1 at speed 1. The speed up is less for A2 because A1 uses 4 threads. The more threads, the higher speed up is expected.

Also, after speed 2, the CCSO search is optimized and the number of searches is significantly reduced. Therefore the speed up for speed 2 and above is less.

Full results:

+-------+-------+-------+-------+-------+-------+-------+-------+
| Speed | Class |     Y |    Cb |    Cr |  wAvg | Enc%  |  Dec% |
+-------+-------+-------+-------+-------+-------+-------+-------+
|   1   | A1    | +0.00 | +0.00 | +0.00 | +0.00 |  79.4 | 101.7 |
|   1   | A2    | +0.00 | +0.00 | +0.00 | +0.00 |  95.7 |  97.8 |
+-------+-------+-------+-------+-------+-------+-------+-------+
|   2   | A1    | +0.00 | +0.00 | +0.00 | +0.00 |  98.5 |  99.2 |
|   2   | A2    | +0.00 | +0.00 | +0.00 | +0.00 |  98.6 | 101.1 |
+-------+-------+-------+-------+-------+-------+-------+-------+
|   3   | A1    | +0.00 | +0.00 | +0.00 | +0.00 |  98.6 |  95.5 |
|   3   | A2    | +0.00 | +0.00 | +0.00 | +0.00 |  98.6 |  96.9 |
+-------+-------+-------+-------+-------+-------+-------+-------+
|   4   | A1    | +0.00 | +0.00 | +0.00 | +0.00 |  98.3 | 101.7 |
|   4   | A2    | +0.00 | +0.00 | +0.00 | +0.00 |  99.1 | 101.3 |
+-------+-------+-------+-------+-------+-------+-------+-------+
|   5   | A1    | +0.00 | +0.00 | +0.00 | +0.00 |  97.8 | 100.0 |
|   5   | A2    | +0.00 | +0.00 | +0.00 | +0.00 |  99.3 |  97.5 |
+-------+-------+-------+-------+-------+-------+-------+-------+
|   6   | A1    | +0.00 | +0.00 | +0.00 | +0.00 |  97.0 |  99.1 |
|   6   | A2    | +0.00 | +0.00 | +0.00 | +0.00 | 100.1 | 103.6 |
+-------+-------+-------+-------+-------+-------+-------+-------+

This is a bit-exact change.

We also add a unit test test/encoder_frame_mt_test.cc. It encodes a clip with 1, 2, 3 and 4 threads and checks that the compressed frames and the decoded frames are identical.