Skip chroma RD evaluation of interpolation filter search

[dikshasingh-sys]

This patch introduces the speed feature 'skip_model_rd_uv' to skip the RD evaluation of chroma planes during the interpolation filter search. The chroma predictors are built later using the best interpolation filter selected by the luma RD. This speed feature is enabled for speed >= 1.

This optimization has been taken from AV1:
https://aomedia-review.googlesource.com/c/aom/+/210801

Results for RA CTC, A1 17 frames, A2 33 frames, speed 1: (Anchor: [12c553a])

+-----+------+-------+------+-------+---------------+
|Class|  Y   |  U    |   V  |  YUV  |EncInstCount(%)|
+-----+------+-------+------+-------+---------------+
| A1  | 0.01 | -0.11 |-0.21 |  0.00 | 99.3          |
| A2  | 0.00 |  0.23 | 0.40 |  0.03 | 98.6          |
+-----+------+-------+------+-------+---------------+

STATS_CHANGED for speed >= 1

[dikshasingh-sys]

@yunqingwang1 @urvangjoshi Please review.
/cc @vikasprasad10 @harishdm @ranjit-tulabandu @cherma-a @deepa-kg @RemyaPrakasan @ram-mohan

[leolzhao]

LGTM.

[yeqing-wu]

LGTM

[urvangjoshi]

@dikshasingh-sys : PSNR U/V do seem to have losses for A2 set, although YUV looks ok.
What is the U/V loss in other testsets?

Was similar loss seen in AV1? I see that speed feature in AV1 is enabled for speed >= 3 only.

[dikshasingh-sys]

@dikshasingh-sys : PSNR U/V do seem to have losses for A2 set, although YUV looks ok. What is the U/V loss in other testsets?

Was similar loss seen in AV1? I see that speed feature in AV1 is enabled for speed >= 3 only.

Please find the BD-Rate impact for the other test sets, tested on commit 12c553a using 33 frames at speed=1:

+-----+------+-------+------+-------+---------------+
|Class|  Y   |   U   |  V   |  YUV  |EncInstCount(%)|
+-----+------+-------+------+-------+---------------+
| A5  | 0.01 |  0.55 | -0.24|  0.02 |          98.6 |
| A4  |-0.01 |  0.28 |  0.43|  0.02 |          98.7 |
| A3  | 0.01 |  0.01 | -0.11|  0.01 |          99.2 |
+-----+------+-------+------+-------+---------------+

The runs for the B1 and B2 test sets are currently in progress. We will update the results once they are completed.

For reference, the speed feature was extended to speed >= 1 in AV1 in 212381 and below are the results for speed 1:

+-------------+-------------------+---------+---------+---------+
| Resolution  | Instruction Count |    Y    |    U    |    V    |
|             | Reduction (%)     |         |         |         |
+-------------+-------------------+---------+---------+---------+
| LOWRES2     |             1.233 |  0.0136 |  0.0122 |  0.0405 |
| MIDRES2     |             0.905 |  0.0024 |  0.1666 |  0.1526 |
| SHORTS608P  |             0.421 | -0.0506 |  0.0243 | -0.0750 |
| SHORTS720P  |             0.227 | -0.0742 | -0.0219 |  0.0442 |
| SHORTS1080P |             0.203 | -0.0628 |  0.0205 |  0.0071 |
| HDRES2      |             0.790 |  0.0007 |  0.2404 |  0.1360 |
| Average     |             0.597 | -0.0319 |  0.0661 |  0.0447 |
+-------------+-------------------+---------+---------+---------+

[urvangjoshi]

@dikshasingh-sys : PSNR U/V do seem to have losses for A2 set, although YUV looks ok. What is the U/V loss in other testsets?
Was similar loss seen in AV1? I see that speed feature in AV1 is enabled for speed >= 3 only.

Please find the BD-Rate impact for the other test sets, tested on commit 12c553a using 33 frames at speed=1:

+-----+------+-------+------+-------+---------------+
|Class|  Y   |   U   |  V   |  YUV  |EncInstCount(%)|
+-----+------+-------+------+-------+---------------+
| A5  | 0.01 |  0.55 | -0.24|  0.02 |          98.6 |
| A4  |-0.01 |  0.28 |  0.43|  0.02 |          98.7 |
| A3  | 0.01 |  0.01 | -0.11|  0.01 |          99.2 |
+-----+------+-------+------+-------+---------------+

The runs for the B1 and B2 test sets are currently in progress. We will update the results once they are completed.

For reference, the speed feature was extended to speed >= 1 in AV1 in 212381 and below are the results for speed 1:

+-------------+-------------------+---------+---------+---------+
| Resolution  | Instruction Count |    Y    |    U    |    V    |
|             | Reduction (%)     |         |         |         |
+-------------+-------------------+---------+---------+---------+
| LOWRES2     |             1.233 |  0.0136 |  0.0122 |  0.0405 |
| MIDRES2     |             0.905 |  0.0024 |  0.1666 |  0.1526 |
| SHORTS608P  |             0.421 | -0.0506 |  0.0243 | -0.0750 |
| SHORTS720P  |             0.227 | -0.0742 | -0.0219 |  0.0442 |
| SHORTS1080P |             0.203 | -0.0628 |  0.0205 |  0.0071 |
| HDRES2      |             0.790 |  0.0007 |  0.2404 |  0.1360 |
| Average     |             0.597 | -0.0319 |  0.0661 |  0.0447 |
+-------------+-------------------+---------+---------+---------+

Thanks @dikshasingh-sys for that info.

So, looks like we see more chroma losses in AV2 (e.g. A2, A4 and A5 sets) than what's seen in AV1 -- with 1% to 1.5% encoder speed-up.

I'm unclear about the trade-off, as we haven't decided the critera for chroma.
So, I'll leave it up to @yunqingwang1 @yeqing-wu @leolzhao to decide.

[yunqingwang1]

Based on AV2's YUV weights, this sounds okay to me.

[leolzhao]

Based on AV2's YUV weights, this sounds okay to me.

The U and V loss look acceptable to me as well.