Orientation partition pruning

[chengchen-google]

Add a new speed feature to analyze the structure orientation of the source signal to prune horz/vert partition types.

The speed feature shows a good tradeoff for speed 5, A1

We turn it on for A1 only.

RA performance, baseline 4a8df58:

    Speed           A1 BD-Rate (YUV)  A1 Speedup    A1 Ratio
    Speed_4         +0.16%             +1.61%       10.1
    Speed_5         +0.14%             +3.09%       22.1
    
    Speed           A2 BD-Rate (YUV)  A2 Speedup    A2 Ratio
    Speed_4         +0.19%             +2.96%       15.6
    Speed_5         +0.21%             +2.89%       13.8

STATS_CHANGED

[leolzhao]

As the loss for class A1 is relatively big and one MR #5253 about block partition speed up was just merged, please rebase the code and re-test it on top of this commit 45dc128.

[urvangjoshi]

As the loss for class A1 is relatively big and one MR #5253 about block partition speed up was just merged, please rebase the code and re-test it on top of this commit 45dc128.

Any update on this @chengchen-google ?

[yeqing-wu]

Let's run this PR on top of latest av2-enc branch, if the tradeoff cannot meet speed 4 minimum bar ratio, but >=15x for speed 5 or >=10x for speed 6, then merge them into speed 5 or speed 6 accordingly.

[chengchen-google]

As the loss for class A1 is relatively big and one MR #5253 about block partition speed up was just merged, please rebase the code and re-test it on top of this commit 45dc128.

Any update on this @chengchen-google ?

Last time I tested on top of two-pass partition, the tradeoff is reduced.
I will rebase to the most recent head and test it.
I will report results as soon as possible.

[chengchen-google]

Let's run this PR on top of latest av2-enc branch, if the tradeoff cannot meet speed 4 minimum bar ratio, but >=15x for speed 5 or >=10x for speed 6, then merge them into speed 5 or speed 6 accordingly.

I tested the PR on the most recent head: 4a8df58

The performance is:
'''

Speed Level	A1 BD-Rate (YUV)	A1 Speedup	A1 Ratio	A2 BD-Rate (YUV)	A2 Speedup	A2 Ratio
Speed 4	+0.16%	+1.61%	10.1	+0.19%	+2.96%	15.6
Speed 5	+0.14%	+3.09%	22.1	+0.21%	+2.89%	13.8
'''

I recommend to turn it on at speed 5.

[chengchen-google]

Let's run this PR on top of latest av2-enc branch, if the tradeoff cannot meet speed 4 minimum bar ratio, but >=15x for speed 5 or >=10x for speed 6, then merge them into speed 5 or speed 6 accordingly.

I tested the PR on the most recent head: 4a8df58

The performance is: '''

Speed Level A1 BD-Rate (YUV) A1 Speedup A1 Ratio A2 BD-Rate (YUV) A2 Speedup A2 Ratio
Speed 4 +0.16% +1.61% 10.1 +0.19% +2.96% 15.6
Speed 5 +0.14% +3.09% 22.1 +0.21% +2.89% 13.8
'''
I recommend to turn it on at speed 5.

In addition, the performance with a relaxed threshold on speed 6 is reported in a separate PR: #5402

[chengchen-google]

Hi @yeqing-wu , could you take a look at this PR? The tradeoff shows it is good for speed 5.
Its follow-up PR as well: #5402

[yeqing-wu]

@urvangjoshi, I believe this is the base PR that you mentioned. For speed 5, the minimum ratio is 15x. A1 exceeds the ratio, but A2 is still below it. Are you all comfortable with this? Alternatively, can we adjust the threshold to ensure that it achieves the minimum ratio? @yunqingwang1 and @leolzhao

[urvangjoshi]

@urvangjoshi, I believe this is the base PR that you mentioned. For speed 5, the minimum ratio is 15x. A1 exceeds the ratio, but A2 is still below it. Are you all comfortable with this? Alternatively, can we adjust the threshold to ensure that it achieves the minimum ratio? @yunqingwang1 and @leolzhao

I agree that trade-off doesn't seem good enough. Do you agree @chengchen-google ?
@yeqing-wu : what I meant is that other PR should not be approved in this case.

[yeqing-wu]

@urvangjoshi, I believe this is the base PR that you mentioned. For speed 5, the minimum ratio is 15x. A1 exceeds the ratio, but A2 is still below it. Are you all comfortable with this? Alternatively, can we adjust the threshold to ensure that it achieves the minimum ratio? @yunqingwang1 and @leolzhao

I agree that trade-off doesn't seem good enough. Do you agree @chengchen-google ? @yeqing-wu : what I meant is that other PR should not be approved in this case.

@urvangjoshi, I see. I noticed that the code between #5275 and #5402 is quite similar, with the only difference being that #5402 only applies to speed 6 with 4k resolution, while #5275 applies to all speeds >= 5. I reviewed #5402 and approved it since the ratio is above the bar, but I did not approve #5275. If the code in #5402 also applies to other speeds besides speed 6 with 4k resolution, @chengchen-google could you please modify the code in #5402 to ensure that it only applies to speed 6 with 4k resolution, so that it aligns with the commit message?

[chengchen-google]

@urvangjoshi, I believe this is the base PR that you mentioned. For speed 5, the minimum ratio is 15x. A1 exceeds the ratio, but A2 is still below it. Are you all comfortable with this? Alternatively, can we adjust the threshold to ensure that it achieves the minimum ratio? @yunqingwang1 and @leolzhao

I agree that trade-off doesn't seem good enough. Do you agree @chengchen-google ? @yeqing-wu : what I meant is that other PR should not be approved in this case.

@urvangjoshi, I see. I noticed that the code between #5275 and #5402 is quite similar, with the only difference being that #5402 only applies to speed 6 with 4k resolution, while #5275 applies to all speeds >= 5. I reviewed #5402 and approved it since the ratio is above the bar, but I did not approve #5275. If the code in #5402 also applies to other speeds besides speed 6 with 4k resolution, @chengchen-google could you please modify the code in #5402 to ensure that it only applies to speed 6 with 4k resolution, so that it aligns with the commit message?

H @yeqing-wu and @urvangjoshi , I agree that this PR's ratio is not good for A2, speed 5.
I modified the PR so that it is only turned on for A1, speed 5.

The follow-up PR #5402 also only works for A1, speed 6.

They are now consistent and both satisfy the bar.

[yeqing-wu]

I think we only need to merge #5402 and discard this PR, right?

[chengchen-google]

I think we only need to merge #5402 and discard this PR, right?

Technically yes. I need to change the PR title and description.