Promote the uneven-4-way partition disable to speed 1

[chengchen-google]

The speed feature was enabled at speed 2.
It turns out that its performance at speeed 1 is also good.

On speed 1, RA, the tradeoff is:

Testset PSNR-Y PSNR-U PSNR-V PSNR-YUV EncTime Ratio
A1 (17 frames) 0.02% 0.30% -0.05% 0.03% 93.77% 207.7
A2 (33 frames) 0.07% -0.30% 0.30% 0.06% 93.30% 111.7
A3 (33 frames) 0.16% 0.07% -0.47% 0.14% 89.52% 74.9
A4 (33 frames) 0.08% -1.66% -2.73% -0.09% 89.29% -----
A5 (33 frames) 0.47% -0.88% -0.75% 0.32% 85.27% 46.0

The speed feature is disabled for resolutions <= 270p.

STATS_CHANGED

[urvangjoshi]

What is the anchor @chengchen-google for these results?
I saw a bigger loss on commit 2abcfff:

• 0.13% average loss w/o B2
• 0.06% on A1
• 0.08% on A2
• 0.19% on B1
• 0.19% on A3
• 0.34% on A4
• Neutralish on A5 but most clips have a loss

If you see something similar, I'd suggest disabling based on resolution (1080+) and also keeping it on for screen content perhaps?

[chengchen-google]

What is the anchor @chengchen-google for these results? I saw a bigger loss on commit 2abcfff:

• 0.13% average loss w/o B2
• 0.06% on A1
• 0.08% on A2
• 0.19% on B1
• 0.19% on A3
• 0.34% on A4
• Neutralish on A5 but most clips have a loss

If you see something similar, I'd suggest disabling based on resolution (1080+) and also keeping it on for screen content perhaps?

The anchor is 5d628d8.
I haven't tested for other resolutions. I do have an impression that the performance change on lower resolutions is larger. Do you think we need the full information to make a decision? If other resolution shows larger loss, perhaps we can make this speed feature dependent on resolution?

[urvangjoshi]

What is the anchor @chengchen-google for these results? I saw a bigger loss on commit 2abcfff:

• 0.13% average loss w/o B2
• 0.06% on A1
• 0.08% on A2
• 0.19% on B1
• 0.19% on A3
• 0.34% on A4
• Neutralish on A5 but most clips have a loss

If you see something similar, I'd suggest disabling based on resolution (1080+) and also keeping it on for screen content perhaps?

The anchor is 5d628d8. I haven't tested for other resolutions. I do have an impression that the performance change on lower resolutions is larger. Do you think we need the full information to make a decision? If other resolution shows larger loss, perhaps we can make this speed feature dependent on resolution?

Yes, pls run other resolutions (RA 33 frames) and let's make a decision based on the result -- resolution dependent was indeed my suggestion.

[chengchen-google]

What is the anchor @chengchen-google for these results? I saw a bigger loss on commit 2abcfff:

• 0.13% average loss w/o B2
• 0.06% on A1
• 0.08% on A2
• 0.19% on B1
• 0.19% on A3
• 0.34% on A4
• Neutralish on A5 but most clips have a loss

If you see something similar, I'd suggest disabling based on resolution (1080+) and also keeping it on for screen content perhaps?

The anchor is 5d628d8. I haven't tested for other resolutions. I do have an impression that the performance change on lower resolutions is larger. Do you think we need the full information to make a decision? If other resolution shows larger loss, perhaps we can make this speed feature dependent on resolution?

Yes, pls run other resolutions (RA 33 frames) and let's make a decision based on the result -- resolution dependent was indeed my suggestion.

Hi Urvang, I tested on A3-A5. The loss is higher on A5, but the tradeoff is still good. Please see the full results here:

Testset PSNR-Y PSNR-U PSNR-V PSNR-YUV EncTime Ratio
A1 (17 frames) 0.02% 0.30% -0.05% 0.03% 93.77% 207.7
A2 (33 frames) 0.07% -0.30% 0.30% 0.06% 93.30% 111.7
A3 (33 frames) 0.16% 0.07% -0.47% 0.14% 89.52% 74.9
A4 (33 frames) 0.08% -1.66% -2.73% -0.09% 89.29% -----
A5 (33 frames) 0.47% -0.88% -0.75% 0.32% 85.27% 46.0

It seems good to apply it for all resolutions. Do you agree?

[urvangjoshi]

Hi Urvang, I tested on A3-A5. The loss is higher on A5, but the tradeoff is still good. Please see the full results here:

Testset PSNR-Y PSNR-U PSNR-V PSNR-YUV EncTime Ratio A1 (17 frames) 0.02% 0.30% -0.05% 0.03% 93.77% 207.7 A2 (33 frames) 0.07% -0.30% 0.30% 0.06% 93.30% 111.7 A3 (33 frames) 0.16% 0.07% -0.47% 0.14% 89.52% 74.9 A4 (33 frames) 0.08% -1.66% -2.73% -0.09% 89.29% ----- A5 (33 frames) 0.47% -0.88% -0.75% 0.32% 85.27% 46.0

It seems good to apply it for all resolutions. Do you agree?

Thanks for testing!

I feel like PSNR-Y for A5 is still borderline ratio. Perhaps just keep it on for <= 270p?

[chengchen-google]

Hi Urvang, I tested on A3-A5. The loss is higher on A5, but the tradeoff is still good. Please see the full results here:
Testset PSNR-Y PSNR-U PSNR-V PSNR-YUV EncTime Ratio A1 (17 frames) 0.02% 0.30% -0.05% 0.03% 93.77% 207.7 A2 (33 frames) 0.07% -0.30% 0.30% 0.06% 93.30% 111.7 A3 (33 frames) 0.16% 0.07% -0.47% 0.14% 89.52% 74.9 A4 (33 frames) 0.08% -1.66% -2.73% -0.09% 89.29% ----- A5 (33 frames) 0.47% -0.88% -0.75% 0.32% 85.27% 46.0
It seems good to apply it for all resolutions. Do you agree?

Thanks for testing!

I feel like PSNR-Y for A5 is still borderline ratio. Perhaps just keep it on for <= 270p?

Done.