Reduce intra mode top count to 3 for speed 3 and 4

[chengchen-google]

By reducing the number of candidates, we achieve a considerable
speed up with accepted coding loss.

Result of reducing the count to 3 for speed 4:

TestSet PSNR-YUV EncTime Ratio
A1 (17 frames) 0.10% 96.70% 34.1
A2 (33 frames) 0.07% 96.81% 47.0

Result of reducing the count to 3 for speed 3:

TestSet PSNR-YUV EncTime Ratio
A1 (17 frames) 0.11% 96.87% 29.4
A2 (33 frames) 0.09% 96.41% 41.3

If we reduce the count to 2
The speedup/loss ratio is good for A2, but not good enough for A1.
On speed 4, RA, the performance is:

TestSet PSNR-YUV EncTime Ratio
A1 (17 frames) 0.41% 92.76% 19.0
A2 (33 frames) 0.19% 93.44% 36.9

STATS_CHANGED

[urvangjoshi]

The ratio for A1 set seems less than the threshold, right?
What if you used value 3 instead of 2?

[chengchen-google]

The ratio for A1 set seems less than the threshold, right? What if you used value 3 instead of 2?

Yes, you are right. I am testing 3.

[leolzhao]

The trade off for class A1 is a little bit worse than the threshold, what about enabling this speed feature for speed 5 and above? @chengchen-google

[leolzhao]

The ratio for A1 set seems less than the threshold, right? What if you used value 3 instead of 2?

Yes, you are right. I am testing 3.

if 3 is good for cpu-used 4, we can set it to 3 for speed 4 and 2 for speed 5 and above.

[chengchen-google]

I tested to reduce the count to 3 and tested on speed 3 and 4. The tradeoff is better. Please see the updated commit message.

[yeqing-wu]

What is the anchor of the results? Meanwhile, should we remove 'Change-Id: Ib9dba881748b9106f295f4d468cdbf0367e28a7d'?

[chengchen-google]

What is the anchor of the results? Meanwhile, should we remove 'Change-Id: Ib9dba881748b9106f295f4d468cdbf0367e28a7d'?

The anchor is ea89c21.