CTU (64x64) vs CTU (32x32)

[rudyb]

Hi,
The default size of CTU in HM model is 64x64.

I was wondering what is the advantage of having a CTU 64x64 vs a CTU of 32x32?

Because to me the maximum transfer size is TU 32x32, then what is the point of having a CTU of 64x64?
The only advantage that I can think of, if maybe all the four CUs inside a CTU have the same motion vectors, then we can save some bits in encoding the motion vector.
Even I don't see a much advantage if we use use the "bypass" mode, where we bypass the TU.

I have a feeling that deciding on a CTU size of 64x64 must have more pronounced advantages over having a CTU size of 32x32, that I may be overlooking !

Can someone tell me the real advantages of having a CTU size of 64x64 vs 32x32 ?


Thanks,
--Rudy

[Sulik]

4x lower overhead (you can have 64x64 skipped CUs)

[rudyb]

Can you elaborate a little more? If Skip corresponds to skipping Transforms, and the biggest Transform Unit size is 32x32, then how can we have a skip size of 64x64, if there is no 64x64 transform size?

[xooyoozoo]

Can you elaborate a little more? If Skip corresponds to skipping Transforms, and the biggest Transform Unit size is 32x32, then how can we have a skip size of 64x64, if there is no 64x64 transform size?

Skipped CUs means there's no residual to transform and therefore no transform to skip.

CTUs are broken down into CUs, with the possibly of a single 64x64 CU. TUs and PUs (which you haven't mentioned, but they are there) operate after that point.

SAO is signaled at the CTU-level, and most things can be signaled at the 64x64 CU-level. Obviously, being forced to redundantly signal the same thing multiple times just because the encoder has a small max-CTU bloats the bitrate and increases the decoder's CABAC load.

[rudyb]

Thanks for the explanation.

[Shevach]

One of benefits of 32x32 CTU against of 64x64 CTU is in WPP parallelization.

The theoretical speed-up ratio of WPP: W*H / (W+2H – 2)

where W, H is width and height of image in CTU units.

For instance, for 3840x1728 resolution with 64x64 CTU size the theoretical speed-up ratio is ≈ 14. However, for 32x32 CTU the speed-up ratio is doubled (=28).