turbo question

[fatxy]

heya i wanna ask if thats right:

quality from bad to good

no qpel/gmc turbo -> no qpel/gmc -> qpel/gmc turbo -> qpel/gmc

dunno if 'no qpel/gmc turbo' is workin tho

[pelle412]

The "Turbo" feature makes QPEL and B-frames encoding a little faster at the expense of a minor quality drop. No matter how closely I look at a video encoded with or w/o Turbo I can't see the difference. My eyes aren't very picky either. There are other encoding pros here who have eyes like hawks.

Turbo has no effect on GMC.

GMC will increase the quality of the video (if bitrate is same as w/o) but the difference is barely if at all noticeable. I never use it because I can't see a difference, just longer encoding times.

QPEL doesn't really increase quality. It gives the encoded video a slightly different appearance. This QPEL "effect" is IMO nice, particularly when used on clean video sources. On noisy sources the effect is not noticeable to me so I don't use it then.

With this information and running your own test encodes, you should be able to answer your own question.

[Sharktooth]

GMC lowers the size of GMCed (S) frames using far less motion vectors (it uses 3 warp point GLOBAL motion vectors) than P or B frames.
That rises the global clip compression (not the quality) by usually less than a few percent, gaining some bits on the average bitrate.
Those bits can be re-distributed (using "target bitrate" or "target size" mode) over the whole encode and get a slightly better overall quality with the same filesize/bitrate.
Dont expect to see a big difference (if there is any).
If i recall correctly, since turbo option affects the motion detection schemes GMC is NOT affected by Turbo option - but i'm not completely sure about that.

EDIT: English is definatly not the language for me. I should give up editing typos and grammar...

[Didée]

Sharktooth:

Regarding GMC, you're right with the effects, but wrong with their cause:

"GMC lowers the size of GMCed (S) frames using far less motion vectors (it uses 3 warp point GLOBAL motion vectors) than P or B frames."

No. It isn't so that all the per-block motion vectors would be simply replaced by the GMC vectors. In fact, the amount of per-block motion vectors stays - basically - the same with or without GMC.

GMC *prepares* a frame in such a way that the aftercoming (normal) motion estimation has "better targets" to find. Imagine there is some sort of zooming-plus-rotation going on from one frame to the next. Without GMC, ME has big problems to find any target for each block, since all texture information has a) altered its size /and/ b) altered its orientation. Extreme example: an 'X'-alike texture element has become '+'-alike in the next frame -- this just cannot be compensated by shifting rectangular blocks on a "cartesian working space". Therefore, since no good targets are to find, the codec has to use many bits for texture information, eventually even through key-blocks.
With GMC, the whole frame is distorted according to the actual zooming-plus-rotation, so that the textures in neighbour frames have again the same size and the same orientation (in ideal case, that is). If this succeeds, then there are good targets for motion estimation available again, and therefore less texture bits have to be coded.
However, after GMC there is still the "normal" ME taking place, where each block gets its own vector: the number of used vectors stays the same. (Exept for some vectors eventually becoming (0,0) through the help of GMC - but when this happens, it's mostly pure luck.)

So, in principle, GMC is more about saving texture bits than about saving vector bits. Vector sizes are of course affected, too - but that's more of a side-effect, since (without GMC) the ME has a hard time to find anything appropriate, and will end up with "random matches".

(That's how I understood the mechanics of GMC - in case I told nonsense here, anyone correct me, please!)

[Sharktooth]

Thanx for the correct explanation Didée.