v0.7: Severe quantizer difference between {...faster} and {fast...} presets

[LigH]

With default options (CRF 28), all presets {'faster' and faster} produce a much larger result and also much better visual quality than all presets {'fast' and slower}.

This means about 5:1 the size and -5 dB SSIM with CUTree enabled, and about 4:1 the size and -4 dB SSIM with CUTree disabled (for the faster group, it is disabled anyway; but it appears to be not the main reason).

Values for a specific CRF value are similar, confirming the groups of presets in general.

Because there is no current Evaluators' Guide for v0.7, I can't be certain which defaults changed between these presets.

BTW: This bug tracker needs more entries in the "Version" selector.

[Selur]

regarding the defaults, see https://bitbucket.org/multicoreware/x265/src/733f6dc4017359268100a52454819ef96a5b44a1/source/common/common.cpp?at=default line 140+

[Steve Borho]

lookahead: bug fix for b-adapt 0 cost estimates (closes #26)

don't overwrite lastNonB reference with current frame, 'lastNonB' state was already replaced with current frame's lowres at this point.

it mostly affected b-adapt=0 but could have affected any of the modes if they had not already estimated this particular P frame cost before this point.

→ <<cset 9eea995a2347>>