976695 - Consider converting ycrcb gfx convert/scale/rotate functions to media/libyuv

Status: RESOLVED DUPLICATE of bug 791941

(Core :: Graphics, defect)

Description

[Randell Jesup [:jesup] (needinfo me)]

The gfx/ycrcb functions were last updated in 2010 from Chromium.  We should consider using media/libyuv functions instead unless there's too much of an interface/functionality mismatch or for some reason the current functions are faster than libyuv (unlikely, but possible).

media/libyuv has all sorts of optimized variations with SSE2/SSE3/AVX2/NEON/etc opts and is up-to-date with upstream (or pretty close).

I mention it because when I was in a test call with profiling in webrtc and I was seeing 6% in FastConvertYUVToRGB32Row().

Alternatively, we should consider updating them from upstream.

Comment 1

[Nathan Froyd [:froydnj]]

It looks like chromium now features SSE*/MMX optimized versions of some of these functions:

http://src.chromium.org/svn/trunk/src/media/base/simd/

Looks like it would be better to use media/libyuv if possible.

Comment 2

[Chris Pearce [:cpearce (Not reading bugmail)]]

Somebody should measure the candidates and determine which is the fastest.

Comment 3

[Timothy B. Terriberry (:derf)]

(In reply to Chris Pearce (:cpearce) from comment #2)
> Somebody should measure the candidates and determine which is the fastest.

I agree.

There are other issues that we will have to consider, particularly with sub-pixel alignment shifts introduced by scaling.