I had a look at the failures in Clip_Tests.cpp .
Actually comparing with what is done on other architectures, the content
of the frame is already wrong before the effect being applied, that is
the content that is imported from the test .mp4 seems false.
I digged into FFmpegReader.cpp and pixels get wrong content after
sws_scale is called in FFmpegReader::ProcessVideoPacket .
This function comes from libswscale (ffmpeg) .
This library has accelerated altivec code on powerpc for colorspace
conversion and forcing use of software processing there makes the
Clip tests works.

So I'd say there is some issue in
ffmpeg-3.4.2/libswscale/ppc/yuv2rgb_altivec.c or around, that I've not
been able to locate so far.


Attachment: pgpBCjtH6AgpF.pgp
Description: PGP signature

Reply via email to