[FFmpeg-devel] [PATCH] swscale/output: Altivec-optimize yuv2plane1_8
cand at gmx.com
Sat Nov 17 10:09:04 EET 2018
On Fri, 16 Nov 2018 22:09:25 +0100
Carl Eugen Hoyos <ceffmpeg at gmail.com> wrote:
> (This is less important atm, but I believe all functions currently
> in libswscale/ppc compile and run fine on - old - 32bit be hardware
> as your new function does.
> My completely inexperienced suspicion is that the instruction that
> you call "VSX" also exists on Altivec.)
VSX functions such as vec_vsx_ld were added in ISA 2.06, aka POWER7.
They shouldn't compile on earlier PPC like Apple G4/G5. Is your machine
at least POWER7?
> I wanted to write that this hunk breaks compilation on big-endian
> (you should be able to test with "#if 0" instead of "#if !HAVE_BIGENDIAN")
> but the good news is that your patch works fine on big-endian,
> just remove the if-endif block. (Tested visually with lena on 32 and 64bit be.)
Thanks, will do.
> Are you aware of the bounty that is offered for this task?
> (and #5569, #5570)
Yes, I admit that's why I started. Looking to make some extra, and
helping IBM is not a bad way to do so. I'm considering getting a Raptor
Blackbird when it comes out next year.
> There is a bug report about one altivec routine that works on
> big-endian but breaks the output visually on little-endian while
> many other functions work on both, could you have a look?
I'll try. This patch was my first time playing with Power vectors.
More information about the ffmpeg-devel