[FFmpeg-devel] [PATCH] vf_overlay: add support to RGBA packed input and output

[Stefano Sabatini]

On date Saturday 2011-10-29 04:10:04 +0200, Michael Niedermayer encoded:
> On Sat, Oct 29, 2011 at 12:56:15AM +0200, Stefano Sabatini wrote:
[...]
> > > please benchmark this with START/STOP_TIMER against the previous code
> > 
> > RGB path was disabled before this one, I split the present patch and
> > did some tests.
> > 
> > * Test with no alpha in the main input
> > 
> > before alpha premultiplication
> > 1287135 dezicycles in first, 2 runs, 0 skips
> > 1335442 dezicycles in first, 4 runs, 0 skips
> > 1245555 dezicycles in first, 8 runs, 0 skips
> > 1162359 dezicycles in first, 16 runs, 0 skips
> > 1144390 dezicycles in first, 32 runs, 0 skips
> > 1134602 dezicycles in first, 64 runs, 0 skips
> > 1133281 dezicycles in first, 128 runs, 0 skips
> > 1114852 dezicycles in first, 256 runs, 0 skips
> > 1108999 dezicycles in first, 512 runs, 0 skips
> > 1101536 dezicycles in first, 1024 runs, 0 skips
> > 1096821 dezicycles in first, 2048 runs, 0 skips
> > 1090508 dezicycles in first, 4096 runs, 0 skips
> > 1085896 dezicycles in first, 8192 runs, 0 skips
> > 1084802 dezicycles in first, 16384 runs, 0 skips
> > 1083604 dezicycles in first, 32768 runs, 0 skips
> > 
> > after alpha premultiplication
> > 1224390 dezicycles in second, 2 runs, 0 skips
> > 1202235 dezicycles in second, 4 runs, 0 skips
> > 1191453 dezicycles in second, 8 runs, 0 skips
> > 1183031 dezicycles in second, 16 runs, 0 skips
> > 1230087 dezicycles in second, 32 runs, 0 skips
> > 1227492 dezicycles in second, 64 runs, 0 skips
> > 1230488 dezicycles in second, 128 runs, 0 skips
> > 1215128 dezicycles in second, 256 runs, 0 skips
> > 1207364 dezicycles in second, 512 runs, 0 skips
> > 1199813 dezicycles in second, 1024 runs, 0 skips
> > 1195857 dezicycles in second, 2048 runs, 0 skips
> > 1193954 dezicycles in second, 4096 runs, 0 skips
> > 1194128 dezicycles in second, 8192 runs, 0 skips
> > 1187481 dezicycles in second, 16384 runs, 0 skips
> > 1181874 dezicycles in second, 32768 runs, 0 skips
> > 
> > * Test with alpha in the main input:
> > 28684935 dezicycles in first, 2 runs, 0 skips
> > 28553902 dezicycles in first, 4 runs, 0 skips
> > 28776015 dezicycles in first, 8 runs, 0 skips
> > 29073680 dezicycles in first, 16 runs, 0 skips
> > 28816918 dezicycles in first, 32 runs, 0 skips
> > 28908704 dezicycles in first, 64 runs, 0 skips
> > 28745401 dezicycles in first, 128 runs, 0 skips
> > 28614980 dezicycles in first, 256 runs, 0 skips
> > 28609710 dezicycles in first, 512 runs, 0 skips
> > 28537037 dezicycles in first, 1024 runs, 0 skips
> > 28517850 dezicycles in first, 2048 runs, 0 skips
> > 28466515 dezicycles in first, 4096 runs, 0 skips
> > 28438388 dezicycles in first, 8192 runs, 0 skips
> > 28440383 dezicycles in first, 16384 runs, 0 skips
> > 28426314 dezicycles in first, 32768 runs, 0 skips
> > 
> > 33347880 dezicycles in second, 2 runs, 0 skips
> > 33131272 dezicycles in second, 4 runs, 0 skips
> > 38018970 dezicycles in second, 8 runs, 0 skips
> > 48715928 dezicycles in second, 16 runs, 0 skips
> > 44290285 dezicycles in second, 32 runs, 0 skips
> > 43696766 dezicycles in second, 64 runs, 0 skips
> > 38599173 dezicycles in second, 128 runs, 0 skips
> > 36112571 dezicycles in second, 256 runs, 0 skips
> > 34737837 dezicycles in second, 512 runs, 0 skips
> > 34066213 dezicycles in second, 1024 runs, 0 skips
> > 33640178 dezicycles in second, 2048 runs, 0 skips
> > 33368757 dezicycles in second, 4096 runs, 0 skips
> > 33233522 dezicycles in second, 8192 runs, 0 skips
> > 33132908 dezicycles in second, 16384 runs, 0 skips
> > 33062949 dezicycles in second, 32768 runs, 0 skips
> > 
> > Results are as expected, alpha pre-multiplication is significantly
> > slower but it may also be what the user wants, so I could make it
> > optional (and preserve the original alpha?, enabled by default?).
> 
> thats not what i meant
> 
> the original code looked like this:
> > -                d[r] = (d[r] * (0xff - s[3]) + s[0] * s[3] + 128) >> 8;
> > -                d[1] = (d[1] * (0xff - s[3]) + s[1] * s[3] + 128) >> 8;
> > -                d[b] = (d[b] * (0xff - s[3]) + s[2] * s[3] + 128) >> 8;
> 
> when i saw what you replaced it by i was ... scared ;)
> 
> if and switch are added in the innermost loop
> constants are replaced by variables
> variables are replaced by reading out of arrays from structures
> a division is added
> 
> all this make the code significantly slower
> 
> Can you explain what equation you are trying to implement ?


Changed the code (second patch), now testbench results changed from:
29891505 dezicycles in first, 2 runs, 0 skips
29780850 dezicycles in first, 4 runs, 0 skips
30056100 dezicycles in first, 8 runs, 0 skips
30378746 dezicycles in first, 16 runs, 0 skips
31263998 dezicycles in first, 32 runs, 0 skips
31422349 dezicycles in first, 64 runs, 0 skips
31441573 dezicycles in first, 128 runs, 0 skips
31319009 dezicycles in first, 256 runs, 0 skips
30925767 dezicycles in first, 512 runs, 0 skips
33965521 dezicycles in first, 1024 runs, 0 skips
32342480 dezicycles in first, 2048 runs, 0 skips
31631954 dezicycles in first, 4096 runs, 0 skips
31252298 dezicycles in first, 8192 runs, 0 skips
31572626 dezicycles in first, 16383 runs, 1 skips
31102288 dezicycles in first, 32767 runs, 1 skips

to:
26084640 dezicycles in first, 2 runs, 0 skips
23856690 dezicycles in first, 4 runs, 0 skips
24238267 dezicycles in first, 8 runs, 0 skips
26151311 dezicycles in first, 16 runs, 0 skips
25807400 dezicycles in first, 32 runs, 0 skips
27391090 dezicycles in first, 64 runs, 0 skips
26028030 dezicycles in first, 128 runs, 0 skips
23729756 dezicycles in first, 256 runs, 0 skips
22114165 dezicycles in first, 512 runs, 0 skips
21465190 dezicycles in first, 1024 runs, 0 skips
20951560 dezicycles in first, 2048 runs, 0 skips
20736770 dezicycles in first, 4096 runs, 0 skips
20573711 dezicycles in first, 8192 runs, 0 skips
20570483 dezicycles in first, 16384 runs, 0 skips
20634111 dezicycles in first, 32768 runs, 0 skips

With the second patch applied (non alpha in input):
24551340 dezicycles in second, 2 runs, 0 skips
23764147 dezicycles in second, 4 runs, 0 skips
23118037 dezicycles in second, 8 runs, 0 skips
22992204 dezicycles in second, 16 runs, 0 skips
22960603 dezicycles in second, 32 runs, 0 skips
23015486 dezicycles in second, 64 runs, 0 skips
23007612 dezicycles in second, 128 runs, 0 skips
22955180 dezicycles in second, 256 runs, 0 skips
23277693 dezicycles in second, 512 runs, 0 skips
23147960 dezicycles in second, 1024 runs, 0 skips
22940401 dezicycles in second, 2048 runs, 0 skips
22811952 dezicycles in second, 4096 runs, 0 skips
22760982 dezicycles in second, 8192 runs, 0 skips
22676573 dezicycles in second, 16384 runs, 0 skips
22622130 dezicycles in second, 32768 runs, 0 skips
(due to the added ifs).

With alpha in the main input/output:
41009130 dezicycles in second, 2 runs, 0 skips
36964740 dezicycles in second, 4 runs, 0 skips
34723803 dezicycles in second, 8 runs, 0 skips
39728604 dezicycles in second, 16 runs, 0 skips
40790327 dezicycles in second, 32 runs, 0 skips
38958495 dezicycles in second, 64 runs, 0 skips
36674410 dezicycles in second, 128 runs, 0 skips
35057610 dezicycles in second, 256 runs, 0 skips
33985402 dezicycles in second, 512 runs, 0 skips
33323452 dezicycles in second, 1024 runs, 0 skips
32870493 dezicycles in second, 2048 runs, 0 skips
32565989 dezicycles in second, 4096 runs, 0 skips
32464448 dezicycles in second, 8192 runs, 0 skips
32574558 dezicycles in second, 16384 runs, 0 skips
32468892 dezicycles in second, 32768 runs, 0 skips

Regarding the second patch, I kept Mark's code but after some time
spent tinkering on it I couldn't figure out the meaning of the
equation:
    d[da] = ( (d[da] << 8) + (256 - d[da]) * s[sa] ) >> 8;

Mark, could you give another look at this patch, check if it's correct
and explain the meaning of this last assignment?
-- 
FFmpeg = Fancy Fast Mythic Ponderous Evanescent Gangster
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0002-vf_overlay-enable-RGB-path.patch
Type: text/x-diff
Size: 9819 bytes
Desc: not available
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20111029/be68d382/attachment.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0003-vf_overlay-add-support-to-alpha-pre-multiplication-i.patch
Type: text/x-diff
Size: 3699 bytes
Desc: not available
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20111029/be68d382/attachment-0001.bin>