gcc 6.x and 7.x generate wrong code for little endian machines
for the vec_lvsl/vec_perm instruction combos in some cases.
The bug was fixed in version 8.x
If these instructions are replaced with vec_xl, the problem goes
away for all versions of the compilers.
Fixes ticket #7124.
The implementation is pretty straight-forward. Most of the existing
NV12 codepaths work regardless of subsampling and are re-used as is.
Where necessary I wrote the slightly different NV24 versions.
Finally, the one thing that confused me for a long time was the
asm specific x86 path that did an explicit exclusion check for NV12.
I replaced that with a semi-planar check and also updated the
equivalent PPC code, which Lauri kindly checked.
./ffmpeg -f lavfi -i yuvtestsrc=duration=1:size=1200x1440 -sws_flags fast_bilinear \
-s 1200x720 -f null -vframes 100 -pix_fmt $i -nostats \
-cpuflags 0 -v error -
32-bit mul, power8 only.
~2x speedup:
rgb24
24431 UNITS in yuv2packed2, 16384 runs, 0 skips
13783 UNITS in yuv2packed2, 16383 runs, 1 skips
bgr24
24396 UNITS in yuv2packed2, 16384 runs, 0 skips
14059 UNITS in yuv2packed2, 16384 runs, 0 skips
rgba
26815 UNITS in yuv2packed2, 16383 runs, 1 skips
12797 UNITS in yuv2packed2, 16383 runs, 1 skips
bgra
27060 UNITS in yuv2packed2, 16384 runs, 0 skips
13138 UNITS in yuv2packed2, 16384 runs, 0 skips
argb
26998 UNITS in yuv2packed2, 16384 runs, 0 skips
12728 UNITS in yuv2packed2, 16381 runs, 3 skips
bgra
26651 UNITS in yuv2packed2, 16384 runs, 0 skips
13124 UNITS in yuv2packed2, 16384 runs, 0 skips
This is a low speedup, but the x86 mmx version also gets only ~2x. The mmx version
is also heavily inaccurate, while the vsx version has high accuracy.
./ffmpeg -f lavfi -i yuvtestsrc=duration=1:size=1200x1440 -sws_flags fast_bilinear \
-s 1200x1440 -f null -vframes 100 -pix_fmt $i -nostats \
-cpuflags 0 -v error -
32-bit mul, power8 only.
1.8-2.3x speedup:
rgb24
18192 UNITS in yuv2packed1, 32767 runs, 1 skips
9983 UNITS in yuv2packed1, 32760 runs, 8 skips
bgr24
18665 UNITS in yuv2packed1, 32766 runs, 2 skips
9925 UNITS in yuv2packed1, 32763 runs, 5 skips
rgba
20239 UNITS in yuv2packed1, 32767 runs, 1 skips
8794 UNITS in yuv2packed1, 32759 runs, 9 skips
bgra
20354 UNITS in yuv2packed1, 32768 runs, 0 skips
8770 UNITS in yuv2packed1, 32761 runs, 7 skips
argb
20185 UNITS in yuv2packed1, 32768 runs, 0 skips
8761 UNITS in yuv2packed1, 32761 runs, 7 skips
bgra
20360 UNITS in yuv2packed1, 32766 runs, 2 skips
8759 UNITS in yuv2packed1, 32764 runs, 4 skips
This is a low speedup, but the x86 mmx version also gets only ~2x. The mmx version
is also heavily inaccurate, while the vsx version has high accuracy.
./ffmpeg -f lavfi -i yuvtestsrc=duration=1:size=1200x1440 \
-s 1200x1440 -f null -vframes 100 -pix_fmt $i -nostats \
-cpuflags 0 -v error -
This uses 32-bit mul, so POWER8 only.
The following output formats get about 4.5x speedup:
rgb24
39980 UNITS in yuv2packed1, 32768 runs, 0 skips
8774 UNITS in yuv2packed1, 32768 runs, 0 skips
bgr24
40069 UNITS in yuv2packed1, 32768 runs, 0 skips
8772 UNITS in yuv2packed1, 32766 runs, 2 skips
rgba
39759 UNITS in yuv2packed1, 32768 runs, 0 skips
8681 UNITS in yuv2packed1, 32767 runs, 1 skips
bgra
39729 UNITS in yuv2packed1, 32768 runs, 0 skips
8696 UNITS in yuv2packed1, 32766 runs, 2 skips
argb
39766 UNITS in yuv2packed1, 32768 runs, 0 skips
8672 UNITS in yuv2packed1, 32766 runs, 2 skips
bgra
39784 UNITS in yuv2packed1, 32768 runs, 0 skips
8659 UNITS in yuv2packed1, 32767 runs, 1 skips
./ffmpeg_g -f rawvideo -pix_fmt rgb24 -s hd1080 -i /dev/zero -pix_fmt yuv420p16be \
-s 1920x1728 -f null -vframes 100 -v error -nostats -
9-14 bit funcs get about 6x speedup, 16-bit gets about 15x.
Fate passes, each format tested with an image to video conversion.
Only POWER8 includes 32-bit vector multiplies, so POWER7 is locked out
of the 16-bit function. This includes the vec_mulo/mule functions too,
not just vmuluwm.
With TIMER_REPORT skips disabled:
yuv420p9le
12412 UNITS in planarX, 131072 runs, 0 skips
73136 UNITS in planarX, 131072 runs, 0 skips
yuv420p9be
12481 UNITS in planarX, 131072 runs, 0 skips
73410 UNITS in planarX, 131072 runs, 0 skips
yuv420p10le
12322 UNITS in planarX, 131072 runs, 0 skips
72546 UNITS in planarX, 131072 runs, 0 skips
yuv420p10be
12291 UNITS in planarX, 131072 runs, 0 skips
72935 UNITS in planarX, 131072 runs, 0 skips
yuv420p12le
12316 UNITS in planarX, 131072 runs, 0 skips
72708 UNITS in planarX, 131072 runs, 0 skips
yuv420p12be
12319 UNITS in planarX, 131072 runs, 0 skips
72577 UNITS in planarX, 131072 runs, 0 skips
yuv420p14le
12259 UNITS in planarX, 131072 runs, 0 skips
72516 UNITS in planarX, 131072 runs, 0 skips
yuv420p14be
12440 UNITS in planarX, 131072 runs, 0 skips
72962 UNITS in planarX, 131072 runs, 0 skips
yuv420p16le
10548 UNITS in planarX, 131072 runs, 0 skips
73429 UNITS in planarX, 131072 runs, 0 skips
yuv420p16be
10634 UNITS in planarX, 131072 runs, 0 skips
150959 UNITS in planarX, 131072 runs, 0 skips
Signed-off-by: Lauri Kasanen <cand@gmx.com>
This function wouldn't benefit from VSX instructions, so I put it
under altivec.
./ffmpeg_g -f rawvideo -pix_fmt rgb24 -s hd1080 -i /dev/zero -pix_fmt grayf32le \
-f null -vframes 100 -v error -nostats -
3743 UNITS in planar1, 65495 runs, 41 skips
-cpuflags 0
23511 UNITS in planar1, 65530 runs, 6 skips
grayf32be
4647 UNITS in planar1, 65449 runs, 87 skips
-cpuflags 0
28608 UNITS in planar1, 65530 runs, 6 skips
The native speedup is 6.28133, and the bswapping one 6.15623.
Fate passes, each format tested with an image to video conversion.
Signed-off-by: Lauri Kasanen <cand@gmx.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Passes fate on LE (with "lavc/jrevdct: Avoid an aliasing violation" applied).
Signed-off-by: Lauri Kasanen <cand@gmx.com>
Tested-by: Michael Kostylev on BE
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
./ffmpeg_g -f rawvideo -pix_fmt rgb24 -s hd1080 -i /dev/zero -pix_fmt yuv420p \
-f null -vframes 100 -v error -nostats -
1158 UNITS in planar1, 65528 runs, 8 skips
-cpuflags 0
19082 UNITS in planar1, 65533 runs, 3 skips
16.48 speedup ratio. On x86, SSE2 is ~7. Curiously, the Power C version
takes as many cycles as the x86 SSE2 version, yikes it's fast.
Note that this function uses VSX instructions, but is not marked so.
This is because several existing functions also make that mistake.
I'll submit a patch moving them once this is reviewed.
Signed-off-by: Lauri Kasanen <cand@gmx.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This patch addresses Trac ticket #5570. The optimized functions are in file
libswscale/ppc/input_vsx.c. Each optimized function name is a concatenation of the
corresponding name in libswscale/input.c with suffix _vsx.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
GCC tool had a bug of PPC intrinsic interpret, which has been fixed in GCC 4.9.1. This bug lead to
errors in two of our previous patches. We found this when we update our GCC tools to 4.9.1 and by
reading the related info on GCC website. We fix our previous error in two separate commits
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
add marcos GET_LS() GET_VF() LOAD_FILTER() LOAD_L1() GET_VF4() FIRST_LOAD() UPDATE_PTR() LOAD_SRCV() LOAD_SRCV8() GET_VFD() for POWER LE
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This gets rid of the variable-length scratch buffer by filtering 16
pixels at a time and writing directly to the destination. The extra
loads this requires to load the source values are compensated by not
doing a round-trip to memory before shifting.
Signed-off-by: Mans Rullgard <mans@mansr.com>
Chip Kerchner
Philip Langdale
Lauri Kasanen
Sergey Lavrushkin
Michael Niedermayer
Michael Niedermayer
Luca Barbato
Ronald S. Bultje
Dan Parrot
Diego Biurrun
Pedro Arthur
Rong Yan
Christophe Gisquet
Sean McGovern
Anton Khirnov
Mans Rullgard