I spotted an interesting pattern that I didn't see before that leads to the implementation being faster.
The bit shifting table I was using before is no longer needed, and was able to remove quite a few lines.
I also add use of FMA on the AVX2 version.
f32 1920x1080 1 thread with prelut
c impl
1434012700 UNITS in lut3d->interp, 1 runs, 0 skips
1434035335 UNITS in lut3d->interp, 2 runs, 0 skips
1423615347 UNITS in lut3d->interp, 4 runs, 0 skips
1426268863 UNITS in lut3d->interp, 8 runs, 0 skips
sse2
905484420 UNITS in lut3d->interp, 1 runs, 0 skips
905659010 UNITS in lut3d->interp, 2 runs, 0 skips
915167140 UNITS in lut3d->interp, 4 runs, 0 skips
915834222 UNITS in lut3d->interp, 8 runs, 0 skips
avx
574794860 UNITS in lut3d->interp, 1 runs, 0 skips
581035090 UNITS in lut3d->interp, 2 runs, 0 skips
584116720 UNITS in lut3d->interp, 4 runs, 0 skips
581460290 UNITS in lut3d->interp, 8 runs, 0 skips
avx2
301698880 UNITS in lut3d->interp, 1 runs, 0 skips
301982880 UNITS in lut3d->interp, 2 runs, 0 skips
306962430 UNITS in lut3d->interp, 4 runs, 0 skips
305472025 UNITS in lut3d->interp, 8 runs, 0 skips
gbrap16 1920x1080 1 thread with prelut
c impl
1480894840 UNITS in lut3d->interp, 1 runs, 0 skips
1502922990 UNITS in lut3d->interp, 2 runs, 0 skips
1496114307 UNITS in lut3d->interp, 4 runs, 0 skips
1492554551 UNITS in lut3d->interp, 8 runs, 0 skips
sse2
980777180 UNITS in lut3d->interp, 1 runs, 0 skips
986121520 UNITS in lut3d->interp, 2 runs, 0 skips
986489840 UNITS in lut3d->interp, 4 runs, 0 skips
998832248 UNITS in lut3d->interp, 8 runs, 0 skips
avx
622212360 UNITS in lut3d->interp, 1 runs, 0 skips
622981160 UNITS in lut3d->interp, 2 runs, 0 skips
645396315 UNITS in lut3d->interp, 4 runs, 0 skips
641057075 UNITS in lut3d->interp, 8 runs, 0 skips
avx2
321336400 UNITS in lut3d->interp, 1 runs, 0 skips
321268920 UNITS in lut3d->interp, 2 runs, 0 skips
323459895 UNITS in lut3d->interp, 4 runs, 0 skips
324949967 UNITS in lut3d->interp, 8 runs, 0 skips
The horizontal pass get ~2x performance with the patch
under single thread.
Tested overall performance using the command(avx2 enabled):
./ffmpeg -i 1080p.mp4 -vf gblur -f null /dev/null
./ffmpeg -i 1080p.mp4 -vf gblur=threads=1 -f null /dev/null
For single thread, the fps improves from 43 to 60, about 40%.
For multi-thread, the fps improves from 110 to 130, about 20%.
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Also add SIMD which works on lines because it is faster then calculating it on
8x8 blocks using pixelutils.
Signed-off-by: Marton Balint <cus@passwd.hu>
Speed of all modes increased by a factor between 7.4 and 19.8 largely depending
on whether bytes are unpacked into words. Modes 2, 3, and 4 have been sped-up
by a factor of 43 (thanks quick sort!)
All modes are available on x86_64 but only modes 1, 10, 11, 12, 13, 14, 19, 20,
21, and 22 are available on x86 due to the number of SIMD registers used.
With a contribution from James Almer <jamrial@gmail.com>
The internal line accumulator for 16bit can overflow, so I changed that
from int to uint64_t in the C code. The matching assembly looks a little
weird but output looks correct.
(avx2 should be trivial to add later.)
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Reviewed-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Both are 2-2.5x faster than their C counterpart.
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Reviewed-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
The only difference with mp=pp7 is that default mode is "medium", as stated
in the MPlayer docs, rather than "hard".
Signed-off-by: Stefano Sabatini <stefasab@gmail.com>
integration by Neil Birkbeck, with help from Vitor Sessak.
core SSE2 loop by Skal (pascal.massimino@gmail.com)
Reviewed-by: Clément Bœsch <u@pkh.me>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This reverts commit a87b17f328.
This reduces the amount of non LGPL code, making a relicensing to LGPL
easier
Conflicts:
libavfilter/vf_yadif.c
libavfilter/x86/yadif.c
libavfilter/x86/yadif_template.c
libavfilter/yadif.h
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
These smaller samples do not need to be unpacked to double words
allowing the code to process more pixels every iteration (still 2 in MMX
but 6 in SSE2). It also avoids emulating the missing double word
instructions on older instruction sets.
Like with the previous code for 16-bit samples this has been tested on
an Athlon64 and a Core2Quad.
Athlon64:
1809275 decicycles in C, 32718 runs, 50 skips
911675 decicycles in mmx, 32727 runs, 41 skips, 2.0x faster
495284 decicycles in sse2, 32747 runs, 21 skips, 3.7x faster
Core2Quad:
921363 decicycles in C, 32756 runs, 12 skips
486537 decicycles in mmx, 32764 runs, 4 skips, 1.9x faster
293296 decicycles in sse2, 32759 runs, 9 skips, 3.1x faster
284910 decicycles in ssse3, 32759 runs, 9 skips, 3.2x faster
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This is a fairly dumb copy of the assembly for 8-bit samples but it
works and produces identical output to the C version. The options have
been tested on an Athlon64 and a Core2Quad.
Athlon64:
1810385 decicycles in C, 32726 runs, 42 skips
1080744 decicycles in mmx, 32744 runs, 24 skips, 1.7x faster
818315 decicycles in sse2, 32735 runs, 33 skips, 2.2x faster
Core2Quad:
924025 decicycles in C, 32750 runs, 18 skips
623995 decicycles in mmx, 32767 runs, 1 skips, 1.5x faster
406223 decicycles in sse2, 32764 runs, 4 skips, 2.3x faster
387842 decicycles in ssse3, 32767 runs, 1 skips, 2.4x faster
307726 decicycles in sse4, 32763 runs, 5 skips, 3.0x faster
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
