SSSE3 CDEF

added 1 commit

• 9e998b0b - wip

Compare with previous version

added 1 commit

• d8932376 - wip

Compare with previous version

added performance label

mentioned in issue #216

added 1 commit

• 9bd64a12 - wip

Compare with previous version

added 1 commit

• 1d4119a0 - wip 32 bit PIC

Compare with previous version

A very impressive performance increase!

decoder	dav1d	dav1d
Build	222bf249	1d4119a0
Build date	2019-02-15	2019-02-18
ISA	SSSE3	SSSE3
Morocco ST	15,89	19,48	122,6%
Morocco MT	66,21	97,48	147,2%
Dua Lipa ST	8,45	12,97	153,5%
Dua Lipa MT	37,66	59,36	157,6%

Intel Xeon w3680 Westmere 6x3.33 GHz (12 logical cores)
Chimera 8-bit 1080p

stock f26bf7fa

frame threads	time
1	13m27.962s
6	4m8.488s

patched 1d4119a0

frame threads	time	diff
1	9m35.266s	40% faster
6	2m49.121s	47% faster

Intel Xeon E5-2699 v4 Broadwell 22x2.2 GHz (44 logical cores)
Liquid cooled with sustained turbo 2.8 GHz
Chimera 8-bit 1080p

stock f26bf7fa (cross GCC 8.2 i686)

frame threads	time
1	15m55.161s
11	4m53.511s
22	4m38.825s

patched 1d4119a0 (cross GCC 8.2 i686)

frame threads	time	diff
1	crash
11	crash
22	crash

It seems when cross-compiled using mingw-w64 and GCC 8.2, dav1d.exe crashes immediately (only with this patchset).

added 6 commits

• 1d4119a0...f26bf7fa - 2 commits from branch videolan:master
• 87333927 - cdef: remove redundant AVX2 code
• 99e4ad64 - not so relevant avx2 changes
• b97f4fc0 - cdef: add SSSE3 implementation
• 3686c068 - dbg

Compare with previous version

added 1 commit

• 9ee8c76e - black_magic_investigation

Compare with previous version

added 10 commits

• 9ee8c76e...3f646a21 - 4 commits from branch videolan:master
• dd47af65 - cdef: remove redundant AVX2 code
• 08101188 - not so relevant avx2 changes
• 450b9a6f - cdef: add SSSE3 implementation
• 14e1ce76 - dbg
• 7285e8d2 - black_magic_investigation
• a4bae912 - test_ci_fix(lzcnt=>bsr+xor31)

Compare with previous version

added 2 commits

• c55589b8 - cdef: improve AVX2 cdef_filter macro consistency
• f038a1d9 - cdef: add SSSE3 implementation

Compare with previous version

mentioned in issue #15 (closed)

added 2 commits

• c2b000fb - cdef: fix potential crash in AVX2
• 2d9544ff - cdef: add SSSE3 implementation

Compare with previous version

added 4 commits

• 33d3dfd6 - cdef: improve AVX2 cdef_filter macro consistency
• 12bd1cf8 - cdef: remove redundant AVX2 code
• bdcdebdc - cdef: fix potential crash in AVX2
• d8df212c - cdef: add SSSE3 implementation

Compare with previous version

changed the description

open for review @rbultje @gramner @jamrial

crashes with the new win32 unaligned stack job (I'll fix that tomorrow)

added 5 commits

• dce4e788 - 1 commit from branch videolan:master
• aaf43c71 - cdef: improve AVX2 cdef_filter macro consistency
• dd368077 - cdef: remove redundant AVX2 code
• fe7429c1 - cdef: fix potential crash in AVX2
• e61d2fa1 - cdef: add SSSE3 implementation

Compare with previous version

added 2 commits

• a54858fb - cdef: add SSSE3 implementation
• d01b00c2 - TODO

Compare with previous version

added 1 commit

• 42f947f5 - TODO

Compare with previous version

added 1 commit

• 4c0e91e9 - wip_check_if_checkasm_is_too_narrow

Compare with previous version

added 1 commit

• cb4317da - wip

Compare with previous version

added 6 commits

• e2f1d869 - checkasm: widen cdef filter strengh max value
• 73fea97a - cdef: improve AVX2 cdef_filter macro consistency
• 364769f5 - cdef: remove redundant AVX2 code
• 0988f07e - cdef: fix potential crash in AVX2
• 110732b1 - cdef: add SSSE3 implementation
• 6d6f6b47 - TODO

Compare with previous version

added 5 commits

• 2f485fc6 - checkasm: widen cdef filter strengh max value
• 30339e8d - cdef: improve AVX2 cdef_filter macro consistency
• 8a6fe183 - cdef: remove redundant AVX2 code
• d29387ed - cdef: fix potential crash in AVX2
• 82834f34 - cdef: add SSSE3 implementation

Compare with previous version

unmarked as a Work In Progress

added 2 commits

• 30357edd - cdef: add SSSE3 implementation
• 60a7dbe7 - cdef: add AVX2 4x8 filter

Compare with previous version

marked as a Work In Progress

@psilokos I see a few AVX2 commits, do you expect significant changes in AVX2 performance?

added 2 commits

• 60238f5a - cdef: add AVX2 4x8 filter
• 76302262 - cdef: add SSSE3 implementation

Compare with previous version

@EwoutH not for what was already there, but I added the YUV 422 case, so you can make a bench of this one yes :)

@psilokos Then I will go on a great quest for 4:2:2 samples :)

@EwoutH you already have one here https://drive.google.com/open?id=18Sg7Kk37mOmYUlk6ycbu4_GCuC7sLvGt

added 1 commit

• e88293fb - cdef: add SSSE3 implementation

Compare with previous version

unmarked as a Work In Progress

resolved all discussions

assigned to @gramner

changed milestone to %0.2.0

Doesn't the "fix potential crash in AVX2" only apply to 4x8? In that case it should be squashed into the commit which adds the 4x8 version.

Intel Xeon w3680 Westmere 6x3.33 GHz (12 logical cores)
Chimera 8-bit 1080p

stock 1ba8423a

frame threads	tile threads	time
1	1	13m21.751s
6	1	4m7.455s
6	2	3m1.221s

patched e88293fb

frame threads	tile threads	time	diff
1	1	9m32.778s	40% faster
6	1	2m48.903s	46% faster
6	2	1m47.225s	69% faster

---

Intel Xeon E5-2699 v4 Broadwell 22x2.2 GHz (44 logical cores)
Liquid cooled with sustained turbo 2.8 GHz
Chimera 8-bit 1080p

stock 1ba8423a (cross GCC 8.2 i686)

frame threads	tile threads	time
1	1	16m22.632s
11	1	5m5.261s
22	1	4m46.58s
11	2	3m59.2s
22	2	3m44.466s

patched e88293fb (cross GCC 8.2 i686)

frame threads	tile threads	time	diff
1	1	10m9.794s	61% faster
11	1	2m39.771s	91% faster
22	1	2m29.10s	92% faster
11	2	1m28.825s	169% faster
22	2	1m23.607s	168% faster

mentioned in issue #78

added 17 commits

• e88293fb...9cb94d29 - 12 commits from branch videolan:master
• 6e774b23 - checkasm: widen cdef filter strengh max value
• 18ea9ba0 - cdef: improve AVX2 cdef_filter macro consistency
• 6b62d725 - cdef: remove redundant AVX2 code
• 018b8db0 - cdef: add AVX2 4x8 filter
• 1e2fc370 - cdef: add SSSE3 implementation

Compare with previous version

added 6 commits

• 3b19f5b6 - checkasm: decrease cdef filter min damping value
• ee3c5aa5 - cdef: improve AVX2 cdef_filter macro consistency
• f9f50601 - cdef: remove redundant AVX2 code
• 50310d80 - cdef: optimize 4 by X filters for HAVE_RIGHT=0
• 040598ba - cdef: add AVX2 4x8 filter
• 2f3ce00c - cdef: add SSSE3 implementation

Compare with previous version