Peter de Rivaz [Mon, 14 Dec 2015 16:35:29 +0000 (16:35 +0000)]
Fix for issue 1114 compile error
In 32-bit build with --enable-shared, there is a lot of
register pressure and register src_strideq is reused.
The code needs to use the stack based version of src_stride,
but this doesn't compile when used in an lea instruction.
This patch also fixes a related segmentation fault caused by the
implementation using src_strideq even though it has been
reused.
This patch also fixes the HBD subpel variance tests that fail
when compiled without disable-optimizations.
These failures were caused by local variables in the assembler
routines colliding with the caller's stack frame.
Zoe Liu [Thu, 3 Dec 2015 00:27:46 +0000 (16:27 -0800)]
Restore the flexibility for the new 3 references
For the experiment of EXT_REFS, removed the previous special handling
on the new last 3 references, i.e. LAST2_FRAME, LAST3_FRAME, and
LAST4_FRAME, at the decoder, so that these new last references are
treated the same way as the other 3 references (LAST_FRAME,
GOLDEN_FRAME, and ALTREF_FRAME). Encoder changes have been made
accordingly to realize this flexibility.
The DST2 is implemented by input alternate sign-flip, followed
by DCT, followed by output reversal.
Results are roughly the same, but it should be easier to optimize
the DST2.
[Interestingly a mtrix multuiply implementation is about 0.1%
better].
paulwilkins [Tue, 15 Dec 2015 14:53:44 +0000 (14:53 +0000)]
1 pass VBR mode bug fix.
(copied from VP9)
The one pass VBR mode selects a Q range based on a
moving average of recent Q values. This calculation
should have been excluding arf overlay frames as these
are usually coded at the highest allowed value. Their
inclusion skews the average and can cause it to drift
upwards even when the clip as a whole is undershooting.
As such it can undermine correct adaptation of the allowed
Q range especially for easy content.
paulwilkins [Tue, 15 Dec 2015 10:27:51 +0000 (10:27 +0000)]
1 pass VBR mode bug fix.
The one pass VBR mode selects a Q range based on a
moving average of recent Q values. This calculation
should have been excluding arf overlay frames as these
are usually coded at the highest allowed value. Their
inclusion skews the average and can cause it to drift
upwards even when the clip as a whole is undershooting.
As such it can undermine correct adaptation of the allowed
Q range especially for easy content.
Marco [Mon, 14 Dec 2015 20:43:32 +0000 (12:43 -0800)]
Non-rd variance partition: Adjust logic for 32->16 force split.
Lower the threshold for splitting 32x32->16x16 based on average variance,
and add lower bound condition for this split to occur. This prevents
unneccassry splitting for areas with very low variance.
Jian Zhou [Fri, 11 Dec 2015 01:25:18 +0000 (17:25 -0800)]
Code clean of tm_predictor_32x32
Reallocate the xmm register usage so that no ARCH_X86_64 required.
Reduce memory access to the left neighbor by half.
Speed up by single digit on big core machine.
Jian Zhou [Thu, 10 Dec 2015 02:12:52 +0000 (18:12 -0800)]
SSE2 based h_predictor_32x32
Relocate the function from SSSE3 to SSE2, Unroll loop from 16 to 8,
and reduce mem access to left.
Speed up by single digit in ./test_intra_pred_speed on big core
machines.
Jingning Han [Wed, 2 Dec 2015 18:59:01 +0000 (10:59 -0800)]
Re-design motion compensated prediction mode entropy coding system
This commit re-works the entropy coding scheme of the motion
compensated prediction modes. It allows more flexible hyperplane
partition for precise classification.
Tom Finegan [Wed, 9 Dec 2015 21:52:06 +0000 (13:52 -0800)]
iosbuild.sh: Support macosx targets in Xcode 7.
Xcode 7 refuses to link to x86 and x86_64 code that's built for
iphone sim, so add an extra command line flag that forces iosbuild
to use darwin15 targets.
paulwilkins [Tue, 8 Dec 2015 15:48:24 +0000 (15:48 +0000)]
Changes to exhaustive motion search.
This change has been imported from VP9 and
alters the nature and use of exhaustive motion search.
Firstly any exhaustive search is preceded by a normal step search.
The exhaustive search is only carried out if the distortion resulting
from the step search is above a threshold value.
Secondly the simple +/- 64 exhaustive search is replaced by a
multi stage mesh based search where each stage has a range
and step/interval size. Subsequent stages use the best position from
the previous stage as the center of the search but use a reduced range
and interval size.
For example:
stage 1: Range +/- 64 interval 4
stage 2: Range +/- 32 interval 2
stage 3: Range +/- 15 interval 1
This process, especially when it follows on from a normal step
search, has shown itself to be almost as effective as a full range
exhaustive search with step 1 but greatly lowers the computational
complexity such that it can be used in some cases for speeds 0-2.
This patch also removes a double exhaustive search for sub 8x8 blocks
which also contained a bug (the two searches used different distortion
metrics).
For best quality in my test animation sequence this patch has almost
no impact on quality but improves encode speed by more than 5X.
Restricted use in good quality speeds 0-2 yields significant quality gains
on the animation test of 0.2 - 0.5 db with only a small impact on encode
speed. On most natural video clips, however, where the step search
is performing well, the quality gain and speed impact are small.
projects / libvpx / log