Jingning Han [Wed, 11 Nov 2015 00:02:33 +0000 (16:02 -0800)]
Fix an encoding failure case when speed features are on
This commit fixes an encoding failure case triggered when early
termination feature is turned on for transform block size search.
It resolves the corresponding enc/dec mismatch issue.
Marco [Mon, 9 Nov 2015 21:36:56 +0000 (13:36 -0800)]
Add bias to zero/small motion for noisy source.
Change is only for real-time mode, speed >= 5, and non-screen content mode.
Add bias to zero/low motion for big blocks, if noise estimation
is enabled and noise level is above threshold.
jackychen [Mon, 9 Nov 2015 19:47:26 +0000 (11:47 -0800)]
VP9 dynamic resize: increase waiting time after key frame.
For 1 pass CBR mode: increase waiting time after key frame
before we start sampling rate control behavior for determining
resize. This change need to disable one internal resize(DownUp)
temporally since it requires a longer clip to do so.
Alex Converse [Tue, 3 Nov 2015 00:28:10 +0000 (16:28 -0800)]
Expand unconstrained nodes in pack_mb_tokens and loop on zeros.
Reduces Linux perf estimated cycle count for pack_mb_tokens on a
lossless encode on my desktop from 61858501855 to 48154040219 or from
26% of the overall profile to 21%.
This breaks 32-bit builds:
runtime error: load of misaligned address 0xf72fdd48 for type 'const
__m128i' (vector of 2 'long long' values), which requires 16 byte
alignment
+ _mm_set1_epi64x is incompatible with some versions of visual studio
Improves entropy coding to send the filter type only if
the motion vectors have subpel components.
Adds one new 8-tap smooth filter, and tweaks the others.
derflr: +0.695%
hevcmr: +0.305%
About 5% encode slowdown. No visible impact for decoding.
Also makes the interpolation framework flexible to support both
strictly interpolating filters as well as non-interpolating
filters that filter integer offsets. This is mainly for
further experimentation and if not found useful the code will
be removed.
Marco [Fri, 6 Nov 2015 00:00:15 +0000 (16:00 -0800)]
vp9: Updates to noise estimation.
Add threshold/condition on spatial_variance and brightness level.
Modification to normalization of block variance.
Change resolution limit below which we disable noise estimation.
Geza Lore [Wed, 28 Oct 2015 14:35:04 +0000 (14:35 +0000)]
Add AVX vectorized vp9_diamond_search_sad
This function now has an AVX intrinsics version which is about 80%
faster compared to the C implementation. This provides a 2-4% total
speed-up for encode, depending on encoding parameters. The function
utilizes 3 properties of the cost function lookup table, constructed
in 'cal_nmvjointsadcost' and 'cal_nmvsadcosts'.
For the joint cost:
- mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3]
For the component costs:
- For all i: mvsadcost[0][i] == mvsadcost[1][i]
(equal per component cost)
- For all i: mvsadcost[0][i] == mvsadcost[0][-i]
(Cost function is even)
These must hold, otherwise the AVX version of the function cannot be used.
Jingning Han [Tue, 3 Nov 2015 20:59:24 +0000 (12:59 -0800)]
Simplify txfm rate-distortion optimization
This commit refactors the rate-distortion optimization scheme for
transform block coding. When both ext-tx and var-tx experiments
are turned on, the encoding time for bus_cif at 1000 kbps goes down
from 706377 ms to 666503 ms (5.6% speed-up). The coding statics
remain unchanged.
Geza Lore [Wed, 4 Nov 2015 14:56:34 +0000 (14:56 +0000)]
Flip the result of the inverse transform for FLIPADST.
When using FLIPADST, the vp10_inv_txfm_add functions used to flip
the destination array, add the result of the inverse transform, to it
and then flip the destination back. This has been replaced by
flipping the result of the inverse transform before adding it to the
destination. Up-Down flipping is done by negating the destination
stride, and staring from the bottom, so it should now be free.
Left-right flipping is done with the usual SSE2 instructions in the
optimized code.
The C functions match the SSE2 functions as expected, so the C functions
now do the flipping as well when required. Adding this cleanly required
some refactoring of the C functions, but there is no measurable
performance impact when ext-tx is not enabled.
hui su [Wed, 7 Oct 2015 16:29:02 +0000 (09:29 -0700)]
ext-intra experiment
Currently there are two parts in this experiment: extra directional intra
prediction modes and the filter intra modes migrated from the nextgen branch.
Several macros are defined in "blockd.h" to provide controls of the experiment
settings. Setting "DR_ONLY" as 1 (default is 0) means we only use directional
modes, and skip the filter-intra modes; "EXT_INTRA_ANGLES" (default is 128)
defines the number of different angles we want to support; setting
"ANGLE_FAST_SEARCH" as 1 (default is 1) means we use fast sub-optimal search
for the best prediction angle, instead of exhaustive search. The fast search
is about 6 times faster than the exhaustive search, while preserving about
60% of the coding gains.
With extra directional prediction modes (fast search), we observe the following
code gains (number in parentheses is for all-key-frame setting):
derflr +0.42% (+1.79%)
hevclr +0.78% (+2.19%)
hevcmr +1.20% (+3.49%)
stdhd +0.56%
Speed-wise, about 110% slower for key frames, and 30% slower overall.
The gains of filter intra modes mostly add up with the gains of directional
modes. The overall coding gain of this experiment:
derflr +0.94%
hevclr +1.46%
hevcmr +1.94%
stdhd +1.58%
Marco [Fri, 30 Oct 2015 18:51:06 +0000 (11:51 -0700)]
Bias against non-zero mv for large blocks.
Change is only for real-time mode, speed > 5, and non-screen content mode.
Bias is based on block size and motion vector level (motion above some threshold).
Helps to improves stability in background from lightning changes.
PSNR/SSIM metrics on RTC set almost no change/neutral (within +/- 0.1).
Geza Lore [Tue, 3 Nov 2015 11:10:20 +0000 (11:10 +0000)]
Eliminate copying for FLIPADST in fwd transforms.
This patch eliminates the copying of data when using FLIPADST forward
transforms, by incorporating the necessary data flipping into the
load_buffer_* functions of the SSE2 optimized forward transforms. The
load_buffer_* functions are normally inlined, so the overhead of copying
the data is removed and the overhead of flipping is minimized. Left to
right flipping is still not free, as the columns need to be shuffled in
registers.
To preserve identity between the C and SSE2 implementations, the
appropriate C implementations now also do the data flipping as part of
the transform, rather than relying on the caller for flipping the input.
Overall speedup is about 1.5-2% in encode on my tests. Note that these
are only the forward transforms. Inverse transforms to come in a later
patch.
There are also a few code hygiene changes:
- Fixed some indents of switch statements.
- DCT_DCT transform now always use vp10_fht* functions, which dispatch
to vpx_fdct* for DCT_DCT (some of them used to call vpx_fdct*
directly, some of them used to call vp10_fht*).
Geza Lore [Tue, 3 Nov 2015 13:53:32 +0000 (13:53 +0000)]
Fix transform tables in C implementations.
These tables were out of sync with the indexing enum since the
refactoring in commit 4f16f119 (change 303389), due to the removal
of the ext_tx_to_txtype lookup table. This patch just puts them
back in order.
Jingning Han [Tue, 3 Nov 2015 16:56:47 +0000 (08:56 -0800)]
Re-work rate-distortion optimization scheme for transform coding
This commit re-works the rate-distortion optimization scheme for
transform coding. It improves the overall compression performance.
For derf set, the ext-tx experiment provides 2.27% coding gains,
and the new scheme that integrates multiple transform type selection
and recursive transform block partitioning provides a total of 3.24%
coding gains.
projects / libvpx / log