This change 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 clips though the quality gain and speed impact are small.
Change-Id: Id22967a840e996e1db273f6ac4ff03f4f52d49aa
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.
Change-Id: I6c2791d43022822a9e6ab43cd124a773946d0bdc
This reverts commit f1342a7b07.
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
Change-Id: I6f6fc3c11403344cef78d1c432cdc9147e5c1673
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.
Change-Id: I184055b864c5a2dc37b2d8c5c9012eb801e9daf6
Re-investigated the second-level sub-pixel motion search. Improved the
way of choosing search points. Rewrote the second-level search code.
At speed 0, the borg tests showed:
1. for stdhd set, Avg PSNR gain: 0.216%; Overall PSNR gain: 0.196%;
SSIM gain: 0.206%. Only 1 out of 15 clips showed PSNR loss.
2. for derf set, Avg PSNR gain: 0.171%; Overall PSNR gain: 0.192%;
SSIM gain: 0.207%. Only 3 out of 30 clips showed PSNR losses.
Added the condition for third-point checking, namely, less points
were checked. Speed tests showed no speed loss(Avg 0.3% speedup at
speed 0).
Change-Id: I6284ebb3fa7ba63be8528184c49e06757211a7f1
Choose a different diagonal point to check when the two costs are
the same, making it consistent with the way we choose the best mv.
This slightly changes the encoding result, and the derflr set borg
test at speed 0 shows 0.027% Overall PSNR gain, 0.024% Avg PSNR
gain, and 0.043% SSIM gain.
Change-Id: Ic8ee3a6767394866d159e4f9e1c777604dd73c17
If the current best mv(namely, the search center) is still the best mv
after the first level search, the second level checks is skipped. This
patch doesn't change the bitstream. At speed 0, it speeds up the encoder
by 1% - 2%.
Change-Id: I054c91b884d3f7aef157436c061744562bd6506d
The only difference between the two was that the vp9 function allowed
for every step in the bilinear filter (16 steps) while vp8 only allowed
for half of those. Since all the call sites in vp9 (<< 1) the input, it
only ever used the same steps as vp8.
This will allow moving the subpel variance to vpx_dsp with the rest of
the variance functions.
Change-Id: I6fa2509350a2dc610c46b3e15bde98a15a084b75
This commit fixes the integral projection motion search crash when
frame resize is used. It fixes issue 994.
Change-Id: Ieeb52619121d7444f7d6b3d0cf09415f990d1506
With the sad functions, and hopefully the variance functions soon,
moving to the vpx_dsp location, place the defines used in the
reference C code in a common location.
Change-Id: I4c8ce7778eb38a0a3ee674d2f1c488eda01cfeca
this macro was used inconsistently and only differs in behavior from
DECLARE_ALIGNED when an alignment attribute is unavailable. this macro
is used with calls to assembly, while generic c-code doesn't rely on it,
so in a c-only build without an alignment attribute the code will
function as expected.
Change-Id: Ie9d06d4028c0de17c63b3a27e6c1b0491cc4ea79
(see I3a05cf1610679fed26e0b2eadd315a9ae91afdd6)
For the test clip used, the decoder performance improved by ~2%.
This is also an intermediate step towards adding back the
mode_info streams.
Change-Id: Idddc4a3f46e4180fbebddc156c4bbf177d5c2e0d
the sse4 code expects 16-byte aligned arrays; vp8 already had a similar
change applied:
b2aa401 Align SAD output array to be 16-byte aligned
Change-Id: I5e902035e5a87e23309e151113f3c0d4a8372226
Move the scaling factor outside column projection. This avoids
repeated calculation of the same scaling factor. Profiling shows
that the percentage of vp9_int_pro_col_sse2 of overall cycles
goes from 2.29% down to 1.88%.
Change-Id: I5ac4e324ab2d7f33ba2de66dd2a12e04e04dfd66
This commit uses a 6-point 1-step refine motion search in the
integral projection based full pixel motion estimation, to replace
the current 9-point search.
It reduces runtime cost of speed -6 on some noisy clips, e.g.,
dark720p single thread
33314 b/f, 40.076 dB, 18231 ms ->
33307 b/f, 40.067 dB, 17768 ms
The compression performance for rtc set remains unchanged.
Change-Id: I194ea5a9ce52e5a10baeee36338633adc22f764c
Make the vp9_int_pro_motion_estimation() function return zero
motion vector if high bit depth is turned on, instead of removing
it from compiled codes.
Change-Id: Ia48f010eb590b2d517d5678c394110b326a1a95e
The original implementation had the following comment:
// Ignore mv costing if mvsadcost is NULL
However the current implementation does not allow for this.
If x exists then nmvsadcost must not be null.
This removes the only warning from -Wpointer-bool-conversion
https://code.google.com/p/webm/issues/detail?id=894
Change-Id: I1a2cee340d7972d41e1bbbe1ec8dfbe917667085
Change 72056 unfolded some macro definitions,
but lost some alternative behaviour required for
high bitdepth encodes.
This causes the encoder to crash, see issue 884.
Change-Id: I8ce4d73c9fe0a3c10ccb86fba210fabc8b2f0ccc
The function pointer in compressor instance does not change, so this
commit changes to call the function directly.
Change-Id: I9c9c460e3475711c384b74c9842f0b4f3d037cc5
This commit unfolds the legacy macro definitions used in the
sub-pixel motion search and refactors the operational flow for
later optimizations.
Change-Id: I3e3f770cad961d03d1a6eb0b2a0186cc77eaf2b8
- Some fixes to surface fit.
- Returns variance function as cost rather than sad in the
pattern search and diamond search functions. Only
vp9_pattern_search_sad function used in bigdia search
uses sad as integer 1-away costs.
- Deploys SUBPEL_TREE_PRUNED_MORE for speed 4+.
Results:
derf [Speed 3]: About +0.036% in coding efficiency without any
discernible speed loss.
derf [Speed 4]: About 2-3% faster at -0.199% loss in coding efficiency.
derf [Speed 5]: About 3-4% faster at -0.149% loss in coding efficiency.
Change-Id: I8462f94f6adb46966ca964f2bd0400977357fd63
One is a more aggressive version of the pruned subpel tree
search where only a single halfpel candidate is searched.
The search candidate is based on a surface fit result.
The other is a method to obtain the subpel position at one
shot based on the same surface fit.
The methods have not been deployed in any speed setting yet.
Change-Id: I34fef3f2e34f11396c9d1ba97f4be8c4ffca62d3
Adds code to return an integer cost list for NSTEP search. Then
uses it for pruned subpel search in speed 3.
derf: -0.06%
Speed on mobcal 720p increaes from 10.28 fps to 10.65 fps.
[Subject to further testing].
Change-Id: Ib591382d25b2c11bcaba9d3a27a93a9d1ab27a96
Improves function to return sad of integer pels by reusing integer
pels already visited in the smallest scale.
Turns on BIGDIA search for speed 4. Also, turns on the
first version of the pruned subpel search at this speed.
derf: -0.32% (speed 4)
Speed seems to improve by at least 5% but subject to verification.
Change-Id: Iaec8eaffd61d6237ac029e6a2a1b0a88b2a35271
Updates the vp9_pattern_search function to return integer one-away
neighbors' sad values, for subsequent use in speeding up the
sub-pel search. Also, removes code for the do_refine option
which is not being used currently.
Updates the integer and subpel functions to pass in a 5-element
sad list for output or input.
A new pruned sub-pel search algorithm is implemented that uses
the sad returned from the integer pel search. But it is not
deployed yet.
Change-Id: Ifa9f5ad024b5b660570366d2bd900343e1891520
Currently, vp9_diamond_search_sadx4() is only called when sse3 is
enabled, which is improper since sse2 optimization of sdx4df
functions are available. Changed to always use
vp9_diamond_search_sadx4().
Change-Id: I4b95d6b7a3c6c645783c373f0ba8d645ece24717
There are sse2 optimization of sdx4df functions. Instead of calling
vp9_refining_search_sadx4 only when sse3 is enabled, call it always.
Change-Id: I24f93818f7d4209d1425039e0eb099ff9ff08fe9
* Replace max_step_search_steps with constant MAX_MVSEARCH_STEPS
* Fold (reduce_first_step_size + speed > 5) into reduce_first_step_size
replacing uses of reduce_first_step_size that don't add the speed
check with zero.
Change-Id: Iae46395dbf3eaca138bf4d18b838a9e364b5a198
