Implemented parallel loopfiltering, which uses existing tile-
decoding threads. Each thread works on one row, and when that row
is loopfiltered, it moves to next unattended row. To ensure the
correct filtering order, threads are synchronized and one
superblock is filtered only if the superblocks it depends on are
filtered already.

To reduce synchronization overhead and speed up the decoder, we use
nsync > 1 for high resolution.

Performance tests:
1. on desktop:
8-tile 4k video using 8 threads, speedup: 70% - 80%
4-tile HD video using 4 threads, speedup: ~35%
2. on mobile device(Nexus 7):
4-tile 1080p video using 4 threads, speedup: 18% - 25%
4-tile 1080p video using 2 threads, speedup: 10% - 15%

Change-Id: If54b4a11960dd706c22d5ad145ad94156031f36a

* Avoid unnecessary type erasure
* Prune unused/duplicate fields from struct rdcost_block_args
* Make struct rdcost_block_args a local

Change-Id: I4f1fd4837ccd028bbfe727191ee8d69f0463b7e5

When showing a previously decoded frame, i.e. when
show_existing_frame=1, the update of the
last_show_frame flag must be disabled.

This is to ensure that the last_show_frame flag
reflects the state of the flag for the immediately
previously decoded frame rather then the value that
was forced to ensure that a previously decoded frame
would be displayed.

This patch also adds a test vector to verify that the
display_existing_frame flag works as expected. Code
for generating the test vector can be found in this
patch:
https://gerrit.chromium.org/gerrit/#/c/68581/

(Bug originally reported by Alexander Voronov
<ru.xalba@gmail.com>).

Change-Id: I731d288fba02088959f7fcc87707137fffc6acf5

use mode instead

Change-Id: I419d7a2dc4b0714ca6ff723c5e824521c150c460

Added a constant to represent the minimum KF boost
rather than using the magic number 2000 in the code.

Change-Id: I9428b61f47d26312caff81c6f9ae8587df004791

So x86_64-win64-vs11 can build successfully.

Change-Id: If354c2ea3921fac8c9b413ed39223e70bc20c535

Some changes in 1ca11865 were mistakenly reverted by a later merge,
this commit re-instated the chanages from values to enums.

Change-Id: Ia6b01c31da584a1f612996e6432612c1295b9eaf

Change-Id: Iea7c9fa0726dbf9792eea79e6a05eb8a3c718d45

Change-Id: I08f45573e0b2195c09fb6aecacb4c57431a711ea

In this new mode, the size range is strictly determined by the min
and max partition size in neighborhood blocks.

Niklas720 encoding time at cpu-used -5 goes from 56250ms to 50676ms,
a 10% reduction.

Change-Id: I316b0e2ac967ff3fad57b28d69c0ec80b7d8b34e

Change-Id: If446225afbb49f6033c2a4516a37c377de6f70f7

Change-Id: Ic2ff6405f01fd43d07c5ee3b5e374909401115cc

Includes a few fixes and clean-ups that adds the ability
to use alt-ref frames in one-pass mode.
Whether alt-refs are actually used or not is controlled by a
macro USE_ALTREF_FOR_ONE_PASS in vp9_firstpass.c.

This first cut seems to improve derf by 15+% in 1-pass mode.
But further experiments with parameters are underway.

Change-Id: I78254421435478003367c788c7930d2dc4ee2816

This patch only works if the video is a width and height that are both
a multiple of 32..   It sets every partition to 16x16, and does INTRADC
only on the first frame and ZEROMV on every other frame.   It always does
does the largest possible transform, and loop filter level is set to 4.

Was ~20% faster than speed -5 of vp8

Now 20% slower but adds motion search ( every block ), nearest, near
and zeromv

The SVC test was changed because - while this realtime mode produces
bad quality albeit quickly, it isn't obeying all the rules it should
about which frames are available.

Change-Id: I235c0b22573957986d41497dfb84568ec1dec8c7

Trap divide by 0 that could occur with a 0 rate target
in aq mode COMPLEXITY_AQ.

Change-Id: I034514f512b2a0db470ae8d37ea395278bf473cf

Added macros to reduce the code duplication.

Change-Id: I1916aa5a386ea07d961d4ec439ab09bb8c45487d

Change-Id: Ia8fa3961eec34545465018281dc022bc6f73869a

Change-Id: I0c286e3d68a4a4ecf6df02e6fd9990327b0ceb22

It seems we don't use it and not going to.

Change-Id: Ie76cd04dafc79b0a5911f8957d4253ca2d787f0c

It is enough to specify (e.g.) idct16, it is obviously different from
idct16x16.

Change-Id: I6b408a37a945de3162429380b59a775b03b95db0

which is 7.8 times faster than C.

Change-Id: I858ef4ec09202a07d445da8db702783d6d9d7321

Change-Id: I8d17867a4772554cbba2bd113cc5b4c99d50146d

Change-Id: I297954b16bce9e23931331520eadfb47540ff660

Change-Id: I832cf83871044bfee7b7e57dbd31bae05cbd53e9

Adds multiple filters in the 0.5-1.0 range in the last stage
of the resize functions to prevent over-smoothing/aliasing

Change-Id: I1a615adb16f0df5095790945c94b28b4d6a6fc48

Change-Id: I31373ad860eb554eb3b03e877e8fba580dc3de07

This avoids fitlering a frame multiple times at the same level.

Change-Id: I1fd54dd7ea257d16da8569f48036b8fad3a3ed61

Factor out the code that tries filtering a frame at a given level.

Change-Id: Ia04507e3ce6b1ad6ae7d05a9d88222fd319f44b7

We don't use different filter kernels for x and y, it is always one kernel
for both directions.

Change-Id: Iefcbb02ec74bf46ea20d9dca672a3efd5d631517

That force the stop of subpel search possibly at full/half/quater pel
stages

Change-Id: Ie50c500417bd78e1a53e6620bd4c2b85f63d9c67

Corresponding renames:
  subpel_kernel              => interp_kernel
  vp9_get_filter_kernel()    => vp9_get_interp_kernel()
  pred_filter_type           => pred_interp_filter
  adaptive_pred_filter_type  => adaptive_pred_interp_filter
  mcomp_filter_type          => interp_filter
  read_interp_filter_type()  => read_interp_filter()
  write_interp_filter_type() => write_interp_filter()
  fix_mcomp_filter_type()    => fix_interp_filter()

Change-Id: I1fa61fa1dc81ebbf043457c3ee2d8d4515bee6d3

Also change its wrongly named dest parameter to reference.

Change-Id: Ide142dead31c9ccda1f09a48b221284369783fb7

Change-Id: I24ff8ab3d2c807906aa86974bcb4c540256206de

SSE for a 64x64 block with 3 planes can go as high as 3*2^28. So left
shift by 4 may overflow 32 bit int.

Change-Id: I63c84aa56894788bb987299badabbd7cc6fd0be6

Use this method with rt at speed -5.

Change-Id: If3bd6fad4c05ddde72131442dad191e4145047e7

The sum of squared mv components can go beyond int range for large
 input resolution. This commit changed the type to int64 to avoid
overflow.

Change-Id: Ib21ea2817845cea1435f893064e6417c79c5bc64

Change-Id: I8cfa5d5eb2c1bbacd9b604cc5dc0a2cd2e5cebb8

Change-Id: I5173f996612e410d9cd495df9414d194b1ab18f3

Change-Id: Ia12aae491202098ff66366145aa0c3da38dc97e5

which is 3.5 times faster than C.

Change-Id: I24439ba7a2971829c11620f34848facf2c916678