Check the bottom neighbor availability with tile boundary for the
intra prediction condition.

BUG=aomedia:1088

Change-Id: I9baa98f8f18da84f95fd83ceca5556cfe9d9d844

Change-Id: Ia790b7f12cf78f628e99bb7fe2e272771c1a68c7

This experiment allows two choices for the frame-level reference
mode flag: SINGLE_REFERENCE and REFERENCE_MODE_SELECT. It removes
the choice of COMPOUND_REFERENCE, as it has been barely used.

Change-Id: I8af18acd2fe3c0d4928f2b05f35aad0ebcb1556a

Changed convolve 1d function names to be more precise.

Change-Id: I9bc1e99f5032ed8a84c61616036735af23d9a587

BUG=aomedia:1065

Change-Id: I35eda84e249fc42e319604079c122df8ab101f90

BUG=aomedia:1065

Change-Id: I0951a276865a5d810eb04bbb5251ed5c1b417ca4

BUG=aomedia:1096

Change-Id: Ibbb448e217ae1dd9a096c23d01b08c9804583003

This is to keep the high range within 32 bits.

AV1FwdTxfm2d.CfgTest passes after this fix.

Change-Id: I2df463c4ec9260c544d68aad445b60cabe2b531b

The purpose of this change is to reduce the cycles needed for warped
motion parameter estimation.

Method 1:
If we remove the 2-bit bit-depth reduction(as in patch set 2), the
downshifting of A, Bx, By is also removed. The borg test result(over
the baseline) is:
             avg_psnr ovr_psnr  ssim
lowres:      0.023     0.020    0.071
cam_lowres: -0.009    -0.017   -0.031

Method 2:
In theory, the above change uses 2 more bits for elements of A, Bx,
By. In patchset 3, we modified LS_STEP to be 8(1 full pixel), and now,
the least 2 bits in A, Bx, By elements are always 0. Namely, 2-bit
bit-depth reduction are achieved without extra operations. The borg
test result(over the baseline) is:
lowres:     -0.004    -0.007   -0.023
cam_lowres: -0.031    -0.033   -0.045
This is a little better than patch set 2 result.

Method 2 is the final choice.

Change-Id: I945aaba412e2ea86b7d67e8a90741fdf395b94cd

Change-Id: Ia8321afd087f99371cdf07f3a03249580e09964d

This patch simplies the checking criteria for the two groups of
compound modes. It also makes the encoder side cdf update inside the
RD loop consistent with that in the bitstream.

Experimental results on Google test sets (30 frames of lowres and
midres) confirm this patch obtains identical coding performance.

Change-Id: I170eea91f7d2be2170df544cfc2c692b09aa82d6

Fix the comments on the precision of quantizers and tx coefficients
during a quantization process for different input depth and tx size.

I think the author really meant "de-quantized/de-coded coefficients" by
"quantized/coded coefficients". So, made it clear to avoid any possible
misunderstandings.

Change-Id: Ib92ac7dcfddcbe58cf3adfb9448497512381c1f5

Load four 8-bit input and process.

Change-Id: I9b3ba58ea3a03c6a8129379afa37c54a57e04501

Also, make them const when appropriate.

Change-Id: I96d544e2cc9a0bce4d52fd33e44a4eaa40edda3c

Can be up to >10% faster with bit exact results

Change-Id: I5f169673fd2d5af96f425f00d862f3c989228d2e

16x64 reuses scan order of 16x32
64x16 reuses scan order of 32x16

max eob is curtailed to 512 (instead of 1024) for both.

Change-Id: Iac2145aa5e3d090009e2a2f5715caa8d84dfb2ee

Change-Id: I17a82393f1b7230119f499e2f9ed8d0b8fe5ba25

Moving CfL to using partition unit DC_PRED requires 4:1 and 1:4 DC_PRED,
which are not currently implemented. A simple solution is to disable CfL
for 4:1 and 1:4 partitions.

CfL is also disabled for luma intra partitions < 4x4. This is inherent
to luma intra prediction partition sizes. We add an assert to enforce
this.

Resulting in the following regression for Subset1
   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |   SSIM | MS SSIM | CIEDE 2000
-0.0093 |  0.1803 |  0.1519 |  -0.0180 | 0.0256 |  0.0226 |     0.0352

https://two.arewecompressedyet.com/?job=CfL%402017-11-30T19%3A05%3A05.639Z&job=CfL-Disable-4to1%402017-11-30T19%3A04%3A00.761Z

Change-Id: Ie2c8b4d9cb6b6f33a103b540209e1a2fb6df74a7

This is for speeding up the testing process

Change-Id: I90866fa239794f14e4801675d471dbf50b779d18

When lv_map_multi is on,
od_ec_encode_q15 is able to handle the situation of fh == fl

Change-Id: I7c837dda561f1d25b0203c018763dadd0cbbc75a

Added a copy function (c version and sse2 version) for full-pixel motion
vectors for jnt_comp experiment following existing av1_convolve_2d_copy
function.

Change-Id: I20fd2219799f9c1451f591574fbe97364f40e0f0

The -Ox check in still useful to avoid the version of nasm distributed
with Apple Xcode.

This reverts commit 29b0c186.

Change-Id: I9237791802267da708c3be8e5a83ca8d71e74afc

This allows for the following options:
 Set 0:
        Inter: All 16 txfms
        Intra: Discrete Trig transforms w/0 flip (4) + Identity (1) +
               1D Hor/vert DCT (2)
 Set 1:
        Inter: Discrete Trig transforms w/ flip (9) + Identity (1) +
               1D Hor/Ver DCT (2)
        Intra: Discrete Trig transforms w/0 flip (4) + Identity (1)
 Set 2:
        Inter: Discrete Trig transforms w/ flip (9) + Identity (1)
        Intra: Discrete Trig transforms w/0 flip (4) + Identity (1)

Results on lowres 40 frames with
disable-ext-partition disable-ext-partition-types

Set 0: 0.03%
Set 1: No change
Set 2: 0.06%

Change-Id: Iec57d8c8fcfa0891528de4ca88f54753dfcb5284

Change-Id: Ieed4ecce63a2a3b2a74c40ccddabe91cb9386632

Constrain 16x64 transform so that the bottom 16x32 is zero;
constrain 64x16 transform so that the right 32x16 is zero;
Also implement 32x64 transform better to reduce intermediate
coefficient range.

Change-Id: Ia9050ee741ed1d5b02a42616635b496d637d932f

Extend context model for comp_group_idx.
Save sending comp_group_idx when masked_compound is not allowed.

Change-Id: Ia7ae53958c9e1c8fe07be4b14a425d9b8648082d

Remove COMPOUND_AVERAGE from compound_type_cdfs since it is now grouped
to compound_idx. However, COMPOUND_AVERAGE is still used elsewhere.

Change-Id: Ie0d460aabf9252e80eb4130cfef9aaf0efc3969d

Divide compound inter prediction modes into two groups:
Group A: jnt_comp, compound_average
Group B: interintra, compound_segment, wedge

Change-Id: I1142da2e3dfadf382d6b8183a87bde95119cf1b7

Change-Id: Ie3e599def556a90c474680567c4537508de2e30a

This 4-point Type-IV asymmetric DST uses the same computation graph as
 the 4-point Type-IV DST.
This change improves the accuracy of the 8-point Type-II DCT:

Old MSE: 1.8927096972341813413041010372151e-06
New MSE: 1.7946367518072710517065436117146e-06

subset-1:

new_dst4@2017-12-04T06:31:41.096Z -> new_dst4a@2017-12-04T06:32:22.698Z

   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
-0.0143 |  0.0410 | -0.2166 |  -0.0556 | -0.0379 | -0.0461 |    -0.0002

Change-Id: Ifde11fca987220130c1657306b0df34ec2f3fe25

This change slightly improves the 16-point DCT round trip accuracy due
 to changes in the rounding.

subset-1:

new_dst2@2017-12-04T01:59:57.412Z -> new_dst4@2017-12-04T06:31:41.096Z

  PSNR | PSNR Cb | PSNR Cr | PSNR HVS |   SSIM | MS SSIM | CIEDE 2000
0.0078 | -0.0001 |  0.0198 |   0.0432 | 0.0408 |  0.0502 |    -0.0057

Change-Id: I75783ace97834af89e70c9ce3002c6f09176e343

This 2-point Type-IV DST uses the same computation graph as the
 asymmetric 2-point Type-IV DST.
Because this transform is embedded, it may be possible to remove the
 initial averaging step by splitting the 2-point Type-IV DST into
 separate forward and inverse transforms.
This change also reduces two multiplication constants (forward and
 inverse transform) so they are less than 1.

subset-1:

new_dst2a@2017-12-04T01:59:12.884Z -> new_dst2@2017-12-04T01:59:57.412Z

   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
-0.0126 |  0.0387 |  0.0441 |   0.0554 | -0.0301 |  0.0034 |    -0.0342

Change-Id: I98568e0c5b97e3a6af27653ddab845ce97d2a53d

This change improves the accuracy of the 4-point Type-II DCT:

Old MSE: 6.2711279572488185887270981198199e-08
New MSE: 6.0281623825882593130347914239103e-08

It also reduces a multiplication constant so it is less than 1.

subset-1:

daala_tx@2017-12-04T01:58:11.321Z -> new_dst2a@2017-12-04T01:59:12.884Z

  PSNR | PSNR Cb | PSNR Cr | PSNR HVS |   SSIM | MS SSIM | CIEDE 2000
0.0274 |  0.0255 |  0.0969 |  -0.0024 | 0.0274 |  0.0027 |     0.0110

Change-Id: I7c9d389af8e98cb39f3bc5923134b5dfe174ba0a

We use the former variant for the 8-point row transforms when the
number of columns exceeds 8, since the scaling can exceed 16 bits.
We ues the latter variant for the 8-point column transforms when
the number of rows exceeds 8, since it allows us to perform twice
as many transforms in parallel.

Change-Id: Ia2595ad827636342f70c3d5b99cf05c278bd1389

On x86 there is no PMULHRSD for use in the 32-bit transform
versions, so the fastest approach is to just do a normal 32-bit
multiply and manually shift and round. This requires keeping the
constants in their reduced precision instead of always promoting
them to Q15.

Change-Id: I76339b5567da3f08f34882a707e0c93122991946

This creates the mechanism by which we can define multiple versions
for different instruction sets and word sizes.

This commit makes no functional changes.

Change-Id: If49ebfc989247692df9c501bea05eb811944d52a

This will aid us in defining multiple versions for different
instruction sets and word sizes without duplicating all of the
code.

This commit makes no functional changes.

Change-Id: I7f240281b0e9edba19c2ee17b9ff3ae36400dcc2

Change-Id: I49e64d3e062c32925abe30118a64a714073fd4d0

This is using the Type IV version with flattened multiplies for
now, since we've identified some potential 16-bit overflows in the
Type VII inverse.

Change-Id: Ib79413ea27efac8b0207602001595ae3ac294eae

Change-Id: Ieb20d64e6531960188feb65296acfa952c858043