Duplicate the ASM custom build rules: build each ASM object
once for each library. This avoids creating a race condition
on the output objects for the assembly custom build rules
when building the static and shared libraries from the same
configuration.

BUG=aomedia:3214

Change-Id: I8c7c44d26ffc1b6c2b9fe2556bf9da800d90c496
(cherry picked from commit 4d5b1222)

In the definition of AOM_ENCODER_ABI_VERSION, replace
AOM_EXT_PART_ABI_VERSION with its value (3) in libaom v3.2.0, so as to
not break the ABI version check in aom_codec_enc_init_ver() when an
application compiled against libaom v3.2.0 passes the old value of
AOM_ENCODER_ABI_VERSION to aom_codec_enc_init_ver().

Mark the external partition API as experimental. This was announced in
the libaom v3.2.0 release notes, but I neglected to add a note to the
source code.

Bug: 3236
Change-Id: Iefbe11ebce94c1ae7ca1e5f05f434eb71824b8fa

Bug: aomedia:3223
Change-Id: I0ca54016d88ec000b94e36b9f9a1266dd365240c

w/gcc-11 in
  av1_build_inter_predictors_for_planes_single_buf
  av1_build_wedge_inter_predictor_from_buf

the array parameters' size may vary based on the number of planes;
replace [3] w/[] and a size comment

Change-Id: I7e9e87bffca4b67ff1e88295383dbf49f5c43a22
(cherry picked from commit 3a862d3d)

While referencing the argument 'search_sites' in functions
av1_make_default_fullpel_ms_params() and av1_set_mv_search_method(),
the size of this array is used as NUM_SEARCH_METHODS instead of
allocated size NUM_DISTINCT_SEARCH_METHODS. This CL corrects the same
for better understanding.

BUG=aomedia:3205

Change-Id: I1a6c20f1755b2e920638941df2cc93268d94cbb2
(cherry picked from commit 67dc8da7)

w/gcc-11
this matches the definition of the function with the declaration

this is similar to the change in libvpx:
7fbcee49d quiet -Warray-parameter warnings

Bug: aomedia:3204
Change-Id: I757b731b9560cb0b0ceec4ec258ec5af5a183b3d
(cherry picked from commit 06e13e81)

BUG=aomedia:3223

Change-Id: I3f3073ed2ae64d6e906396a9f1f06829553028c8
(cherry picked from commit 82175730)

Bug: aomedia:3223
Change-Id: I012913b4bcffdd7bf5abc9e3e445df20bfa2eeee
(cherry picked from commit e3cb1a31)

Bug: aomedia:3223
Change-Id: Ia3a351899bbab04fb506ed10f038a6b8f8ec20f4
(cherry picked from commit 8ca6f2f2)

Bug: aomedia:3223
Change-Id: I29d392670c8d53cc1ecbab8a0b9fc2f2c3147a3c

The parameter type for AOMD_GET_SHOW_FRAME_FLAG and AOMD_GET_BASE_Q_IDX
should be int*.

The parameter type for AOMD_GET_ORDER_HINT should be unsigned int*.

The AOM_CTRL_USE_TYPE macro assumes the control ID takes one parameter.
Since AV1D_GET_MI_INFO takes three parameters, the AOM_CTRL_USE_TYPE
macro can't be used with AV1D_GET_MI_INFO.

Bug: aomedia:3223
Change-Id: I9866aed3551b56e017281a3db71c6f05639922fe
(cherry picked from commit 8bc4488c)

Bug: aomedia:3223
Change-Id: I26864d6e6353c234fbcb6b62f98576aae408f15a
(cherry picked from commit f563bf5d)

Some of the controls were missing the type-checked ids.

Bug: aomedia:3223
Change-Id: I1e0fc088c99af147143cb040251aaa463452df67
(cherry picked from commit 109d8cf6)

armv7 build failure was seen with gcc 11. armv7-a target no longer
includes an FPU by default, so we need to specify -march=armv7-a+fp.
However, armv7-a+fp isn't supported in old gcc version(e.g. gcc-5.4).
To accommodate that, use -mfpu=vfpv3 instead.

BUG=aomedia:3226,aomedia:3223

Change-Id: I13afa284d881c4a5d83675e1b9fb1af8c45480af
(cherry picked from commit 1908972c)

https://chromium.googlesource.com/webm/libwebm/+log/206d268d4d8066e5a37c49025325b80c95c771dd..ee0bab576c338c9807249b99588e352b7268cb62

only one commit affects this snapshot:
ee0bab5 Revert "mkvmuxer,Cluster::Size: make uint64 conversion explicit"

Bug: aomedia:3223
Change-Id: Ib1f21fc5589098af346d110ff88c94bb1ba0a027
(cherry picked from commit dc7efaa5)

Document that the ctrl_id argument of aom_codec_control() must be
nonzero. This is enforced at the beginning of aom_codec_control():

  // Control ID must be non-zero.
  if (!ctrl_id) {
    ctx->err = AOM_CODEC_INVALID_PARAM;
    return AOM_CODEC_INVALID_PARAM;
  }

Bug: aomedia:3223
Change-Id: Iec9b0fd19c02fe2013682be1f0e5834e6be488d2
(cherry picked from commit f03b45ec)

Bug: aomedia:3223
Change-Id: I7c69be0cc2b7058cb50649d6bce0a37d8e0f4f33
(cherry picked from commit cc7d10c8)

Bug: aomedia:3223
Change-Id: Iae307273d57930b42f1cc0d4a1422c84c8e2c5a6
(cherry picked from commit d4dfe02d)

this avoids compilation failures when projects enable ASM language
support in project(). this file was never included in libaom_srcs.*

aom & aom_static were missed in the removal from aom_ports in:
676f00c61 aom_ports.cmake: don't add x86_abi_support.asm to source list

Bug: aomedia:3201
Bug: aomedia:3223
Change-Id: Id26e7a02955eab43c83ca26bd414fe34d4b34e50
(cherry picked from commit 8cba6495)

this avoids compilation failures when projects enable ASM language
support in project(). this file was never included in libaom_srcs.*

Bug: aomedia:3201
Bug: aomedia:3223
Change-Id: Ifc11ea9d1ef568c8695c0b3e305030d686b8ab9f
(cherry picked from commit 07deb29e)

for aomedia.googlesource.com refs; main is the preferred branch now and
pushing to master will be blocked from submitting

Bug: b/204460717
Bug: aomedia:3223
Change-Id: I5b306bc2b4bf15c30689ee8e8e0cb5c42b7b76ef
(cherry picked from commit 77165f5d)

Bug: aomedia:3223
Change-Id: Idb6c28b3651c01bfad379f6793abaee6d8d8bada
(cherry picked from commit 56a666e4)

the calculation of the aligned allocation size could rollover if the
requested size is near SIZE_MAX

+ add a missing overflow check for aom_calloc()

Bug: aomedia:3195
Bug: aomedia:3223
Change-Id: I385de3344dadf9af56f0673b1e1a69b254e73c92
(cherry picked from commit 0929aa0a)

Disabled FPMT when unsupported configurations are set through control
calls. Recomputed maximum number of frame contexts at the beginning of
each sub-gop to account for updates to supported configurations.
Cleanup in is_fpmt_config function.

Change-Id: I1c343110e0b8cebc1584895cd43ff1ba1779768e

This allows to speed up intra frame encoding to 35-40% with overall
quality impact of <0.5% (rtc and rtc_derf sets speed 9 and 10)

Change-Id: Ic29e2287103c13b79fa241405abfcd00b69685f2

This flag is added for rate control library for av1, to disable content
dependency so the test can run the encoder without content related
adjustment to qp.

BD rate loss:
           aq 0      aq 3
rtc        1%        2.2%
rtc_derf   1.7%      3.2%

Change-Id: I3c88f000b44e43ec537096232647776c44eda6e3

The sf top_intra_model_count_allowed indicates the number of top
model rds to be stored during intra mode decision and these stored
model rds are used to prune the luma intra modes. This CL
introduces aggressive pruning of modes by reducing the number of
candidates used to store the model rd and by adapting the index of
the candidate for pruning based on neighbor blocks and quantizer
information. This change is applicable for speed >= 6.

For AVIF still-image encode,

          Instruction Count   BD-Rate Loss(%)
cpu-used     Reduction(%)     psnr       ssim
    6           5.481         0.0222    -0.0281
    7           4.170         0.2591     0.2361
    8           2.437         0.1108     0.1157
    9           0.000         0.0000     0.0000

STATS_CHANGED

Change-Id: Ib1bf8cae2f78c640c387050e6a2d4dfb988e742e

avoids -Wunused-result warnings

Change-Id: I7fe83b67b12e0c4c6f0d22bc0dc68d374f448035

This will improve key frame encoding on speed 7 in 1.7 times with
neglidgible effect on overall video quality.

Change-Id: Iee9321f1ed787ace01dfc7de66a4111058ea8cb4

Add code to estimate the noise level in a still image and choose
a noise synthesis level.

The auto selection of the noise synthesis level will only occur where
--allintra is set on the command line and --denoise-noise-level > 0.

Also on the command line:

if --enable-dnl-denoising=1 then the source image will also be filtered.
If --enable-dnl-denoising=0  then the source image will not be filtered.

Filtering of the source image is much more likely to cause some loss
fine detail if a strong filter level is chosen and is not recommended.

Adjustment from previous patch increases filtering on some images
where subjective evaluation said it was too low and reduced it on
some high complexity images where it was too high (by lowering the
edge threshold).

Change-Id: I90c99c3c12ddb8f5155f0ad7d87c35767ffccb38

Before this CL, apply_filtering will be triggered with overlay
frame. This will make
cpi->source = av1_realloc_and_scale_if_required() happen in the
end of denoise_and_encode().

BUG=aomedia:3194

Change-Id: I38a7dd63fd0dc4708931bc01f5707744e914ad16

enable_keyframe_filtering > 1 will introduce an extra overlay frame at
key frame location. However when
baseline_gf_interval == MAX_STATIC_GF_GROUP_LENGTH, we can't
afford to have an extra overlay frame.
Otherwise, the gf_group->size will exceed MAX_STATIC_GF_GROUP_LENGTH,
which leads to memory error.

A cheap solution is to turn off kf_decomp when
baseline_gf_interval == MAX_STATIC_GF_GROUP_LENGTH.

BUG=aomedia:3163

Change-Id: I05c604065496d698611f58ceb5c6ac7c7bbf5ebe

we only go into this function if short_circuit_low_temp_var set to 1.
There is no need to check MVs as left part of || operator is always true

Change-Id: I98e325f3ffb81981fdeef35df735f00778ba0892

Prior to this commit, the condition was added to prevent a regression.
The regression was due to an incorrect segment id, which led to the
wrong q index being used.

This commit fixes the issue by replacing
av1_set_offsets_without_segment_id with av1_set_offsets.

Performance:
 SPD_SET | TESTSET | AVG_PSNR | OVR_PSNR |  SSIM   |  SPD
---------+---------+----------+----------+---------+-------
    5    |   rtc   | -0.617%  | -0.487%  | -0.923% | +0.9%
    5    |rtc_derf | -0.256%  | -0.243%  | -0.178% | -2.0%
---------+---------+----------+----------+---------+-------
    6    |   rtc   | -1.283%  | -1.147%  | -1.593% | -0.3%
    6    |rtc_derf | -0.127%  | -0.199%  | -0.120% | -1.6%
---------+---------+----------+----------+---------+-------

STATS_CHANGED

Change-Id: I643cde296344bfa68e0f9c655bfed0d85cd7e404

The mbmi buffers are used to store mode information at block level.
In allintra encode though minimum partition sf setting can be 8x8,
these buffers were allocated at init time by assuming minimum
partition size as 4x4.

As the minimum partition size is known in the encode stage, this CL
moves the allocation of mi buffers from init stage to encode stage
and the buffers are allocated based on minimum partition size.

For AVIF still-image encode with speed = 9,

             HEAP Memory reduction(%)
Resolution   threads=1    threads=4
640x360        19.96        15.78
768x512        19.14        15.54
832x480        23.59        20.20
1280x720       24.15        21.50

For threads=4, an average encode time reduction of ~13.75% is
observed for 360p-720p resolutions.

HEAP memory reduction was measured using the following command.
$valgrind --tool=massif ./avifenc ...

Change-Id: I2cfb2a594a6bc3a7f748910a05d88d2f7316f336

When row-mt is enabled, a thread waits for the encoding of
the top-right superblock to be complete before encoding the
current superblock. This wait time is reduced in allintra
mode, by delaying the sync wait until the data from top-right
superblock is needed. This change is applicable for
resolutions < 4k in speed >= 9 as cost update frequencies
are set to COST_UPD_OFF for these resolutions.

For AVIF still-image encode, an average encode time reduction
of ~6.12% is observed for 360p-720p resolutions with speed=9
and 4 threads.

Change-Id: I3cdeb7ad58df976788cc6bd058184dfc65b63593

Removed unused edge detection and related unit tests. So we
don't need to maintain and support it.

BUG=aomedia:3188

Change-Id: I92fcee5a34d0949b49ed98ee01176078b68a6418

and update them to the CMake equivalent

Bug: aomedia:2673
Change-Id: Ib1e950e4481552460314a567451999e1e2a89d11

On superframes whose base is key: fix the reference
predicton structure so that the SL layer (for SL > 0)
only does the inter-layer prediction from the SL-1 layer.
This requires fixes to the sample encoder and in the
nonrd_pickmode.

In the current code SL2 was either predicting only from
SL0 (for ksvc), or it was predicting from both SL0 & SL1
(in non-ksvc mode), which was not the intended behavior.
So this fix improves the key frame quality for ksvc and
reduces the encode_time for non-ksvc.

Change-Id: Ifde750c741b53b42e4f27d493945036fb7e0e334

This patch extends SF prune_zero_mv_with_sse to speed 3
with reduced aggressiveness.

cpu-used   Instruction Count       BD-Rate Loss(%)
              Reduction(%)    avg.psnr   ovr.psnr   ssim
    3           1.093         -0.0162    -0.0093   -0.0038

STATS_CHANGED

Change-Id: I95bfe546653722db870201578642e74fca9ac415