Bug: 394596176
Change-Id: Ib4e56b433fadab1d4a8d5cb4c8fd5b870f53ac6a

Bug: b:394563983
Change-Id: Iaa7148962df5afcbfa554bad42026bae361ca150
(cherry picked from commit 96c8ce61)

These parameters are causing the tsan race condition
in the issue below.

Removing them for now. This may cause some slowdown in
some cases of webrtc usage, so we may add them back in
later.

Add screen 2TL test with 4 threads to catch the
issue below.

Bug: b:394563983
Change-Id: I9d1abc58ecc57921ddbbd79fb1da1dc114f974ac
(cherry picked from commit 2454213a)

Bug: 394596176
Change-Id: I8e3c4194090a0d2941ab936065441e298f1ea906

The comment in aom/aomcx.h was missed in
https://aomedia-review.googlesource.com/c/aom/+/194641.

Also change "on frame qindex" to "based on frame qindex".

Bug: aomedia:375221136
Change-Id: I110951c3f59661659ce32821620df9103cd7c7e0
(cherry picked from commit 39e039b2)

Bug: 394596176
Change-Id: Ifb8ab56f868e0a7123a4b9b4443f4ad8c8c17a4b

We want tune=ssimulacra2 to mean a mode that maximizes the SSIMULACRA 2
metric. The current tune=ssimulacra2 mode increases image quality by
using the SSIMULACRA 2 metric as a guidance and validating with
subjective visual quality checks. Rename it tune=iq, which stands for
image quality or intra quality.

Bug: b:390668380
Change-Id: I6f1943395ffc186e9cc773e96d0d69c80d857641

This was missed due to `/* clang-format off */`, which was meant for the
`av1_warped_filter[]` array entries.

Change-Id: I9abb43349ed98b119da06d2dc307575eafde47c9

Interleaving is not necessary and interleaving load/stores can be very
slow. Write new helper functions for old compiler versions that don't
support these _x2/_x4 variants.

Change-Id: I9d865acc68528e3ea192d56e2cceab5bad87d14b

Fixes a `-Wsign-compare` warning seen with clang 18.1.8 shipped with
Microsoft Visual Studio 2022 17.12.4.

aomenc.c:2321:76: warning: comparison of integers of different signs:
  'unsigned int' and 'aom_bit_depth_t' (aka 'enum aom_bit_depth')
  [-Wsign-compare]
| if ((global.show_psnr == 2) && (stream->config.cfg.g_input_bit_depth ==
|                                 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^
|                                 stream->config.cfg.g_bit_depth)) {
|                                 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Change-Id: Ifb30405efe6b7e4c9cba72922ef79b2e635c4084

Change-Id: Ic9184eb4a58256fbbe204835593c3d07bc8564fa

Port the optimizations and cleanups found while porting these functions
to SVT-AV1.

Change-Id: I894460ea2b1790d3c1312a0a38c7c5ab2c445aa6

Port the optimizations and cleanups found while porting these functions
to SVT-AV1.

Change-Id: I71037e2a4c6df5a291ceaac6ff4de7c354c78442

Port the optimizations and cleanups found while porting these functions
to SVT-AV1.

Change-Id: I5a07c67bbf393536a8d85b760d2ade19112400f1

Some minor optimizations and cleanup were found while porting these
functions to SVT-AV1, so backport them here.

Change-Id: Ib31b66de90438c6b3de918bfa61d4b69b1dc3a0f

Use the existing helper to compute filter4() and return if filtering
isn't needed.

This is a port from 487fa384eeb542e3cd885b161b212578c97dc908 in SVT-AV1.

Change-Id: I557ed6a9ecc367243858ab4ab71e781d8b4ae9e8

The Neon implementations of aom_lpf_vertical_14 and
aom_lpf_horizontal_14 compute every filter (filter4(), filter8() and
filter14()) before selecting for each element which filter is actually
needed. In practice, however, a lot of cases only need one of the
filters, so specialize for these scenarios, computing only the filters
that are needed and eliminate bitwise select.  This makes the case where
all filters are needed slightly slower, but as it is far from the most
common case this is ok.

Also move the actual filter computation to separate functions to avoid
code duplication.

This is a port from 4d6f560d1f4a35a47e5dfc9b7c2dc4f5483d49dc in SVT-AV1.

Change-Id: Id608edad6fee69fc263eb6ee9e0956ca6311d77f

The Neon implementations of aom_lpf_vertical_8 and aom_lpf_horizontal_8
compute both filter4() and filter8() before selecting for each element
which filter is actually needed. In practice, however, a lot of cases
only need one of the filters, so specialize for these scenarios,
computing only the filters that are needed and eliminate bitwise select.
This makes the case where all filters are needed slightly slower, but as
it is far from the most common case this is ok.

Also move the actual filter computation to separate functions to avoid
code duplication.

This is a port from d8ebdeec1edc1df12b2b19a6ffe477aea4bc158e in SVT-AV1.

Change-Id: I2e7157f26515f9ec51e82ad32434117adc4be928

The Neon implementations of aom_lpf_vertical_6 and aom_lpf_horizontal_6
compute both filter4() and filter6() before selecting for each element
which filter is actually needed. In practice, however, a lot of cases
only need one of the filters, so specialize for these scenarios,
computing only the filters that are needed and eliminate bitwise select.
This makes the case where all filters are needed slightly slower, but as
it is far from the most common case this is ok.

Also move the actual filter computation to separate functions to avoid
code duplication.

This is a port from d29673138e044fa7dec420f74e1451c53a9ea8ca in SVT-AV1.

Change-Id: Ibda2567c3781e51a9d43c59480284072115ceab6

When building with `BUILD_SHARED_LIBS=1` the aom_av1_rc library will
link against a static version of libaom as it requires some internal
functions. This causes the library to contain duplicate versions of some
of the symbols in libaom which causes link errors when targeting cygwin
or msys2/mingw-w64. `--allow-multiple-definition` is added to the linker
flags as a workaround. A better fix will be to limit the symbols
exported by the library to only the public interface, as is done with
libaom.

Fixes errors of the form:
.../x86_64-pc-cygwin/bin/ld:
  libaom.a(aom_codec.c.o):aom_codec.c:(.text+0x30): multiple definition
  of 'aom_codec_iface_name';
  libaom_av1_rc.dll.a(cygaom_av1_rc_dll_d000055.o):(.text+0x0): first
  defined here

Bug: aomedia:391715078
Change-Id: Ib6c7431de6b7d6f867371c9518ac4b5fdd80667d

fixes:
.../x86_64-pc-cygwin/bin/ld: unrecognized option '-z'
.../x86_64-pc-cygwin/bin/ld: use the --help option for usage information
collect2: error: ld returned 1 exit status

Bug: aomedia:391549038
Change-Id: I263fa211eb676029448a88ec1cdaedc44528ea32

Previously the values of rc_buf_sz, rc_buf_initial_sz and
rc_buf_optimal_sz would be ignored when rc_end_usage was set to AOM_VBR
and the targets set to quite large values.

Change-Id: I68b302d702e389f0341968bfcc0019fb40da7066

Add more conditions for increasing the horiz/col scroll
motion window, to reduce speed regression. Condition
it on the high motion frame-level flag and whether
the last TL0 had horiz scroll.

Leave the condition as is for now for the vert/row
scroll since its more common.

Bug: b:383306740
Change-Id: Idf54694ef53f024da2e9a7561f25db77f9348bf0

Set table data type from 16-bit to 8-bit when not using ARM/X86
intrinsics. Changing to narrower 8-bit type matches how data is
used naturally in the code; all values are 8-bit anyway. Narrower
width allows compiler to use consistent type, enabling better
vectorization and reducing memory bandwidth.

aarch64 and x86-64 intrinsics paths have similar optimizations.

Change-Id: I03fb0e016143b38a0b7197f8b21577b470578a61

Complete the work started in commit eac382e1.

Change-Id: I2bfd03fe18276d97e96c29569f3f2d0dbf77371d

The int16_t * casts were apparently added to fix the following compiler
warning:
  warning: incompatible pointer types assigning to 'const short *'
  from 'const int16_t (*)[8]' (aka 'const short (*)[8]')
  [-Wincompatible-pointer-types]

Using av1_warped_filter[...] (instead of av1_warped_filter + ...) has
the correct type. The av1_warped_filter[...] form is used in
av1/common/warped_motion.c.

Note: The int16_t * casts also cast away the const inadvertently because
av1_warped_filter is a const array.

Change-Id: Iabb564cd0eb1f62a75e8f0283d292e6c64a3088d

These binaries also work with Visual Studio and the yasm version may
work better with GitHub actions based builds.

Bug: webm:355182136
Change-Id: I4fa79618e8c7696ce09a1a6da84473f571b5c70c

Change-Id: Ibe20e6932632e8167948982d177760e28f8459ff

The functions are private, declared in aom_mem/aom_mem.h.

These functions are used by test_libaom only when test_libaom is linked
with the libaom static library.

Bug: 370599806
Change-Id: I9cf12429e1866adacd8ea1fa3f3316193950de00

Add screendata.1920_1080.y4m clip.

Change-Id: I4b8141286289bad366edfa32c2ce129a8e1ccc95

When testing sample superblocks off-center
within the frame, make sure the row/col position
is within max search window. This fixes the issue
below and brings back the gains from the revert:
https://aomedia-review.googlesource.com/c/aom/+/196401

Bug: b:387021199, b:383306740
Change-Id: Ie259a1cbe665f28f6b324f5222904acef6f527fd

This reverts commit f74eae54.

Reason for revert: Culprit CL for issue b:387021199, revert for now until issue is fixed.

Bug: b:387021199, b:383306740

Change-Id: Ia44eabe764a4430c168995c465322beb85f9ba30

For scene changes that occur on non_reference
frames (highest temporal layer) the encoder will do
a skip encoding. The logic to do this is in the
variance_partitioning, but this path may be bypassed
under high motion flag and sb source sad, which is not
correct. This can lead to some superblocks not being skip
encoded while others are.

This fixes the artifact observed in comment#18 of the
issue below.

Bug: b:383306740
Change-Id: I35ca017edf352efe4e66a5cfbd2879ab73d8a0f6

Bug: b:383744068
Change-Id: Ibb44d790f14ae990b5d041d06dda26edcd0cc9f6

Tested with gcc-12/13/14, clang-14/16.

Change-Id: Ifc841db9e3c78531d313aa4b346b982c7f932436

DECODE_WIDTH_LIMIT and DECODE_HEIGHT_LIMIT (used when CONFIG_SIZE_LIMIT
is defined as 1) are intended for the decoder only. Since
aom_realloc_frame_buffer() is shared by the decoder and the encoder, it
should not check DECODE_WIDTH_LIMIT and DECODE_HEIGHT_LIMIT. Move the
check to the aom_realloc_frame_buffer() call in the
read_uncompressed_header() function.

Bug: 385367183
Change-Id: Icdb3660e3631fefb690ef9f28a4d782e4c9a939d

Scroll detection occurs in two places:
1) at the frame level in estimate_scroll_motion(). This is only done under certain conditions and only for a few superblocks.
2) at the superblock level in the variance partitioning, this is done on every frame.

This change only increases the horiz/col search window (makes it equal to the vert/row motion for high frame_source_sad) in the estimate_scroll_motion(), to avoid speed slowdown for now.

Reduces artifacts in the issue below for horiz scrolling.

Stats changed for speed 11 rtc_screen:
negligible change in speed/IC.

Bug: b:383306740
Change-Id: I54f329cff424299252cf5bd46d1329eb68009f59

Change-Id: I7419d3ca97e4596f675e3375635a34b60f6e5508

Bug: b:383744068
Change-Id: Ide61a2f99075ea0b35fd2e8f9283fa0f57ca5ca5

Create the temp output files as text files in
multilayer_metadata_test.cc because we parse them as text files using
std::ifstream. This works around a bug in the MSVC Standard C++ Library
(https://github.com/microsoft/STL/issues/1784).

Add a `text_mode` boolean parameter to the libaom_test::TempOutFile
constructor and the libaom_test::GetTempOutFile() function, default to
true.

Bug: 383816673
Change-Id: Ic070105482c86efb13ad20451da7f3be93f113c5