BUG=chromium:339877165

Change-Id: I69dcc2cda098ec96a34e1e5f7ef557ee8caf5521
(cherry picked from commit 01467cdb)
(cherry picked from commit 3a15245e)

This ensures the encoder state/allocations stay in sync with scaling and
svc layer changes. In the SVC case, depending on the resolution,
differences in the chosen superblock size among layers may have caused a
crash. This was reproducible in WebRTC in screen content mode.

The fix is based on a change by Yuan Tong (tongyuan200097) [1]. It
refreshes the encoder config after AOME_SET_SCALEMODE,
AOME_SET_NUMBER_SPATIAL_LAYERS and AV1E_SET_SVC_PARAMS if no frames have
been encoded. AV1E_SET_SVC_PARAMS was missed in the original change.

[1]: https://aomedia-review.googlesource.com/c/aom/+/171941/2

Bug: chromium:339877165
Change-Id: Ib3d2a123b159898d7c7e19c81e89ff148920e1f1
(cherry picked from commit e42f4b19)
(cherry picked from commit 75971361)

Change-Id: I6a9dc3100277dafae9f982f97f108811aba60876

Superblock size is unrelated to DSP.

Change-Id: I74c5f3254db856c4944036152e4c3f9156421292

In this CL, 'error_info' is passed to 'av1_identify_regions()'
to report the error encountered during local buffer allocations.
Also, abstracted the code to free the local buffers into a
function 'free_firstpass_stats_buffers()'.

Bug: aomedia:3276
Change-Id: I3410f3089c9ce9fff5d8023e2b269ec4f7d009ea

This CL addresses allocation failures for buffers allocated in
'tf_alloc_and_reset_data()'. Also replaces 'malloc()' and 'free()'
with 'aom_malloc()' and 'aom_free()' for the allocation and
deallocation of 'tf_data->tmp_mbmi' buffer.

Bug: aomedia:3276
Change-Id: I24460c4e7daae15703ed4f8a7e31247d1cf16271

Add Neon implementation of av1_highbd_estimate_noise_from_single_plane
as well as the corresponding tests.

Change-Id: Iefa204b497f6195241326e56ba74e4cc96ad609c

Add a test suite for the high bitdepth version of
av1_estimate_noise_from_single_plane.

Change-Id: I3f94bd08071ee333e85083f9cb9dc8cbea4e2d94

Consume less stack memory by moving the 64 KiB local array this_dqcoeff
in search_tx_type() to the MACROBLOCK struct. Rename it dqcoeff_buf.

Bug: aomedia:2703
Change-Id: I520b759fb1ea23db356191e8e596c4e16fc204c9

Change-Id: I39843a8fc1046badaedaac2557ed93aad25c7d3d

In this CL, 'sb_stats_cache' and 'sb_fp_stats' (pointers to
'SB_FIRST_PASS_STATS') are added in MACROBLOCK structure to
facilitate its freeing during encoder_destroy() to avoid memory
leaks in case of an early termination.

Bug: aomedia:3276
Change-Id: Ibd5c5e3b4addd5bd96f6f60ee8ab10d52b060435

The helper functions are already inlined in the generated assembly
however the existing code uses a mixture of INLINE and AOM_FORCE_INLINE.
To be consistent we convert all helpers to be AOM_FORCE_INLINE.

Change-Id: I61b8f85ed716b551b5cff3c3f446749664ef6d4c

The second half of the butterfly is formed by subtraction, not addition.

Change-Id: Ia3e1400071748a40200302be081febb51c22a21f

Introduce new `butterfly_dct_{pre,post}` helpers similar to the ones
that already exist in the lowbd code and use them across the component
kernels to avoid duplication of the add/sub steps.

Also make more heavy use of the existing butterfly helpers rather than
having the sequence hard-coded in the component kernels themselves.

Benchmarking this change on a Neoverse N2 machine with Clang 16 and GCC
12, the highbd txfm2d speed tests show a geomean 1.1% reduction in times
reported.

Change-Id: I2221bbb714ffe0ffa499449a4b33da584b9278ef

The existing ADST kernels have a few `vnegq_s32` operations, however we
can instead propagate these negations down until we find a butterfly
multiply operation to merge the negation into. Since we already need the
negated versions of the cospi constants we can then simply adjust the
butterfly to match and remove the negation completely.

Benchmarking this change on a Neoverse N2 machine with Clang 16 and GCC
12, the highbd txfm2d speed tests show a geomean 0.4% reduction in times
reported.

Change-Id: I1c67ecfa4f36cf2050231af16e8fa889ea1b6d6c

The macro parameter in0/1 was unused, we just happened to always have
the right variable around anyway so it never made a difference.

Change-Id: Ie8c13e7f3ab61d05b905cb33a82b93e2a87431e6

Add a high bitpdeth version of the EncodeAPI.SetSFrameOnFirstFrame test
suite.

Change-Id: I2055c3454acef3980cd30a60737c9477eb05b58b

Add a high bitdepth version of all the tests from the
EncodeSmallWidthHeight test suite.

Change-Id: I5c7b833aa3276bfa45508744c12b76c953ec901a

The Neon implementation of av1_estimate_noise_from_single_plane was
causing a segfault due to some bad loop conditions. Fix the bad
conditions and re-enable the Neon path.

Bug: aomedia:3501
Change-Id: I2ad74e0b2315d8569760a3f2886618ccb6cb49bf

Speed and Extreme test suites were testing for only one set of
parameters, as a result only one helper function used by
av1_calc_proj_params out of the three was tested. Set suitable
parameters to test all three helper functions, and remove the setting of
r.s as it is unused.

Change-Id: I317d29e1ae0759286f07edb1fec25ff63a21deb3

Add Neon implementation of av1_calc_proj_params_high_bd as well as the
corresponding tests.

Change-Id: I8da658a9dbba152e27c07bc031c3cfc78334d16d

Currently in cases where the txfm2d component kernels are called via an
array lookup the compiler has no way of inlining these. For the
`load_buffer_*` helper function this is an issue since we load data from
memory and then must immediately store it back to the stack for passing
to the component kernel, and similarly for the `store_buffer_*` calls.

To get around this issue we can inline the `load_buffer_*` and
`store_buffer_*` helpers into the component kernels themselves, creating
specific "col" and "row" kernels for their respective transforms and
avoiding the unnecessary stack spills between calls. This also matches a
similar change previously done for the lowbd Neon code.

Specifically we introduce new `TRANSFORM_COL_{ONE,MANY}` and
`TRANSFORM_ROW{,_RECT}_{ONE,MANY}` macros to define the new wrapper
kernels. The "many" kernels also handle looping over the component
kernels and helpers where we were previously passing a "howmany"
parameter to the component kernel.

To allow using the `store_buffer_*` helpers in the new wrapper kernels,
we also need to adjust the helper to expose a stride parameter rather
than re-using the height from the macro.

Benchmarking this change on a Neoverse N2 machine with Clang 16 and GCC
12, the highbd txfm2d speed tests show a geomean 5.2% reduction in times
reported.

Change-Id: I8c17f9f1bb5f6931593e9f990ec45ce7ca225519

Originally reviewed in
https://aomedia-review.googlesource.com/c/aom/+/136501.

Change-Id: Iccecc10ea8bcbf926f53df35f66354c06eee8be4

Assert !ctx->img.self_allocd before calling aom_img_free(&ctx->img),
because aom_img_free() must not free ctx->img.

Related to the Coverity defect:
CID 323776: In decoder_destroy: Free of an address-of expression,
which can never be heap allocated

Change-Id: I59ef09c3c33da2dde429009ddea2b331baff59b6

The function doesn't exist.

Change-Id: Ic307b08e3b4699bf7f5bfb85ea011819ceda608b

Currently we only have a "_xn" kernel with a "howmany" parameter for the
16 kernel however this does not match any of the other kernels where the
primary building-block is the simpler "_x4" kernel.

This change is primarily intended as a precursor to introducing
TRANSFORM_{COL,ROW} macros matching the lowbd code in a future commit,
since the macros can then use the "_x4" kernels universally.

Benchmarking this change on a Neoverse N2 machine with Clang 16 and GCC
12, the highbd txfm2d speed tests show a geomean 0.2% reduction in times
reported.

Change-Id: I7f4d40e4ffd6efd2752d96f38ba87eaa512d24e1

Originally suggested here:
https://aomedia-review.googlesource.com/c/aom/+/182721/1/av1/encoder/arm/neon/highbd_fwd_txfm_neon.c#303

Change-Id: I67d35d1a6dfe901780a167b33eb201ab4898e94f

This seems to be beneficial for 4x4 and 8x8 txfm2d sizes but not for
16x16, so only apply it to the two smaller sizes.

Benchmarking this change on a Neoverse N2 machine with Clang 16 and GCC
12, the highbd txfm2d speed tests show a geomean 0.8% reduction in times
reported.

Change-Id: I6fc91f85946a37ac0f22f9766caca9a9263b4297

And adjust typedefs for kernel array types to match.

Change-Id: Ia2acd2714429c6e03b0ec22c81d9a68f40a0e9fd

This function is only ever called with one of two sets of parameters, so
duplicate the definition and remove the now-redundant parameters.

Benchmarking this change on a Neoverse N2 machine with Clang 16 and GCC
12, the highbd txfm2d speed tests show a geomean 1.4% reduction in times
reported.

Change-Id: I9182173630f0a4dc1daadd75c0ee8fafe7be6e57

Add a unit test with comment about setting max frame sizes
for random input.

BUG=aomedia:3349

Change-Id: Idf0a1caab9aaf302ce44c164ceba9e93df232e0b

In prepare_features_after_part_ab(), assert bsize is allowed for 4-way
partitions. av1_ml_prune_4_partition() already has such an assertion, so
just add a comment for that. However, these two assertions did not fix
the Coverity defects listed below.

Add a comment for the assertion in prune_4_way_partition_search() added
in commit c9f3d681

The variables ab_min_bsize_allowed and part4_min_bsize_allowed are
renamed ab_bsize_thresh and part4_bsize_thresh, respectively, because
these two variables are actually one less than the minimum bsize
allowed for ab and 4-way partitions.

Assert that horz_4_bs and vert_4_bs are not equal to BLOCK_INVALID in
av1_ml_prune_4_partition() and prepare_features_after_part_ab(). These
assertions actually fixed the following Coverity defects:
* CID 323648: In av1_ml_prune_4_partition: Out-of-bounds read from a
  buffer
* CID 323757: In av1_ml_prune_4_partition: Out-of-bounds read from a
  buffer (apparently a duplicate of the above)
* CID 323761: In prepare_features_after_part_ab: Out-of-bounds read from
  a buffer
* CID 323828: In prepare_features_after_part_ab: Out-of-bounds read from
  a buffer (apparently a duplicate of the above)

Bug: aomedia:3499
Change-Id: Idd4ff8cf3e7e6a3656885472af426b6f98489197

Add Neon implementation for av1_wedge_compute_delta_squares and the
corresponding test as well.

Change-Id: I33de356718a1e135b083ed1a13614f97b20391ab

Declare the speed feature ext_partition_eval_thresh as BLOCK_SIZE. Also
declare `bsize` (function parameters, local variables, or data members)
as BLOCK_SIZE.

The first change comes from
https://aomedia-review.googlesource.com/c/aom/+/136501.

Change-Id: I72b741c81950b9f36b9d22b794e89cce3f229662

Add Neon implementation of aom_highbd_blend_a64_d16_mask as well as
the corresponding test.

Change-Id: I33525b59a0459cc6f8bd8c7e573cec4948122bad

Bug: aomedia:3499
Change-Id: I0909319ba886dce11f1d88eded0020a90be3c4ec

Change-Id: I14b09dae59dd1743f4de5c37e6ae16d734648f36

This CL frees 'mv_costs' and 'dv_costs' in
free_thread_data() to prevent memory leaks in case of
early termination. In order to avoid double free error
in case of 'x->mv_costs' and 'x->dv_costs', the allocation
and freeing of the same is done using newly introduced
'mv_costs_alloc' and 'dv_costs_alloc' pointers in ThreadData
structure. 'x->mv_costs'and 'x->dv_costs' are initialized
appropriately.

Bug: aomedia:3276
Change-Id: I0954513c9c0bdb216cd4fbeaf75533552472132f

This currently crashes in the EncodeAPI.SetSFrameOnFirstFrame and
EncodeSmallWidthHeight.1x1 tests.

Bug: aomedia:3501
Change-Id: I28ce6470abefb9900cc9e833afe91f246386abed

BUG=aomedia:3349

Change-Id: I72d482f320774e05be8e88e1e9ff98d4f6904c5e