platform/upstream/libvpx.git
16 months agoOptimize vp9_block_error_fp_neon
Salome Thirot [Wed, 1 Mar 2023 10:06:01 +0000 (10:06 +0000)]
Optimize vp9_block_error_fp_neon

Currently vp9_block_error_fp_neon is only used when
CONFIG_VP9_HIGHBITDEPTH is set to false. This patch optimizes the
implementation and uses tran_low_t instead of int16_t so that the
function can also be used in builds where vp9_highbitdepth is enabled.

Change-Id: Ibab7ec5f74b7652fa2ae5edf328f9ec587088fd3

16 months agovpx_convolve_copy_neon: fix unaligned loads w/w==4
James Zern [Mon, 6 Mar 2023 21:56:17 +0000 (13:56 -0800)]
vpx_convolve_copy_neon: fix unaligned loads w/w==4

Fixes a -fsanitize=undefined warning:

vpx_dsp/arm/vpx_convolve_copy_neon.c:29:26: runtime error: load of
misaligned address 0xffffa8242bea for type 'const uint32_t' (aka 'const
unsigned int'), which requires 4 byte alignment
0xffffa8242bea: note: pointer points here
 88 81  7d 7d 7d 7d 7d 81 81 7d  81 80 87 97 a8 ab a0 91 ...
              ^
    #0 0xb0447c in vpx_convolve_copy_neon
       vpx_dsp/arm/vpx_convolve_copy_neon.c:29:26
    #1 0x12285c8 in inter_predictor vp9/common/vp9_reconinter.h:29:3
    #2 0x1228430 in dec_build_inter_predictors
       vp9/decoder/vp9_decodeframe.c
    ...

Change-Id: Iaec4ac2a400b6e6db72d12e5a7acb316262b12a7

16 months agodisable vp8_sixtap_predict16x16_neon
James Zern [Fri, 3 Mar 2023 23:33:16 +0000 (15:33 -0800)]
disable vp8_sixtap_predict16x16_neon

This causes various buffer overflows in the tests:

[ RUN      ] NEON/SixtapPredictTest.TestWithPresetData/0
=================================================================
==22346==ERROR: AddressSanitizer: global-buffer-overflow on address
0x0000012b4a5b at pc 0x000000df0f60 bp 0xffffcf6e64b0 sp 0xffffcf6e64a8
READ of size 8 at 0x0000012b4a5b thread T0
    #0 0xdf0f5c in vp8_sixtap_predict16x16_neon
       vp8/common/arm/neon/sixtappredict_neon.c:1507:13
    #1 0x8819e4 in (anonymous
        namespace)::SixtapPredictTest_TestWithPresetData_Test::TestBody()
       test/predict_test.cc:293:3
    ...

0x0000012b4a5b is located 2 bytes to the right of global variable
'kTestData' defined in '../test/predict_test.cc:237:24' (0x12b48a0) of
size 441

[ RUN      ] NEON/SixtapPredictTest.TestWithRandomData/0
=================================================================
==22338==ERROR: AddressSanitizer: heap-buffer-overflow on address
0xffff8b5321fb at pc 0x000000df0f60 bp 0xfffff7e0cf30 sp 0xfffff7e0cf28
READ of size 8 at 0xffff8b5321fb thread T0
    #0 0xdf0f5c in vp8_sixtap_predict16x16_neon
       vp8/common/arm/neon/sixtappredict_neon.c:1507:13
    #1 0x87d4c0 in (anonymous
       namespace)::PredictTestBase::TestWithRandomData(void (*)(unsigned
       char*, int, int, int, unsigned char*, int))
       test/predict_test.cc:170:9
    ...

0xffff8b5321fb is located 2 bytes to the right of 441-byte region
[0xffff8b532040,0xffff8b5321f9)
allocated by thread T0 here:
    #0 0x5fd4f0 in operator new[](unsigned long) (test_libvpx+0x5fd4f0)
    #1 0x87c2e0 in (anonymous namespace)::PredictTestBase::SetUp()
       test/predict_test.cc:47:12
    #2 0x87d074 in non-virtual thunk to (anonymous
       namespace)::PredictTestBase::SetUp() test/predict_test.cc
    ...

Bug: webm:1795
Change-Id: I32213a381eef91547d00f88acf90f1cf2ec2ea75

16 months agodisable vpx_get4x4sse_cs_neon
James Zern [Fri, 3 Mar 2023 20:56:29 +0000 (20:56 +0000)]
disable vpx_get4x4sse_cs_neon

This function causes a heap overflow in the tests:
[ RUN      ] NEON/VpxSseTest.RefSse/0
=================================================================
==876922==ERROR: AddressSanitizer: heap-buffer-overflow on address
0xffff8949d903 at pc 0x000000dd95d4 bp 0xfffffdd7f260 sp 0xfffffdd7f258
READ of size 8 at 0xffff8949d903 thread T0
    #0 0xdd95d0 in vpx_get4x4sse_cs_neon
       vpx_dsp/arm/variance_neon.c:556:10
    #1 0x9d4894 in (anonymous namespace)::MainTestClass<unsigned int
       (*)(unsigned char const*, int, unsigned char const*,
           int)>::RefTestSse() test/variance_test.cc:531:5
    #2 0x9d4894 in (anonymous
       namespace)::VpxSseTest_RefSse_Test::TestBody()
           test/variance_test.cc:772:30
    ...

0xffff8949d903 is located 3 bytes to the right of 16-byte region
[0xffff8949d8f0,0xffff8949d900)
allocated by thread T0 here:
    #0 0x5fd050 in operator new[](unsigned long) (test_libvpx+0x5fd050)
    #1 0x9d3e04 in (anonymous namespace)::MainTestClass<unsigned int
       (*)(unsigned char const*, int, unsigned char const*,
           int)>::SetUp() test/variance_test.cc:299:12

Bug: webm:1794
Change-Id: I4bc681eb9a436743ef8bfe2a2abae59ce754309c

16 months agoRevert "Implement d117_predictor using Neon"
James Zern [Fri, 3 Mar 2023 20:34:36 +0000 (12:34 -0800)]
Revert "Implement d117_predictor using Neon"

This reverts commit 360e9069b6cc1dd3a004728b876fb923413f4b11.

This causes ASan errors:
[ RUN      ] VP9/TestVectorTest.MD5Match/1
=================================================================
==837858==ERROR: AddressSanitizer: stack-buffer-overflow on address
0xffff82ecad40 at pc 0x000000c494d4 bp 0xffffe1695800 sp 0xffffe16957f8
READ of size 16 at 0xffff82ecad40 thread T0
    #0 0xc494d0 in vpx_d117_predictor_32x32_neon (test_libvpx+0xc494d0)
    #1 0x1040b34 in vp9_predict_intra_block (test_libvpx+0x1040b34)
    #2 0xf8feec in decode_block (test_libvpx+0xf8feec)
    #3 0xf8f588 in decode_partition (test_libvpx+0xf8f588)
    #4 0xf7be5c in vp9_decode_frame (test_libvpx+0xf7be5c)
    ...
Address 0xffff82ecad40 is located in stack of thread T0 at offset 64 in
frame
    #0 0x103fd3c in vp9_predict_intra_block (test_libvpx+0x103fd3c)

  This frame has 2 object(s):
    [32, 64) 'left_col.i' <== Memory access at offset 64 overflows this
                              variable
    [96, 176) 'above_data.i'

Change-Id: I058213364617dfe1036126c33a3307f8288d9ae0

16 months agoRevert "Allow macroblock_plane to have its own rounding buffer"
Johann [Fri, 3 Mar 2023 03:46:01 +0000 (12:46 +0900)]
Revert "Allow macroblock_plane to have its own rounding buffer"

This reverts commit 5359ae810cdbb974060297ecf935183baf7b009b.

Reason for revert: Blocks quantize cleanups

Original change's description:
> Allow macroblock_plane to have its own rounding buffer
>
> Add 8 bytes buffer to macroblock_plane to support rounding factor.
>
> Change-Id: I3751689e4449c0caea28d3acf6cd17d7f39508ed

Change-Id: Ia2424d2114207370f0b45350313a5ff8521d25a8

16 months agoRevert "quantize: simplify 32x32_b args"
James Zern [Wed, 1 Mar 2023 23:53:18 +0000 (15:53 -0800)]
Revert "quantize: simplify 32x32_b args"

This reverts commit 848f6e733789c627b6606baf1c85e32be997e36f.

This has alignment issues, causing crashes in the tests:
SSSE3/VP9QuantizeTest.EOBCheck/*

Change-Id: Ic12014ab0a78ed3cde02d642509061552cdc8fc9

16 months agoRevert "quantize: simplifly highbd 32x32_b args"
James Zern [Wed, 1 Mar 2023 23:53:14 +0000 (15:53 -0800)]
Revert "quantize: simplifly highbd 32x32_b args"

This reverts commit 573f5e662b544dbc553d73fa2b61055c30dfe8cc.

This has alignment issues, causing crashes in the tests:
SSSE3/VP9QuantizeTest.EOBCheck/*

Change-Id: Ibf05e6b116c46f6e2c11187b3e3578bbd2d2c227

16 months agoRevert "quantize: use scan_order instead of passing scan/iscan"
James Zern [Wed, 1 Mar 2023 23:52:20 +0000 (15:52 -0800)]
Revert "quantize: use scan_order instead of passing scan/iscan"

This reverts commit 14fc40040ff30486c45111056db44ee18590a24a.

This has alignment issues, causing crashes in the tests:
SSSE3/VP9QuantizeTest.EOBCheck/*

Change-Id: I934f9a4c3ce3db33058a65180fa645c8649c3670

16 months agoMerge "Optimize Neon implementation of high bitdepth MSE functions" into main
James Zern [Wed, 1 Mar 2023 23:13:34 +0000 (23:13 +0000)]
Merge "Optimize Neon implementation of high bitdepth MSE functions" into main

16 months agoRevert "Implement highbd_d63_predictor using Neon"
James Zern [Wed, 1 Mar 2023 20:14:51 +0000 (12:14 -0800)]
Revert "Implement highbd_d63_predictor using Neon"

This reverts commit 7cdf139e3d6237386e0f93bdb0bdc1b459c663bf.

This causes failures in the VP9/ExternalFrameBufferMD5Test and
VP9/TestVectorTest.MD5Match tests in both armv7 and aarch64 builds.

Change-Id: I7ac4ba0ddc70e7e7860df9f962e6658defe1cdd5

16 months agoOptimize Neon implementation of high bitdepth MSE functions
Salome Thirot [Mon, 27 Feb 2023 17:58:18 +0000 (17:58 +0000)]
Optimize Neon implementation of high bitdepth MSE functions

Currently MSE functions just call the variance helpers but don't
actually use the computed sum. This patch adds dedicated helpers to
perform the computation of sse.

Add the corresponding tests as well.

Change-Id: I96a8590e3410e84d77f7187344688e02efe03902

16 months agoquantize: use scan_order instead of passing scan/iscan
Johann [Mon, 14 Nov 2022 07:47:33 +0000 (16:47 +0900)]
quantize: use scan_order instead of passing scan/iscan

further reduces the arguments for the 32x32. This will be applied to the base
version as well.

Change-Id: I25a162b5248b14af53d9e20c6a7fa2a77028a6d1

16 months agoquantize: simplifly highbd 32x32_b args
Johann [Fri, 11 Nov 2022 23:23:17 +0000 (08:23 +0900)]
quantize: simplifly highbd 32x32_b args

Change-Id: I431a41279c4c4193bc70cfe819da6ea7e1d2fba1

16 months agoMerge changes I892fbd2c,Ic59df16c,I7228327b,Ib4a1a2cb into main
James Zern [Tue, 28 Feb 2023 21:50:11 +0000 (21:50 +0000)]
Merge changes I892fbd2c,Ic59df16c,I7228327b,Ib4a1a2cb into main

* changes:
  Implement highbd_d117_predictor using Neon
  Implement highbd_d63_predictor using Neon
  Implement d117_predictor using Neon
  Implement d63_predictor using Neon

16 months agoMerge "quantize: simplify 32x32_b args" into main
James Zern [Tue, 28 Feb 2023 21:40:26 +0000 (21:40 +0000)]
Merge "quantize: simplify 32x32_b args" into main

16 months agoImplement highbd_d117_predictor using Neon
George Steed [Tue, 21 Feb 2023 11:17:10 +0000 (11:17 +0000)]
Implement highbd_d117_predictor using Neon

Add Neon implementations of the highbd d117 predictor for 4x4, 8x8,
16x16 and 32x32 block sizes. Also update tests to add new corresponding
cases.

An explanation of the general implementation strategy is given in the
8x8 implementation body, and is mostly identical to the non-highbd
version.

Speedups over the C code (higher is better):

Microarch.  | Compiler | Block | Speedup
Neoverse N1 |  LLVM 15 |   4x4 |    1.99
Neoverse N1 |  LLVM 15 |   8x8 |    4.37
Neoverse N1 |  LLVM 15 | 16x16 |    6.81
Neoverse N1 |  LLVM 15 | 32x32 |    6.49
Neoverse N1 |   GCC 12 |   4x4 |    2.49
Neoverse N1 |   GCC 12 |   8x8 |    4.10
Neoverse N1 |   GCC 12 | 16x16 |    5.58
Neoverse N1 |   GCC 12 | 32x32 |    2.16
Neoverse V1 |  LLVM 15 |   4x4 |    1.99
Neoverse V1 |  LLVM 15 |   8x8 |    5.03
Neoverse V1 |  LLVM 15 | 16x16 |    6.61
Neoverse V1 |  LLVM 15 | 32x32 |    6.01
Neoverse V1 |   GCC 12 |   4x4 |    2.09
Neoverse V1 |   GCC 12 |   8x8 |    4.52
Neoverse V1 |   GCC 12 | 16x16 |    4.23
Neoverse V1 |   GCC 12 | 32x32 |    2.70

Change-Id: I892fbd2c17ac527ddc22b91acca907ffc84c5cd2

16 months agoImplement highbd_d63_predictor using Neon
George Steed [Mon, 20 Feb 2023 11:41:40 +0000 (11:41 +0000)]
Implement highbd_d63_predictor using Neon

Add Neon implementations of the highbd d63 predictor for 4x4, 8x8, 16x16
and 32x32 block sizes. Also update tests to add new corresponding cases.

Speedups over the C code (higher is better):

Microarch.  | Compiler | Block | Speedup
Neoverse N1 |  LLVM 15 |   4x4 |    2.43
Neoverse N1 |  LLVM 15 |   8x8 |    4.03
Neoverse N1 |  LLVM 15 | 16x16 |    3.07
Neoverse N1 |  LLVM 15 | 32x32 |    4.11
Neoverse N1 |   GCC 12 |   4x4 |    2.92
Neoverse N1 |   GCC 12 |   8x8 |    7.20
Neoverse N1 |   GCC 12 | 16x16 |    4.43
Neoverse N1 |   GCC 12 | 32x32 |    3.18
Neoverse V1 |  LLVM 15 |   4x4 |    1.99
Neoverse V1 |  LLVM 15 |   8x8 |    3.66
Neoverse V1 |  LLVM 15 | 16x16 |    3.60
Neoverse V1 |  LLVM 15 | 32x32 |    3.29
Neoverse V1 |   GCC 12 |   4x4 |    2.39
Neoverse V1 |   GCC 12 |   8x8 |    4.76
Neoverse V1 |   GCC 12 | 16x16 |    3.29
Neoverse V1 |   GCC 12 | 32x32 |    2.43

Change-Id: Ic59df16ceeb468003754b4374be2f4d9af6589e4

16 months agoImplement d117_predictor using Neon
George Steed [Tue, 7 Feb 2023 12:16:00 +0000 (12:16 +0000)]
Implement d117_predictor using Neon

Add Neon implementations of the d117 predictor for 4x4, 8x8, 16x16 and
32x32 block sizes. Also update tests to add new corresponding cases.

An explanation of the general implementation strategy is given in the
8x8 implementation body.

Speedups over the C code (higher is better):

Microarch.  | Compiler | Block | Speedup
Neoverse N1 |  LLVM 15 |   4x4 |    1.73
Neoverse N1 |  LLVM 15 |   8x8 |    5.24
Neoverse N1 |  LLVM 15 | 16x16 |    9.77
Neoverse N1 |  LLVM 15 | 32x32 |   14.13
Neoverse N1 |   GCC 12 |   4x4 |    2.04
Neoverse N1 |   GCC 12 |   8x8 |    4.70
Neoverse N1 |   GCC 12 | 16x16 |    8.64
Neoverse N1 |   GCC 12 | 32x32 |    4.57
Neoverse V1 |  LLVM 15 |   4x4 |    1.75
Neoverse V1 |  LLVM 15 |   8x8 |    6.79
Neoverse V1 |  LLVM 15 | 16x16 |    9.16
Neoverse V1 |  LLVM 15 | 32x32 |   14.47
Neoverse V1 |   GCC 12 |   4x4 |    1.75
Neoverse V1 |   GCC 12 |   8x8 |    6.00
Neoverse V1 |   GCC 12 | 16x16 |    7.63
Neoverse V1 |   GCC 12 | 32x32 |    4.32

Change-Id: I7228327b5be27ee7a68deecafa05be0bd2a40ff4

16 months agoImplement d63_predictor using Neon
George Steed [Fri, 3 Feb 2023 17:12:46 +0000 (17:12 +0000)]
Implement d63_predictor using Neon

Add Neon implementations of the d63 predictor for 4x4, 8x8, 16x16 and
32x32 block sizes. Also update tests to add new corresponding cases.

Speedups over the C code (higher is better):

Microarch.  | Compiler | Block | Speedup
Neoverse N1 |  LLVM 15 |   4x4 |    2.10
Neoverse N1 |  LLVM 15 |   8x8 |    4.45
Neoverse N1 |  LLVM 15 | 16x16 |    4.74
Neoverse N1 |  LLVM 15 | 32x32 |    2.27
Neoverse N1 |   GCC 12 |   4x4 |    2.46
Neoverse N1 |   GCC 12 |   8x8 |   10.37
Neoverse N1 |   GCC 12 | 16x16 |   11.46
Neoverse N1 |   GCC 12 | 32x32 |    6.57
Neoverse V1 |  LLVM 15 |   4x4 |    2.24
Neoverse V1 |  LLVM 15 |   8x8 |    3.53
Neoverse V1 |  LLVM 15 | 16x16 |    4.44
Neoverse V1 |  LLVM 15 | 32x32 |    2.17
Neoverse V1 |   GCC 12 |   4x4 |    2.25
Neoverse V1 |   GCC 12 |   8x8 |    7.67
Neoverse V1 |   GCC 12 | 16x16 |    8.97
Neoverse V1 |   GCC 12 | 32x32 |    4.77

Change-Id: Ib4a1a2cb5a5c4495ae329529f8847664cbd0dfe0

16 months agoquantize: simplify 32x32_b args
Johann [Sat, 5 Nov 2022 00:53:07 +0000 (09:53 +0900)]
quantize: simplify 32x32_b args

Now that all the implementations of the 32x32 quantize are in
intrinsics we can reference struct members directly. Saves
pushing them to the stack.

n_coeffs is not used at all for this function.

Change-Id: I2104fea3fa20c455087e21b347d6abd7ea1f3e1e

16 months agoMerge "Add Neon implementations of standard bitdepth MSE functions" into main
James Zern [Tue, 28 Feb 2023 02:44:28 +0000 (02:44 +0000)]
Merge "Add Neon implementations of standard bitdepth MSE functions" into main

16 months agoMerge "Optimize transpose_neon.h helper functions" into main
James Zern [Tue, 28 Feb 2023 02:36:41 +0000 (02:36 +0000)]
Merge "Optimize transpose_neon.h helper functions" into main

16 months agotools_common,VpxInterface: remove unneeded const
James Zern [Mon, 27 Feb 2023 21:48:47 +0000 (13:48 -0800)]
tools_common,VpxInterface: remove unneeded const

Change-Id: Ic309aab2ff1750bdbcc36e8aafe05d52930ba694

16 months agoMerge "tools_common,VpxInterface: fix interface fn ptr proto" into main
James Zern [Mon, 27 Feb 2023 19:52:18 +0000 (19:52 +0000)]
Merge "tools_common,VpxInterface: fix interface fn ptr proto" into main

16 months agoAdd Neon implementations of standard bitdepth MSE functions
Salome Thirot [Fri, 24 Feb 2023 18:05:43 +0000 (18:05 +0000)]
Add Neon implementations of standard bitdepth MSE functions

Currently only vpx_mse16x16 has a Neon implementation. This patch adds
optimized Armv8.0 and Armv8.4 dot-product paths for all block sizes:
8x8, 8x16, 16x8 and 16x16.

Add the corresponding tests as well.

Change-Id: Ib0357fdcdeb05860385fec89633386e34395e260

16 months agoOptimize transpose_neon.h helper functions
Jonathan Wright [Sat, 25 Feb 2023 00:43:46 +0000 (00:43 +0000)]
Optimize transpose_neon.h helper functions

1) Use vtrn[12]q_[su]64 in vpx_vtrnq_[su]64* helpers on AArch64
   targets. This produces half as many TRN1/2 instructions compared to
   the number of MOVs that result from vcombine.

2) Use vpx_vtrnq_[su]64* helpers wherever applicable.

3) Refactor transpose_4x8_s16 to operate on 128-bit vectors.

Change-Id: I9a8b1c1fe2a98a429e0c5f39def5eb2f65759127

16 months agotools_common,VpxInterface: fix interface fn ptr proto
James Zern [Sat, 25 Feb 2023 03:25:39 +0000 (19:25 -0800)]
tools_common,VpxInterface: fix interface fn ptr proto

Use (void) to indicate an empty parameter list and match the declaration
of vpx_codec_vp[89]_[cd]x. This fixes a cfi sanitizer error.

Change-Id: I190f432eea4d1765afffd84c7458ec44d863f90c

16 months agoMerge changes I65d86038,If3299fe5,I3ef1ff19 into main
James Zern [Fri, 24 Feb 2023 17:58:15 +0000 (17:58 +0000)]
Merge changes I65d86038,If3299fe5,I3ef1ff19 into main

* changes:
  Add Neon implementation of high bitdepth 32x32 hadamard transform
  Add Neon implementation of high bitdepth 16x16 hadamard transform
  Add Neon implementation of high bitdepth 8x8 hadamard transform

16 months agoMerge changes Ia64d175a,Ie4ea8f0a into main
James Zern [Fri, 24 Feb 2023 17:49:25 +0000 (17:49 +0000)]
Merge changes Ia64d175a,Ie4ea8f0a into main

* changes:
  vp9_loop_filter_alloc: clear -Wshadow warnings
  vp9_adapt_mode_probs: clear -Wshadow warning

16 months agoAdd Neon implementation of high bitdepth 32x32 hadamard transform
Salome Thirot [Thu, 23 Feb 2023 12:05:30 +0000 (12:05 +0000)]
Add Neon implementation of high bitdepth 32x32 hadamard transform

Add Neon implementation of vpx_highbd_hadamard_32x32 as well as the
corresponding tests.

Change-Id: I65d8603896649de1996b353aa79eee54824b4708

16 months agoAdd Neon implementation of high bitdepth 16x16 hadamard transform
Salome Thirot [Wed, 22 Feb 2023 17:27:56 +0000 (17:27 +0000)]
Add Neon implementation of high bitdepth 16x16 hadamard transform

Add Neon implementation of vpx_highbd_hadamard_16x16 as well as the
corresponding tests.

Change-Id: If3299fe556351dfe3db994ac171d83a95ea1504b

16 months agoMerge "vp9 rc test: change param type to bool" into main
Jerome Jiang [Fri, 24 Feb 2023 01:45:54 +0000 (01:45 +0000)]
Merge "vp9 rc test: change param type to bool" into main

16 months agovp9 rc test: change param type to bool
Jerome Jiang [Thu, 23 Feb 2023 19:28:30 +0000 (14:28 -0500)]
vp9 rc test: change param type to bool

Change-Id: Ib45522e32d9137678da9062830044e9dd87537e5

16 months agoMerge "Disable some intra modes for TX_32X32" into main
Chi Yo Tsai [Thu, 23 Feb 2023 18:01:05 +0000 (18:01 +0000)]
Merge "Disable some intra modes for TX_32X32" into main

16 months agoAdd Neon implementation of high bitdepth 8x8 hadamard transform
Salome Thirot [Tue, 21 Feb 2023 17:40:20 +0000 (17:40 +0000)]
Add Neon implementation of high bitdepth 8x8 hadamard transform

Add Neon implementation of vpx_highbd_hadamard_8x8 as well as the
corresponding tests.

Change-Id: I3ef1ff199d76b6b010591ef15a81b0f36c9ded03

16 months agovp9_loop_filter_alloc: clear -Wshadow warnings
James Zern [Wed, 22 Feb 2023 21:25:29 +0000 (13:25 -0800)]
vp9_loop_filter_alloc: clear -Wshadow warnings

Bug: webm:1793
Change-Id: Ia64d175aa69dc2ecde2babf64bde04f02b32795b

16 months agovp9_adapt_mode_probs: clear -Wshadow warning
James Zern [Wed, 22 Feb 2023 21:21:27 +0000 (13:21 -0800)]
vp9_adapt_mode_probs: clear -Wshadow warning

Bug: webm:1793
Change-Id: Ie4ea8f0a3295e6f58dc6f7d5c61d46700c539d40

16 months agoMerge "vp9_block.h: rename diff struct to Diff" into main
James Zern [Thu, 23 Feb 2023 06:07:25 +0000 (06:07 +0000)]
Merge "vp9_block.h: rename diff struct to Diff" into main

16 months agoDisable some intra modes for TX_32X32
chiyotsai [Wed, 22 Feb 2023 20:44:47 +0000 (12:44 -0800)]
Disable some intra modes for TX_32X32

Performance:
| SPD_SET | TESTSET | AVG_PSNR | OVR_PSNR |  SSIM   | ENC_T |
|---------|---------|----------|----------|---------|-------|
|    0    | hdres2  | +0.036%  | +0.032%  | +0.014% | -3.9% |
|    0    | lowres2 | -0.002%  | -0.011%  | +0.020% | -3.6% |
|    0    | midres2 | +0.045%  | +0.025%  | -0.007% | -4.0% |

STATS_CHANGED

Change-Id: I75a927333d26f2a37f0dda57a641b455b845f5b9

16 months agovpx_subpixel_8t_intrin_avx2: clear -Wshadow warnings
James Zern [Wed, 22 Feb 2023 20:54:21 +0000 (12:54 -0800)]
vpx_subpixel_8t_intrin_avx2: clear -Wshadow warnings

no changes to assembly

Bug: webm:1793
Change-Id: I6a82290cafee7f4a7909d497ccfdefd5a78fb8ed

16 months agovp9_block.h: rename diff struct to Diff
James Zern [Wed, 22 Feb 2023 19:34:30 +0000 (11:34 -0800)]
vp9_block.h: rename diff struct to Diff

This matches the style guide and fixes some -Wshadow warnings related to
variables with the same name. Something similar was done in libaom in:
863b04994b Fix warnings reported by -Wshadow: Part2: av1 directory

Bug: webm:1793
Change-Id: I4df1bbc8d079a3174d75f0d35d54c200ffdbb677

16 months agoMerge "Skip redundant iterations in joint motion search " into main
Yunqing Wang [Wed, 22 Feb 2023 19:28:17 +0000 (19:28 +0000)]
Merge "Skip redundant iterations in joint motion search " into main

16 months agoMerge "vp9 rc: Make it work for SVC parallel encoding" into main
Jerome Jiang [Wed, 22 Feb 2023 14:59:49 +0000 (14:59 +0000)]
Merge "vp9 rc: Make it work for SVC parallel encoding" into main

17 months agoOptimize Neon implementation of high bitpdeth variance functions
Salome Thirot [Mon, 13 Feb 2023 16:11:31 +0000 (16:11 +0000)]
Optimize Neon implementation of high bitpdeth variance functions

Specialize implementation of high bitdepth variance functions such that
we only widen data processing element types when absolutely necessary.

Change-Id: If4cc3fea7b5ab0821e3129ebd79ff63706a512bf

17 months agoSkip redundant iterations in joint motion search 
Deepa K G [Thu, 16 Feb 2023 16:17:24 +0000 (21:47 +0530)]
Skip redundant iterations in joint motion search 

In joint_motion_search, there are four iterations.
Even iterations search in the first reference frame
and odd iterations search in the second. The last two
iterations use the search result of the first two
iterations as the start point. If the search result does
not change,last two iterations are not necessary and can
be skipped.

          Instruction Count
cpu-used   Reduction(%)
  0          1.411

Change-Id: Ie583c9f75dd0a22bbdfb432ccdd62eea6ec4fce8

17 months agovp9 rc: Make it work for SVC parallel encoding
Jerome Jiang [Fri, 11 Nov 2022 19:21:27 +0000 (14:21 -0500)]
vp9 rc: Make it work for SVC parallel encoding

Added unit test.

Keep track of spatial layer id and frame type in case where spatial
layers are encoded parallel by the hardware encoder.

ComputeQP() / PostEncodeUpdate() doesn't need to be called sequentially
when there is no inter layer prediction.

Bug: b/257368998
Change-Id: I50beaefcfc205d3f9a9d3dbe11fead5bfdc71489

17 months agoMerge "vp9 rc: Verify QP for all spatial layers" into main
Jerome Jiang [Fri, 17 Feb 2023 02:11:31 +0000 (02:11 +0000)]
Merge "vp9 rc: Verify QP for all spatial layers" into main

17 months agovp9 rc: Verify QP for all spatial layers
Jerome Jiang [Thu, 16 Feb 2023 22:48:49 +0000 (17:48 -0500)]
vp9 rc: Verify QP for all spatial layers

Change-Id: Ic669c96d25d7c039d370e9acd00dc45e09054552

17 months agoRelax frame recode tolerance on speed 0 to 1 above 480p
chiyotsai [Tue, 14 Feb 2023 22:29:29 +0000 (14:29 -0800)]
Relax frame recode tolerance on speed 0 to 1 above 480p

Performance:
| SPD_SET | TESTSET | AVG_PSNR | OVR_PSNR |  SSIM   | ENC_T |
|---------|---------|----------|----------|---------|-------|
|    0    | hdres2  | -0.028%  | +0.030%  | -0.408% | -2.0% |
|    0    | lowres2 | +0.000%  | +0.000%  | +0.000% | +0.0% |
|    0    | midres2 | -0.138%  | +0.042%  | -0.427% | -2.5% |
|---------|---------|----------|----------|---------|-------|
|    1    | hdres2  | -0.032%  | +0.018%  | -0.342% | -1.1% |
|    1    | lowres2 | +0.000%  | +0.000%  | +0.000% | +0.0% |
|    1    | midres2 | +0.050%  | +0.060%  | -0.257% | -1.6% |

Rate Error:
|         |         |     AVG_RC_ERROR    |     MAX_RC_ERROR    |
|         |         |---------------------|---------------------|
| SPD_SET | TESTSET |   BASE   |   TEST   |   BASE   |   TEST   |
|---------|---------|----------|----------|----------|----------|
|    0    | hdres2  |  33.044% |  33.065% | 149.903% | 149.903% |
|    0    | midres2 |  59.632% |  59.566% |  79.091% |  79.249% |
|---------|---------|----------|----------|----------|----------|
|    1    | hdres2  |  33.050% |  33.057% | 151.278% | 151.278% |
|    1    | midres2 |  59.640% |  59.614% |  78.707% |  78.842% |

STATS_CHANGED

Change-Id: I5d09601fede3912d5173717ce9dd070df3a97ec8

17 months agoEnable some more speed features on speed 0 to 2
chiyotsai [Tue, 14 Feb 2023 01:57:26 +0000 (17:57 -0800)]
Enable some more speed features on speed 0 to 2

Performance:
| SPD_SET | TESTSET | AVG_PSNR | OVR_PSNR |  SSIM   | ENC_T |
|---------|---------|----------|----------|---------|-------|
|    0    | hdres2  | +0.034%  | +0.030%  | +0.033% | -3.7% |
|    0    | lowres2 | +0.012%  | +0.017%  | +0.044% | -2.1% |
|    0    | midres2 | +0.030%  | +0.035%  | +0.060% | -1.9% |
|---------|---------|----------|----------|---------|-------|
|    1    | hdres2  | +0.027%  | +0.036%  | +0.030% | -2.7% |
|    1    | lowres2 | -0.006%  | -0.002%  | +0.006% | -1.0% |
|    1    | midres2 | -0.006%  | -0.012%  | -0.010% | -1.0% |
|---------|---------|----------|----------|---------|-------|
|    2    | hdres2  | -0.006%  | -0.001%  | -0.020% | -2.4% |
|    2    | lowres2 | -0.010%  | -0.015%  | -0.001% | -0.9% |
|    2    | midres2 | +0.006%  | -0.005%  | +0.009% | -1.0% |

STATS_CHANGED

Change-Id: I1431ac07215bb844739a410697387b9aead82792

17 months agoMerge changes Id74a6d9c,I5c31e0e9,Id5a2b2d9,I73182c97,I2f5916d5, ... into main
James Zern [Tue, 14 Feb 2023 02:46:51 +0000 (02:46 +0000)]
Merge changes Id74a6d9c,I5c31e0e9,Id5a2b2d9,I73182c97,I2f5916d5, ... into main

* changes:
  Optimize vpx_highbd_comp_avg_pred_neon
  Add Neon AvgPredTestHBD test suite
  Specialize Neon high bitdepth avg subpel variance by filter value
  Specialize Neon high bitdepth subpel variance by filter value
  Refactor Neon high bitdepth avg subpel variance functions
  Optimize Neon high bitdepth subpel variance functions

17 months agoOptimize vpx_highbd_comp_avg_pred_neon
Salome Thirot [Fri, 10 Feb 2023 10:50:47 +0000 (10:50 +0000)]
Optimize vpx_highbd_comp_avg_pred_neon

Optimize the implementation of vpx_highbd_comp_avg_pred_neon by making
use of the URHADD instruction to compute the average.

Change-Id: Id74a6d9c33e89bc548c3c7ecace59af69051b4a7

17 months agoAdd Neon AvgPredTestHBD test suite
Salome Thirot [Fri, 10 Feb 2023 10:29:24 +0000 (10:29 +0000)]
Add Neon AvgPredTestHBD test suite

Add test suite for vpx_highbd_comp_avg_pred_neon.

Change-Id: I5c31e0e990661ee3b8030bb517829c088fceae4d

17 months agoSpecialize Neon high bitdepth avg subpel variance by filter value
Salome Thirot [Thu, 9 Feb 2023 16:45:01 +0000 (16:45 +0000)]
Specialize Neon high bitdepth avg subpel variance by filter value

Use the same specialization as for standard bitdepth. The rationale for
the specialization is as follows:

The optimal implementation of the bilinear interpolation depends on the
filter values being used. For both horizontal and vertical interpolation
this can simplify to just taking the source values, or averaging the
source and reference values - which can be computed more easily than a
bilinear interpolation with arbitrary filter values.

This patch introduces tests to find the most optimal bilinear
interpolation implementation based on the filter values being used.
This new specialization is only used for larger block sizes.

Change-Id: Id5a2b2d9fac6f878795a6ed9de2bc27d9e62d661

17 months agoSpecialize Neon high bitdepth subpel variance by filter value
Salome Thirot [Thu, 9 Feb 2023 14:16:30 +0000 (14:16 +0000)]
Specialize Neon high bitdepth subpel variance by filter value

Use the same specialization as for standard bitdepth. The rationale for
the specialization is as follows:

The optimal implementation of the bilinear interpolation depends on the
filter values being used. For both horizontal and vertical interpolation
this can simplify to just taking the source values, or averaging the
source and reference values - which can be computed more easily than a
bilinear interpolation with arbitrary filter values.

This patch introduces tests to find the most optimal bilinear
interpolation implementation based on the filter values being used.
This new specialization is only used for larger block sizes.

Change-Id: I73182c979255f0332a274f2e5907df7f38c9eeb3

17 months agoRefactor Neon high bitdepth avg subpel variance functions
Salome Thirot [Wed, 8 Feb 2023 16:50:59 +0000 (16:50 +0000)]
Refactor Neon high bitdepth avg subpel variance functions

Use the same general code style as in the standard bitdepth Neon
implementation - merging the computation of vpx_highbd_comp_avg_pred
with the second pass of the bilinear filter to avoid storing and loading
the block again.

Also move vpx_highbd_comp_avg_pred_neon to its own file (like the
standard bitdepth implementation) since we're no longer using it for
averaging sub-pixel variance.

Change-Id: I2f5916d5b397db44b3247b478ef57046797dae6c

17 months agoOptimize Neon high bitdepth subpel variance functions
Salome Thirot [Tue, 7 Feb 2023 14:08:33 +0000 (14:08 +0000)]
Optimize Neon high bitdepth subpel variance functions

Use the same general code style as in the standard bitdepth Neon
implementation. Additionally, do not unnecessarily widen to 32-bit data
types when doing bilinear filtering - allowing us to process twice as
many elements per instruction.

Change-Id: I1e178991d2aa71f5f77a376e145d19257481e90f

17 months agoREADME: update release version to 1.13.0
James Zern [Sat, 11 Feb 2023 03:04:41 +0000 (19:04 -0800)]
README: update release version to 1.13.0

this was missed in the v1.13.0 tag

Bug: webm:1780
Change-Id: I3044534123bf67861174970e6241f6586055358e

17 months agoMerge "Remove CONFIG_CONSISTENT_RECODE flag" into main
Chi Yo Tsai [Fri, 10 Feb 2023 22:13:50 +0000 (22:13 +0000)]
Merge "Remove CONFIG_CONSISTENT_RECODE flag" into main

17 months agoRemove CONFIG_CONSISTENT_RECODE flag
chiyotsai [Wed, 8 Feb 2023 21:54:46 +0000 (13:54 -0800)]
Remove CONFIG_CONSISTENT_RECODE flag

Currently, libvpx does not properly clear and re-initialize the memories
when it re-encodes a frame. As a result, out-of-date values are used in
the encoding process, and re-encoding a frame with the same parameter
will give different outputs.

This commit enables the code under CONFIG_CONSISTENT_RECODE to correct
this behavior. This change has minor effect on the coding performance,
but it ensures valid values are used in the encoding process.

Furthermore, the flag is removed as it is now always turned on.

Performance:
| SPD_SET | TESTSET | AVG_PSNR | OVR_PSNR |  SSIM   | ENC_T |
|---------|---------|----------|----------|---------|-------|
|    0    | hdres2  | -0.012%  | -0.021%  | -0.030% | +0.1% |
|    0    | lowres2 | +0.029%  | +0.019%  | +0.047% | +0.1% |
|    0    | midres2 | -0.004%  | +0.009%  | +0.026% | +0.1% |
|---------|---------|----------|----------|---------|-------|
|    1    | hdres2  | +0.032%  | +0.032%  | -0.000% | -0.0% |
|    1    | lowres2 | -0.005%  | -0.011%  | -0.014% | +0.0% |
|    1    | midres2 | +0.004%  | +0.020%  | +0.027% | +0.2% |
|---------|---------|----------|----------|---------|-------|
|    2    | hdres2  | +0.048%  | +0.056%  | +0.057% | +0.1% |
|    2    | lowres2 | +0.007%  | +0.002%  | -0.016% | -0.0% |
|    2    | midres2 | -0.015%  | -0.008%  | -0.002% | +0.1% |
|---------|---------|----------|----------|---------|-------|
|    3    | hdres2  | +0.010%  | +0.014%  | +0.004% | -0.0% |
|    3    | lowres2 | +0.000%  | -0.021%  | -0.001% | +0.0% |
|    3    | midres2 | +0.007%  | -0.038%  | +0.012% | -0.2% |
|---------|---------|----------|----------|---------|-------|
|    4    | hdres2  | +0.107%  | +0.136%  | +0.124% | -0.0% |
|    4    | lowres2 | -0.012%  | -0.024%  | -0.020% | -0.0% |
|    4    | midres2 | +0.055%  | -0.004%  | +0.048% | -0.1% |
|---------|---------|----------|----------|---------|-------|
|    5    | hdres2  | +0.026%  | +0.027%  | +0.020% | -0.0% |
|    5    | lowres2 | +0.009%  | -0.008%  | +0.028% | +0.1% |
|    5    | midres2 | -0.025%  | +0.021%  | -0.020% | -0.1% |

STATS_CHANGED

Change-Id: I3967aee8c8e4d0608a492e07f99ab8de9744ba57

17 months agoMerge "Optimize Neon high bitdepth convolve copy" into main
James Zern [Fri, 10 Feb 2023 03:35:22 +0000 (03:35 +0000)]
Merge "Optimize Neon high bitdepth convolve copy" into main

17 months agoMerge "Merge tag 'v1.13.0'" into main
Jerome Jiang [Thu, 9 Feb 2023 22:07:28 +0000 (22:07 +0000)]
Merge "Merge tag 'v1.13.0'" into main

17 months agoMerge "Remove onyx_int.h from vp8 rc header" into main
Jerome Jiang [Thu, 9 Feb 2023 21:27:59 +0000 (21:27 +0000)]
Merge "Remove onyx_int.h from vp8 rc header" into main

17 months agoRemove onyx_int.h from vp8 rc header
Jerome Jiang [Tue, 7 Feb 2023 22:22:12 +0000 (17:22 -0500)]
Remove onyx_int.h from vp8 rc header

Also move the FRAME_TYPE declaration to common.h

Bug: webm:1766

Change-Id: Ic3016bd16548a5d2e0ae828a7fd7ad8adda8b8f6

17 months agoMerge tag 'v1.13.0'
Jerome Jiang [Thu, 9 Feb 2023 19:37:33 +0000 (14:37 -0500)]
Merge tag 'v1.13.0'

Release v1.13.0 Ugly Duckling

2023-01-31 v1.13.0 "Ugly Duckling"

  This release includes more Neon and AVX2 optimizations, adds a new codec
  control to set per frame QP, upgrades GoogleTest to v1.12.1, and includes
  numerous bug fixes.

- Upgrading:
    This release is ABI incompatible with the previous release.

    New codec control VP9E_SET_QUANTIZER_ONE_PASS to set per frame QP.

    GoogleTest is upgraded to v1.12.1.

    .clang-format is upgraded to clang-format-11.

    VPX_EXT_RATECTRL_ABI_VERSION was bumped due to incompatible changes to the
    feature of using external rate control models for vp9.

- Enhancement:
    Numerous improvements on Neon optimizations.
    Numerous improvements on AVX2 optimizations.
    Additional ARM targets added for Visual Studio.

- Bug fixes:
    Fix to calculating internal stats when frame dropped.
    Fix to segfault for external resize test in vp9.
    Fix to build system with replacing egrep with grep -E.
    Fix to a few bugs with external RTC rate control library.
    Fix to make SVC work with VBR.
    Fix to key frame setting in VP9 external RC.
    Fix to -Wimplicit-int (Clang 16).
    Fix to VP8 external RC for buffer levels.
    Fix to VP8 external RC for dynamic update of layers.
    Fix to VP9 auto level.
    Fix to off-by-one error of max w/h in validate_config.
    Fix to make SVC work for Profile 1.

Bug: webm:1780

Change-Id: I371fc1444ead56f8d7fc510e05582b6415c3ddb1

17 months agoOptimize Neon high bitdepth convolve copy
Jonathan Wright [Thu, 9 Feb 2023 11:57:10 +0000 (11:57 +0000)]
Optimize Neon high bitdepth convolve copy

Use standard loads and stores instead of the significantly slower
interleaving/de-interleaving variants. Also move all loads in loop
bodies above all stores as a mitigation against the compiler thinking
that the src and dst pointers alias (since we can't use restrict in
C89.)

Change-Id: Idd59dca51387f553f8db27144a2b8f2377c937d3

17 months agoMerge "Copy BLOCK_8X8's mi to PICK_MODE_CONTEXT::mic" into main
Chi Yo Tsai [Wed, 8 Feb 2023 23:16:48 +0000 (23:16 +0000)]
Merge "Copy BLOCK_8X8's mi to PICK_MODE_CONTEXT::mic" into main

17 months agoCopy BLOCK_8X8's mi to PICK_MODE_CONTEXT::mic
chiyotsai [Wed, 8 Feb 2023 22:01:19 +0000 (14:01 -0800)]
Copy BLOCK_8X8's mi to PICK_MODE_CONTEXT::mic

STATS_CHANGED

BUG=webm:1789

Change-Id: I74efe28bdf90a179c59fe3d1f5a15d497f57080d

17 months agoAdd missing high bitdepth Neon subpel variance tests
Salome Thirot [Wed, 8 Feb 2023 17:05:25 +0000 (17:05 +0000)]
Add missing high bitdepth Neon subpel variance tests

Add missing 4x4 and 4x8 tests for both high bitdepth sub-pixel variance
and high bitdepth averaging sub-pixel variance.

Change-Id: I042752c5b7ccc14f58075694d0bb1d36f144ad06

17 months agoFix unsigned integer overflow in sse computation
Cheng Chen [Mon, 30 Jan 2023 19:51:58 +0000 (11:51 -0800)]
Fix unsigned integer overflow in sse computation

Basically port the fix from libaom:
https://aomedia-review.googlesource.com/c/aom/+/169361

Change-Id: Id06a5db91372037832399200ded75d514e096726
(cherry picked from commit a94cdd57ffd95ee7beb48d2794dae538f25da46c)

17 months agoMerge "Enable some speed features on speed 0" into main
Chi Yo Tsai [Wed, 8 Feb 2023 00:44:46 +0000 (00:44 +0000)]
Merge "Enable some speed features on speed 0" into main

17 months agoEnable some speed features on speed 0
chiyotsai [Tue, 7 Feb 2023 19:11:35 +0000 (11:11 -0800)]
Enable some speed features on speed 0

Performance:
| SPD_SET | TESTSET | AVG_PSNR | OVR_PSNR |  SSIM   | ENC_T |
|---------|---------|----------|----------|---------|-------|
|    0    | hdres2  | +0.069%  | +0.067%  | +0.100% | -8.6% |
|    0    | midres2 | +0.116%  | +0.103%  | +0.062% | -9.6% |
|    0    | lowres2 | +0.276%  | +0.283%  | +0.214% |-11.9% |

STATS_CHANGED

Change-Id: I8b26c0be2312fcd0f8c9e889367682e80ea8de4b

17 months agoUse 4D reduction Neon helper for standard bitdepth SAD4D
Salome Thirot [Tue, 7 Feb 2023 11:28:15 +0000 (11:28 +0000)]
Use 4D reduction Neon helper for standard bitdepth SAD4D

Move the 4D reduction helper function to sum_neon.h and use this for
both standard and high bitdepth SAD4D paths. This also removes the
AArch64 requirement for using the UDOT Neon SAD4D paths.

Change-Id: I207f76b3d42aa541809b0672c3b3d86e54d133ff

17 months agoMerge "Move TPL to a new file" into main
Yunqing Wang [Tue, 7 Feb 2023 04:22:40 +0000 (04:22 +0000)]
Merge "Move TPL to a new file" into main

17 months agoMerge changes Ica45c44f,I75c5f099,I9e626d7f into main
James Zern [Tue, 7 Feb 2023 01:32:03 +0000 (01:32 +0000)]
Merge changes Ica45c44f,I75c5f099,I9e626d7f into main

* changes:
  Optimize Neon implementation of high bitdepth SAD4D functions
  Optimize Neon implementation of high bitdepth avg SAD functions
  Optimize Neon implementation of high bitdepth SAD functions

17 months agoMove TPL to a new file
Yunqing Wang [Mon, 6 Feb 2023 22:48:34 +0000 (14:48 -0800)]
Move TPL to a new file

This is a refactoring CL.

Change-Id: Ic8c1575601d27f14ecd1b1bf0a038e447eaae458

17 months agoMerge "Remove duplicated VPX_SCALING declaration" into main
Jerome Jiang [Mon, 6 Feb 2023 22:16:41 +0000 (22:16 +0000)]
Merge "Remove duplicated VPX_SCALING declaration" into main

17 months agoOptimize Neon implementation of high bitdepth SAD4D functions
Salome Thirot [Thu, 2 Feb 2023 16:06:38 +0000 (16:06 +0000)]
Optimize Neon implementation of high bitdepth SAD4D functions

Optimizations take a similar form to those implemented for Armv8.0
standard bitdepth SAD4D:

- Use ABD, UADALP instead of ABAL, ABAL2 (double the throughput on
  modern out-of-order Arm-designed cores.)
- Use more accumulator registers to make better use of Neon pipeline
  resources on Arm CPUs that have four Neon pipes.
- Compute the four SAD sums in parallel so that we only load the source
  block once - instead of four times.

Change-Id: Ica45c44fd167e5fcc83871d8c138fc72ed3a9723

17 months agoRemove duplicated VPX_SCALING declaration
Jerome Jiang [Mon, 6 Feb 2023 18:29:58 +0000 (13:29 -0500)]
Remove duplicated VPX_SCALING declaration

Use VPX_SCALING_MODE instead

Change-Id: Iab9d29f20838703e00bd9f7641035d8ebd69af53

17 months agoOptimize Neon implementation of high bitdepth avg SAD functions
Salome Thirot [Fri, 3 Feb 2023 11:00:19 +0000 (11:00 +0000)]
Optimize Neon implementation of high bitdepth avg SAD functions

Optimizations take a similar form to those implemented for standard
bitdepth averaging SAD:

- Use ABD, UADALP instead of ABAL, ABAL2 (double the throughput on
  modern out-of-order Arm-designed cores.)
- Use more accumulator registers to make better use of Neon pipeline
  resources on Arm CPUs that have four Neon pipes.

Change-Id: I75c5f09948f6bf17200f82e00e7a827a80451108

17 months agoOptimize Neon implementation of high bitdepth SAD functions
Salome Thirot [Wed, 1 Feb 2023 16:37:24 +0000 (16:37 +0000)]
Optimize Neon implementation of high bitdepth SAD functions

Optimizations take a similar form to those implemented for standard
bitdepth SAD:

- Use ABD, UADALP instead of ABAL, ABAL2 (double the throughput on
  modern out-of-order Arm-designed cores.)
- Use more accumulator registers to make better use of Neon pipeline
  resources on Arm CPUs that have four Neon pipes.

Change-Id: I9e626d7fa0e271908dc43448405a7985b80e6230

17 months agoMerge "Fix uninitialized mesh feature for BEST mode" into main
Yunqing Wang [Fri, 3 Feb 2023 23:22:58 +0000 (23:22 +0000)]
Merge "Fix uninitialized mesh feature for BEST mode" into main

17 months agoSet _img->bit_depth in y4m_input_fetch_frame()
Wan-Teh Chang [Fri, 3 Feb 2023 22:07:09 +0000 (14:07 -0800)]
Set _img->bit_depth in y4m_input_fetch_frame()

This is a port of
https://aomedia-review.googlesource.com/c/aom/+/169961.

Change-Id: I2aa0d12cafde0c73448bf8c57eab0cd92e846468

17 months agoFix uninitialized mesh feature for BEST mode
Yunqing Wang [Fri, 3 Feb 2023 00:30:09 +0000 (16:30 -0800)]
Fix uninitialized mesh feature for BEST mode

At BEST encoding mode, the mesh search range wasn't initialized for
non FC_GRAPHICS_ANIMATION content type, which actually/mistakenly
used speed 0's setting. Fixed it by adding the initialization.

There were 2 ways to fix this. Patchset 1 set to use speed 0's setting
for non FC_GRAPHICS_ANIMATION type. This didn't change BEST mode's
encoding results much, and only a couple of clips' results were changed.

Borg result for BEST mode:
         avg_psnr:  ovr_psnr:  ssim:  encoding_spdup:
lowres2:  -0.004     -0.003   -0.000    0.030
midres2:  -0.006     -0.009   -0.012    0.033
hdres2:    0.002      0.002    0.004    0.015

Patchset 2 set to use BEST's setting for non FC_GRAPHICS_ANIMATION type.
However, the majority of test clips' BDrate got changed up to
~0.5% (gain or loss), and overall it didn't give better performance
than patchset 1. So, we chose to use patchset 1.

Change-Id: Ibbf578dad04420e6ba22cb9a3ddec137a7e4deef

17 months agovp9_diamond_search_sad_neon: use DECLARE_ALIGNED
James Zern [Wed, 1 Feb 2023 21:27:06 +0000 (13:27 -0800)]
vp9_diamond_search_sad_neon: use DECLARE_ALIGNED

rather than the gcc specific __attribute__((aligned())); fixes build
targeting ARM64 windows.

Bug: webm:1788
Change-Id: I2210fc215f44d90c1ce9dee9b54888eb1b78c99e

17 months agoUpdate AUTHORS .mailmap and version
Jerome Jiang [Wed, 1 Feb 2023 16:38:42 +0000 (11:38 -0500)]
Update AUTHORS .mailmap and version

Bug: webm:1780
Change-Id: I75a24bdd076dc1746b23bababfaafccbce3b4214

17 months agoFix per frame qp for temporal layers
Jerome Jiang [Thu, 26 Jan 2023 00:25:12 +0000 (19:25 -0500)]
Fix per frame qp for temporal layers

Also add tests with fixed temporal layering mode.

Change-Id: If516fe94e3fb7f5a745821d1788bfe6cf90edaac
(cherry picked from commit db69ce6aea278bee88668fd9cc2af2e544516fdb)

17 months agoUpdate CHANGELOG
Jerome Jiang [Tue, 31 Jan 2023 17:16:38 +0000 (12:16 -0500)]
Update CHANGELOG

Bug: webm:1780
Change-Id: I3ab4729bff1d27ef7127ef26e780a469e9278c21

17 months agoMerge "Use load_unaligned mem_neon.h helpers in SAD and SAD4D" into main
James Zern [Tue, 31 Jan 2023 21:20:16 +0000 (21:20 +0000)]
Merge "Use load_unaligned mem_neon.h helpers in SAD and SAD4D" into main

17 months agoUse load_unaligned mem_neon.h helpers in SAD and SAD4D
Jonathan Wright [Tue, 31 Jan 2023 13:32:33 +0000 (13:32 +0000)]
Use load_unaligned mem_neon.h helpers in SAD and SAD4D

Use the load_unaligned helper functions in mem_neon.h to load strided
sequences of 4 bytes where alignment is not guaranteed in the Neon
SAD and SAD4D paths.

Change-Id: I941d226ef94fd7a633b09fc92165a00ba68a1501

17 months agoFix unsigned integer overflow in sse computation
Cheng Chen [Mon, 30 Jan 2023 19:51:58 +0000 (11:51 -0800)]
Fix unsigned integer overflow in sse computation

Basically port the fix from libaom:
https://aomedia-review.googlesource.com/c/aom/+/169361

Change-Id: Id06a5db91372037832399200ded75d514e096726

17 months agoMerge "Refactor 8x8 16-bit Neon transpose functions" into main
James Zern [Mon, 30 Jan 2023 19:30:45 +0000 (19:30 +0000)]
Merge "Refactor 8x8 16-bit Neon transpose functions" into main

17 months agoRefactor Neon implementation of SAD4D functions
Salome Thirot [Fri, 27 Jan 2023 16:16:16 +0000 (16:16 +0000)]
Refactor Neon implementation of SAD4D functions

Refactor and optimize the Neon implementation of SAD4D functions -
effectively backporting these libaom changes[1,2].

[1] https://aomedia-review.googlesource.com/c/aom/+/162181
[2] https://aomedia-review.googlesource.com/c/aom/+/162183

Change-Id: Icb04bd841d86f2d0e2596aa7ba86b74f8d2d360b

17 months agoMerge "Add encoder component timing information" into main
Yunqing Wang [Sat, 28 Jan 2023 00:27:57 +0000 (00:27 +0000)]
Merge "Add encoder component timing information" into main

17 months agoAdd encoder component timing information
Yunqing Wang [Fri, 27 Jan 2023 01:20:54 +0000 (17:20 -0800)]
Add encoder component timing information

Change-Id: Iaa5b73a9593ecfd74b6426ed47d2b529ec7ae2b5

17 months agoRefactor 8x8 16-bit Neon transpose functions
Gerda Zsejke More [Thu, 26 Jan 2023 15:12:55 +0000 (16:12 +0100)]
Refactor 8x8 16-bit Neon transpose functions

Refactor the Neon implementation of transpose_s16_8x8(q) and
transpose_u16_8x8 so that the final step compiles to 8 ZIP1/ZIP2
instructions as opposed to 8 EXT, MOV pairs. This change removes 8
instructions per call to transpose_s16_8x8(q), transpose_u16_8x8
where the result stays in registers for further processing - rather
than being stored to memory - like in vpx_hadamard_8x8_neon, for
example.

This is a backport of this libaom patch[1].
[1] https://aomedia-review.googlesource.com/c/aom/+/169426

Change-Id: Icef3e51d40efeca7008e1c4fc701bf39bd319c88

17 months agoMerge "Fix per frame qp for temporal layers" into main
Jerome Jiang [Thu, 26 Jan 2023 21:31:14 +0000 (21:31 +0000)]
Merge "Fix per frame qp for temporal layers" into main

17 months agoFix per frame qp for temporal layers
Jerome Jiang [Thu, 26 Jan 2023 00:25:12 +0000 (19:25 -0500)]
Fix per frame qp for temporal layers

Also add tests with fixed temporal layering mode.

Change-Id: If516fe94e3fb7f5a745821d1788bfe6cf90edaac

17 months agoMerge "Refactor Neon implementation of SAD functions" into main
James Zern [Thu, 26 Jan 2023 03:26:38 +0000 (03:26 +0000)]
Merge "Refactor Neon implementation of SAD functions" into main