[FFmpeg-devel,1/4] avcodec/aom_film_grain: add AOM film grain synthesis

From: Niklas Haas <git@haasn.dev>

Implementation copied wholesale from dav1d, sans SIMD, under permissive
license. This implementation was extensively verified to be bit-exact,
so it serves as a much better starting point than trying to re-engineer
this from scratch for no reason. (I also authored the original
implementation in dav1d, so any "clean room" implementation would end up
looking much the same, anyway)

The notable changes I had to make while adapting this from the dav1d
code-base to the FFmpeg codebase include:

- replacing bitdepth parameters by preprocessor symbol (BITDEPTH)
- reordering variable declarations to avoid triggering warnings
- replacing several inline helpers by avutil equivalents
- changing code that accesses frame metadata
- replacing raw plane copying logic by av_image_copy_plane

Apart from this, the implementation is basically unmodified.
---
 libavcodec/aom_film_grain.c          | 288 +++++++++++++
 libavcodec/aom_film_grain.h          |  38 ++
 libavcodec/aom_film_grain_template.c | 599 +++++++++++++++++++++++++++
 3 files changed, 925 insertions(+)
 create mode 100644 libavcodec/aom_film_grain.c
 create mode 100644 libavcodec/aom_film_grain.h
 create mode 100644 libavcodec/aom_film_grain_template.c

Niklas Haas:
> From: Niklas Haas <git@haasn.dev>
> 
> Implementation copied wholesale from dav1d, sans SIMD, under permissive
> license. This implementation was extensively verified to be bit-exact,
> so it serves as a much better starting point than trying to re-engineer
> this from scratch for no reason. (I also authored the original
> implementation in dav1d, so any "clean room" implementation would end up
> looking much the same, anyway)
> 
> The notable changes I had to make while adapting this from the dav1d
> code-base to the FFmpeg codebase include:
> 
> - replacing bitdepth parameters by preprocessor symbol (BITDEPTH)
> - reordering variable declarations to avoid triggering warnings
> - replacing several inline helpers by avutil equivalents
> - changing code that accesses frame metadata
> - replacing raw plane copying logic by av_image_copy_plane
> 
> Apart from this, the implementation is basically unmodified.
> ---
>  libavcodec/aom_film_grain.c          | 288 +++++++++++++
>  libavcodec/aom_film_grain.h          |  38 ++
>  libavcodec/aom_film_grain_template.c | 599 +++++++++++++++++++++++++++
>  3 files changed, 925 insertions(+)
>  create mode 100644 libavcodec/aom_film_grain.c
>  create mode 100644 libavcodec/aom_film_grain.h
>  create mode 100644 libavcodec/aom_film_grain_template.c
> 
> diff --git a/libavcodec/aom_film_grain.c b/libavcodec/aom_film_grain.c
> new file mode 100644
> index 00000000000..1bdaafc39f7
> --- /dev/null
> +++ b/libavcodec/aom_film_grain.c
> @@ -0,0 +1,288 @@
> +/*
> + * AOM film grain synthesis
> + * Copyright (c) 2023 Niklas Haas <ffmpeg@haasn.xyz>
> + *
> + * This file is part of FFmpeg.
> + *
> + * FFmpeg is free software; you can redistribute it and/or
> + * modify it under the terms of the GNU Lesser General Public
> + * License as published by the Free Software Foundation; either
> + * version 2.1 of the License, or (at your option) any later version.
> + *
> + * FFmpeg is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> + * Lesser General Public License for more details.
> + *
> + * You should have received a copy of the GNU Lesser General Public
> + * License along with FFmpeg; if not, write to the Free Software
> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
> + */
> +
> +/**
> + * @file
> + * AOM film grain synthesis.
> + * @author Niklas Haas <ffmpeg@haasn.xyz>
> + */
> +
> +#include "libavutil/avassert.h"
> +#include "libavutil/imgutils.h"
> +
> +#include "aom_film_grain.h"
> +
> +// Common/shared helpers (not dependent on BITDEPTH)
> +static inline int get_random_number(const int bits, unsigned *const state) {
> +    const int r = *state;
> +    unsigned bit = ((r >> 0) ^ (r >> 1) ^ (r >> 3) ^ (r >> 12)) & 1;
> +    *state = (r >> 1) | (bit << 15);
> +
> +    return (*state >> (16 - bits)) & ((1 << bits) - 1);
> +}
> +
> +static inline int round2(const int x, const uint64_t shift) {
> +    return (x + ((1 << shift) >> 1)) >> shift;
> +}
> +
> +static const int16_t gaussian_sequence[2048];
> +
> +#define BITDEPTH 12
> +#include "aom_film_grain_template.c"
> +#undef BITDEPTH
> +
> +#define BITDEPTH 10
> +#include "aom_film_grain_template.c"
> +#undef BITDEPTH
> +
> +#define BITDEPTH 9
> +#include "aom_film_grain_template.c"
> +#undef BITDEPTH
> +
> +#define BITDEPTH 8
> +#include "aom_film_grain_template.c"
> +#undef BITDEPTH

I don't think this should be duplicated for every bitdepth.

- Andreas

On Wed, 28 Feb 2024 16:45:50 +0100 Andreas Rheinhardt <andreas.rheinhardt@outlook.com> wrote:
> I don't think this should be duplicated for every bitdepth.

To clarify, you think the bit depth should be just treated as a runtime
parameter instead of this being a template at all? How would you suggest
to handle the diference in pixel size between 8-bit and 10/12-bit?

Maybe a middle ground would be to compile this template only twice, once
for HBD and once for 8-bit, with the HBD code-path taking the bitdepth
as an extra parameter.

(That's actually what dav1d does internally)

Niklas Haas:
> On Wed, 28 Feb 2024 16:45:50 +0100 Andreas Rheinhardt <andreas.rheinhardt@outlook.com> wrote:
>> I don't think this should be duplicated for every bitdepth.
> 
> To clarify, you think the bit depth should be just treated as a runtime
> parameter instead of this being a template at all? How would you suggest
> to handle the diference in pixel size between 8-bit and 10/12-bit?
> 
> Maybe a middle ground would be to compile this template only twice, once
> for HBD and once for 8-bit, with the HBD code-path taking the bitdepth
> as an extra parameter.

That's my preferred way.

> 
> (That's actually what dav1d does internally)

On Wed, 28 Feb 2024 17:01:13 +0100 Andreas Rheinhardt <andreas.rheinhardt@outlook.com> wrote:
> Niklas Haas:
> > On Wed, 28 Feb 2024 16:45:50 +0100 Andreas Rheinhardt <andreas.rheinhardt@outlook.com> wrote:
> >> I don't think this should be duplicated for every bitdepth.
> > 
> > To clarify, you think the bit depth should be just treated as a runtime
> > parameter instead of this being a template at all? How would you suggest
> > to handle the diference in pixel size between 8-bit and 10/12-bit?
> > 
> > Maybe a middle ground would be to compile this template only twice, once
> > for HBD and once for 8-bit, with the HBD code-path taking the bitdepth
> > as an extra parameter.
> 
> That's my preferred way.
> 
> > 
> > (That's actually what dav1d does internally)

I gave it a try, do you think it's better?
From 81a32dd8df29f2e00b9be39e9cf79f7ff58e33d8 Mon Sep 17 00:00:00 2001
From: Niklas Haas <git@haasn.dev>
Date: Wed, 4 Oct 2023 13:43:51 +0200
Subject: [PATCH 1/4] avcodec/aom_film_grain: add AOM film grain synthesis

Implementation copied wholesale from dav1d, sans SIMD, under permissive
license. This implementation was extensively verified to be bit-exact,
so it serves as a much better starting point than trying to re-engineer
this from scratch for no reason. (I also authored the original
implementation in dav1d, so any "clean room" implementation would end up
looking much the same, anyway)

The notable changes I had to make while adapting this from the dav1d
code-base to the FFmpeg codebase include:

- reordering variable declarations to avoid triggering warnings
- replacing several inline helpers by avutil equivalents
- changing code that accesses frame metadata
- replacing raw plane copying logic by av_image_copy_plane

Apart from this, the implementation is basically unmodified.
---
 libavcodec/aom_film_grain.c          | 280 +++++++++++++
 libavcodec/aom_film_grain.h          |  38 ++
 libavcodec/aom_film_grain_template.c | 606 +++++++++++++++++++++++++++
 3 files changed, 924 insertions(+)
 create mode 100644 libavcodec/aom_film_grain.c
 create mode 100644 libavcodec/aom_film_grain.h
 create mode 100644 libavcodec/aom_film_grain_template.c

diff --git a/libavcodec/aom_film_grain.c b/libavcodec/aom_film_grain.c
new file mode 100644
index 00000000000..32cb8661f60
--- /dev/null
+++ b/libavcodec/aom_film_grain.c
@@ -0,0 +1,280 @@
+/*
+ * AOM film grain synthesis
+ * Copyright (c) 2023 Niklas Haas <ffmpeg@haasn.xyz>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file
+ * AOM film grain synthesis.
+ * @author Niklas Haas <ffmpeg@haasn.xyz>
+ */
+
+#include "libavutil/avassert.h"
+#include "libavutil/imgutils.h"
+
+#include "aom_film_grain.h"
+
+// Common/shared helpers (not dependent on BITDEPTH)
+static inline int get_random_number(const int bits, unsigned *const state) {
+    const int r = *state;
+    unsigned bit = ((r >> 0) ^ (r >> 1) ^ (r >> 3) ^ (r >> 12)) & 1;
+    *state = (r >> 1) | (bit << 15);
+
+    return (*state >> (16 - bits)) & ((1 << bits) - 1);
+}
+
+static inline int round2(const int x, const uint64_t shift) {
+    return (x + ((1 << shift) >> 1)) >> shift;
+}
+
+static const int16_t gaussian_sequence[2048];
+
+#define BIT_DEPTH 16
+#include "aom_film_grain_template.c"
+#undef BIT_DEPTH
+
+#define BIT_DEPTH 8
+#include "aom_film_grain_template.c"
+#undef BIT_DEPTH
+
+
+int ff_aom_apply_film_grain(AVFrame *out, const AVFrame *in,
+                            const AVFilmGrainParams *params)
+{
+    av_assert0(out->format == in->format);
+    switch (in->format) {
+    case AV_PIX_FMT_GRAY8:
+    case AV_PIX_FMT_YUV420P:
+    case AV_PIX_FMT_YUV422P:
+    case AV_PIX_FMT_YUV444P:
+    case AV_PIX_FMT_YUVJ420P:
+    case AV_PIX_FMT_YUVJ422P:
+    case AV_PIX_FMT_YUVJ444P:
+        return apply_film_grain_8(out, in, params);
+    case AV_PIX_FMT_GRAY9:
+    case AV_PIX_FMT_YUV420P9:
+    case AV_PIX_FMT_YUV422P9:
+    case AV_PIX_FMT_YUV444P9:
+        return apply_film_grain_16(out, in, params, 9);
+    case AV_PIX_FMT_GRAY10:
+    case AV_PIX_FMT_YUV420P10:
+    case AV_PIX_FMT_YUV422P10:
+    case AV_PIX_FMT_YUV444P10:
+        return apply_film_grain_16(out, in, params, 10);
+    case AV_PIX_FMT_GRAY12:
+    case AV_PIX_FMT_YUV420P12:
+    case AV_PIX_FMT_YUV422P12:
+    case AV_PIX_FMT_YUV444P12:
+        return apply_film_grain_16(out, in, params, 12);
+    }
+
+    /* The AV1 spec only defines film grain synthesis for these formats */
+    return AVERROR_INVALIDDATA;
+}
+
+// Taken from the AV1 spec. Range is [-2048, 2047], mean is 0 and stddev is 512
+static const int16_t gaussian_sequence[2048] = {
+    56,    568,   -180,  172,   124,   -84,   172,   -64,   -900,  24,   820,
+    224,   1248,  996,   272,   -8,    -916,  -388,  -732,  -104,  -188, 800,
+    112,   -652,  -320,  -376,  140,   -252,  492,   -168,  44,    -788, 588,
+    -584,  500,   -228,  12,    680,   272,   -476,  972,   -100,  652,  368,
+    432,   -196,  -720,  -192,  1000,  -332,  652,   -136,  -552,  -604, -4,
+    192,   -220,  -136,  1000,  -52,   372,   -96,   -624,  124,   -24,  396,
+    540,   -12,   -104,  640,   464,   244,   -208,  -84,   368,   -528, -740,
+    248,   -968,  -848,  608,   376,   -60,   -292,  -40,   -156,  252,  -292,
+    248,   224,   -280,  400,   -244,  244,   -60,   76,    -80,   212,  532,
+    340,   128,   -36,   824,   -352,  -60,   -264,  -96,   -612,  416,  -704,
+    220,   -204,  640,   -160,  1220,  -408,  900,   336,   20,    -336, -96,
+    -792,  304,   48,    -28,   -1232, -1172, -448,  104,   -292,  -520, 244,
+    60,    -948,  0,     -708,  268,   108,   356,   -548,  488,   -344, -136,
+    488,   -196,  -224,  656,   -236,  -1128, 60,    4,     140,   276,  -676,
+    -376,  168,   -108,  464,   8,     564,   64,    240,   308,   -300, -400,
+    -456,  -136,  56,    120,   -408,  -116,  436,   504,   -232,  328,  844,
+    -164,  -84,   784,   -168,  232,   -224,  348,   -376,  128,   568,  96,
+    -1244, -288,  276,   848,   832,   -360,  656,   464,   -384,  -332, -356,
+    728,   -388,  160,   -192,  468,   296,   224,   140,   -776,  -100, 280,
+    4,     196,   44,    -36,   -648,  932,   16,    1428,  28,    528,  808,
+    772,   20,    268,   88,    -332,  -284,  124,   -384,  -448,  208,  -228,
+    -1044, -328,  660,   380,   -148,  -300,  588,   240,   540,   28,   136,
+    -88,   -436,  256,   296,   -1000, 1400,  0,     -48,   1056,  -136, 264,
+    -528,  -1108, 632,   -484,  -592,  -344,  796,   124,   -668,  -768, 388,
+    1296,  -232,  -188,  -200,  -288,  -4,    308,   100,   -168,  256,  -500,
+    204,   -508,  648,   -136,  372,   -272,  -120,  -1004, -552,  -548, -384,
+    548,   -296,  428,   -108,  -8,    -912,  -324,  -224,  -88,   -112, -220,
+    -100,  996,   -796,  548,   360,   -216,  180,   428,   -200,  -212, 148,
+    96,    148,   284,   216,   -412,  -320,  120,   -300,  -384,  -604, -572,
+    -332,  -8,    -180,  -176,  696,   116,   -88,   628,   76,    44,   -516,
+    240,   -208,  -40,   100,   -592,  344,   -308,  -452,  -228,  20,   916,
+    -1752, -136,  -340,  -804,  140,   40,    512,   340,   248,   184,  -492,
+    896,   -156,  932,   -628,  328,   -688,  -448,  -616,  -752,  -100, 560,
+    -1020, 180,   -800,  -64,   76,    576,   1068,  396,   660,   552,  -108,
+    -28,   320,   -628,  312,   -92,   -92,   -472,  268,   16,    560,  516,
+    -672,  -52,   492,   -100,  260,   384,   284,   292,   304,   -148, 88,
+    -152,  1012,  1064,  -228,  164,   -376,  -684,  592,   -392,  156,  196,
+    -524,  -64,   -884,  160,   -176,  636,   648,   404,   -396,  -436, 864,
+    424,   -728,  988,   -604,  904,   -592,  296,   -224,  536,   -176, -920,
+    436,   -48,   1176,  -884,  416,   -776,  -824,  -884,  524,   -548, -564,
+    -68,   -164,  -96,   692,   364,   -692,  -1012, -68,   260,   -480, 876,
+    -1116, 452,   -332,  -352,  892,   -1088, 1220,  -676,  12,    -292, 244,
+    496,   372,   -32,   280,   200,   112,   -440,  -96,   24,    -644, -184,
+    56,    -432,  224,   -980,  272,   -260,  144,   -436,  420,   356,  364,
+    -528,  76,    172,   -744,  -368,  404,   -752,  -416,  684,   -688, 72,
+    540,   416,   92,    444,   480,   -72,   -1416, 164,   -1172, -68,  24,
+    424,   264,   1040,  128,   -912,  -524,  -356,  64,    876,   -12,  4,
+    -88,   532,   272,   -524,  320,   276,   -508,  940,   24,    -400, -120,
+    756,   60,    236,   -412,  100,   376,   -484,  400,   -100,  -740, -108,
+    -260,  328,   -268,  224,   -200,  -416,  184,   -604,  -564,  -20,  296,
+    60,    892,   -888,  60,    164,   68,    -760,  216,   -296,  904,  -336,
+    -28,   404,   -356,  -568,  -208,  -1480, -512,  296,   328,   -360, -164,
+    -1560, -776,  1156,  -428,  164,   -504,  -112,  120,   -216,  -148, -264,
+    308,   32,    64,    -72,   72,    116,   176,   -64,   -272,  460,  -536,
+    -784,  -280,  348,   108,   -752,  -132,  524,   -540,  -776,  116,  -296,
+    -1196, -288,  -560,  1040,  -472,  116,   -848,  -1116, 116,   636,  696,
+    284,   -176,  1016,  204,   -864,  -648,  -248,  356,   972,   -584, -204,
+    264,   880,   528,   -24,   -184,  116,   448,   -144,  828,   524,  212,
+    -212,  52,    12,    200,   268,   -488,  -404,  -880,  824,   -672, -40,
+    908,   -248,  500,   716,   -576,  492,   -576,  16,    720,   -108, 384,
+    124,   344,   280,   576,   -500,  252,   104,   -308,  196,   -188, -8,
+    1268,  296,   1032,  -1196, 436,   316,   372,   -432,  -200,  -660, 704,
+    -224,  596,   -132,  268,   32,    -452,  884,   104,   -1008, 424,  -1348,
+    -280,  4,     -1168, 368,   476,   696,   300,   -8,    24,    180,  -592,
+    -196,  388,   304,   500,   724,   -160,  244,   -84,   272,   -256, -420,
+    320,   208,   -144,  -156,  156,   364,   452,   28,    540,   316,  220,
+    -644,  -248,  464,   72,    360,   32,    -388,  496,   -680,  -48,  208,
+    -116,  -408,  60,    -604,  -392,  548,   -840,  784,   -460,  656,  -544,
+    -388,  -264,  908,   -800,  -628,  -612,  -568,  572,   -220,  164,  288,
+    -16,   -308,  308,   -112,  -636,  -760,  280,   -668,  432,   364,  240,
+    -196,  604,   340,   384,   196,   592,   -44,   -500,  432,   -580, -132,
+    636,   -76,   392,   4,     -412,  540,   508,   328,   -356,  -36,  16,
+    -220,  -64,   -248,  -60,   24,    -192,  368,   1040,  92,    -24,  -1044,
+    -32,   40,    104,   148,   192,   -136,  -520,  56,    -816,  -224, 732,
+    392,   356,   212,   -80,   -424,  -1008, -324,  588,   -1496, 576,  460,
+    -816,  -848,  56,    -580,  -92,   -1372, -112,  -496,  200,   364,  52,
+    -140,  48,    -48,   -60,   84,    72,    40,    132,   -356,  -268, -104,
+    -284,  -404,  732,   -520,  164,   -304,  -540,  120,   328,   -76,  -460,
+    756,   388,   588,   236,   -436,  -72,   -176,  -404,  -316,  -148, 716,
+    -604,  404,   -72,   -88,   -888,  -68,   944,   88,    -220,  -344, 960,
+    472,   460,   -232,  704,   120,   832,   -228,  692,   -508,  132,  -476,
+    844,   -748,  -364,  -44,   1116,  -1104, -1056, 76,    428,   552,  -692,
+    60,    356,   96,    -384,  -188,  -612,  -576,  736,   508,   892,  352,
+    -1132, 504,   -24,   -352,  324,   332,   -600,  -312,  292,   508,  -144,
+    -8,    484,   48,    284,   -260,  -240,  256,   -100,  -292,  -204, -44,
+    472,   -204,  908,   -188,  -1000, -256,  92,    1164,  -392,  564,  356,
+    652,   -28,   -884,  256,   484,   -192,  760,   -176,  376,   -524, -452,
+    -436,  860,   -736,  212,   124,   504,   -476,  468,   76,    -472, 552,
+    -692,  -944,  -620,  740,   -240,  400,   132,   20,    192,   -196, 264,
+    -668,  -1012, -60,   296,   -316,  -828,  76,    -156,  284,   -768, -448,
+    -832,  148,   248,   652,   616,   1236,  288,   -328,  -400,  -124, 588,
+    220,   520,   -696,  1032,  768,   -740,  -92,   -272,  296,   448,  -464,
+    412,   -200,  392,   440,   -200,  264,   -152,  -260,  320,   1032, 216,
+    320,   -8,    -64,   156,   -1016, 1084,  1172,  536,   484,   -432, 132,
+    372,   -52,   -256,  84,    116,   -352,  48,    116,   304,   -384, 412,
+    924,   -300,  528,   628,   180,   648,   44,    -980,  -220,  1320, 48,
+    332,   748,   524,   -268,  -720,  540,   -276,  564,   -344,  -208, -196,
+    436,   896,   88,    -392,  132,   80,    -964,  -288,  568,   56,   -48,
+    -456,  888,   8,     552,   -156,  -292,  948,   288,   128,   -716, -292,
+    1192,  -152,  876,   352,   -600,  -260,  -812,  -468,  -28,   -120, -32,
+    -44,   1284,  496,   192,   464,   312,   -76,   -516,  -380,  -456, -1012,
+    -48,   308,   -156,  36,    492,   -156,  -808,  188,   1652,  68,   -120,
+    -116,  316,   160,   -140,  352,   808,   -416,  592,   316,   -480, 56,
+    528,   -204,  -568,  372,   -232,  752,   -344,  744,   -4,    324,  -416,
+    -600,  768,   268,   -248,  -88,   -132,  -420,  -432,  80,    -288, 404,
+    -316,  -1216, -588,  520,   -108,  92,    -320,  368,   -480,  -216, -92,
+    1688,  -300,  180,   1020,  -176,  820,   -68,   -228,  -260,  436,  -904,
+    20,    40,    -508,  440,   -736,  312,   332,   204,   760,   -372, 728,
+    96,    -20,   -632,  -520,  -560,  336,   1076,  -64,   -532,  776,  584,
+    192,   396,   -728,  -520,  276,   -188,  80,    -52,   -612,  -252, -48,
+    648,   212,   -688,  228,   -52,   -260,  428,   -412,  -272,  -404, 180,
+    816,   -796,  48,    152,   484,   -88,   -216,  988,   696,   188,  -528,
+    648,   -116,  -180,  316,   476,   12,    -564,  96,    476,   -252, -364,
+    -376,  -392,  556,   -256,  -576,  260,   -352,  120,   -16,   -136, -260,
+    -492,  72,    556,   660,   580,   616,   772,   436,   424,   -32,  -324,
+    -1268, 416,   -324,  -80,   920,   160,   228,   724,   32,    -516, 64,
+    384,   68,    -128,  136,   240,   248,   -204,  -68,   252,   -932, -120,
+    -480,  -628,  -84,   192,   852,   -404,  -288,  -132,  204,   100,  168,
+    -68,   -196,  -868,  460,   1080,  380,   -80,   244,   0,     484,  -888,
+    64,    184,   352,   600,   460,   164,   604,   -196,  320,   -64,  588,
+    -184,  228,   12,    372,   48,    -848,  -344,  224,   208,   -200, 484,
+    128,   -20,   272,   -468,  -840,  384,   256,   -720,  -520,  -464, -580,
+    112,   -120,  644,   -356,  -208,  -608,  -528,  704,   560,   -424, 392,
+    828,   40,    84,    200,   -152,  0,     -144,  584,   280,   -120, 80,
+    -556,  -972,  -196,  -472,  724,   80,    168,   -32,   88,    160,  -688,
+    0,     160,   356,   372,   -776,  740,   -128,  676,   -248,  -480, 4,
+    -364,  96,    544,   232,   -1032, 956,   236,   356,   20,    -40,  300,
+    24,    -676,  -596,  132,   1120,  -104,  532,   -1096, 568,   648,  444,
+    508,   380,   188,   -376,  -604,  1488,  424,   24,    756,   -220, -192,
+    716,   120,   920,   688,   168,   44,    -460,  568,   284,   1144, 1160,
+    600,   424,   888,   656,   -356,  -320,  220,   316,   -176,  -724, -188,
+    -816,  -628,  -348,  -228,  -380,  1012,  -452,  -660,  736,   928,  404,
+    -696,  -72,   -268,  -892,  128,   184,   -344,  -780,  360,   336,  400,
+    344,   428,   548,   -112,  136,   -228,  -216,  -820,  -516,  340,  92,
+    -136,  116,   -300,  376,   -244,  100,   -316,  -520,  -284,  -12,  824,
+    164,   -548,  -180,  -128,  116,   -924,  -828,  268,   -368,  -580, 620,
+    192,   160,   0,     -1676, 1068,  424,   -56,   -360,  468,   -156, 720,
+    288,   -528,  556,   -364,  548,   -148,  504,   316,   152,   -648, -620,
+    -684,  -24,   -376,  -384,  -108,  -920,  -1032, 768,   180,   -264, -508,
+    -1268, -260,  -60,   300,   -240,  988,   724,   -376,  -576,  -212, -736,
+    556,   192,   1092,  -620,  -880,  376,   -56,   -4,    -216,  -32,  836,
+    268,   396,   1332,  864,   -600,  100,   56,    -412,  -92,   356,  180,
+    884,   -468,  -436,  292,   -388,  -804,  -704,  -840,  368,   -348, 140,
+    -724,  1536,  940,   372,   112,   -372,  436,   -480,  1136,  296,  -32,
+    -228,  132,   -48,   -220,  868,   -1016, -60,   -1044, -464,  328,  916,
+    244,   12,    -736,  -296,  360,   468,   -376,  -108,  -92,   788,  368,
+    -56,   544,   400,   -672,  -420,  728,   16,    320,   44,    -284, -380,
+    -796,  488,   132,   204,   -596,  -372,  88,    -152,  -908,  -636, -572,
+    -624,  -116,  -692,  -200,  -56,   276,   -88,   484,   -324,  948,  864,
+    1000,  -456,  -184,  -276,  292,   -296,  156,   676,   320,   160,  908,
+    -84,   -1236, -288,  -116,  260,   -372,  -644,  732,   -756,  -96,  84,
+    344,   -520,  348,   -688,  240,   -84,   216,   -1044, -136,  -676, -396,
+    -1500, 960,   -40,   176,   168,   1516,  420,   -504,  -344,  -364, -360,
+    1216,  -940,  -380,  -212,  252,   -660,  -708,  484,   -444,  -152, 928,
+    -120,  1112,  476,   -260,  560,   -148,  -344,  108,   -196,  228,  -288,
+    504,   560,   -328,  -88,   288,   -1008, 460,   -228,  468,   -836, -196,
+    76,    388,   232,   412,   -1168, -716,  -644,  756,   -172,  -356, -504,
+    116,   432,   528,   48,    476,   -168,  -608,  448,   160,   -532, -272,
+    28,    -676,  -12,   828,   980,   456,   520,   104,   -104,  256,  -344,
+    -4,    -28,   -368,  -52,   -524,  -572,  -556,  -200,  768,   1124, -208,
+    -512,  176,   232,   248,   -148,  -888,  604,   -600,  -304,  804,  -156,
+    -212,  488,   -192,  -804,  -256,  368,   -360,  -916,  -328,  228,  -240,
+    -448,  -472,  856,   -556,  -364,  572,   -12,   -156,  -368,  -340, 432,
+    252,   -752,  -152,  288,   268,   -580,  -848,  -592,  108,   -76,  244,
+    312,   -716,  592,   -80,   436,   360,   4,     -248,  160,   516,  584,
+    732,   44,    -468,  -280,  -292,  -156,  -588,  28,    308,   912,  24,
+    124,   156,   180,   -252,  944,   -924,  -772,  -520,  -428,  -624, 300,
+    -212,  -1144, 32,    -724,  800,   -1128, -212,  -1288, -848,  180,  -416,
+    440,   192,   -576,  -792,  -76,   -1080, 80,    -532,  -352,  -132, 380,
+    -820,  148,   1112,  128,   164,   456,   700,   -924,  144,   -668, -384,
+    648,   -832,  508,   552,   -52,   -100,  -656,  208,   -568,  748,  -88,
+    680,   232,   300,   192,   -408,  -1012, -152,  -252,  -268,  272,  -876,
+    -664,  -648,  -332,  -136,  16,    12,    1152,  -28,   332,   -536, 320,
+    -672,  -460,  -316,  532,   -260,  228,   -40,   1052,  -816,  180,  88,
+    -496,  -556,  -672,  -368,  428,   92,    356,   404,   -408,  252,  196,
+    -176,  -556,  792,   268,   32,    372,   40,    96,    -332,  328,  120,
+    372,   -900,  -40,   472,   -264,  -592,  952,   128,   656,   112,  664,
+    -232,  420,   4,     -344,  -464,  556,   244,   -416,  -32,   252,  0,
+    -412,  188,   -696,  508,   -476,  324,   -1096, 656,   -312,  560,  264,
+    -136,  304,   160,   -64,   -580,  248,   336,   -720,  560,   -348, -288,
+    -276,  -196,  -500,  852,   -544,  -236,  -1128, -992,  -776,  116,  56,
+    52,    860,   884,   212,   -12,   168,   1020,  512,   -552,  924,  -148,
+    716,   188,   164,   -340,  -520,  -184,  880,   -152,  -680,  -208, -1156,
+    -300,  -528,  -472,  364,   100,   -744,  -1056, -32,   540,   280,  144,
+    -676,  -32,   -232,  -280,  -224,  96,    568,   -76,   172,   148,  148,
+    104,   32,    -296,  -32,   788,   -80,   32,    -16,   280,   288,  944,
+    428,   -484
+};
diff --git a/libavcodec/aom_film_grain.h b/libavcodec/aom_film_grain.h
new file mode 100644
index 00000000000..5d772bd7d17
--- /dev/null
+++ b/libavcodec/aom_film_grain.h
@@ -0,0 +1,38 @@
+/*
+ * AOM film grain synthesis
+ * Copyright (c) 2021 Niklas Haas <ffmpeg@haasn.xyz>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file
+ * AOM film grain synthesis.
+ * @author Niklas Haas <ffmpeg@haasn.xyz>
+ */
+
+#ifndef AVCODEC_AOM_FILM_GRAIN_H
+#define AVCODEC_AOM_FILM_GRAIN_H
+
+#include "libavutil/film_grain_params.h"
+
+// Synthesizes film grain on top of `in` and stores the result to `out`. `out`
+// must already have been allocated and set to the same size and format as `in`.
+int ff_aom_apply_film_grain(AVFrame *out, const AVFrame *in,
+                            const AVFilmGrainParams *params);
+
+#endif /* AVCODEC_AOM_FILM_GRAIN_H */
diff --git a/libavcodec/aom_film_grain_template.c b/libavcodec/aom_film_grain_template.c
new file mode 100644
index 00000000000..39ee674bc15
--- /dev/null
+++ b/libavcodec/aom_film_grain_template.c
@@ -0,0 +1,606 @@
+/*
+ * AOM film grain synthesis
+ * Copyright (c) 2023 Niklas Haas <ffmpeg@haasn.xyz>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/*
+ * Copyright © 2018, Niklas Haas
+ * Copyright © 2018, VideoLAN and dav1d authors
+ * Copyright © 2018, Two Orioles, LLC
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright notice, this
+ *    list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright notice,
+ *    this list of conditions and the following disclaimer in the documentation
+ *    and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+ * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "bit_depth_template.c"
+
+#undef entry
+#undef bitdepth
+#undef bitdepth_max
+#undef HBD_DECL
+#undef HBD_CALL
+#undef SCALING_SIZE
+
+#if BIT_DEPTH > 8
+# define entry int16_t
+# define bitdepth_max ((1 << bitdepth) - 1)
+# define HBD_DECL , const int bitdepth
+# define HBD_CALL , bitdepth
+# define SCALING_SIZE 4096
+#else
+# define entry int8_t
+# define bitdepth 8
+# define bitdepth_max UINT8_MAX
+# define HBD_DECL
+# define HBD_CALL
+# define SCALING_SIZE 256
+#endif
+
+// Symbols that do not depend on bit depth
+#ifndef PXSTRIDE
+# define PXSTRIDE(x) ((x) / sizeof(pixel))
+# define GRAIN_WIDTH 82
+# define GRAIN_HEIGHT 73
+# define SUB_GRAIN_WIDTH 44
+# define SUB_GRAIN_HEIGHT 38
+# define FG_BLOCK_SIZE 32
+#endif
+
+static void FUNC(generate_grain_y_c)(entry buf[][GRAIN_WIDTH],
+                                     const AVFilmGrainParams *const params
+                                     HBD_DECL)
+{
+    const AVFilmGrainAOMParams *const data = &params->codec.aom;
+    const int bitdepth_min_8 = bitdepth - 8;
+    unsigned seed = params->seed;
+    const int shift = 4 - bitdepth_min_8 + data->grain_scale_shift;
+    const int grain_ctr = 128 << bitdepth_min_8;
+    const int grain_min = -grain_ctr, grain_max = grain_ctr - 1;
+
+    const int ar_pad = 3;
+    const int ar_lag = data->ar_coeff_lag;
+
+    for (int y = 0; y < GRAIN_HEIGHT; y++) {
+        for (int x = 0; x < GRAIN_WIDTH; x++) {
+            const int value = get_random_number(11, &seed);
+            buf[y][x] = round2(gaussian_sequence[ value ], shift);
+        }
+    }
+
+    for (int y = ar_pad; y < GRAIN_HEIGHT; y++) {
+        for (int x = ar_pad; x < GRAIN_WIDTH - ar_pad; x++) {
+            const int8_t *coeff = data->ar_coeffs_y;
+            int sum = 0, grain;
+            for (int dy = -ar_lag; dy <= 0; dy++) {
+                for (int dx = -ar_lag; dx <= ar_lag; dx++) {
+                    if (!dx && !dy)
+                        break;
+                    sum += *(coeff++) * buf[y + dy][x + dx];
+                }
+            }
+
+            grain = buf[y][x] + round2(sum, data->ar_coeff_shift);
+            buf[y][x] = av_clip(grain, grain_min, grain_max);
+        }
+    }
+}
+
+static void
+FUNC(generate_grain_uv_c)(entry buf[][GRAIN_WIDTH],
+                          const entry buf_y[][GRAIN_WIDTH],
+                          const AVFilmGrainParams *const params, const intptr_t uv,
+                          const int subx, const int suby HBD_DECL)
+{
+    const AVFilmGrainAOMParams *const data = &params->codec.aom;
+    const int bitdepth_min_8 = bitdepth - 8;
+    unsigned seed = params->seed ^ (uv ? 0x49d8 : 0xb524);
+    const int shift = 4 - bitdepth_min_8 + data->grain_scale_shift;
+    const int grain_ctr = 128 << bitdepth_min_8;
+    const int grain_min = -grain_ctr, grain_max = grain_ctr - 1;
+
+    const int chromaW = subx ? SUB_GRAIN_WIDTH  : GRAIN_WIDTH;
+    const int chromaH = suby ? SUB_GRAIN_HEIGHT : GRAIN_HEIGHT;
+
+    const int ar_pad = 3;
+    const int ar_lag = data->ar_coeff_lag;
+
+    for (int y = 0; y < chromaH; y++) {
+        for (int x = 0; x < chromaW; x++) {
+            const int value = get_random_number(11, &seed);
+            buf[y][x] = round2(gaussian_sequence[ value ], shift);
+        }
+    }
+
+    for (int y = ar_pad; y < chromaH; y++) {
+        for (int x = ar_pad; x < chromaW - ar_pad; x++) {
+            const int8_t *coeff = data->ar_coeffs_uv[uv];
+            int sum = 0, grain;
+            for (int dy = -ar_lag; dy <= 0; dy++) {
+                for (int dx = -ar_lag; dx <= ar_lag; dx++) {
+                    // For the final (current) pixel, we need to add in the
+                    // contribution from the luma grain texture
+                    if (!dx && !dy) {
+                        const int lumaX = ((x - ar_pad) << subx) + ar_pad;
+                        const int lumaY = ((y - ar_pad) << suby) + ar_pad;
+                        int luma = 0;
+                        if (!data->num_y_points)
+                            break;
+                        for (int i = 0; i <= suby; i++) {
+                            for (int j = 0; j <= subx; j++) {
+                                luma += buf_y[lumaY + i][lumaX + j];
+                            }
+                        }
+                        luma = round2(luma, subx + suby);
+                        sum += luma * (*coeff);
+                        break;
+                    }
+
+                    sum += *(coeff++) * buf[y + dy][x + dx];
+                }
+            }
+
+            grain = buf[y][x] + round2(sum, data->ar_coeff_shift);
+            buf[y][x] = av_clip(grain, grain_min, grain_max);
+        }
+    }
+}
+
+// samples from the correct block of a grain LUT, while taking into account the
+// offsets provided by the offsets cache
+static inline entry FUNC(sample_lut)(const entry grain_lut[][GRAIN_WIDTH],
+                                     const int offsets[2][2],
+                                     const int subx, const int suby,
+                                     const int bx, const int by,
+                                     const int x, const int y)
+{
+    const int randval = offsets[bx][by];
+    const int offx = 3 + (2 >> subx) * (3 + (randval >> 4));
+    const int offy = 3 + (2 >> suby) * (3 + (randval & 0xF));
+    return grain_lut[offy + y + (FG_BLOCK_SIZE >> suby) * by]
+                    [offx + x + (FG_BLOCK_SIZE >> subx) * bx];
+}
+
+static void FUNC(fgy_32x32xn_c)(pixel *const dst_row, const pixel *const src_row,
+                                const ptrdiff_t stride,
+                                const AVFilmGrainParams *const params, const size_t pw,
+                                const uint8_t scaling[SCALING_SIZE],
+                                const entry grain_lut[][GRAIN_WIDTH],
+                                const int bh, const int row_num HBD_DECL)
+{
+    const AVFilmGrainAOMParams *const data = &params->codec.aom;
+    const int rows = 1 + (data->overlap_flag && row_num > 0);
+    const int bitdepth_min_8 = bitdepth - 8;
+    const int grain_ctr = 128 << bitdepth_min_8;
+    const int grain_min = -grain_ctr, grain_max = grain_ctr - 1;
+    unsigned seed[2];
+    int offsets[2 /* col offset */][2 /* row offset */];
+
+    int min_value, max_value;
+    if (data->limit_output_range) {
+        min_value = 16 << bitdepth_min_8;
+        max_value = 235 << bitdepth_min_8;
+    } else {
+        min_value = 0;
+        max_value = bitdepth_max;
+    }
+
+    // seed[0] contains the current row, seed[1] contains the previous
+    for (int i = 0; i < rows; i++) {
+        seed[i] = params->seed;
+        seed[i] ^= (((row_num - i) * 37  + 178) & 0xFF) << 8;
+        seed[i] ^= (((row_num - i) * 173 + 105) & 0xFF);
+    }
+
+    av_assert1(stride % (FG_BLOCK_SIZE * sizeof(pixel)) == 0);
+
+    // process this row in FG_BLOCK_SIZE^2 blocks
+    for (unsigned bx = 0; bx < pw; bx += FG_BLOCK_SIZE) {
+        const int bw = FFMIN(FG_BLOCK_SIZE, (int) pw - bx);
+        const pixel *src;
+        pixel *dst;
+        int noise;
+
+        // x/y block offsets to compensate for overlapped regions
+        const int ystart = data->overlap_flag && row_num ? FFMIN(2, bh) : 0;
+        const int xstart = data->overlap_flag && bx      ? FFMIN(2, bw) : 0;
+
+        static const int w[2][2] = { { 27, 17 }, { 17, 27 } };
+
+        if (data->overlap_flag && bx) {
+            // shift previous offsets left
+            for (int i = 0; i < rows; i++)
+                offsets[1][i] = offsets[0][i];
+        }
+
+        // update current offsets
+        for (int i = 0; i < rows; i++)
+            offsets[0][i] = get_random_number(8, &seed[i]);
+
+#define add_noise_y(x, y, grain)                                        \
+        src = src_row + (y) * PXSTRIDE(stride) + (x) + bx;              \
+        dst = dst_row + (y) * PXSTRIDE(stride) + (x) + bx;              \
+        noise = round2(scaling[ *src ] * (grain), data->scaling_shift); \
+        *dst = av_clip(*src + noise, min_value, max_value);
+
+        for (int y = ystart; y < bh; y++) {
+            // Non-overlapped image region (straightforward)
+            for (int x = xstart; x < bw; x++) {
+                int grain = FUNC(sample_lut)(grain_lut, offsets, 0, 0, 0, 0, x, y);
+                add_noise_y(x, y, grain);
+            }
+
+            // Special case for overlapped column
+            for (int x = 0; x < xstart; x++) {
+                int grain = FUNC(sample_lut)(grain_lut, offsets, 0, 0, 0, 0, x, y);
+                int old   = FUNC(sample_lut)(grain_lut, offsets, 0, 0, 1, 0, x, y);
+                grain = round2(old * w[x][0] + grain * w[x][1], 5);
+                grain = av_clip(grain, grain_min, grain_max);
+                add_noise_y(x, y, grain);
+            }
+        }
+
+        for (int y = 0; y < ystart; y++) {
+            // Special case for overlapped row (sans corner)
+            for (int x = xstart; x < bw; x++) {
+                int grain = FUNC(sample_lut)(grain_lut, offsets, 0, 0, 0, 0, x, y);
+                int old   = FUNC(sample_lut)(grain_lut, offsets, 0, 0, 0, 1, x, y);
+                grain = round2(old * w[y][0] + grain * w[y][1], 5);
+                grain = av_clip(grain, grain_min, grain_max);
+                add_noise_y(x, y, grain);
+            }
+
+            // Special case for doubly-overlapped corner
+            for (int x = 0; x < xstart; x++) {
+                int grain = FUNC(sample_lut)(grain_lut, offsets, 0, 0, 0, 0, x, y);
+                int top = FUNC(sample_lut)(grain_lut, offsets, 0, 0, 0, 1, x, y);
+                int old = FUNC(sample_lut)(grain_lut, offsets, 0, 0, 1, 1, x, y);
+
+                // Blend the top pixel with the top left block
+                top = round2(old * w[x][0] + top * w[x][1], 5);
+                top = av_clip(top, grain_min, grain_max);
+
+                // Blend the current pixel with the left block
+                old = FUNC(sample_lut)(grain_lut, offsets, 0, 0, 1, 0, x, y);
+                grain = round2(old * w[x][0] + grain * w[x][1], 5);
+                grain = av_clip(grain, grain_min, grain_max);
+
+                // Mix the row rows together and apply grain
+                grain = round2(top * w[y][0] + grain * w[y][1], 5);
+                grain = av_clip(grain, grain_min, grain_max);
+                add_noise_y(x, y, grain);
+            }
+        }
+    }
+}
+
+static void
+FUNC(fguv_32x32xn_c)(pixel *const dst_row, const pixel *const src_row,
+                     const ptrdiff_t stride, const AVFilmGrainParams *const params,
+                     const size_t pw, const uint8_t scaling[SCALING_SIZE],
+                     const entry grain_lut[][GRAIN_WIDTH], const int bh,
+                     const int row_num, const pixel *const luma_row,
+                     const ptrdiff_t luma_stride, const int uv, const int is_id,
+                     const int sx, const int sy HBD_DECL)
+{
+    const AVFilmGrainAOMParams *const data = &params->codec.aom;
+    const int rows = 1 + (data->overlap_flag && row_num > 0);
+    const int bitdepth_min_8 = bitdepth - 8;
+    const int grain_ctr = 128 << bitdepth_min_8;
+    const int grain_min = -grain_ctr, grain_max = grain_ctr - 1;
+    unsigned seed[2];
+    int offsets[2 /* col offset */][2 /* row offset */];
+
+    int min_value, max_value;
+    if (data->limit_output_range) {
+        min_value = 16 << bitdepth_min_8;
+        max_value = (is_id ? 235 : 240) << bitdepth_min_8;
+    } else {
+        min_value = 0;
+        max_value = bitdepth_max;
+    }
+
+    // seed[0] contains the current row, seed[1] contains the previous
+    for (int i = 0; i < rows; i++) {
+        seed[i] = params->seed;
+        seed[i] ^= (((row_num - i) * 37  + 178) & 0xFF) << 8;
+        seed[i] ^= (((row_num - i) * 173 + 105) & 0xFF);
+    }
+
+    av_assert1(stride % (FG_BLOCK_SIZE * sizeof(pixel)) == 0);
+
+    // process this row in FG_BLOCK_SIZE^2 blocks (subsampled)
+    for (unsigned bx = 0; bx < pw; bx += FG_BLOCK_SIZE >> sx) {
+        const int bw = FFMIN(FG_BLOCK_SIZE >> sx, (int)(pw - bx));
+        int val, lx, ly, noise;
+        const pixel *src, *luma;
+        pixel *dst, avg;
+
+        // x/y block offsets to compensate for overlapped regions
+        const int ystart = data->overlap_flag && row_num ? FFMIN(2 >> sy, bh) : 0;
+        const int xstart = data->overlap_flag && bx      ? FFMIN(2 >> sx, bw) : 0;
+
+        static const int w[2 /* sub */][2 /* off */][2] = {
+            { { 27, 17 }, { 17, 27 } },
+            { { 23, 22 } },
+        };
+
+        if (data->overlap_flag && bx) {
+            // shift previous offsets left
+            for (int i = 0; i < rows; i++)
+                offsets[1][i] = offsets[0][i];
+        }
+
+        // update current offsets
+        for (int i = 0; i < rows; i++)
+            offsets[0][i] = get_random_number(8, &seed[i]);
+
+#define add_noise_uv(x, y, grain)                                               \
+            lx = (bx + x) << sx;                                                \
+            ly = y << sy;                                                       \
+            luma = luma_row + ly * PXSTRIDE(luma_stride) + lx;                  \
+            avg = luma[0];                                                      \
+            if (sx)                                                             \
+                avg = (avg + luma[1] + 1) >> 1;                                 \
+            src = src_row + (y) * PXSTRIDE(stride) + (bx + (x));                \
+            dst = dst_row + (y) * PXSTRIDE(stride) + (bx + (x));                \
+            val = avg;                                                          \
+            if (!data->chroma_scaling_from_luma) {                              \
+                const int combined = avg * data->uv_mult_luma[uv] +             \
+                                    *src * data->uv_mult[uv];                   \
+                val = av_clip( (combined >> 6) +                                \
+                               (data->uv_offset[uv] * (1 << bitdepth_min_8)),   \
+                               0, bitdepth_max );                               \
+            }                                                                   \
+            noise = round2(scaling[ val ] * (grain), data->scaling_shift);      \
+            *dst = av_clip(*src + noise, min_value, max_value);
+
+        for (int y = ystart; y < bh; y++) {
+            // Non-overlapped image region (straightforward)
+            for (int x = xstart; x < bw; x++) {
+                int grain = FUNC(sample_lut)(grain_lut, offsets, sx, sy, 0, 0, x, y);
+                add_noise_uv(x, y, grain);
+            }
+
+            // Special case for overlapped column
+            for (int x = 0; x < xstart; x++) {
+                int grain = FUNC(sample_lut)(grain_lut, offsets, sx, sy, 0, 0, x, y);
+                int old   = FUNC(sample_lut)(grain_lut, offsets, sx, sy, 1, 0, x, y);
+                grain = round2(old * w[sx][x][0] + grain * w[sx][x][1], 5);
+                grain = av_clip(grain, grain_min, grain_max);
+                add_noise_uv(x, y, grain);
+            }
+        }
+
+        for (int y = 0; y < ystart; y++) {
+            // Special case for overlapped row (sans corner)
+            for (int x = xstart; x < bw; x++) {
+                int grain = FUNC(sample_lut)(grain_lut, offsets, sx, sy, 0, 0, x, y);
+                int old   = FUNC(sample_lut)(grain_lut, offsets, sx, sy, 0, 1, x, y);
+                grain = round2(old * w[sy][y][0] + grain * w[sy][y][1], 5);
+                grain = av_clip(grain, grain_min, grain_max);
+                add_noise_uv(x, y, grain);
+            }
+
+            // Special case for doubly-overlapped corner
+            for (int x = 0; x < xstart; x++) {
+                int top = FUNC(sample_lut)(grain_lut, offsets, sx, sy, 0, 1, x, y);
+                int old = FUNC(sample_lut)(grain_lut, offsets, sx, sy, 1, 1, x, y);
+                int grain = FUNC(sample_lut)(grain_lut, offsets, sx, sy, 0, 0, x, y);
+
+                // Blend the top pixel with the top left block
+                top = round2(old * w[sx][x][0] + top * w[sx][x][1], 5);
+                top = av_clip(top, grain_min, grain_max);
+
+                // Blend the current pixel with the left block
+                old = FUNC(sample_lut)(grain_lut, offsets, sx, sy, 1, 0, x, y);
+                grain = round2(old * w[sx][x][0] + grain * w[sx][x][1], 5);
+                grain = av_clip(grain, grain_min, grain_max);
+
+                // Mix the row rows together and apply to image
+                grain = round2(top * w[sy][y][0] + grain * w[sy][y][1], 5);
+                grain = av_clip(grain, grain_min, grain_max);
+                add_noise_uv(x, y, grain);
+            }
+        }
+    }
+}
+
+static void FUNC(generate_scaling)(const uint8_t points[][2], const int num,
+                                   uint8_t scaling[SCALING_SIZE] HBD_DECL)
+{
+    const int shift_x = bitdepth - 8;
+    const int scaling_size = 1 << bitdepth;
+    const int max_value = points[num - 1][0] << shift_x;
+    av_assert0(scaling_size <= SCALING_SIZE);
+
+    if (num == 0) {
+        memset(scaling, 0, scaling_size);
+        return;
+    }
+
+    // Fill up the preceding entries with the initial value
+    memset(scaling, points[0][1], points[0][0] << shift_x);
+
+    // Linearly interpolate the values in the middle
+    for (int i = 0; i < num - 1; i++) {
+        const int bx = points[i][0];
+        const int by = points[i][1];
+        const int ex = points[i+1][0];
+        const int ey = points[i+1][1];
+        const int dx = ex - bx;
+        const int dy = ey - by;
+        const int delta = dy * ((0x10000 + (dx >> 1)) / dx);
+        av_assert1(dx > 0);
+        for (int x = 0, d = 0x8000; x < dx; x++) {
+            scaling[(bx + x) << shift_x] = by + (d >> 16);
+            d += delta;
+        }
+    }
+
+    // Fill up the remaining entries with the final value
+    memset(&scaling[max_value], points[num - 1][1], scaling_size - max_value);
+
+#if BIT_DEPTH != 8
+    for (int i = 0; i < num - 1; i++) {
+        const int pad = 1 << shift_x, rnd = pad >> 1;
+        const int bx = points[i][0] << shift_x;
+        const int ex = points[i+1][0] << shift_x;
+        const int dx = ex - bx;
+        for (int x = 0; x < dx; x += pad) {
+            const int range = scaling[bx + x + pad] - scaling[bx + x];
+            for (int n = 1, r = rnd; n < pad; n++) {
+                r += range;
+                scaling[bx + x + n] = scaling[bx + x] + (r >> shift_x);
+            }
+        }
+    }
+#endif
+}
+
+static av_always_inline void
+FUNC(apply_grain_row)(AVFrame *out, const AVFrame *in,
+                      const int ss_x, const int ss_y,
+                      const uint8_t scaling[3][SCALING_SIZE],
+                      const entry grain_lut[3][GRAIN_HEIGHT+1][GRAIN_WIDTH],
+                      const AVFilmGrainParams *params,
+                      const int row HBD_DECL)
+{
+    // Synthesize grain for the affected planes
+    const AVFilmGrainAOMParams *const data = &params->codec.aom;
+    const int cpw = (out->width + ss_x) >> ss_x;
+    const int is_id = out->colorspace == AVCOL_SPC_RGB;
+    const int bh = (FFMIN(out->height - row * FG_BLOCK_SIZE, FG_BLOCK_SIZE) + ss_y) >> ss_y;
+    const ptrdiff_t uv_off = row * FG_BLOCK_SIZE * PXSTRIDE(out->linesize[1]) >> ss_y;
+    pixel *const luma_src =
+        ((pixel *) in->data[0]) + row * FG_BLOCK_SIZE * PXSTRIDE(in->linesize[0]);
+
+    if (data->num_y_points) {
+        const int bh = FFMIN(out->height - row * FG_BLOCK_SIZE, FG_BLOCK_SIZE);
+        const ptrdiff_t off = row * FG_BLOCK_SIZE * PXSTRIDE(out->linesize[0]);
+        FUNC(fgy_32x32xn_c)(((pixel *) out->data[0]) + off, luma_src,
+                            out->linesize[0], params, out->width, scaling[0],
+                            grain_lut[0], bh, row HBD_CALL);
+    }
+
+    if (!data->num_uv_points[0] && !data->num_uv_points[1] &&
+        !data->chroma_scaling_from_luma)
+    {
+        return;
+    }
+
+    // extend padding pixels
+    if (out->width & ss_x) {
+        pixel *ptr = luma_src;
+        for (int y = 0; y < bh; y++) {
+            ptr[out->width] = ptr[out->width - 1];
+            ptr += PXSTRIDE(in->linesize[0]) << ss_y;
+        }
+    }
+
+    if (data->chroma_scaling_from_luma) {
+        for (int pl = 0; pl < 2; pl++)
+            FUNC(fguv_32x32xn_c)(((pixel *) out->data[1 + pl]) + uv_off,
+                                 ((const pixel *) in->data[1 + pl]) + uv_off,
+                                 in->linesize[1], params, cpw, scaling[0],
+                                 grain_lut[1 + pl], bh, row, luma_src,
+                                 in->linesize[0], pl, is_id, ss_x, ss_y HBD_CALL);
+    } else {
+        for (int pl = 0; pl < 2; pl++) {
+            if (data->num_uv_points[pl]) {
+                FUNC(fguv_32x32xn_c)(((pixel *) out->data[1 + pl]) + uv_off,
+                                     ((const pixel *) in->data[1 + pl]) + uv_off,
+                                     in->linesize[1], params, cpw, scaling[1 + pl],
+                                     grain_lut[1 + pl], bh, row, luma_src,
+                                     in->linesize[0], pl, is_id, ss_x, ss_y HBD_CALL);
+            }
+        }
+    }
+}
+
+static int FUNC(apply_film_grain)(AVFrame *out_frame, const AVFrame *in_frame,
+                                  const AVFilmGrainParams *params HBD_DECL)
+{
+    entry grain_lut[3][GRAIN_HEIGHT + 1][GRAIN_WIDTH];
+    uint8_t scaling[3][SCALING_SIZE];
+
+    const AVFilmGrainAOMParams *const data = &params->codec.aom;
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(out_frame->format);
+    const int rows = AV_CEIL_RSHIFT(out_frame->height, 5); /* log2(FG_BLOCK_SIZE) */
+    const int subx = desc->log2_chroma_w, suby = desc->log2_chroma_h;
+    av_assert0(params->type == AV_FILM_GRAIN_PARAMS_AV1);
+    av_assert0(out_frame->format == in_frame->format);
+    if (in_frame->format != AV_PIX_FMT_YUV420P10LE)
+        return AVERROR_PATCHWELCOME;
+
+    // Generate grain LUTs as needed
+    FUNC(generate_grain_y_c)(grain_lut[0], params HBD_CALL);
+    if (data->num_uv_points[0] || data->chroma_scaling_from_luma)
+        FUNC(generate_grain_uv_c)(grain_lut[1], grain_lut[0], params, 0, subx, suby HBD_CALL);
+    if (data->num_uv_points[1] || data->chroma_scaling_from_luma)
+        FUNC(generate_grain_uv_c)(grain_lut[2], grain_lut[0], params, 1, subx, suby HBD_CALL);
+
+    // Generate scaling LUTs as needed
+    if (data->num_y_points || data->chroma_scaling_from_luma)
+        FUNC(generate_scaling)(data->y_points, data->num_y_points, scaling[0] HBD_CALL);
+    if (data->num_uv_points[0])
+        FUNC(generate_scaling)(data->uv_points[0], data->num_uv_points[0], scaling[1] HBD_CALL);
+    if (data->num_uv_points[1])
+        FUNC(generate_scaling)(data->uv_points[1], data->num_uv_points[1], scaling[2] HBD_CALL);
+
+    // Copy over the non-modified planes
+    if (!data->num_y_points) {
+        av_image_copy_plane(out_frame->data[0], out_frame->linesize[0],
+                            in_frame->data[0], in_frame->linesize[0],
+                            out_frame->width * sizeof(pixel), out_frame->height);
+    }
+    for (int uv = 0; uv < 2; uv++) {
+        if (!data->num_uv_points[uv]) {
+            av_image_copy_plane(out_frame->data[1+uv], out_frame->linesize[1+uv],
+                                in_frame->data[1+uv], in_frame->linesize[1+uv],
+                                AV_CEIL_RSHIFT(out_frame->width, subx) * sizeof(pixel),
+                                AV_CEIL_RSHIFT(out_frame->height, suby));
+        }
+    }
+
+    for (int row = 0; row < rows; row++) {
+        FUNC(apply_grain_row)(out_frame, in_frame, subx, suby, scaling, grain_lut,
+                              params, row HBD_CALL);
+    }
+
+    return 0;
+}

Niklas Haas:
> +#if BIT_DEPTH > 8
> +# define entry int16_t
> +# define bitdepth_max ((1 << bitdepth) - 1)
> +# define HBD_DECL , const int bitdepth
> +# define HBD_CALL , bitdepth
> +# define SCALING_SIZE 4096
> +#else
> +# define entry int8_t
> +# define bitdepth 8
> +# define bitdepth_max UINT8_MAX

Why are you using signed types, but unsigned max values?

> +# define HBD_DECL
> +# define HBD_CALL
> +# define SCALING_SIZE 256
> +#endif
> +
> +// Symbols that do not depend on bit depth
> +#ifndef PXSTRIDE
> +# define PXSTRIDE(x) ((x) / sizeof(pixel))

There are several things that are wrong with this pattern:
1. sizeof is size_t and this has a conversion rank >= int/ptrdiff_t on
all systems I am aware of, therefore the division will be performed with
unsigned types, i.e. it is a logical right shift. And this will just not
work with negative linesizes (but given that pointer arithmetic via a
pixel* commonly involves a multiplication by sizeof(pixel), it often
happens to work in practice, but e.g. UBSan warns about this).
2. It presumes that linesize is always a multiple of the sizeof of the
underlying type. This need not be so. Imagine a system where there are
no alignment requirements whatsoever (i.e. the required alignment of
every type is 1 and vector instructions (if existing) also have no
alignment requirement). Then it is legal for our callers to use frames
where linesize can be anything.
3. Even if linesize is always a multiple of sizeof(pixel), the compiler
does not know this and therefore will have to mask the low bits. So
instead of
src_row + (y) * PXSTRIDE(stride) + (x) + bx
use
(const pixel*)((const char*)src_row + (y) * stride) + (x) + bx

(Please don't cast const away on intermediate pointers when changing this.)

> +# define GRAIN_WIDTH 82
> +# define GRAIN_HEIGHT 73
> +# define SUB_GRAIN_WIDTH 44
> +# define SUB_GRAIN_HEIGHT 38
> +# define FG_BLOCK_SIZE 32
> +#endif
> +

...

> +    // Copy over the non-modified planes
> +    if (!data->num_y_points) {
> +        av_image_copy_plane(out_frame->data[0], out_frame->linesize[0],
> +                            in_frame->data[0], in_frame->linesize[0],
> +                            out_frame->width * sizeof(pixel), out_frame->height);
> +    }
> +    for (int uv = 0; uv < 2; uv++) {
> +        if (!data->num_uv_points[uv]) {
> +            av_image_copy_plane(out_frame->data[1+uv], out_frame->linesize[1+uv],
> +                                in_frame->data[1+uv], in_frame->linesize[1+uv],
> +                                AV_CEIL_RSHIFT(out_frame->width, subx) * sizeof(pixel),
> +                                AV_CEIL_RSHIFT(out_frame->height, suby));
> +        }
> +    }
> +

This is generic and can be moved out of the template.

Niklas Haas:
> +    av_assert0(out_frame->format == in_frame->format);

You already asserted that generically at the start of
ff_aom_apply_film_grain().

> +    if (in_frame->format != AV_PIX_FMT_YUV420P10LE)
> +        return AVERROR_PATCHWELCOME;

wtf! This makes the whole code pointless on BE and makes the 8 bit
version pointless, too.

- Andreas

On Thu, 29 Feb 2024 12:36:36 +0100 Andreas Rheinhardt <andreas.rheinhardt@outlook.com> wrote:
> Niklas Haas:
> > +#if BIT_DEPTH > 8
> > +# define entry int16_t
> > +# define bitdepth_max ((1 << bitdepth) - 1)
> > +# define HBD_DECL , const int bitdepth
> > +# define HBD_CALL , bitdepth
> > +# define SCALING_SIZE 4096
> > +#else
> > +# define entry int8_t
> > +# define bitdepth 8
> > +# define bitdepth_max UINT8_MAX
> 
> Why are you using signed types, but unsigned max values?

`entry` is the type of a (signed) film grain LUT entry, which is
distinct from the maximum value for the image bit-depth.

Consider the canonical usage as:

  entry grain_lut[];
  pixel = av_clip( src_pixel + grain_lut[idx], 0, bitdepth_max );

> > +# define HBD_DECL
> > +# define HBD_CALL
> > +# define SCALING_SIZE 256
> > +#endif
> > +
> > +// Symbols that do not depend on bit depth
> > +#ifndef PXSTRIDE
> > +# define PXSTRIDE(x) ((x) / sizeof(pixel))
> 
> There are several things that are wrong with this pattern:
> 1. sizeof is size_t and this has a conversion rank >= int/ptrdiff_t on
> all systems I am aware of, therefore the division will be performed with
> unsigned types, i.e. it is a logical right shift. And this will just not
> work with negative linesizes (but given that pointer arithmetic via a
> pixel* commonly involves a multiplication by sizeof(pixel), it often
> happens to work in practice, but e.g. UBSan warns about this).
> 2. It presumes that linesize is always a multiple of the sizeof of the
> underlying type. This need not be so. Imagine a system where there are
> no alignment requirements whatsoever (i.e. the required alignment of
> every type is 1 and vector instructions (if existing) also have no
> alignment requirement). Then it is legal for our callers to use frames
> where linesize can be anything.
> 3. Even if linesize is always a multiple of sizeof(pixel), the compiler
> does not know this and therefore will have to mask the low bits. So
> instead of
> src_row + (y) * PXSTRIDE(stride) + (x) + bx
> use
> (const pixel*)((const char*)src_row + (y) * stride) + (x) + bx
> 
> (Please don't cast const away on intermediate pointers when changing this.)

Thanks. The code was copied as-is from dav1d (which also supports
negative strides, but I guess they don't care about this particular UB?)

I will change it to the pattern you recommended.

> > +    // Copy over the non-modified planes
> > +    if (!data->num_y_points) {
> > +        av_image_copy_plane(out_frame->data[0], out_frame->linesize[0],
> > +                            in_frame->data[0], in_frame->linesize[0],
> > +                            out_frame->width * sizeof(pixel), out_frame->height);
> > +    }
> > +    for (int uv = 0; uv < 2; uv++) {
> > +        if (!data->num_uv_points[uv]) {
> > +            av_image_copy_plane(out_frame->data[1+uv], out_frame->linesize[1+uv],
> > +                                in_frame->data[1+uv], in_frame->linesize[1+uv],
> > +                                AV_CEIL_RSHIFT(out_frame->width, subx) * sizeof(pixel),
> > +                                AV_CEIL_RSHIFT(out_frame->height, suby));
> > +        }
> > +    }
> > +
> 
> This is generic and can be moved out of the template.

Done.

On Thu, 29 Feb 2024 12:52:58 +0100 Andreas Rheinhardt <andreas.rheinhardt@outlook.com> wrote:
> Niklas Haas:
> > +    av_assert0(out_frame->format == in_frame->format);
> 
> You already asserted that generically at the start of
> ff_aom_apply_film_grain().

Fixed.

> > +    if (in_frame->format != AV_PIX_FMT_YUV420P10LE)
> > +        return AVERROR_PATCHWELCOME;
> 
> wtf! This makes the whole code pointless on BE and makes the 8 bit
> version pointless, too.

Yikes, this was clearly left over from before I made this code generic!
Fixed.

(Doesn't help that I have no other test files for this code..)