1 // Copyright 2020 The libgav1 Authors
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 // http://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14
15 #include "src/film_grain.h"
16
17 #include <algorithm>
18 #include <cassert>
19 #include <cstddef>
20 #include <cstdint>
21 #include <cstring>
22 #include <new>
23
24 #include "src/dsp/common.h"
25 #include "src/dsp/constants.h"
26 #include "src/dsp/dsp.h"
27 #include "src/dsp/film_grain_common.h"
28 #include "src/utils/array_2d.h"
29 #include "src/utils/blocking_counter.h"
30 #include "src/utils/common.h"
31 #include "src/utils/compiler_attributes.h"
32 #include "src/utils/constants.h"
33 #include "src/utils/logging.h"
34 #include "src/utils/threadpool.h"
35
36 namespace libgav1 {
37
38 namespace {
39
40 // The kGaussianSequence array contains random samples from a Gaussian
41 // distribution with zero mean and standard deviation of about 512 clipped to
42 // the range of [-2048, 2047] (representable by a signed integer using 12 bits
43 // of precision) and rounded to the nearest multiple of 4.
44 //
45 // Note: It is important that every element in the kGaussianSequence array be
46 // less than 2040, so that RightShiftWithRounding(kGaussianSequence[i], 4) is
47 // less than 128 for bitdepth=8 (GrainType=int8_t).
48 constexpr int16_t kGaussianSequence[/*2048*/] = {
49 56, 568, -180, 172, 124, -84, 172, -64, -900, 24, 820,
50 224, 1248, 996, 272, -8, -916, -388, -732, -104, -188, 800,
51 112, -652, -320, -376, 140, -252, 492, -168, 44, -788, 588,
52 -584, 500, -228, 12, 680, 272, -476, 972, -100, 652, 368,
53 432, -196, -720, -192, 1000, -332, 652, -136, -552, -604, -4,
54 192, -220, -136, 1000, -52, 372, -96, -624, 124, -24, 396,
55 540, -12, -104, 640, 464, 244, -208, -84, 368, -528, -740,
56 248, -968, -848, 608, 376, -60, -292, -40, -156, 252, -292,
57 248, 224, -280, 400, -244, 244, -60, 76, -80, 212, 532,
58 340, 128, -36, 824, -352, -60, -264, -96, -612, 416, -704,
59 220, -204, 640, -160, 1220, -408, 900, 336, 20, -336, -96,
60 -792, 304, 48, -28, -1232, -1172, -448, 104, -292, -520, 244,
61 60, -948, 0, -708, 268, 108, 356, -548, 488, -344, -136,
62 488, -196, -224, 656, -236, -1128, 60, 4, 140, 276, -676,
63 -376, 168, -108, 464, 8, 564, 64, 240, 308, -300, -400,
64 -456, -136, 56, 120, -408, -116, 436, 504, -232, 328, 844,
65 -164, -84, 784, -168, 232, -224, 348, -376, 128, 568, 96,
66 -1244, -288, 276, 848, 832, -360, 656, 464, -384, -332, -356,
67 728, -388, 160, -192, 468, 296, 224, 140, -776, -100, 280,
68 4, 196, 44, -36, -648, 932, 16, 1428, 28, 528, 808,
69 772, 20, 268, 88, -332, -284, 124, -384, -448, 208, -228,
70 -1044, -328, 660, 380, -148, -300, 588, 240, 540, 28, 136,
71 -88, -436, 256, 296, -1000, 1400, 0, -48, 1056, -136, 264,
72 -528, -1108, 632, -484, -592, -344, 796, 124, -668, -768, 388,
73 1296, -232, -188, -200, -288, -4, 308, 100, -168, 256, -500,
74 204, -508, 648, -136, 372, -272, -120, -1004, -552, -548, -384,
75 548, -296, 428, -108, -8, -912, -324, -224, -88, -112, -220,
76 -100, 996, -796, 548, 360, -216, 180, 428, -200, -212, 148,
77 96, 148, 284, 216, -412, -320, 120, -300, -384, -604, -572,
78 -332, -8, -180, -176, 696, 116, -88, 628, 76, 44, -516,
79 240, -208, -40, 100, -592, 344, -308, -452, -228, 20, 916,
80 -1752, -136, -340, -804, 140, 40, 512, 340, 248, 184, -492,
81 896, -156, 932, -628, 328, -688, -448, -616, -752, -100, 560,
82 -1020, 180, -800, -64, 76, 576, 1068, 396, 660, 552, -108,
83 -28, 320, -628, 312, -92, -92, -472, 268, 16, 560, 516,
84 -672, -52, 492, -100, 260, 384, 284, 292, 304, -148, 88,
85 -152, 1012, 1064, -228, 164, -376, -684, 592, -392, 156, 196,
86 -524, -64, -884, 160, -176, 636, 648, 404, -396, -436, 864,
87 424, -728, 988, -604, 904, -592, 296, -224, 536, -176, -920,
88 436, -48, 1176, -884, 416, -776, -824, -884, 524, -548, -564,
89 -68, -164, -96, 692, 364, -692, -1012, -68, 260, -480, 876,
90 -1116, 452, -332, -352, 892, -1088, 1220, -676, 12, -292, 244,
91 496, 372, -32, 280, 200, 112, -440, -96, 24, -644, -184,
92 56, -432, 224, -980, 272, -260, 144, -436, 420, 356, 364,
93 -528, 76, 172, -744, -368, 404, -752, -416, 684, -688, 72,
94 540, 416, 92, 444, 480, -72, -1416, 164, -1172, -68, 24,
95 424, 264, 1040, 128, -912, -524, -356, 64, 876, -12, 4,
96 -88, 532, 272, -524, 320, 276, -508, 940, 24, -400, -120,
97 756, 60, 236, -412, 100, 376, -484, 400, -100, -740, -108,
98 -260, 328, -268, 224, -200, -416, 184, -604, -564, -20, 296,
99 60, 892, -888, 60, 164, 68, -760, 216, -296, 904, -336,
100 -28, 404, -356, -568, -208, -1480, -512, 296, 328, -360, -164,
101 -1560, -776, 1156, -428, 164, -504, -112, 120, -216, -148, -264,
102 308, 32, 64, -72, 72, 116, 176, -64, -272, 460, -536,
103 -784, -280, 348, 108, -752, -132, 524, -540, -776, 116, -296,
104 -1196, -288, -560, 1040, -472, 116, -848, -1116, 116, 636, 696,
105 284, -176, 1016, 204, -864, -648, -248, 356, 972, -584, -204,
106 264, 880, 528, -24, -184, 116, 448, -144, 828, 524, 212,
107 -212, 52, 12, 200, 268, -488, -404, -880, 824, -672, -40,
108 908, -248, 500, 716, -576, 492, -576, 16, 720, -108, 384,
109 124, 344, 280, 576, -500, 252, 104, -308, 196, -188, -8,
110 1268, 296, 1032, -1196, 436, 316, 372, -432, -200, -660, 704,
111 -224, 596, -132, 268, 32, -452, 884, 104, -1008, 424, -1348,
112 -280, 4, -1168, 368, 476, 696, 300, -8, 24, 180, -592,
113 -196, 388, 304, 500, 724, -160, 244, -84, 272, -256, -420,
114 320, 208, -144, -156, 156, 364, 452, 28, 540, 316, 220,
115 -644, -248, 464, 72, 360, 32, -388, 496, -680, -48, 208,
116 -116, -408, 60, -604, -392, 548, -840, 784, -460, 656, -544,
117 -388, -264, 908, -800, -628, -612, -568, 572, -220, 164, 288,
118 -16, -308, 308, -112, -636, -760, 280, -668, 432, 364, 240,
119 -196, 604, 340, 384, 196, 592, -44, -500, 432, -580, -132,
120 636, -76, 392, 4, -412, 540, 508, 328, -356, -36, 16,
121 -220, -64, -248, -60, 24, -192, 368, 1040, 92, -24, -1044,
122 -32, 40, 104, 148, 192, -136, -520, 56, -816, -224, 732,
123 392, 356, 212, -80, -424, -1008, -324, 588, -1496, 576, 460,
124 -816, -848, 56, -580, -92, -1372, -112, -496, 200, 364, 52,
125 -140, 48, -48, -60, 84, 72, 40, 132, -356, -268, -104,
126 -284, -404, 732, -520, 164, -304, -540, 120, 328, -76, -460,
127 756, 388, 588, 236, -436, -72, -176, -404, -316, -148, 716,
128 -604, 404, -72, -88, -888, -68, 944, 88, -220, -344, 960,
129 472, 460, -232, 704, 120, 832, -228, 692, -508, 132, -476,
130 844, -748, -364, -44, 1116, -1104, -1056, 76, 428, 552, -692,
131 60, 356, 96, -384, -188, -612, -576, 736, 508, 892, 352,
132 -1132, 504, -24, -352, 324, 332, -600, -312, 292, 508, -144,
133 -8, 484, 48, 284, -260, -240, 256, -100, -292, -204, -44,
134 472, -204, 908, -188, -1000, -256, 92, 1164, -392, 564, 356,
135 652, -28, -884, 256, 484, -192, 760, -176, 376, -524, -452,
136 -436, 860, -736, 212, 124, 504, -476, 468, 76, -472, 552,
137 -692, -944, -620, 740, -240, 400, 132, 20, 192, -196, 264,
138 -668, -1012, -60, 296, -316, -828, 76, -156, 284, -768, -448,
139 -832, 148, 248, 652, 616, 1236, 288, -328, -400, -124, 588,
140 220, 520, -696, 1032, 768, -740, -92, -272, 296, 448, -464,
141 412, -200, 392, 440, -200, 264, -152, -260, 320, 1032, 216,
142 320, -8, -64, 156, -1016, 1084, 1172, 536, 484, -432, 132,
143 372, -52, -256, 84, 116, -352, 48, 116, 304, -384, 412,
144 924, -300, 528, 628, 180, 648, 44, -980, -220, 1320, 48,
145 332, 748, 524, -268, -720, 540, -276, 564, -344, -208, -196,
146 436, 896, 88, -392, 132, 80, -964, -288, 568, 56, -48,
147 -456, 888, 8, 552, -156, -292, 948, 288, 128, -716, -292,
148 1192, -152, 876, 352, -600, -260, -812, -468, -28, -120, -32,
149 -44, 1284, 496, 192, 464, 312, -76, -516, -380, -456, -1012,
150 -48, 308, -156, 36, 492, -156, -808, 188, 1652, 68, -120,
151 -116, 316, 160, -140, 352, 808, -416, 592, 316, -480, 56,
152 528, -204, -568, 372, -232, 752, -344, 744, -4, 324, -416,
153 -600, 768, 268, -248, -88, -132, -420, -432, 80, -288, 404,
154 -316, -1216, -588, 520, -108, 92, -320, 368, -480, -216, -92,
155 1688, -300, 180, 1020, -176, 820, -68, -228, -260, 436, -904,
156 20, 40, -508, 440, -736, 312, 332, 204, 760, -372, 728,
157 96, -20, -632, -520, -560, 336, 1076, -64, -532, 776, 584,
158 192, 396, -728, -520, 276, -188, 80, -52, -612, -252, -48,
159 648, 212, -688, 228, -52, -260, 428, -412, -272, -404, 180,
160 816, -796, 48, 152, 484, -88, -216, 988, 696, 188, -528,
161 648, -116, -180, 316, 476, 12, -564, 96, 476, -252, -364,
162 -376, -392, 556, -256, -576, 260, -352, 120, -16, -136, -260,
163 -492, 72, 556, 660, 580, 616, 772, 436, 424, -32, -324,
164 -1268, 416, -324, -80, 920, 160, 228, 724, 32, -516, 64,
165 384, 68, -128, 136, 240, 248, -204, -68, 252, -932, -120,
166 -480, -628, -84, 192, 852, -404, -288, -132, 204, 100, 168,
167 -68, -196, -868, 460, 1080, 380, -80, 244, 0, 484, -888,
168 64, 184, 352, 600, 460, 164, 604, -196, 320, -64, 588,
169 -184, 228, 12, 372, 48, -848, -344, 224, 208, -200, 484,
170 128, -20, 272, -468, -840, 384, 256, -720, -520, -464, -580,
171 112, -120, 644, -356, -208, -608, -528, 704, 560, -424, 392,
172 828, 40, 84, 200, -152, 0, -144, 584, 280, -120, 80,
173 -556, -972, -196, -472, 724, 80, 168, -32, 88, 160, -688,
174 0, 160, 356, 372, -776, 740, -128, 676, -248, -480, 4,
175 -364, 96, 544, 232, -1032, 956, 236, 356, 20, -40, 300,
176 24, -676, -596, 132, 1120, -104, 532, -1096, 568, 648, 444,
177 508, 380, 188, -376, -604, 1488, 424, 24, 756, -220, -192,
178 716, 120, 920, 688, 168, 44, -460, 568, 284, 1144, 1160,
179 600, 424, 888, 656, -356, -320, 220, 316, -176, -724, -188,
180 -816, -628, -348, -228, -380, 1012, -452, -660, 736, 928, 404,
181 -696, -72, -268, -892, 128, 184, -344, -780, 360, 336, 400,
182 344, 428, 548, -112, 136, -228, -216, -820, -516, 340, 92,
183 -136, 116, -300, 376, -244, 100, -316, -520, -284, -12, 824,
184 164, -548, -180, -128, 116, -924, -828, 268, -368, -580, 620,
185 192, 160, 0, -1676, 1068, 424, -56, -360, 468, -156, 720,
186 288, -528, 556, -364, 548, -148, 504, 316, 152, -648, -620,
187 -684, -24, -376, -384, -108, -920, -1032, 768, 180, -264, -508,
188 -1268, -260, -60, 300, -240, 988, 724, -376, -576, -212, -736,
189 556, 192, 1092, -620, -880, 376, -56, -4, -216, -32, 836,
190 268, 396, 1332, 864, -600, 100, 56, -412, -92, 356, 180,
191 884, -468, -436, 292, -388, -804, -704, -840, 368, -348, 140,
192 -724, 1536, 940, 372, 112, -372, 436, -480, 1136, 296, -32,
193 -228, 132, -48, -220, 868, -1016, -60, -1044, -464, 328, 916,
194 244, 12, -736, -296, 360, 468, -376, -108, -92, 788, 368,
195 -56, 544, 400, -672, -420, 728, 16, 320, 44, -284, -380,
196 -796, 488, 132, 204, -596, -372, 88, -152, -908, -636, -572,
197 -624, -116, -692, -200, -56, 276, -88, 484, -324, 948, 864,
198 1000, -456, -184, -276, 292, -296, 156, 676, 320, 160, 908,
199 -84, -1236, -288, -116, 260, -372, -644, 732, -756, -96, 84,
200 344, -520, 348, -688, 240, -84, 216, -1044, -136, -676, -396,
201 -1500, 960, -40, 176, 168, 1516, 420, -504, -344, -364, -360,
202 1216, -940, -380, -212, 252, -660, -708, 484, -444, -152, 928,
203 -120, 1112, 476, -260, 560, -148, -344, 108, -196, 228, -288,
204 504, 560, -328, -88, 288, -1008, 460, -228, 468, -836, -196,
205 76, 388, 232, 412, -1168, -716, -644, 756, -172, -356, -504,
206 116, 432, 528, 48, 476, -168, -608, 448, 160, -532, -272,
207 28, -676, -12, 828, 980, 456, 520, 104, -104, 256, -344,
208 -4, -28, -368, -52, -524, -572, -556, -200, 768, 1124, -208,
209 -512, 176, 232, 248, -148, -888, 604, -600, -304, 804, -156,
210 -212, 488, -192, -804, -256, 368, -360, -916, -328, 228, -240,
211 -448, -472, 856, -556, -364, 572, -12, -156, -368, -340, 432,
212 252, -752, -152, 288, 268, -580, -848, -592, 108, -76, 244,
213 312, -716, 592, -80, 436, 360, 4, -248, 160, 516, 584,
214 732, 44, -468, -280, -292, -156, -588, 28, 308, 912, 24,
215 124, 156, 180, -252, 944, -924, -772, -520, -428, -624, 300,
216 -212, -1144, 32, -724, 800, -1128, -212, -1288, -848, 180, -416,
217 440, 192, -576, -792, -76, -1080, 80, -532, -352, -132, 380,
218 -820, 148, 1112, 128, 164, 456, 700, -924, 144, -668, -384,
219 648, -832, 508, 552, -52, -100, -656, 208, -568, 748, -88,
220 680, 232, 300, 192, -408, -1012, -152, -252, -268, 272, -876,
221 -664, -648, -332, -136, 16, 12, 1152, -28, 332, -536, 320,
222 -672, -460, -316, 532, -260, 228, -40, 1052, -816, 180, 88,
223 -496, -556, -672, -368, 428, 92, 356, 404, -408, 252, 196,
224 -176, -556, 792, 268, 32, 372, 40, 96, -332, 328, 120,
225 372, -900, -40, 472, -264, -592, 952, 128, 656, 112, 664,
226 -232, 420, 4, -344, -464, 556, 244, -416, -32, 252, 0,
227 -412, 188, -696, 508, -476, 324, -1096, 656, -312, 560, 264,
228 -136, 304, 160, -64, -580, 248, 336, -720, 560, -348, -288,
229 -276, -196, -500, 852, -544, -236, -1128, -992, -776, 116, 56,
230 52, 860, 884, 212, -12, 168, 1020, 512, -552, 924, -148,
231 716, 188, 164, -340, -520, -184, 880, -152, -680, -208, -1156,
232 -300, -528, -472, 364, 100, -744, -1056, -32, 540, 280, 144,
233 -676, -32, -232, -280, -224, 96, 568, -76, 172, 148, 148,
234 104, 32, -296, -32, 788, -80, 32, -16, 280, 288, 944,
235 428, -484};
236 static_assert(sizeof(kGaussianSequence) / sizeof(kGaussianSequence[0]) == 2048,
237 "");
238
239 // The number of rows in a contiguous group computed by a single worker thread
240 // before checking for the next available group.
241 constexpr int kFrameChunkHeight = 8;
242
243 // |width| and |height| refer to the plane, not the frame, meaning any
244 // subsampling should be applied by the caller.
245 template <typename Pixel>
CopyImagePlane(const uint8_t * source_plane,ptrdiff_t source_stride,int width,int height,uint8_t * dest_plane,ptrdiff_t dest_stride)246 inline void CopyImagePlane(const uint8_t* source_plane, ptrdiff_t source_stride,
247 int width, int height, uint8_t* dest_plane,
248 ptrdiff_t dest_stride) {
249 // If it's the same buffer there's nothing to do.
250 if (source_plane == dest_plane) return;
251
252 int y = 0;
253 do {
254 memcpy(dest_plane, source_plane, width * sizeof(Pixel));
255 source_plane += source_stride;
256 dest_plane += dest_stride;
257 } while (++y < height);
258 }
259
260 } // namespace
261
262 template <int bitdepth>
FilmGrain(const FilmGrainParams & params,bool is_monochrome,bool color_matrix_is_identity,int subsampling_x,int subsampling_y,int width,int height,ThreadPool * thread_pool)263 FilmGrain<bitdepth>::FilmGrain(const FilmGrainParams& params,
264 bool is_monochrome,
265 bool color_matrix_is_identity, int subsampling_x,
266 int subsampling_y, int width, int height,
267 ThreadPool* thread_pool)
268 : params_(params),
269 is_monochrome_(is_monochrome),
270 color_matrix_is_identity_(color_matrix_is_identity),
271 subsampling_x_(subsampling_x),
272 subsampling_y_(subsampling_y),
273 width_(width),
274 height_(height),
275 template_uv_width_((subsampling_x != 0) ? kMinChromaWidth
276 : kMaxChromaWidth),
277 template_uv_height_((subsampling_y != 0) ? kMinChromaHeight
278 : kMaxChromaHeight),
279 thread_pool_(thread_pool) {}
280
281 template <int bitdepth>
Init()282 bool FilmGrain<bitdepth>::Init() {
283 // Section 7.18.3.3. Generate grain process.
284 const dsp::Dsp& dsp = *dsp::GetDspTable(bitdepth);
285 // If params_.num_y_points is 0, luma_grain_ will never be read, so we don't
286 // need to generate it.
287 const bool use_luma = params_.num_y_points > 0;
288 if (use_luma) {
289 GenerateLumaGrain(params_, luma_grain_);
290 // If params_.auto_regression_coeff_lag is 0, the filter is the identity
291 // filter and therefore can be skipped.
292 if (params_.auto_regression_coeff_lag > 0) {
293 dsp.film_grain
294 .luma_auto_regression[params_.auto_regression_coeff_lag - 1](
295 params_, luma_grain_);
296 }
297 } else {
298 // Have AddressSanitizer warn if luma_grain_ is used.
299 ASAN_POISON_MEMORY_REGION(luma_grain_, sizeof(luma_grain_));
300 }
301 if (!is_monochrome_) {
302 GenerateChromaGrains(params_, template_uv_width_, template_uv_height_,
303 u_grain_, v_grain_);
304 if (params_.auto_regression_coeff_lag > 0 || use_luma) {
305 dsp.film_grain.chroma_auto_regression[static_cast<int>(
306 use_luma)][params_.auto_regression_coeff_lag](
307 params_, luma_grain_, subsampling_x_, subsampling_y_, u_grain_,
308 v_grain_);
309 }
310 }
311
312 // Section 7.18.3.4. Scaling lookup initialization process.
313
314 // Initialize scaling_lut_y_. If params_.num_y_points > 0, scaling_lut_y_
315 // is used for the Y plane. If params_.chroma_scaling_from_luma is true,
316 // scaling_lut_u_ and scaling_lut_v_ are the same as scaling_lut_y_ and are
317 // set up as aliases. So we need to initialize scaling_lut_y_ under these
318 // two conditions.
319 //
320 // Note: Although it does not seem to make sense, there are test vectors
321 // with chroma_scaling_from_luma=true and params_.num_y_points=0.
322 #if LIBGAV1_MSAN
323 // Quiet film grain / md5 msan warnings.
324 memset(scaling_lut_y_, 0, sizeof(scaling_lut_y_));
325 #endif
326 if (use_luma || params_.chroma_scaling_from_luma) {
327 dsp.film_grain.initialize_scaling_lut(
328 params_.num_y_points, params_.point_y_value, params_.point_y_scaling,
329 scaling_lut_y_, kScalingLutLength);
330 } else {
331 ASAN_POISON_MEMORY_REGION(scaling_lut_y_, sizeof(scaling_lut_y_));
332 }
333 if (!is_monochrome_) {
334 if (params_.chroma_scaling_from_luma) {
335 scaling_lut_u_ = scaling_lut_y_;
336 scaling_lut_v_ = scaling_lut_y_;
337 } else if (params_.num_u_points > 0 || params_.num_v_points > 0) {
338 const size_t buffer_size =
339 kScalingLutLength * (static_cast<int>(params_.num_u_points > 0) +
340 static_cast<int>(params_.num_v_points > 0));
341 scaling_lut_chroma_buffer_.reset(new (std::nothrow) int16_t[buffer_size]);
342 if (scaling_lut_chroma_buffer_ == nullptr) return false;
343
344 int16_t* buffer = scaling_lut_chroma_buffer_.get();
345 #if LIBGAV1_MSAN
346 // Quiet film grain / md5 msan warnings.
347 memset(buffer, 0, buffer_size * 2);
348 #endif
349 if (params_.num_u_points > 0) {
350 scaling_lut_u_ = buffer;
351 dsp.film_grain.initialize_scaling_lut(
352 params_.num_u_points, params_.point_u_value,
353 params_.point_u_scaling, scaling_lut_u_, kScalingLutLength);
354 buffer += kScalingLutLength;
355 }
356 if (params_.num_v_points > 0) {
357 scaling_lut_v_ = buffer;
358 dsp.film_grain.initialize_scaling_lut(
359 params_.num_v_points, params_.point_v_value,
360 params_.point_v_scaling, scaling_lut_v_, kScalingLutLength);
361 }
362 }
363 }
364 return true;
365 }
366
367 template <int bitdepth>
GenerateLumaGrain(const FilmGrainParams & params,GrainType * luma_grain)368 void FilmGrain<bitdepth>::GenerateLumaGrain(const FilmGrainParams& params,
369 GrainType* luma_grain) {
370 // If params.num_y_points is equal to 0, Section 7.18.3.3 specifies we set
371 // the luma_grain array to all zeros. But the Note at the end of Section
372 // 7.18.3.3 says luma_grain "will never be read in this case". So we don't
373 // call GenerateLumaGrain if params.num_y_points is equal to 0.
374 assert(params.num_y_points > 0);
375 const int shift = kBitdepth12 - bitdepth + params.grain_scale_shift;
376 uint16_t seed = params.grain_seed;
377 GrainType* luma_grain_row = luma_grain;
378 for (int y = 0; y < kLumaHeight; ++y) {
379 for (int x = 0; x < kLumaWidth; ++x) {
380 luma_grain_row[x] = RightShiftWithRounding(
381 kGaussianSequence[GetFilmGrainRandomNumber(11, &seed)], shift);
382 }
383 luma_grain_row += kLumaWidth;
384 }
385 }
386
387 template <int bitdepth>
GenerateChromaGrains(const FilmGrainParams & params,int chroma_width,int chroma_height,GrainType * u_grain,GrainType * v_grain)388 void FilmGrain<bitdepth>::GenerateChromaGrains(const FilmGrainParams& params,
389 int chroma_width,
390 int chroma_height,
391 GrainType* u_grain,
392 GrainType* v_grain) {
393 const int shift = kBitdepth12 - bitdepth + params.grain_scale_shift;
394 if (params.num_u_points == 0 && !params.chroma_scaling_from_luma) {
395 memset(u_grain, 0, chroma_height * chroma_width * sizeof(*u_grain));
396 } else {
397 uint16_t seed = params.grain_seed ^ 0xb524;
398 GrainType* u_grain_row = u_grain;
399 assert(chroma_width > 0);
400 assert(chroma_height > 0);
401 int y = 0;
402 do {
403 int x = 0;
404 do {
405 u_grain_row[x] = RightShiftWithRounding(
406 kGaussianSequence[GetFilmGrainRandomNumber(11, &seed)], shift);
407 } while (++x < chroma_width);
408
409 u_grain_row += chroma_width;
410 } while (++y < chroma_height);
411 }
412 if (params.num_v_points == 0 && !params.chroma_scaling_from_luma) {
413 memset(v_grain, 0, chroma_height * chroma_width * sizeof(*v_grain));
414 } else {
415 GrainType* v_grain_row = v_grain;
416 uint16_t seed = params.grain_seed ^ 0x49d8;
417 int y = 0;
418 do {
419 int x = 0;
420 do {
421 v_grain_row[x] = RightShiftWithRounding(
422 kGaussianSequence[GetFilmGrainRandomNumber(11, &seed)], shift);
423 } while (++x < chroma_width);
424
425 v_grain_row += chroma_width;
426 } while (++y < chroma_height);
427 }
428 }
429
430 template <int bitdepth>
AllocateNoiseStripes()431 bool FilmGrain<bitdepth>::AllocateNoiseStripes() {
432 const int half_height = DivideBy2(height_ + 1);
433 assert(half_height > 0);
434 // ceil(half_height / 16.0)
435 const int max_luma_num = DivideBy16(half_height + 15);
436 constexpr int kNoiseStripeHeight = 34;
437 size_t noise_buffer_size = kNoiseStripePadding;
438 if (params_.num_y_points > 0) {
439 noise_buffer_size += max_luma_num * kNoiseStripeHeight * width_;
440 }
441 if (!is_monochrome_) {
442 noise_buffer_size += 2 * max_luma_num *
443 (kNoiseStripeHeight >> subsampling_y_) *
444 SubsampledValue(width_, subsampling_x_);
445 }
446 noise_buffer_.reset(new (std::nothrow) GrainType[noise_buffer_size]);
447 if (noise_buffer_ == nullptr) return false;
448 GrainType* noise_buffer = noise_buffer_.get();
449 if (params_.num_y_points > 0) {
450 noise_stripes_[kPlaneY].Reset(max_luma_num, kNoiseStripeHeight * width_,
451 noise_buffer);
452 noise_buffer += max_luma_num * kNoiseStripeHeight * width_;
453 }
454 if (!is_monochrome_) {
455 noise_stripes_[kPlaneU].Reset(max_luma_num,
456 (kNoiseStripeHeight >> subsampling_y_) *
457 SubsampledValue(width_, subsampling_x_),
458 noise_buffer);
459 noise_buffer += max_luma_num * (kNoiseStripeHeight >> subsampling_y_) *
460 SubsampledValue(width_, subsampling_x_);
461 noise_stripes_[kPlaneV].Reset(max_luma_num,
462 (kNoiseStripeHeight >> subsampling_y_) *
463 SubsampledValue(width_, subsampling_x_),
464 noise_buffer);
465 }
466 return true;
467 }
468
469 template <int bitdepth>
AllocateNoiseImage()470 bool FilmGrain<bitdepth>::AllocateNoiseImage() {
471 // When LIBGAV1_MSAN is enabled, zero initialize to quiet optimized film grain
472 // msan warnings.
473 constexpr bool zero_initialize = LIBGAV1_MSAN == 1;
474 if (params_.num_y_points > 0 &&
475 !noise_image_[kPlaneY].Reset(height_, width_ + kNoiseImagePadding,
476 zero_initialize)) {
477 return false;
478 }
479 if (!is_monochrome_) {
480 if (!noise_image_[kPlaneU].Reset(
481 (height_ + subsampling_y_) >> subsampling_y_,
482 ((width_ + subsampling_x_) >> subsampling_x_) + kNoiseImagePadding,
483 zero_initialize)) {
484 return false;
485 }
486 if (!noise_image_[kPlaneV].Reset(
487 (height_ + subsampling_y_) >> subsampling_y_,
488 ((width_ + subsampling_x_) >> subsampling_x_) + kNoiseImagePadding,
489 zero_initialize)) {
490 return false;
491 }
492 }
493 return true;
494 }
495
496 // Uses |overlap_flag| to skip rows that are covered by the overlap computation.
497 template <int bitdepth>
ConstructNoiseImage(const Array2DView<GrainType> * noise_stripes,int width,int height,int subsampling_x,int subsampling_y,int stripe_start_offset,Array2D<GrainType> * noise_image)498 void FilmGrain<bitdepth>::ConstructNoiseImage(
499 const Array2DView<GrainType>* noise_stripes, int width, int height,
500 int subsampling_x, int subsampling_y, int stripe_start_offset,
501 Array2D<GrainType>* noise_image) {
502 const int plane_width = (width + subsampling_x) >> subsampling_x;
503 const int plane_height = (height + subsampling_y) >> subsampling_y;
504 const int stripe_height = 32 >> subsampling_y;
505 const int stripe_mask = stripe_height - 1;
506 int y = 0;
507 // |luma_num| = y >> (5 - |subsampling_y|). Hence |luma_num| == 0 for all y up
508 // to either 16 or 32.
509 const GrainType* first_noise_stripe = (*noise_stripes)[0];
510 do {
511 memcpy((*noise_image)[y], first_noise_stripe + y * plane_width,
512 plane_width * sizeof(first_noise_stripe[0]));
513 } while (++y < std::min(stripe_height, plane_height));
514 // End special iterations for luma_num == 0.
515
516 int luma_num = 1;
517 for (; y < (plane_height & ~stripe_mask); ++luma_num, y += stripe_height) {
518 const GrainType* noise_stripe = (*noise_stripes)[luma_num];
519 int i = stripe_start_offset;
520 do {
521 memcpy((*noise_image)[y + i], noise_stripe + i * plane_width,
522 plane_width * sizeof(noise_stripe[0]));
523 } while (++i < stripe_height);
524 }
525
526 // If there is a partial stripe, copy any rows beyond the overlap rows.
527 const int remaining_height = plane_height - y;
528 if (remaining_height > stripe_start_offset) {
529 assert(luma_num < noise_stripes->rows());
530 const GrainType* noise_stripe = (*noise_stripes)[luma_num];
531 int i = stripe_start_offset;
532 do {
533 memcpy((*noise_image)[y + i], noise_stripe + i * plane_width,
534 plane_width * sizeof(noise_stripe[0]));
535 } while (++i < remaining_height);
536 }
537 }
538
539 template <int bitdepth>
BlendNoiseChromaWorker(const dsp::Dsp & dsp,const Plane * planes,int num_planes,std::atomic<int> * job_counter,int min_value,int max_chroma,const uint8_t * source_plane_y,ptrdiff_t source_stride_y,const uint8_t * source_plane_u,const uint8_t * source_plane_v,ptrdiff_t source_stride_uv,uint8_t * dest_plane_u,uint8_t * dest_plane_v,ptrdiff_t dest_stride_uv)540 void FilmGrain<bitdepth>::BlendNoiseChromaWorker(
541 const dsp::Dsp& dsp, const Plane* planes, int num_planes,
542 std::atomic<int>* job_counter, int min_value, int max_chroma,
543 const uint8_t* source_plane_y, ptrdiff_t source_stride_y,
544 const uint8_t* source_plane_u, const uint8_t* source_plane_v,
545 ptrdiff_t source_stride_uv, uint8_t* dest_plane_u, uint8_t* dest_plane_v,
546 ptrdiff_t dest_stride_uv) {
547 assert(num_planes > 0);
548 const int full_jobs_per_plane = height_ / kFrameChunkHeight;
549 const int remainder_job_height = height_ & (kFrameChunkHeight - 1);
550 const int total_full_jobs = full_jobs_per_plane * num_planes;
551 // If the frame height is not a multiple of kFrameChunkHeight, one job with
552 // a smaller number of rows is necessary at the end of each plane.
553 const int total_jobs =
554 total_full_jobs + ((remainder_job_height == 0) ? 0 : num_planes);
555 int job_index;
556 // Each job corresponds to a slice of kFrameChunkHeight rows in the luma
557 // plane. dsp->blend_noise_chroma handles subsampling.
558 // This loop body handles a slice of one plane or the other, depending on
559 // which are active. That way, threads working on consecutive jobs will keep
560 // the same region of luma source in working memory.
561 while ((job_index = job_counter->fetch_add(1, std::memory_order_relaxed)) <
562 total_jobs) {
563 const Plane plane = planes[job_index % num_planes];
564 const int slice_index = job_index / num_planes;
565 const int start_height = slice_index * kFrameChunkHeight;
566 const int job_height = std::min(height_ - start_height, kFrameChunkHeight);
567
568 const auto* source_cursor_y = reinterpret_cast<const Pixel*>(
569 source_plane_y + start_height * source_stride_y);
570 const int16_t* scaling_lut_uv;
571 const uint8_t* source_plane_uv;
572 uint8_t* dest_plane_uv;
573
574 if (plane == kPlaneU) {
575 scaling_lut_uv = scaling_lut_u_;
576 source_plane_uv = source_plane_u;
577 dest_plane_uv = dest_plane_u;
578 } else {
579 assert(plane == kPlaneV);
580 scaling_lut_uv = scaling_lut_v_;
581 source_plane_uv = source_plane_v;
582 dest_plane_uv = dest_plane_v;
583 }
584 const auto* source_cursor_uv = reinterpret_cast<const Pixel*>(
585 source_plane_uv + (start_height >> subsampling_y_) * source_stride_uv);
586 auto* dest_cursor_uv = reinterpret_cast<Pixel*>(
587 dest_plane_uv + (start_height >> subsampling_y_) * dest_stride_uv);
588 dsp.film_grain.blend_noise_chroma[params_.chroma_scaling_from_luma](
589 plane, params_, noise_image_, min_value, max_chroma, width_, job_height,
590 start_height, subsampling_x_, subsampling_y_, scaling_lut_uv,
591 source_cursor_y, source_stride_y, source_cursor_uv, source_stride_uv,
592 dest_cursor_uv, dest_stride_uv);
593 }
594 }
595
596 template <int bitdepth>
BlendNoiseLumaWorker(const dsp::Dsp & dsp,std::atomic<int> * job_counter,int min_value,int max_luma,const uint8_t * source_plane_y,ptrdiff_t source_stride_y,uint8_t * dest_plane_y,ptrdiff_t dest_stride_y)597 void FilmGrain<bitdepth>::BlendNoiseLumaWorker(
598 const dsp::Dsp& dsp, std::atomic<int>* job_counter, int min_value,
599 int max_luma, const uint8_t* source_plane_y, ptrdiff_t source_stride_y,
600 uint8_t* dest_plane_y, ptrdiff_t dest_stride_y) {
601 const int total_full_jobs = height_ / kFrameChunkHeight;
602 const int remainder_job_height = height_ & (kFrameChunkHeight - 1);
603 const int total_jobs =
604 total_full_jobs + static_cast<int>(remainder_job_height > 0);
605 int job_index;
606 // Each job is some number of rows in a plane.
607 while ((job_index = job_counter->fetch_add(1, std::memory_order_relaxed)) <
608 total_jobs) {
609 const int start_height = job_index * kFrameChunkHeight;
610 const int job_height = std::min(height_ - start_height, kFrameChunkHeight);
611
612 const auto* source_cursor_y = reinterpret_cast<const Pixel*>(
613 source_plane_y + start_height * source_stride_y);
614 auto* dest_cursor_y =
615 reinterpret_cast<Pixel*>(dest_plane_y + start_height * dest_stride_y);
616 dsp.film_grain.blend_noise_luma(
617 noise_image_, min_value, max_luma, params_.chroma_scaling, width_,
618 job_height, start_height, scaling_lut_y_, source_cursor_y,
619 source_stride_y, dest_cursor_y, dest_stride_y);
620 }
621 }
622
623 template <int bitdepth>
AddNoise(const uint8_t * source_plane_y,ptrdiff_t source_stride_y,const uint8_t * source_plane_u,const uint8_t * source_plane_v,ptrdiff_t source_stride_uv,uint8_t * dest_plane_y,ptrdiff_t dest_stride_y,uint8_t * dest_plane_u,uint8_t * dest_plane_v,ptrdiff_t dest_stride_uv)624 bool FilmGrain<bitdepth>::AddNoise(
625 const uint8_t* source_plane_y, ptrdiff_t source_stride_y,
626 const uint8_t* source_plane_u, const uint8_t* source_plane_v,
627 ptrdiff_t source_stride_uv, uint8_t* dest_plane_y, ptrdiff_t dest_stride_y,
628 uint8_t* dest_plane_u, uint8_t* dest_plane_v, ptrdiff_t dest_stride_uv) {
629 if (!Init()) {
630 LIBGAV1_DLOG(ERROR, "Init() failed.");
631 return false;
632 }
633 if (!AllocateNoiseStripes()) {
634 LIBGAV1_DLOG(ERROR, "AllocateNoiseStripes() failed.");
635 return false;
636 }
637
638 const dsp::Dsp& dsp = *dsp::GetDspTable(bitdepth);
639 const bool use_luma = params_.num_y_points > 0;
640
641 // Construct noise stripes.
642 if (use_luma) {
643 // The luma plane is never subsampled.
644 dsp.film_grain
645 .construct_noise_stripes[static_cast<int>(params_.overlap_flag)](
646 luma_grain_, params_.grain_seed, width_, height_,
647 /*subsampling_x=*/0, /*subsampling_y=*/0, &noise_stripes_[kPlaneY]);
648 }
649 if (!is_monochrome_) {
650 dsp.film_grain
651 .construct_noise_stripes[static_cast<int>(params_.overlap_flag)](
652 u_grain_, params_.grain_seed, width_, height_, subsampling_x_,
653 subsampling_y_, &noise_stripes_[kPlaneU]);
654 dsp.film_grain
655 .construct_noise_stripes[static_cast<int>(params_.overlap_flag)](
656 v_grain_, params_.grain_seed, width_, height_, subsampling_x_,
657 subsampling_y_, &noise_stripes_[kPlaneV]);
658 }
659
660 if (!AllocateNoiseImage()) {
661 LIBGAV1_DLOG(ERROR, "AllocateNoiseImage() failed.");
662 return false;
663 }
664
665 // Construct noise image.
666 if (use_luma) {
667 ConstructNoiseImage(
668 &noise_stripes_[kPlaneY], width_, height_, /*subsampling_x=*/0,
669 /*subsampling_y=*/0, static_cast<int>(params_.overlap_flag) << 1,
670 &noise_image_[kPlaneY]);
671 if (params_.overlap_flag) {
672 dsp.film_grain.construct_noise_image_overlap(
673 &noise_stripes_[kPlaneY], width_, height_, /*subsampling_x=*/0,
674 /*subsampling_y=*/0, &noise_image_[kPlaneY]);
675 }
676 }
677 if (!is_monochrome_) {
678 ConstructNoiseImage(&noise_stripes_[kPlaneU], width_, height_,
679 subsampling_x_, subsampling_y_,
680 static_cast<int>(params_.overlap_flag)
681 << (1 - subsampling_y_),
682 &noise_image_[kPlaneU]);
683 ConstructNoiseImage(&noise_stripes_[kPlaneV], width_, height_,
684 subsampling_x_, subsampling_y_,
685 static_cast<int>(params_.overlap_flag)
686 << (1 - subsampling_y_),
687 &noise_image_[kPlaneV]);
688 if (params_.overlap_flag) {
689 dsp.film_grain.construct_noise_image_overlap(
690 &noise_stripes_[kPlaneU], width_, height_, subsampling_x_,
691 subsampling_y_, &noise_image_[kPlaneU]);
692 dsp.film_grain.construct_noise_image_overlap(
693 &noise_stripes_[kPlaneV], width_, height_, subsampling_x_,
694 subsampling_y_, &noise_image_[kPlaneV]);
695 }
696 }
697
698 // Blend noise image.
699 int min_value;
700 int max_luma;
701 int max_chroma;
702 if (params_.clip_to_restricted_range) {
703 min_value = 16 << (bitdepth - kBitdepth8);
704 max_luma = 235 << (bitdepth - kBitdepth8);
705 if (color_matrix_is_identity_) {
706 max_chroma = max_luma;
707 } else {
708 max_chroma = 240 << (bitdepth - kBitdepth8);
709 }
710 } else {
711 min_value = 0;
712 max_luma = (256 << (bitdepth - kBitdepth8)) - 1;
713 max_chroma = max_luma;
714 }
715
716 // Handle all chroma planes first because luma source may be altered in place.
717 if (!is_monochrome_) {
718 // This is done in a strange way but Vector can't be passed by copy to the
719 // lambda capture that spawns the thread.
720 Plane planes_to_blend[2];
721 int num_planes = 0;
722 if (params_.chroma_scaling_from_luma) {
723 // Both noise planes are computed from the luma scaling lookup table.
724 planes_to_blend[num_planes++] = kPlaneU;
725 planes_to_blend[num_planes++] = kPlaneV;
726 } else {
727 const int height_uv = SubsampledValue(height_, subsampling_y_);
728 const int width_uv = SubsampledValue(width_, subsampling_x_);
729
730 // Noise is applied according to a lookup table defined by pieceiwse
731 // linear "points." If the lookup table is empty, that corresponds to
732 // outputting zero noise.
733 if (params_.num_u_points == 0) {
734 CopyImagePlane<Pixel>(source_plane_u, source_stride_uv, width_uv,
735 height_uv, dest_plane_u, dest_stride_uv);
736 } else {
737 planes_to_blend[num_planes++] = kPlaneU;
738 }
739 if (params_.num_v_points == 0) {
740 CopyImagePlane<Pixel>(source_plane_v, source_stride_uv, width_uv,
741 height_uv, dest_plane_v, dest_stride_uv);
742 } else {
743 planes_to_blend[num_planes++] = kPlaneV;
744 }
745 }
746 if (thread_pool_ != nullptr && num_planes > 0) {
747 const int num_workers = thread_pool_->num_threads();
748 BlockingCounter pending_workers(num_workers);
749 std::atomic<int> job_counter(0);
750 for (int i = 0; i < num_workers; ++i) {
751 thread_pool_->Schedule([this, dsp, &pending_workers, &planes_to_blend,
752 num_planes, &job_counter, min_value, max_chroma,
753 source_plane_y, source_stride_y, source_plane_u,
754 source_plane_v, source_stride_uv, dest_plane_u,
755 dest_plane_v, dest_stride_uv]() {
756 BlendNoiseChromaWorker(dsp, planes_to_blend, num_planes, &job_counter,
757 min_value, max_chroma, source_plane_y,
758 source_stride_y, source_plane_u,
759 source_plane_v, source_stride_uv, dest_plane_u,
760 dest_plane_v, dest_stride_uv);
761 pending_workers.Decrement();
762 });
763 }
764 BlendNoiseChromaWorker(
765 dsp, planes_to_blend, num_planes, &job_counter, min_value, max_chroma,
766 source_plane_y, source_stride_y, source_plane_u, source_plane_v,
767 source_stride_uv, dest_plane_u, dest_plane_v, dest_stride_uv);
768
769 pending_workers.Wait();
770 } else {
771 // Single threaded.
772 if (params_.num_u_points > 0 || params_.chroma_scaling_from_luma) {
773 dsp.film_grain.blend_noise_chroma[params_.chroma_scaling_from_luma](
774 kPlaneU, params_, noise_image_, min_value, max_chroma, width_,
775 height_, /*start_height=*/0, subsampling_x_, subsampling_y_,
776 scaling_lut_u_, source_plane_y, source_stride_y, source_plane_u,
777 source_stride_uv, dest_plane_u, dest_stride_uv);
778 }
779 if (params_.num_v_points > 0 || params_.chroma_scaling_from_luma) {
780 dsp.film_grain.blend_noise_chroma[params_.chroma_scaling_from_luma](
781 kPlaneV, params_, noise_image_, min_value, max_chroma, width_,
782 height_, /*start_height=*/0, subsampling_x_, subsampling_y_,
783 scaling_lut_v_, source_plane_y, source_stride_y, source_plane_v,
784 source_stride_uv, dest_plane_v, dest_stride_uv);
785 }
786 }
787 }
788 if (use_luma) {
789 if (thread_pool_ != nullptr) {
790 const int num_workers = thread_pool_->num_threads();
791 BlockingCounter pending_workers(num_workers);
792 std::atomic<int> job_counter(0);
793 for (int i = 0; i < num_workers; ++i) {
794 thread_pool_->Schedule(
795 [this, dsp, &pending_workers, &job_counter, min_value, max_luma,
796 source_plane_y, source_stride_y, dest_plane_y, dest_stride_y]() {
797 BlendNoiseLumaWorker(dsp, &job_counter, min_value, max_luma,
798 source_plane_y, source_stride_y,
799 dest_plane_y, dest_stride_y);
800 pending_workers.Decrement();
801 });
802 }
803
804 BlendNoiseLumaWorker(dsp, &job_counter, min_value, max_luma,
805 source_plane_y, source_stride_y, dest_plane_y,
806 dest_stride_y);
807 pending_workers.Wait();
808 } else {
809 dsp.film_grain.blend_noise_luma(
810 noise_image_, min_value, max_luma, params_.chroma_scaling, width_,
811 height_, /*start_height=*/0, scaling_lut_y_, source_plane_y,
812 source_stride_y, dest_plane_y, dest_stride_y);
813 }
814 } else {
815 CopyImagePlane<Pixel>(source_plane_y, source_stride_y, width_, height_,
816 dest_plane_y, dest_stride_y);
817 }
818
819 return true;
820 }
821
822 // Explicit instantiations.
823 template class FilmGrain<kBitdepth8>;
824 #if LIBGAV1_MAX_BITDEPTH >= 10
825 template class FilmGrain<kBitdepth10>;
826 #endif
827 #if LIBGAV1_MAX_BITDEPTH == 12
828 template class FilmGrain<kBitdepth12>;
829 #endif
830
831 } // namespace libgav1
832