/src/aom/av1/common/av1_txfm.c
Line | Count | Source |
1 | | /* |
2 | | * Copyright (c) 2017, Alliance for Open Media. All rights reserved. |
3 | | * |
4 | | * This source code is subject to the terms of the BSD 2 Clause License and |
5 | | * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License |
6 | | * was not distributed with this source code in the LICENSE file, you can |
7 | | * obtain it at www.aomedia.org/license/software. If the Alliance for Open |
8 | | * Media Patent License 1.0 was not distributed with this source code in the |
9 | | * PATENTS file, you can obtain it at www.aomedia.org/license/patent. |
10 | | */ |
11 | | |
12 | | #include <inttypes.h> |
13 | | |
14 | | #include "config/aom_dsp_rtcd.h" |
15 | | #include "config/av1_rtcd.h" |
16 | | |
17 | | #include "av1/common/av1_txfm.h" |
18 | | |
19 | | // av1_cospi_arr[i][j] = (int)round(cos(PI*j/128) * (1<<(cos_bit_min+i))); |
20 | | const int32_t av1_cospi_arr_data[4][64] = { |
21 | | { 1024, 1024, 1023, 1021, 1019, 1016, 1013, 1009, 1004, 999, 993, 987, 980, |
22 | | 972, 964, 955, 946, 936, 926, 915, 903, 891, 878, 865, 851, 837, |
23 | | 822, 807, 792, 775, 759, 742, 724, 706, 688, 669, 650, 630, 610, |
24 | | 590, 569, 548, 526, 505, 483, 460, 438, 415, 392, 369, 345, 321, |
25 | | 297, 273, 249, 224, 200, 175, 150, 125, 100, 75, 50, 25 }, |
26 | | { 2048, 2047, 2046, 2042, 2038, 2033, 2026, 2018, 2009, 1998, 1987, |
27 | | 1974, 1960, 1945, 1928, 1911, 1892, 1872, 1851, 1829, 1806, 1782, |
28 | | 1757, 1730, 1703, 1674, 1645, 1615, 1583, 1551, 1517, 1483, 1448, |
29 | | 1412, 1375, 1338, 1299, 1260, 1220, 1179, 1138, 1096, 1053, 1009, |
30 | | 965, 921, 876, 830, 784, 737, 690, 642, 595, 546, 498, |
31 | | 449, 400, 350, 301, 251, 201, 151, 100, 50 }, |
32 | | { 4096, 4095, 4091, 4085, 4076, 4065, 4052, 4036, 4017, 3996, 3973, |
33 | | 3948, 3920, 3889, 3857, 3822, 3784, 3745, 3703, 3659, 3612, 3564, |
34 | | 3513, 3461, 3406, 3349, 3290, 3229, 3166, 3102, 3035, 2967, 2896, |
35 | | 2824, 2751, 2675, 2598, 2520, 2440, 2359, 2276, 2191, 2106, 2019, |
36 | | 1931, 1842, 1751, 1660, 1567, 1474, 1380, 1285, 1189, 1092, 995, |
37 | | 897, 799, 700, 601, 501, 401, 301, 201, 101 }, |
38 | | { 8192, 8190, 8182, 8170, 8153, 8130, 8103, 8071, 8035, 7993, 7946, |
39 | | 7895, 7839, 7779, 7713, 7643, 7568, 7489, 7405, 7317, 7225, 7128, |
40 | | 7027, 6921, 6811, 6698, 6580, 6458, 6333, 6203, 6070, 5933, 5793, |
41 | | 5649, 5501, 5351, 5197, 5040, 4880, 4717, 4551, 4383, 4212, 4038, |
42 | | 3862, 3683, 3503, 3320, 3135, 2948, 2760, 2570, 2378, 2185, 1990, |
43 | | 1795, 1598, 1401, 1202, 1003, 803, 603, 402, 201 } |
44 | | }; |
45 | | |
46 | | // av1_sinpi_arr_data[i][j] = (int)round((sqrt(2) * sin(j*Pi/9) * 2 / 3) * (1 |
47 | | // << (cos_bit_min + i))) modified so that elements j=1,2 sum to element j=4. |
48 | | const int32_t av1_sinpi_arr_data[4][5] = { { 0, 330, 621, 836, 951 }, |
49 | | { 0, 660, 1241, 1672, 1901 }, |
50 | | { 0, 1321, 2482, 3344, 3803 }, |
51 | | { 0, 2642, 4964, 6689, 7606 } }; |
52 | | |
53 | | // The reduced bit-width arrays are only used in the Arm Neon implementations |
54 | | // in av1_fwd_txfm2d_neon.c for now. |
55 | | #if HAVE_NEON |
56 | | // Constants are stored in groups of four, where symmetrical constants in the |
57 | | // cospi array are stored adjacent in memory, followed immediately by the same |
58 | | // constants but negated, i.e.: |
59 | | // f(i,j) = (int)round(cos(PI*j/128) * (1<<(cos_bit_min+i))) << (3-i) |
60 | | // and then in memory we store 4-tuples of constants together as: |
61 | | // f4(i,j) = [ f(i,j), f(i,64-j), -f(i,j), -f(i,64-j) ] |
62 | | // |
63 | | // Constants are stored in Q2.13 format, see: |
64 | | // https://en.wikipedia.org/wiki/Q_(number_format) |
65 | | // |
66 | | // The order of the constants is such that increasing subdivisions of 64 store |
67 | | // f4 tuples contiguously: |
68 | | // av1_cospi_arr_q13_data[i] = { |
69 | | // f4(i,32), // f(i,32) twice |
70 | | // f4(i,16), // f(i,16) and f(i,48), f4(i,32) skipped since present above. |
71 | | // f4(i,8), f(i,24), // f4(i,16) and f4(i,32) skipped since present above. |
72 | | // f4(i,4), f(i,12), f4(i,20), f4(i,28), |
73 | | // f4(i,2), f4(i,6), f4(i,10), f4(i,14), f4(i,18), ... |
74 | | // f4(i,1), f4(i,3), f4(i,5), f4(i,7), f4(i,9), f4(i,11), ... |
75 | | // } |
76 | | const int16_t av1_cospi_arr_q13_data[4][128] = { |
77 | | { |
78 | | 5792, 5792, -5792, -5792, 7568, 3136, -7568, -3136, 8032, 1600, |
79 | | -8032, -1600, 6808, 4552, -6808, -4552, 8152, 800, -8152, -800, |
80 | | 7840, 2376, -7840, -2376, 7224, 3864, -7224, -3864, 6336, 5200, |
81 | | -6336, -5200, 8184, 400, -8184, -400, 8104, 1200, -8104, -1200, |
82 | | 7944, 1992, -7944, -1992, 7712, 2760, -7712, -2760, 7408, 3504, |
83 | | -7408, -3504, 7024, 4208, -7024, -4208, 6576, 4880, -6576, -4880, |
84 | | 6072, 5504, -6072, -5504, 8192, 200, -8192, -200, 8168, 600, |
85 | | -8168, -600, 8128, 1000, -8128, -1000, 8072, 1400, -8072, -1400, |
86 | | 7992, 1792, -7992, -1792, 7896, 2184, -7896, -2184, 7776, 2568, |
87 | | -7776, -2568, 7640, 2952, -7640, -2952, 7488, 3320, -7488, -3320, |
88 | | 7320, 3680, -7320, -3680, 7128, 4040, -7128, -4040, 6920, 4384, |
89 | | -6920, -4384, 6696, 4720, -6696, -4720, 6456, 5040, -6456, -5040, |
90 | | 6200, 5352, -6200, -5352, 5936, 5648, -5936, -5648, |
91 | | }, |
92 | | { |
93 | | 5792, 5792, -5792, -5792, 7568, 3136, -7568, -3136, 8036, 1600, |
94 | | -8036, -1600, 6812, 4552, -6812, -4552, 8152, 804, -8152, -804, |
95 | | 7840, 2380, -7840, -2380, 7224, 3860, -7224, -3860, 6332, 5196, |
96 | | -6332, -5196, 8184, 400, -8184, -400, 8104, 1204, -8104, -1204, |
97 | | 7948, 1992, -7948, -1992, 7712, 2760, -7712, -2760, 7404, 3504, |
98 | | -7404, -3504, 7028, 4212, -7028, -4212, 6580, 4880, -6580, -4880, |
99 | | 6068, 5500, -6068, -5500, 8188, 200, -8188, -200, 8168, 604, |
100 | | -8168, -604, 8132, 1004, -8132, -1004, 8072, 1400, -8072, -1400, |
101 | | 7992, 1796, -7992, -1796, 7896, 2184, -7896, -2184, 7780, 2568, |
102 | | -7780, -2568, 7644, 2948, -7644, -2948, 7488, 3320, -7488, -3320, |
103 | | 7316, 3684, -7316, -3684, 7128, 4036, -7128, -4036, 6920, 4384, |
104 | | -6920, -4384, 6696, 4716, -6696, -4716, 6460, 5040, -6460, -5040, |
105 | | 6204, 5352, -6204, -5352, 5932, 5648, -5932, -5648, |
106 | | }, |
107 | | { |
108 | | 5792, 5792, -5792, -5792, 7568, 3134, -7568, -3134, 8034, 1598, |
109 | | -8034, -1598, 6812, 4552, -6812, -4552, 8152, 802, -8152, -802, |
110 | | 7840, 2378, -7840, -2378, 7224, 3862, -7224, -3862, 6332, 5196, |
111 | | -6332, -5196, 8182, 402, -8182, -402, 8104, 1202, -8104, -1202, |
112 | | 7946, 1990, -7946, -1990, 7714, 2760, -7714, -2760, 7406, 3502, |
113 | | -7406, -3502, 7026, 4212, -7026, -4212, 6580, 4880, -6580, -4880, |
114 | | 6070, 5502, -6070, -5502, 8190, 202, -8190, -202, 8170, 602, |
115 | | -8170, -602, 8130, 1002, -8130, -1002, 8072, 1400, -8072, -1400, |
116 | | 7992, 1794, -7992, -1794, 7896, 2184, -7896, -2184, 7778, 2570, |
117 | | -7778, -2570, 7644, 2948, -7644, -2948, 7490, 3320, -7490, -3320, |
118 | | 7318, 3684, -7318, -3684, 7128, 4038, -7128, -4038, 6922, 4382, |
119 | | -6922, -4382, 6698, 4718, -6698, -4718, 6458, 5040, -6458, -5040, |
120 | | 6204, 5350, -6204, -5350, 5934, 5648, -5934, -5648, |
121 | | }, |
122 | | { |
123 | | 5793, 5793, -5793, -5793, 7568, 3135, -7568, -3135, 8035, 1598, |
124 | | -8035, -1598, 6811, 4551, -6811, -4551, 8153, 803, -8153, -803, |
125 | | 7839, 2378, -7839, -2378, 7225, 3862, -7225, -3862, 6333, 5197, |
126 | | -6333, -5197, 8182, 402, -8182, -402, 8103, 1202, -8103, -1202, |
127 | | 7946, 1990, -7946, -1990, 7713, 2760, -7713, -2760, 7405, 3503, |
128 | | -7405, -3503, 7027, 4212, -7027, -4212, 6580, 4880, -6580, -4880, |
129 | | 6070, 5501, -6070, -5501, 8190, 201, -8190, -201, 8170, 603, |
130 | | -8170, -603, 8130, 1003, -8130, -1003, 8071, 1401, -8071, -1401, |
131 | | 7993, 1795, -7993, -1795, 7895, 2185, -7895, -2185, 7779, 2570, |
132 | | -7779, -2570, 7643, 2948, -7643, -2948, 7489, 3320, -7489, -3320, |
133 | | 7317, 3683, -7317, -3683, 7128, 4038, -7128, -4038, 6921, 4383, |
134 | | -6921, -4383, 6698, 4717, -6698, -4717, 6458, 5040, -6458, -5040, |
135 | | 6203, 5351, -6203, -5351, 5933, 5649, -5933, -5649, |
136 | | } |
137 | | }; |
138 | | |
139 | | // av1_sinpi_arr_q13_data[i][j] = |
140 | | // round((sqrt2 * sin((j+1)*Pi/9) * 2/3) * (1 << (cos_bit_min + i))) << (3-i) |
141 | | // modified so that elements j=0,1 sum to element j=3. |
142 | | // See also: https://en.wikipedia.org/wiki/Q_(number_format) |
143 | | const int16_t av1_sinpi_arr_q13_data[4][4] = { { 2640, 4968, 6688, 7608 }, |
144 | | { 2640, 4964, 6688, 7604 }, |
145 | | { 2642, 4964, 6688, 7606 }, |
146 | | { 2642, 4964, 6689, 7606 } }; |
147 | | |
148 | | // Constants are stored in pairs, where symmetrical constants in the |
149 | | // cospi array are stored adjacent in memory, i.e.: |
150 | | // f(i,j) = (int)round(cos(PI*j/128) * (1<<(cos_bit_min+i))) |
151 | | // and then in memory we store 4-tuples of constants together as: |
152 | | // f2(i,j) = [ f(i,j), f(i,64-j) ] |
153 | | const int32_t av1_cospi_arr_s32_data[4][66] = { |
154 | | { |
155 | | 1024, 0, 1024, 25, 1023, 50, 1021, 75, 1019, 100, 1016, |
156 | | 125, 1013, 150, 1009, 175, 1004, 200, 999, 224, 993, 249, |
157 | | 987, 273, 980, 297, 972, 321, 964, 345, 955, 369, 946, |
158 | | 392, 936, 415, 926, 438, 915, 460, 903, 483, 891, 505, |
159 | | 878, 526, 865, 548, 851, 569, 837, 590, 822, 610, 807, |
160 | | 630, 792, 650, 775, 669, 759, 688, 742, 706, 724, 724, |
161 | | }, |
162 | | { |
163 | | 2048, 0, 2047, 50, 2046, 100, 2042, 151, 2038, 201, 2033, |
164 | | 251, 2026, 301, 2018, 350, 2009, 400, 1998, 449, 1987, 498, |
165 | | 1974, 546, 1960, 595, 1945, 642, 1928, 690, 1911, 737, 1892, |
166 | | 784, 1872, 830, 1851, 876, 1829, 921, 1806, 965, 1782, 1009, |
167 | | 1757, 1053, 1730, 1096, 1703, 1138, 1674, 1179, 1645, 1220, 1615, |
168 | | 1260, 1583, 1299, 1551, 1338, 1517, 1375, 1483, 1412, 1448, 1448, |
169 | | }, |
170 | | { |
171 | | 4096, 0, 4095, 101, 4091, 201, 4085, 301, 4076, 401, 4065, |
172 | | 501, 4052, 601, 4036, 700, 4017, 799, 3996, 897, 3973, 995, |
173 | | 3948, 1092, 3920, 1189, 3889, 1285, 3857, 1380, 3822, 1474, 3784, |
174 | | 1567, 3745, 1660, 3703, 1751, 3659, 1842, 3612, 1931, 3564, 2019, |
175 | | 3513, 2106, 3461, 2191, 3406, 2276, 3349, 2359, 3290, 2440, 3229, |
176 | | 2520, 3166, 2598, 3102, 2675, 3035, 2751, 2967, 2824, 2896, 2896, |
177 | | }, |
178 | | { |
179 | | 8192, 0, 8190, 201, 8182, 402, 8170, 603, 8153, 803, 8130, |
180 | | 1003, 8103, 1202, 8071, 1401, 8035, 1598, 7993, 1795, 7946, 1990, |
181 | | 7895, 2185, 7839, 2378, 7779, 2570, 7713, 2760, 7643, 2948, 7568, |
182 | | 3135, 7489, 3320, 7405, 3503, 7317, 3683, 7225, 3862, 7128, 4038, |
183 | | 7027, 4212, 6921, 4383, 6811, 4551, 6698, 4717, 6580, 4880, 6458, |
184 | | 5040, 6333, 5197, 6203, 5351, 6070, 5501, 5933, 5649, 5793, 5793, |
185 | | } |
186 | | }; |
187 | | |
188 | | #endif // HAVE_NEON |
189 | | |
190 | 0 | void av1_round_shift_array_c(int32_t *arr, int size, int bit) { |
191 | 0 | int i; |
192 | 0 | if (bit == 0) { |
193 | 0 | return; |
194 | 0 | } else { |
195 | 0 | if (bit > 0) { |
196 | 0 | for (i = 0; i < size; i++) { |
197 | 0 | arr[i] = round_shift(arr[i], bit); |
198 | 0 | } |
199 | 0 | } else { |
200 | 0 | for (i = 0; i < size; i++) { |
201 | 0 | arr[i] = (int32_t)clamp64(((int64_t)1 << (-bit)) * arr[i], INT32_MIN, |
202 | 0 | INT32_MAX); |
203 | 0 | } |
204 | 0 | } |
205 | 0 | } |
206 | 0 | } |
207 | | |
208 | | const TXFM_TYPE av1_txfm_type_ls[5][TX_TYPES_1D] = { |
209 | | { TXFM_TYPE_DCT4, TXFM_TYPE_ADST4, TXFM_TYPE_ADST4, TXFM_TYPE_IDENTITY4 }, |
210 | | { TXFM_TYPE_DCT8, TXFM_TYPE_ADST8, TXFM_TYPE_ADST8, TXFM_TYPE_IDENTITY8 }, |
211 | | { TXFM_TYPE_DCT16, TXFM_TYPE_ADST16, TXFM_TYPE_ADST16, TXFM_TYPE_IDENTITY16 }, |
212 | | { TXFM_TYPE_DCT32, TXFM_TYPE_INVALID, TXFM_TYPE_INVALID, |
213 | | TXFM_TYPE_IDENTITY32 }, |
214 | | { TXFM_TYPE_DCT64, TXFM_TYPE_INVALID, TXFM_TYPE_INVALID, TXFM_TYPE_INVALID } |
215 | | }; |
216 | | |
217 | | const int8_t av1_txfm_stage_num_list[TXFM_TYPES] = { |
218 | | 4, // TXFM_TYPE_DCT4 |
219 | | 6, // TXFM_TYPE_DCT8 |
220 | | 8, // TXFM_TYPE_DCT16 |
221 | | 10, // TXFM_TYPE_DCT32 |
222 | | 12, // TXFM_TYPE_DCT64 |
223 | | 7, // TXFM_TYPE_ADST4 |
224 | | 8, // TXFM_TYPE_ADST8 |
225 | | 10, // TXFM_TYPE_ADST16 |
226 | | 1, // TXFM_TYPE_IDENTITY4 |
227 | | 1, // TXFM_TYPE_IDENTITY8 |
228 | | 1, // TXFM_TYPE_IDENTITY16 |
229 | | 1, // TXFM_TYPE_IDENTITY32 |
230 | | }; |
231 | | |
232 | | void av1_range_check_buf(int32_t stage, const int32_t *input, |
233 | 0 | const int32_t *buf, int32_t size, int8_t bit) { |
234 | | #if CONFIG_COEFFICIENT_RANGE_CHECKING |
235 | | const int64_t max_value = (1LL << (bit - 1)) - 1; |
236 | | const int64_t min_value = -(1LL << (bit - 1)); |
237 | | |
238 | | int in_range = 1; |
239 | | |
240 | | for (int i = 0; i < size; ++i) { |
241 | | if (buf[i] < min_value || buf[i] > max_value) { |
242 | | in_range = 0; |
243 | | } |
244 | | } |
245 | | |
246 | | if (!in_range) { |
247 | | fprintf(stderr, "Error: coeffs contain out-of-range values\n"); |
248 | | fprintf(stderr, "size: %d\n", size); |
249 | | fprintf(stderr, "stage: %d\n", stage); |
250 | | fprintf(stderr, "allowed range: [%" PRId64 ";%" PRId64 "]\n", min_value, |
251 | | max_value); |
252 | | |
253 | | fprintf(stderr, "coeffs: "); |
254 | | |
255 | | fprintf(stderr, "["); |
256 | | for (int j = 0; j < size; j++) { |
257 | | if (j > 0) fprintf(stderr, ", "); |
258 | | fprintf(stderr, "%d", input[j]); |
259 | | } |
260 | | fprintf(stderr, "]\n"); |
261 | | |
262 | | fprintf(stderr, " buf: "); |
263 | | |
264 | | fprintf(stderr, "["); |
265 | | for (int j = 0; j < size; j++) { |
266 | | if (j > 0) fprintf(stderr, ", "); |
267 | | fprintf(stderr, "%d", buf[j]); |
268 | | } |
269 | | fprintf(stderr, "]\n\n"); |
270 | | } |
271 | | |
272 | | assert(in_range); |
273 | | #else |
274 | 0 | (void)stage; |
275 | 0 | (void)input; |
276 | 0 | (void)buf; |
277 | 0 | (void)size; |
278 | 0 | (void)bit; |
279 | 0 | #endif |
280 | 0 | } |