blob: 8b2863033b6a502732c56a7921e2dfb87896ab65
1 | /* |
2 | * Texture block compression |
3 | * Copyright (C) 2015 Vittorio Giovara <vittorio.giovara@gmail.com> |
4 | * Based on public domain code by Fabian Giesen, Sean Barrett and Yann Collet. |
5 | * |
6 | * This file is part of FFmpeg |
7 | * |
8 | * Permission is hereby granted, free of charge, to any person obtaining a |
9 | * copy of this software and associated documentation files (the "Software"), |
10 | * to deal in the Software without restriction, including without limitation |
11 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, |
12 | * and/or sell copies of the Software, and to permit persons to whom the |
13 | * Software is furnished to do so, subject to the following conditions: |
14 | * The above copyright notice and this permission notice shall be included |
15 | * in all copies or substantial portions of the Software. |
16 | * |
17 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
18 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
19 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
20 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
21 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
22 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS |
23 | * IN THE SOFTWARE. |
24 | */ |
25 | |
26 | #include <stddef.h> |
27 | #include <stdint.h> |
28 | |
29 | #include "libavutil/attributes.h" |
30 | #include "libavutil/common.h" |
31 | #include "libavutil/intreadwrite.h" |
32 | |
33 | #include "texturedsp.h" |
34 | |
35 | static const uint8_t expand5[32] = { |
36 | 0, 8, 16, 24, 33, 41, 49, 57, 66, 74, 82, 90, |
37 | 99, 107, 115, 123, 132, 140, 148, 156, 165, 173, 181, 189, |
38 | 198, 206, 214, 222, 231, 239, 247, 255, |
39 | }; |
40 | |
41 | static const uint8_t expand6[64] = { |
42 | 0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44, |
43 | 48, 52, 56, 60, 65, 69, 73, 77, 81, 85, 89, 93, |
44 | 97, 101, 105, 109, 113, 117, 121, 125, 130, 134, 138, 142, |
45 | 146, 150, 154, 158, 162, 166, 170, 174, 178, 182, 186, 190, |
46 | 195, 199, 203, 207, 211, 215, 219, 223, 227, 231, 235, 239, |
47 | 243, 247, 251, 255, |
48 | }; |
49 | |
50 | static const uint8_t match5[256][2] = { |
51 | { 0, 0 }, { 0, 0 }, { 0, 1 }, { 0, 1 }, { 1, 0 }, { 1, 0 }, |
52 | { 1, 0 }, { 1, 1 }, { 1, 1 }, { 2, 0 }, { 2, 0 }, { 0, 4 }, |
53 | { 2, 1 }, { 2, 1 }, { 2, 1 }, { 3, 0 }, { 3, 0 }, { 3, 0 }, |
54 | { 3, 1 }, { 1, 5 }, { 3, 2 }, { 3, 2 }, { 4, 0 }, { 4, 0 }, |
55 | { 4, 1 }, { 4, 1 }, { 4, 2 }, { 4, 2 }, { 4, 2 }, { 3, 5 }, |
56 | { 5, 1 }, { 5, 1 }, { 5, 2 }, { 4, 4 }, { 5, 3 }, { 5, 3 }, |
57 | { 5, 3 }, { 6, 2 }, { 6, 2 }, { 6, 2 }, { 6, 3 }, { 5, 5 }, |
58 | { 6, 4 }, { 6, 4 }, { 4, 8 }, { 7, 3 }, { 7, 3 }, { 7, 3 }, |
59 | { 7, 4 }, { 7, 4 }, { 7, 4 }, { 7, 5 }, { 5, 9 }, { 7, 6 }, |
60 | { 7, 6 }, { 8, 4 }, { 8, 4 }, { 8, 5 }, { 8, 5 }, { 8, 6 }, |
61 | { 8, 6 }, { 8, 6 }, { 7, 9 }, { 9, 5 }, { 9, 5 }, { 9, 6 }, |
62 | { 8, 8 }, { 9, 7 }, { 9, 7 }, { 9, 7 }, { 10, 6 }, { 10, 6 }, |
63 | { 10, 6 }, { 10, 7 }, { 9, 9 }, { 10, 8 }, { 10, 8 }, { 8, 12 }, |
64 | { 11, 7 }, { 11, 7 }, { 11, 7 }, { 11, 8 }, { 11, 8 }, { 11, 8 }, |
65 | { 11, 9 }, { 9, 13 }, { 11, 10 }, { 11, 10 }, { 12, 8 }, { 12, 8 }, |
66 | { 12, 9 }, { 12, 9 }, { 12, 10 }, { 12, 10 }, { 12, 10 }, { 11, 13 }, |
67 | { 13, 9 }, { 13, 9 }, { 13, 10 }, { 12, 12 }, { 13, 11 }, { 13, 11 }, |
68 | { 13, 11 }, { 14, 10 }, { 14, 10 }, { 14, 10 }, { 14, 11 }, { 13, 13 }, |
69 | { 14, 12 }, { 14, 12 }, { 12, 16 }, { 15, 11 }, { 15, 11 }, { 15, 11 }, |
70 | { 15, 12 }, { 15, 12 }, { 15, 12 }, { 15, 13 }, { 13, 17 }, { 15, 14 }, |
71 | { 15, 14 }, { 16, 12 }, { 16, 12 }, { 16, 13 }, { 16, 13 }, { 16, 14 }, |
72 | { 16, 14 }, { 16, 14 }, { 15, 17 }, { 17, 13 }, { 17, 13 }, { 17, 14 }, |
73 | { 16, 16 }, { 17, 15 }, { 17, 15 }, { 17, 15 }, { 18, 14 }, { 18, 14 }, |
74 | { 18, 14 }, { 18, 15 }, { 17, 17 }, { 18, 16 }, { 18, 16 }, { 16, 20 }, |
75 | { 19, 15 }, { 19, 15 }, { 19, 15 }, { 19, 16 }, { 19, 16 }, { 19, 16 }, |
76 | { 19, 17 }, { 17, 21 }, { 19, 18 }, { 19, 18 }, { 20, 16 }, { 20, 16 }, |
77 | { 20, 17 }, { 20, 17 }, { 20, 18 }, { 20, 18 }, { 20, 18 }, { 19, 21 }, |
78 | { 21, 17 }, { 21, 17 }, { 21, 18 }, { 20, 20 }, { 21, 19 }, { 21, 19 }, |
79 | { 21, 19 }, { 22, 18 }, { 22, 18 }, { 22, 18 }, { 22, 19 }, { 21, 21 }, |
80 | { 22, 20 }, { 22, 20 }, { 20, 24 }, { 23, 19 }, { 23, 19 }, { 23, 19 }, |
81 | { 23, 20 }, { 23, 20 }, { 23, 20 }, { 23, 21 }, { 21, 25 }, { 23, 22 }, |
82 | { 23, 22 }, { 24, 20 }, { 24, 20 }, { 24, 21 }, { 24, 21 }, { 24, 22 }, |
83 | { 24, 22 }, { 24, 22 }, { 23, 25 }, { 25, 21 }, { 25, 21 }, { 25, 22 }, |
84 | { 24, 24 }, { 25, 23 }, { 25, 23 }, { 25, 23 }, { 26, 22 }, { 26, 22 }, |
85 | { 26, 22 }, { 26, 23 }, { 25, 25 }, { 26, 24 }, { 26, 24 }, { 24, 28 }, |
86 | { 27, 23 }, { 27, 23 }, { 27, 23 }, { 27, 24 }, { 27, 24 }, { 27, 24 }, |
87 | { 27, 25 }, { 25, 29 }, { 27, 26 }, { 27, 26 }, { 28, 24 }, { 28, 24 }, |
88 | { 28, 25 }, { 28, 25 }, { 28, 26 }, { 28, 26 }, { 28, 26 }, { 27, 29 }, |
89 | { 29, 25 }, { 29, 25 }, { 29, 26 }, { 28, 28 }, { 29, 27 }, { 29, 27 }, |
90 | { 29, 27 }, { 30, 26 }, { 30, 26 }, { 30, 26 }, { 30, 27 }, { 29, 29 }, |
91 | { 30, 28 }, { 30, 28 }, { 30, 28 }, { 31, 27 }, { 31, 27 }, { 31, 27 }, |
92 | { 31, 28 }, { 31, 28 }, { 31, 28 }, { 31, 29 }, { 31, 29 }, { 31, 30 }, |
93 | { 31, 30 }, { 31, 30 }, { 31, 31 }, { 31, 31 }, |
94 | }; |
95 | |
96 | static const uint8_t match6[256][2] = { |
97 | { 0, 0 }, { 0, 1 }, { 1, 0 }, { 1, 0 }, { 1, 1 }, { 2, 0 }, |
98 | { 2, 1 }, { 3, 0 }, { 3, 0 }, { 3, 1 }, { 4, 0 }, { 4, 0 }, |
99 | { 4, 1 }, { 5, 0 }, { 5, 1 }, { 6, 0 }, { 6, 0 }, { 6, 1 }, |
100 | { 7, 0 }, { 7, 0 }, { 7, 1 }, { 8, 0 }, { 8, 1 }, { 8, 1 }, |
101 | { 8, 2 }, { 9, 1 }, { 9, 2 }, { 9, 2 }, { 9, 3 }, { 10, 2 }, |
102 | { 10, 3 }, { 10, 3 }, { 10, 4 }, { 11, 3 }, { 11, 4 }, { 11, 4 }, |
103 | { 11, 5 }, { 12, 4 }, { 12, 5 }, { 12, 5 }, { 12, 6 }, { 13, 5 }, |
104 | { 13, 6 }, { 8, 16 }, { 13, 7 }, { 14, 6 }, { 14, 7 }, { 9, 17 }, |
105 | { 14, 8 }, { 15, 7 }, { 15, 8 }, { 11, 16 }, { 15, 9 }, { 15, 10 }, |
106 | { 16, 8 }, { 16, 9 }, { 16, 10 }, { 15, 13 }, { 17, 9 }, { 17, 10 }, |
107 | { 17, 11 }, { 15, 16 }, { 18, 10 }, { 18, 11 }, { 18, 12 }, { 16, 16 }, |
108 | { 19, 11 }, { 19, 12 }, { 19, 13 }, { 17, 17 }, { 20, 12 }, { 20, 13 }, |
109 | { 20, 14 }, { 19, 16 }, { 21, 13 }, { 21, 14 }, { 21, 15 }, { 20, 17 }, |
110 | { 22, 14 }, { 22, 15 }, { 25, 10 }, { 22, 16 }, { 23, 15 }, { 23, 16 }, |
111 | { 26, 11 }, { 23, 17 }, { 24, 16 }, { 24, 17 }, { 27, 12 }, { 24, 18 }, |
112 | { 25, 17 }, { 25, 18 }, { 28, 13 }, { 25, 19 }, { 26, 18 }, { 26, 19 }, |
113 | { 29, 14 }, { 26, 20 }, { 27, 19 }, { 27, 20 }, { 30, 15 }, { 27, 21 }, |
114 | { 28, 20 }, { 28, 21 }, { 28, 21 }, { 28, 22 }, { 29, 21 }, { 29, 22 }, |
115 | { 24, 32 }, { 29, 23 }, { 30, 22 }, { 30, 23 }, { 25, 33 }, { 30, 24 }, |
116 | { 31, 23 }, { 31, 24 }, { 27, 32 }, { 31, 25 }, { 31, 26 }, { 32, 24 }, |
117 | { 32, 25 }, { 32, 26 }, { 31, 29 }, { 33, 25 }, { 33, 26 }, { 33, 27 }, |
118 | { 31, 32 }, { 34, 26 }, { 34, 27 }, { 34, 28 }, { 32, 32 }, { 35, 27 }, |
119 | { 35, 28 }, { 35, 29 }, { 33, 33 }, { 36, 28 }, { 36, 29 }, { 36, 30 }, |
120 | { 35, 32 }, { 37, 29 }, { 37, 30 }, { 37, 31 }, { 36, 33 }, { 38, 30 }, |
121 | { 38, 31 }, { 41, 26 }, { 38, 32 }, { 39, 31 }, { 39, 32 }, { 42, 27 }, |
122 | { 39, 33 }, { 40, 32 }, { 40, 33 }, { 43, 28 }, { 40, 34 }, { 41, 33 }, |
123 | { 41, 34 }, { 44, 29 }, { 41, 35 }, { 42, 34 }, { 42, 35 }, { 45, 30 }, |
124 | { 42, 36 }, { 43, 35 }, { 43, 36 }, { 46, 31 }, { 43, 37 }, { 44, 36 }, |
125 | { 44, 37 }, { 44, 37 }, { 44, 38 }, { 45, 37 }, { 45, 38 }, { 40, 48 }, |
126 | { 45, 39 }, { 46, 38 }, { 46, 39 }, { 41, 49 }, { 46, 40 }, { 47, 39 }, |
127 | { 47, 40 }, { 43, 48 }, { 47, 41 }, { 47, 42 }, { 48, 40 }, { 48, 41 }, |
128 | { 48, 42 }, { 47, 45 }, { 49, 41 }, { 49, 42 }, { 49, 43 }, { 47, 48 }, |
129 | { 50, 42 }, { 50, 43 }, { 50, 44 }, { 48, 48 }, { 51, 43 }, { 51, 44 }, |
130 | { 51, 45 }, { 49, 49 }, { 52, 44 }, { 52, 45 }, { 52, 46 }, { 51, 48 }, |
131 | { 53, 45 }, { 53, 46 }, { 53, 47 }, { 52, 49 }, { 54, 46 }, { 54, 47 }, |
132 | { 57, 42 }, { 54, 48 }, { 55, 47 }, { 55, 48 }, { 58, 43 }, { 55, 49 }, |
133 | { 56, 48 }, { 56, 49 }, { 59, 44 }, { 56, 50 }, { 57, 49 }, { 57, 50 }, |
134 | { 60, 45 }, { 57, 51 }, { 58, 50 }, { 58, 51 }, { 61, 46 }, { 58, 52 }, |
135 | { 59, 51 }, { 59, 52 }, { 62, 47 }, { 59, 53 }, { 60, 52 }, { 60, 53 }, |
136 | { 60, 53 }, { 60, 54 }, { 61, 53 }, { 61, 54 }, { 61, 54 }, { 61, 55 }, |
137 | { 62, 54 }, { 62, 55 }, { 62, 55 }, { 62, 56 }, { 63, 55 }, { 63, 56 }, |
138 | { 63, 56 }, { 63, 57 }, { 63, 58 }, { 63, 59 }, { 63, 59 }, { 63, 60 }, |
139 | { 63, 61 }, { 63, 62 }, { 63, 62 }, { 63, 63 }, |
140 | }; |
141 | |
142 | /* Multiplication over 8 bit emulation */ |
143 | #define mul8(a, b) (((a) * (b) + 128 + (((a) * (b) + 128) >> 8)) >> 8) |
144 | |
145 | /* Conversion from rgb24 to rgb565 */ |
146 | #define rgb2rgb565(r, g, b) \ |
147 | ((mul8(r, 31) << 11) | (mul8(g, 63) << 5) | (mul8(b, 31) << 0)) |
148 | |
149 | /* Linear interpolation at 1/3 point between a and b */ |
150 | #define lerp13(a, b) ((2 * (a) + (b)) / 3) |
151 | |
152 | /* Linear interpolation on an RGB pixel */ |
153 | static inline void lerp13rgb(uint8_t *out, uint8_t *p1, uint8_t *p2) |
154 | { |
155 | out[0] = lerp13(p1[0], p2[0]); |
156 | out[1] = lerp13(p1[1], p2[1]); |
157 | out[2] = lerp13(p1[2], p2[2]); |
158 | } |
159 | |
160 | /* Conversion from rgb565 to rgb24 */ |
161 | static inline void rgb5652rgb(uint8_t *out, uint16_t v) |
162 | { |
163 | int rv = (v & 0xf800) >> 11; |
164 | int gv = (v & 0x07e0) >> 5; |
165 | int bv = (v & 0x001f) >> 0; |
166 | |
167 | out[0] = expand5[rv]; |
168 | out[1] = expand6[gv]; |
169 | out[2] = expand5[bv]; |
170 | out[3] = 0; |
171 | } |
172 | |
173 | /* Color matching function */ |
174 | static unsigned int match_colors(const uint8_t *block, ptrdiff_t stride, |
175 | uint16_t c0, uint16_t c1) |
176 | { |
177 | uint32_t mask = 0; |
178 | int dirr, dirg, dirb; |
179 | int dots[16]; |
180 | int stops[4]; |
181 | int x, y, k = 0; |
182 | int c0_point, half_point, c3_point; |
183 | uint8_t color[16]; |
184 | static const int indexMap[8] = { |
185 | 0 << 30, 2 << 30, 0 << 30, 2 << 30, |
186 | 3 << 30, 3 << 30, 1 << 30, 1 << 30, |
187 | }; |
188 | |
189 | /* Fill color and compute direction for each component */ |
190 | rgb5652rgb(color + 0, c0); |
191 | rgb5652rgb(color + 4, c1); |
192 | lerp13rgb(color + 8, color + 0, color + 4); |
193 | lerp13rgb(color + 12, color + 4, color + 0); |
194 | |
195 | dirr = color[0 * 4 + 0] - color[1 * 4 + 0]; |
196 | dirg = color[0 * 4 + 1] - color[1 * 4 + 1]; |
197 | dirb = color[0 * 4 + 2] - color[1 * 4 + 2]; |
198 | |
199 | for (y = 0; y < 4; y++) { |
200 | for (x = 0; x < 4; x++) |
201 | dots[k++] = block[0 + x * 4 + y * stride] * dirr + |
202 | block[1 + x * 4 + y * stride] * dirg + |
203 | block[2 + x * 4 + y * stride] * dirb; |
204 | |
205 | stops[y] = color[0 + y * 4] * dirr + |
206 | color[1 + y * 4] * dirg + |
207 | color[2 + y * 4] * dirb; |
208 | } |
209 | |
210 | /* Think of the colors as arranged on a line; project point onto that line, |
211 | * then choose next color out of available ones. we compute the crossover |
212 | * points for 'best color in top half'/'best in bottom half' and then |
213 | * the same inside that subinterval. |
214 | * |
215 | * Relying on this 1d approximation isn't always optimal in terms of |
216 | * Euclidean distance, but it's very close and a lot faster. |
217 | * |
218 | * http://cbloomrants.blogspot.com/2008/12/12-08-08-dxtc-summary.html */ |
219 | c0_point = (stops[1] + stops[3]) >> 1; |
220 | half_point = (stops[3] + stops[2]) >> 1; |
221 | c3_point = (stops[2] + stops[0]) >> 1; |
222 | |
223 | for (x = 0; x < 16; x++) { |
224 | int dot = dots[x]; |
225 | int bits = (dot < half_point ? 4 : 0) | |
226 | (dot < c0_point ? 2 : 0) | |
227 | (dot < c3_point ? 1 : 0); |
228 | |
229 | mask >>= 2; |
230 | mask |= indexMap[bits]; |
231 | } |
232 | |
233 | return mask; |
234 | } |
235 | |
236 | /* Color optimization function */ |
237 | static void optimize_colors(const uint8_t *block, ptrdiff_t stride, |
238 | uint16_t *pmax16, uint16_t *pmin16) |
239 | { |
240 | const uint8_t *minp; |
241 | const uint8_t *maxp; |
242 | const int iter_power = 4; |
243 | double magn; |
244 | int v_r, v_g, v_b; |
245 | float covf[6], vfr, vfg, vfb; |
246 | int mind, maxd; |
247 | int cov[6] = { 0 }; |
248 | int mu[3], min[3], max[3]; |
249 | int ch, iter, x, y; |
250 | |
251 | /* Determine color distribution */ |
252 | for (ch = 0; ch < 3; ch++) { |
253 | const uint8_t *bp = &block[ch]; |
254 | int muv, minv, maxv; |
255 | |
256 | muv = minv = maxv = bp[0]; |
257 | for (y = 0; y < 4; y++) { |
258 | for (x = 4; x < 4; x += 4) { |
259 | muv += bp[x * 4 + y * stride]; |
260 | if (bp[x] < minv) |
261 | minv = bp[x * 4 + y * stride]; |
262 | else if (bp[x] > maxv) |
263 | maxv = bp[x * 4 + y * stride]; |
264 | } |
265 | } |
266 | |
267 | mu[ch] = (muv + 8) >> 4; |
268 | min[ch] = minv; |
269 | max[ch] = maxv; |
270 | } |
271 | |
272 | /* Determine covariance matrix */ |
273 | for (y = 0; y < 4; y++) { |
274 | for (x = 0; x < 4; x++) { |
275 | int r = block[x * 4 + stride * y + 0] - mu[0]; |
276 | int g = block[x * 4 + stride * y + 1] - mu[1]; |
277 | int b = block[x * 4 + stride * y + 2] - mu[2]; |
278 | |
279 | cov[0] += r * r; |
280 | cov[1] += r * g; |
281 | cov[2] += r * b; |
282 | cov[3] += g * g; |
283 | cov[4] += g * b; |
284 | cov[5] += b * b; |
285 | } |
286 | } |
287 | |
288 | /* Convert covariance matrix to float, find principal axis via power iter */ |
289 | for (x = 0; x < 6; x++) |
290 | covf[x] = cov[x] / 255.0f; |
291 | |
292 | vfr = (float) (max[0] - min[0]); |
293 | vfg = (float) (max[1] - min[1]); |
294 | vfb = (float) (max[2] - min[2]); |
295 | |
296 | for (iter = 0; iter < iter_power; iter++) { |
297 | float r = vfr * covf[0] + vfg * covf[1] + vfb * covf[2]; |
298 | float g = vfr * covf[1] + vfg * covf[3] + vfb * covf[4]; |
299 | float b = vfr * covf[2] + vfg * covf[4] + vfb * covf[5]; |
300 | |
301 | vfr = r; |
302 | vfg = g; |
303 | vfb = b; |
304 | } |
305 | |
306 | magn = fabs(vfr); |
307 | if (fabs(vfg) > magn) |
308 | magn = fabs(vfg); |
309 | if (fabs(vfb) > magn) |
310 | magn = fabs(vfb); |
311 | |
312 | /* if magnitude is too small, default to luminance */ |
313 | if (magn < 4.0f) { |
314 | /* JPEG YCbCr luma coefs, scaled by 1000 */ |
315 | v_r = 299; |
316 | v_g = 587; |
317 | v_b = 114; |
318 | } else { |
319 | magn = 512.0 / magn; |
320 | v_r = (int) (vfr * magn); |
321 | v_g = (int) (vfg * magn); |
322 | v_b = (int) (vfb * magn); |
323 | } |
324 | |
325 | /* Pick colors at extreme points */ |
326 | mind = maxd = block[0] * v_r + block[1] * v_g + block[2] * v_b; |
327 | minp = maxp = block; |
328 | for (y = 0; y < 4; y++) { |
329 | for (x = 0; x < 4; x++) { |
330 | int dot = block[x * 4 + y * stride + 0] * v_r + |
331 | block[x * 4 + y * stride + 1] * v_g + |
332 | block[x * 4 + y * stride + 2] * v_b; |
333 | |
334 | if (dot < mind) { |
335 | mind = dot; |
336 | minp = block + x * 4 + y * stride; |
337 | } else if (dot > maxd) { |
338 | maxd = dot; |
339 | maxp = block + x * 4 + y * stride; |
340 | } |
341 | } |
342 | } |
343 | |
344 | *pmax16 = rgb2rgb565(maxp[0], maxp[1], maxp[2]); |
345 | *pmin16 = rgb2rgb565(minp[0], minp[1], minp[2]); |
346 | } |
347 | |
348 | /* Try to optimize colors to suit block contents better, by solving |
349 | * a least squares system via normal equations + Cramer's rule. */ |
350 | static int refine_colors(const uint8_t *block, ptrdiff_t stride, |
351 | uint16_t *pmax16, uint16_t *pmin16, uint32_t mask) |
352 | { |
353 | uint32_t cm = mask; |
354 | uint16_t oldMin = *pmin16; |
355 | uint16_t oldMax = *pmax16; |
356 | uint16_t min16, max16; |
357 | int x, y; |
358 | |
359 | /* Additional magic to save a lot of multiplies in the accumulating loop. |
360 | * The tables contain precomputed products of weights for least squares |
361 | * system, accumulated inside one 32-bit register */ |
362 | static const int w1tab[4] = { 3, 0, 2, 1 }; |
363 | static const int prods[4] = { 0x090000, 0x000900, 0x040102, 0x010402 }; |
364 | |
365 | /* Check if all pixels have the same index */ |
366 | if ((mask ^ (mask << 2)) < 4) { |
367 | /* If so, linear system would be singular; solve using optimal |
368 | * single-color match on average color. */ |
369 | int r = 8, g = 8, b = 8; |
370 | for (y = 0; y < 4; y++) { |
371 | for (x = 0; x < 4; x++) { |
372 | r += block[0 + x * 4 + y * stride]; |
373 | g += block[1 + x * 4 + y * stride]; |
374 | b += block[2 + x * 4 + y * stride]; |
375 | } |
376 | } |
377 | |
378 | r >>= 4; |
379 | g >>= 4; |
380 | b >>= 4; |
381 | |
382 | max16 = (match5[r][0] << 11) | (match6[g][0] << 5) | match5[b][0]; |
383 | min16 = (match5[r][1] << 11) | (match6[g][1] << 5) | match5[b][1]; |
384 | } else { |
385 | float fr, fg, fb; |
386 | int at1_r = 0, at1_g = 0, at1_b = 0; |
387 | int at2_r = 0, at2_g = 0, at2_b = 0; |
388 | int akku = 0; |
389 | int xx, xy, yy; |
390 | |
391 | for (y = 0; y < 4; y++) { |
392 | for (x = 0; x < 4; x++) { |
393 | int step = cm & 3; |
394 | int w1 = w1tab[step]; |
395 | int r = block[0 + x * 4 + y * stride]; |
396 | int g = block[1 + x * 4 + y * stride]; |
397 | int b = block[2 + x * 4 + y * stride]; |
398 | |
399 | akku += prods[step]; |
400 | at1_r += w1 * r; |
401 | at1_g += w1 * g; |
402 | at1_b += w1 * b; |
403 | at2_r += r; |
404 | at2_g += g; |
405 | at2_b += b; |
406 | |
407 | cm >>= 2; |
408 | } |
409 | } |
410 | |
411 | at2_r = 3 * at2_r - at1_r; |
412 | at2_g = 3 * at2_g - at1_g; |
413 | at2_b = 3 * at2_b - at1_b; |
414 | |
415 | /* Extract solutions and decide solvability */ |
416 | xx = akku >> 16; |
417 | yy = (akku >> 8) & 0xFF; |
418 | xy = (akku >> 0) & 0xFF; |
419 | |
420 | fr = 3.0f * 31.0f / 255.0f / (xx * yy - xy * xy); |
421 | fg = fr * 63.0f / 31.0f; |
422 | fb = fr; |
423 | |
424 | /* Solve */ |
425 | max16 = av_clip_uintp2((at1_r * yy - at2_r * xy) * fr + 0.5f, 5) << 11; |
426 | max16 |= av_clip_uintp2((at1_g * yy - at2_g * xy) * fg + 0.5f, 6) << 5; |
427 | max16 |= av_clip_uintp2((at1_b * yy - at2_b * xy) * fb + 0.5f, 5) << 0; |
428 | |
429 | min16 = av_clip_uintp2((at2_r * xx - at1_r * xy) * fr + 0.5f, 5) << 11; |
430 | min16 |= av_clip_uintp2((at2_g * xx - at1_g * xy) * fg + 0.5f, 6) << 5; |
431 | min16 |= av_clip_uintp2((at2_b * xx - at1_b * xy) * fb + 0.5f, 5) << 0; |
432 | } |
433 | |
434 | *pmin16 = min16; |
435 | *pmax16 = max16; |
436 | return oldMin != min16 || oldMax != max16; |
437 | } |
438 | |
439 | /* Check if input block is a constant color */ |
440 | static int constant_color(const uint8_t *block, ptrdiff_t stride) |
441 | { |
442 | int x, y; |
443 | uint32_t first = AV_RL32(block); |
444 | |
445 | for (y = 0; y < 4; y++) |
446 | for (x = 0; x < 4; x++) |
447 | if (first != AV_RL32(block + x * 4 + y * stride)) |
448 | return 0; |
449 | return 1; |
450 | } |
451 | |
452 | /* Main color compression function */ |
453 | static void compress_color(uint8_t *dst, ptrdiff_t stride, const uint8_t *block) |
454 | { |
455 | uint32_t mask; |
456 | uint16_t max16, min16; |
457 | int constant = constant_color(block, stride); |
458 | |
459 | /* Constant color will load values from tables */ |
460 | if (constant) { |
461 | int r = block[0]; |
462 | int g = block[1]; |
463 | int b = block[2]; |
464 | mask = 0xAAAAAAAA; |
465 | max16 = (match5[r][0] << 11) | (match6[g][0] << 5) | match5[b][0]; |
466 | min16 = (match5[r][1] << 11) | (match6[g][1] << 5) | match5[b][1]; |
467 | } else { |
468 | int refine; |
469 | |
470 | /* Otherwise find pca and map along principal axis */ |
471 | optimize_colors(block, stride, &max16, &min16); |
472 | if (max16 != min16) |
473 | mask = match_colors(block, stride, max16, min16); |
474 | else |
475 | mask = 0; |
476 | |
477 | /* One pass refinement */ |
478 | refine = refine_colors(block, stride, &max16, &min16, mask); |
479 | if (refine) { |
480 | if (max16 != min16) |
481 | mask = match_colors(block, stride, max16, min16); |
482 | else |
483 | mask = 0; |
484 | } |
485 | } |
486 | |
487 | /* Finally write the color block */ |
488 | if (max16 < min16) { |
489 | FFSWAP(uint16_t, min16, max16); |
490 | mask ^= 0x55555555; |
491 | } |
492 | |
493 | AV_WL16(dst + 0, max16); |
494 | AV_WL16(dst + 2, min16); |
495 | AV_WL32(dst + 4, mask); |
496 | } |
497 | |
498 | /* Alpha compression function */ |
499 | static void compress_alpha(uint8_t *dst, ptrdiff_t stride, const uint8_t *block) |
500 | { |
501 | int x, y; |
502 | int dist, bias, dist4, dist2; |
503 | int mn, mx; |
504 | int bits = 0; |
505 | int mask = 0; |
506 | |
507 | memset(dst, 0, 8); |
508 | |
509 | /* Find min/max color */ |
510 | mn = mx = block[3]; |
511 | for (y = 0; y < 4; y++) { |
512 | for (x = 0; x < 4; x++) { |
513 | int val = block[3 + x * 4 + y * stride]; |
514 | if (val < mn) |
515 | mn = val; |
516 | else if (val > mx) |
517 | mx = val; |
518 | } |
519 | } |
520 | |
521 | /* Encode them */ |
522 | dst[0] = (uint8_t) mx; |
523 | dst[1] = (uint8_t) mn; |
524 | dst += 2; |
525 | |
526 | /* Mono-alpha shortcut */ |
527 | if (mn == mx) |
528 | return; |
529 | |
530 | /* Determine bias and emit color indices. |
531 | * Given the choice of mx/mn, these indices are optimal: |
532 | * fgiesen.wordpress.com/2009/12/15/dxt5-alpha-block-index-determination */ |
533 | dist = mx - mn; |
534 | |
535 | dist4 = dist * 4; |
536 | dist2 = dist * 2; |
537 | if (dist < 8) |
538 | bias = dist - 1 - mn * 7; |
539 | else |
540 | bias = dist / 2 + 2 - mn * 7; |
541 | |
542 | for (y = 0; y < 4; y++) { |
543 | for (x = 0; x < 4; x++) { |
544 | int alp = block[3 + x * 4 + y * stride] * 7 + bias; |
545 | int ind, tmp; |
546 | |
547 | /* This is a "linear scale" lerp factor between 0 (val=min) |
548 | * and 7 (val=max) to select index. */ |
549 | tmp = (alp >= dist4) ? -1 : 0; |
550 | ind = tmp & 4; |
551 | alp -= dist4 & tmp; |
552 | tmp = (alp >= dist2) ? -1 : 0; |
553 | ind += tmp & 2; |
554 | alp -= dist2 & tmp; |
555 | ind += (alp >= dist); |
556 | |
557 | /* Turn linear scale into DXT index (0/1 are extreme points) */ |
558 | ind = -ind & 7; |
559 | ind ^= (2 > ind); |
560 | |
561 | /* Write index */ |
562 | mask |= ind << bits; |
563 | bits += 3; |
564 | if (bits >= 8) { |
565 | *dst++ = mask; |
566 | mask >>= 8; |
567 | bits -= 8; |
568 | } |
569 | } |
570 | } |
571 | } |
572 | |
573 | /** |
574 | * Convert a RGBA buffer to unscaled YCoCg. |
575 | * Scale is usually introduced to avoid banding over a certain range of colors, |
576 | * but this version of the algorithm does not introduce it as much as other |
577 | * implementations, allowing for a simpler and faster conversion. |
578 | */ |
579 | static void rgba2ycocg(uint8_t *dst, const uint8_t *pixel) |
580 | { |
581 | int r = pixel[0]; |
582 | int g = (pixel[1] + 1) >> 1; |
583 | int b = pixel[2]; |
584 | int t = (2 + r + b) >> 2; |
585 | |
586 | dst[0] = av_clip_uint8(128 + ((r - b + 1) >> 1)); /* Co */ |
587 | dst[1] = av_clip_uint8(128 + g - t); /* Cg */ |
588 | dst[2] = 0; |
589 | dst[3] = av_clip_uint8(g + t); /* Y */ |
590 | } |
591 | |
592 | /** |
593 | * Compress one block of RGBA pixels in a DXT1 texture and store the |
594 | * resulting bytes in 'dst'. Alpha is not preserved. |
595 | * |
596 | * @param dst output buffer. |
597 | * @param stride scanline in bytes. |
598 | * @param block block to compress. |
599 | * @return how much texture data has been written. |
600 | */ |
601 | static int dxt1_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block) |
602 | { |
603 | compress_color(dst, stride, block); |
604 | |
605 | return 8; |
606 | } |
607 | |
608 | /** |
609 | * Compress one block of RGBA pixels in a DXT5 texture and store the |
610 | * resulting bytes in 'dst'. Alpha is preserved. |
611 | * |
612 | * @param dst output buffer. |
613 | * @param stride scanline in bytes. |
614 | * @param block block to compress. |
615 | * @return how much texture data has been written. |
616 | */ |
617 | static int dxt5_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block) |
618 | { |
619 | compress_alpha(dst, stride, block); |
620 | compress_color(dst + 8, stride, block); |
621 | |
622 | return 16; |
623 | } |
624 | |
625 | /** |
626 | * Compress one block of RGBA pixels in a DXT5-YCoCg texture and store the |
627 | * resulting bytes in 'dst'. Alpha is not preserved. |
628 | * |
629 | * @param dst output buffer. |
630 | * @param stride scanline in bytes. |
631 | * @param block block to compress. |
632 | * @return how much texture data has been written. |
633 | */ |
634 | static int dxt5ys_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block) |
635 | { |
636 | int x, y; |
637 | uint8_t reorder[64]; |
638 | |
639 | /* Reorder the components and then run a normal DXT5 compression. */ |
640 | for (y = 0; y < 4; y++) |
641 | for (x = 0; x < 4; x++) |
642 | rgba2ycocg(reorder + x * 4 + y * 16, block + x * 4 + y * stride); |
643 | |
644 | compress_alpha(dst + 0, 16, reorder); |
645 | compress_color(dst + 8, 16, reorder); |
646 | |
647 | return 16; |
648 | } |
649 | |
650 | av_cold void ff_texturedspenc_init(TextureDSPContext *c) |
651 | { |
652 | c->dxt1_block = dxt1_block; |
653 | c->dxt5_block = dxt5_block; |
654 | c->dxt5ys_block = dxt5ys_block; |
655 | } |
656 |