FFmpeg coverage

Directory: ../../../ffmpeg/
File: src/libavcodec/vvc/filter_template.c
Date: 2024-05-03 15:42:48
Exec Total Coverage
Lines: 524 533 98.3%
Functions: 38 72 52.8%
Branches: 323 344 93.9%
Line Branch Exec Source
1 /*
2 * VVC filters DSP
3 *
4 * Copyright (C) 2022 Nuo Mi
5 *
6 * This file is part of FFmpeg.
7 *
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22
23 #include "libavcodec/h26x/h2656_sao_template.c"
24
25 380220 static void FUNC(lmcs_filter_luma)(uint8_t *_dst, ptrdiff_t dst_stride, const int width, const int height, const void *_lut)
26 {
27 380220 const pixel *lut = _lut;
28 380220 pixel *dst = (pixel*)_dst;
29 380220 dst_stride /= sizeof(pixel);
30
31
2/2
✓ Branch 0 taken 6825656 times.
✓ Branch 1 taken 190110 times.
 14031532 for (int y = 0; y < height; y++) {
32
2/2
✓ Branch 0 taken 536579296 times.
✓ Branch 1 taken 6825656 times.
 1086809904 for (int x = 0; x < width; x++)
33 1073158592 dst[x] = lut[dst[x]];
34 13651312 dst += dst_stride;
35 }
36 380220 }
37
38 8017559040 static av_always_inline int16_t FUNC(alf_clip)(pixel curr, pixel v0, pixel v1, int16_t clip)
39 {
40 8017559040 return av_clip(v0 - curr, -clip, clip) + av_clip(v1 - curr, -clip, clip);
41 }
42
43 34962 static void FUNC(alf_filter_luma)(uint8_t *_dst, ptrdiff_t dst_stride, const uint8_t *_src, ptrdiff_t src_stride,
44 const int width, const int height, const int16_t *filter, const int16_t *clip, const int vb_pos)
45 {
46 34962 const pixel *src = (pixel *)_src;
47 34962 const int shift = 7;
48 34962 const int offset = 1 << ( shift - 1 );
49 34962 const int vb_above = vb_pos - 4;
50 34962 const int vb_below = vb_pos + 3;
51
52 34962 dst_stride /= sizeof(pixel);
53 34962 src_stride /= sizeof(pixel);
54
55
2/2
✓ Branch 0 taken 520604 times.
✓ Branch 1 taken 17481 times.
 1076170 for (int y = 0; y < height; y += ALF_BLOCK_SIZE) {
56
2/2
✓ Branch 0 taken 16146592 times.
✓ Branch 1 taken 520604 times.
 33334392 for (int x = 0; x < width; x += ALF_BLOCK_SIZE) {
57 32293184 const pixel *s0 = src + y * src_stride + x;
58 32293184 const pixel *s1 = s0 + src_stride;
59 32293184 const pixel *s2 = s0 - src_stride;
60 32293184 const pixel *s3 = s1 + src_stride;
61 32293184 const pixel *s4 = s2 - src_stride;
62 32293184 const pixel *s5 = s3 + src_stride;
63 32293184 const pixel *s6 = s4 - src_stride;
64
65
2/2
✓ Branch 0 taken 64586368 times.
✓ Branch 1 taken 16146592 times.
 161465920 for (int i = 0; i < ALF_BLOCK_SIZE; i++) {
66 129172736 pixel *dst = (pixel *)_dst + (y + i) * dst_stride + x;
67
68 129172736 const pixel *p0 = s0 + i * src_stride;
69 129172736 const pixel *p1 = s1 + i * src_stride;
70 129172736 const pixel *p2 = s2 + i * src_stride;
71 129172736 const pixel *p3 = s3 + i * src_stride;
72 129172736 const pixel *p4 = s4 + i * src_stride;
73 129172736 const pixel *p5 = s5 + i * src_stride;
74 129172736 const pixel *p6 = s6 + i * src_stride;
75
76
4/4
✓ Branch 0 taken 62832000 times.
✓ Branch 1 taken 1754368 times.
✓ Branch 2 taken 438592 times.
✓ Branch 3 taken 62393408 times.
 129172736 const int is_near_vb_above = (y + i < vb_pos) && (y + i >= vb_pos - 1);
77
4/4
✓ Branch 0 taken 1754368 times.
✓ Branch 1 taken 62832000 times.
✓ Branch 2 taken 438592 times.
✓ Branch 3 taken 1315776 times.
 129172736 const int is_near_vb_below = (y + i >= vb_pos) && (y + i <= vb_pos);
78
4/4
✓ Branch 0 taken 64147776 times.
✓ Branch 1 taken 438592 times.
✓ Branch 2 taken 438592 times.
✓ Branch 3 taken 63709184 times.
 129172736 const int is_near_vb = is_near_vb_above || is_near_vb_below;
79
80
4/4
✓ Branch 0 taken 62832000 times.
✓ Branch 1 taken 1754368 times.
✓ Branch 2 taken 1754368 times.
✓ Branch 3 taken 61077632 times.
 129172736 if ((y + i < vb_pos) && ((y + i) >= vb_above)) {
81
2/2
✓ Branch 0 taken 438592 times.
✓ Branch 1 taken 1315776 times.
 3508736 p1 = (y + i == vb_pos - 1) ? p0 : p1;
82
2/2
✓ Branch 0 taken 877184 times.
✓ Branch 1 taken 877184 times.
 3508736 p3 = (y + i >= vb_pos - 2) ? p1 : p3;
83
2/2
✓ Branch 0 taken 1315776 times.
✓ Branch 1 taken 438592 times.
 3508736 p5 = (y + i >= vb_pos - 3) ? p3 : p5;
84
85
2/2
✓ Branch 0 taken 438592 times.
✓ Branch 1 taken 1315776 times.
 3508736 p2 = (y + i == vb_pos - 1) ? p0 : p2;
86
2/2
✓ Branch 0 taken 877184 times.
✓ Branch 1 taken 877184 times.
 3508736 p4 = (y + i >= vb_pos - 2) ? p2 : p4;
87
2/2
✓ Branch 0 taken 1315776 times.
✓ Branch 1 taken 438592 times.
 3508736 p6 = (y + i >= vb_pos - 3) ? p4 : p6;
88
3/4
✓ Branch 0 taken 1754368 times.
✓ Branch 1 taken 61077632 times.
✓ Branch 2 taken 1754368 times.
✗ Branch 3 not taken.
 125664000 } else if ((y + i >= vb_pos) && ((y + i) <= vb_below)) {
89
2/2
✓ Branch 0 taken 438592 times.
✓ Branch 1 taken 1315776 times.
 3508736 p2 = (y + i == vb_pos ) ? p0 : p2;
90
2/2
✓ Branch 0 taken 877184 times.
✓ Branch 1 taken 877184 times.
 3508736 p4 = (y + i <= vb_pos + 1) ? p2 : p4;
91
2/2
✓ Branch 0 taken 1315776 times.
✓ Branch 1 taken 438592 times.
 3508736 p6 = (y + i <= vb_pos + 2) ? p4 : p6;
92
93
2/2
✓ Branch 0 taken 438592 times.
✓ Branch 1 taken 1315776 times.
 3508736 p1 = (y + i == vb_pos ) ? p0 : p1;
94
2/2
✓ Branch 0 taken 877184 times.
✓ Branch 1 taken 877184 times.
 3508736 p3 = (y + i <= vb_pos + 1) ? p1 : p3;
95
2/2
✓ Branch 0 taken 1315776 times.
✓ Branch 1 taken 438592 times.
 3508736 p5 = (y + i <= vb_pos + 2) ? p3 : p5;
96 }
97
98
2/2
✓ Branch 0 taken 258345472 times.
✓ Branch 1 taken 64586368 times.
 645863680 for (int j = 0; j < ALF_BLOCK_SIZE; j++) {
99 516690944 int sum = 0;
100 516690944 const pixel curr = *p0;
101
102 516690944 sum += filter[0] * FUNC(alf_clip)(curr, p5[+0], p6[+0], clip[0]);
103 516690944 sum += filter[1] * FUNC(alf_clip)(curr, p3[+1], p4[-1], clip[1]);
104 516690944 sum += filter[2] * FUNC(alf_clip)(curr, p3[+0], p4[+0], clip[2]);
105 516690944 sum += filter[3] * FUNC(alf_clip)(curr, p3[-1], p4[+1], clip[3]);
106 516690944 sum += filter[4] * FUNC(alf_clip)(curr, p1[+2], p2[-2], clip[4]);
107 516690944 sum += filter[5] * FUNC(alf_clip)(curr, p1[+1], p2[-1], clip[5]);
108 516690944 sum += filter[6] * FUNC(alf_clip)(curr, p1[+0], p2[+0], clip[6]);
109 516690944 sum += filter[7] * FUNC(alf_clip)(curr, p1[-1], p2[+1], clip[7]);
110 516690944 sum += filter[8] * FUNC(alf_clip)(curr, p1[-2], p2[+2], clip[8]);
111 516690944 sum += filter[9] * FUNC(alf_clip)(curr, p0[+3], p0[-3], clip[9]);
112 516690944 sum += filter[10] * FUNC(alf_clip)(curr, p0[+2], p0[-2], clip[10]);
113 516690944 sum += filter[11] * FUNC(alf_clip)(curr, p0[+1], p0[-1], clip[11]);
114
115
2/2
✓ Branch 0 taken 254836736 times.
✓ Branch 1 taken 3508736 times.
 516690944 if (!is_near_vb)
116 509673472 sum = (sum + offset) >> shift;
117 else
118 7017472 sum = (sum + (1 << ((shift + 3) - 1))) >> (shift + 3);
119 516690944 sum += curr;
120 516690944 dst[j] = CLIP(sum);
121
122 516690944 p0++;
123 516690944 p1++;
124 516690944 p2++;
125 516690944 p3++;
126 516690944 p4++;
127 516690944 p5++;
128 516690944 p6++;
129 }
130 }
131 32293184 filter += ALF_NUM_COEFF_LUMA;
132 32293184 clip += ALF_NUM_COEFF_LUMA;
133 }
134 }
135 34962 }
136
137 41162 static void FUNC(alf_filter_chroma)(uint8_t* _dst, ptrdiff_t dst_stride, const uint8_t* _src, ptrdiff_t src_stride,
138 const int width, const int height, const int16_t* filter, const int16_t* clip, const int vb_pos)
139 {
140 41162 const pixel *src = (pixel *)_src;
141 41162 const int shift = 7;
142 41162 const int offset = 1 << ( shift - 1 );
143 41162 const int vb_above = vb_pos - 2;
144 41162 const int vb_below = vb_pos + 1;
145
146 41162 dst_stride /= sizeof(pixel);
147 41162 src_stride /= sizeof(pixel);
148
149
2/2
✓ Branch 0 taken 405381 times.
✓ Branch 1 taken 20581 times.
 851924 for (int y = 0; y < height; y += ALF_BLOCK_SIZE) {
150
2/2
✓ Branch 0 taken 9464936 times.
✓ Branch 1 taken 405381 times.
 19740634 for (int x = 0; x < width; x += ALF_BLOCK_SIZE) {
151 18929872 const pixel *s0 = src + y * src_stride + x;
152 18929872 const pixel *s1 = s0 + src_stride;
153 18929872 const pixel *s2 = s0 - src_stride;
154 18929872 const pixel *s3 = s1 + src_stride;
155 18929872 const pixel *s4 = s2 - src_stride;
156 18929872 const pixel *s5 = s3 + src_stride;
157 18929872 const pixel *s6 = s4 - src_stride;
158
159
2/2
✓ Branch 0 taken 37859744 times.
✓ Branch 1 taken 9464936 times.
 94649360 for (int i = 0; i < ALF_BLOCK_SIZE; i++) {
160 75719488 pixel *dst = (pixel *)_dst + (y + i) * dst_stride + x;
161
162 75719488 const pixel *p0 = s0 + i * src_stride;
163 75719488 const pixel *p1 = s1 + i * src_stride;
164 75719488 const pixel *p2 = s2 + i * src_stride;
165 75719488 const pixel *p3 = s3 + i * src_stride;
166 75719488 const pixel *p4 = s4 + i * src_stride;
167 75719488 const pixel *p5 = s5 + i * src_stride;
168 75719488 const pixel *p6 = s6 + i * src_stride;
169
170
4/4
✓ Branch 0 taken 37148100 times.
✓ Branch 1 taken 711644 times.
✓ Branch 2 taken 355822 times.
✓ Branch 3 taken 36792278 times.
 75719488 const int is_near_vb_above = (y + i < vb_pos) && (y + i >= vb_pos - 1);
171
4/4
✓ Branch 0 taken 711644 times.
✓ Branch 1 taken 37148100 times.
✓ Branch 2 taken 355822 times.
✓ Branch 3 taken 355822 times.
 75719488 const int is_near_vb_below = (y + i >= vb_pos) && (y + i <= vb_pos);
172
4/4
✓ Branch 0 taken 37503922 times.
✓ Branch 1 taken 355822 times.
✓ Branch 2 taken 355822 times.
✓ Branch 3 taken 37148100 times.
 75719488 const int is_near_vb = is_near_vb_above || is_near_vb_below;
173
174
4/4
✓ Branch 0 taken 37148100 times.
✓ Branch 1 taken 711644 times.
✓ Branch 2 taken 711644 times.
✓ Branch 3 taken 36436456 times.
 75719488 if ((y + i < vb_pos) && ((y + i) >= vb_above)) {
175
2/2
✓ Branch 0 taken 355822 times.
✓ Branch 1 taken 355822 times.
 1423288 p1 = (y + i == vb_pos - 1) ? p0 : p1;
176
1/2
✓ Branch 0 taken 711644 times.
✗ Branch 1 not taken.
 1423288 p3 = (y + i >= vb_pos - 2) ? p1 : p3;
177
1/2
✓ Branch 0 taken 711644 times.
✗ Branch 1 not taken.
 1423288 p5 = (y + i >= vb_pos - 3) ? p3 : p5;
178
179
2/2
✓ Branch 0 taken 355822 times.
✓ Branch 1 taken 355822 times.
 1423288 p2 = (y + i == vb_pos - 1) ? p0 : p2;
180
1/2
✓ Branch 0 taken 711644 times.
✗ Branch 1 not taken.
 1423288 p4 = (y + i >= vb_pos - 2) ? p2 : p4;
181
1/2
✓ Branch 0 taken 711644 times.
✗ Branch 1 not taken.
 1423288 p6 = (y + i >= vb_pos - 3) ? p4 : p6;
182
3/4
✓ Branch 0 taken 711644 times.
✓ Branch 1 taken 36436456 times.
✓ Branch 2 taken 711644 times.
✗ Branch 3 not taken.
 74296200 } else if ((y + i >= vb_pos) && ((y + i) <= vb_below)) {
183
2/2
✓ Branch 0 taken 355822 times.
✓ Branch 1 taken 355822 times.
 1423288 p2 = (y + i == vb_pos ) ? p0 : p2;
184
1/2
✓ Branch 0 taken 711644 times.
✗ Branch 1 not taken.
 1423288 p4 = (y + i <= vb_pos + 1) ? p2 : p4;
185
1/2
✓ Branch 0 taken 711644 times.
✗ Branch 1 not taken.
 1423288 p6 = (y + i <= vb_pos + 2) ? p4 : p6;
186
187
2/2
✓ Branch 0 taken 355822 times.
✓ Branch 1 taken 355822 times.
 1423288 p1 = (y + i == vb_pos ) ? p0 : p1;
188
1/2
✓ Branch 0 taken 711644 times.
✗ Branch 1 not taken.
 1423288 p3 = (y + i <= vb_pos + 1) ? p1 : p3;
189
1/2
✓ Branch 0 taken 711644 times.
✗ Branch 1 not taken.
 1423288 p5 = (y + i <= vb_pos + 2) ? p3 : p5;
190 }
191
192
2/2
✓ Branch 0 taken 151438976 times.
✓ Branch 1 taken 37859744 times.
 378597440 for (int j = 0; j < ALF_BLOCK_SIZE; j++) {
193 302877952 int sum = 0;
194 302877952 const pixel curr = *p0;
195
196 302877952 sum += filter[0] * FUNC(alf_clip)(curr, p3[+0], p4[+0], clip[0]);
197 302877952 sum += filter[1] * FUNC(alf_clip)(curr, p1[+1], p2[-1], clip[1]);
198 302877952 sum += filter[2] * FUNC(alf_clip)(curr, p1[+0], p2[+0], clip[2]);
199 302877952 sum += filter[3] * FUNC(alf_clip)(curr, p1[-1], p2[+1], clip[3]);
200 302877952 sum += filter[4] * FUNC(alf_clip)(curr, p0[+2], p0[-2], clip[4]);
201 302877952 sum += filter[5] * FUNC(alf_clip)(curr, p0[+1], p0[-1], clip[5]);
202
203
2/2
✓ Branch 0 taken 148592400 times.
✓ Branch 1 taken 2846576 times.
 302877952 if (!is_near_vb)
204 297184800 sum = (sum + offset) >> shift;
205 else
206 5693152 sum = (sum + (1 << ((shift + 3) - 1))) >> (shift + 3);
207 302877952 sum += curr;
208 302877952 dst[j] = CLIP(sum);
209
210 302877952 p0++;
211 302877952 p1++;
212 302877952 p2++;
213 302877952 p3++;
214 302877952 p4++;
215 302877952 p5++;
216 302877952 p6++;
217 }
218 }
219 }
220 }
221 41162 }
222
223 16704 static void FUNC(alf_filter_cc)(uint8_t *_dst, ptrdiff_t dst_stride, const uint8_t *_luma, const ptrdiff_t luma_stride,
224 const int width, const int height, const int hs, const int vs, const int16_t *filter, const int vb_pos)
225 {
226 16704 const ptrdiff_t stride = luma_stride / sizeof(pixel);
227
228 16704 dst_stride /= sizeof(pixel);
229
230
2/2
✓ Branch 0 taken 652124 times.
✓ Branch 1 taken 8352 times.
 1320952 for (int y = 0; y < height; y++) {
231
2/2
✓ Branch 0 taken 60583680 times.
✓ Branch 1 taken 652124 times.
 122471608 for (int x = 0; x < width; x++) {
232 121167360 int sum = 0;
233 121167360 pixel *dst = (pixel *)_dst + y * dst_stride + x;
234 121167360 const pixel *src = (pixel *)_luma + (y << vs) * stride + (x << hs);
235
236 121167360 const pixel *s0 = src - stride;
237 121167360 const pixel *s1 = src;
238 121167360 const pixel *s2 = src + stride;
239 121167360 const pixel *s3 = src + 2 * stride;
240
241 121167360 const int pos = y << vs;
242
6/6
✓ Branch 0 taken 38562816 times.
✓ Branch 1 taken 22020864 times.
✓ Branch 2 taken 38280192 times.
✓ Branch 3 taken 282624 times.
✓ Branch 4 taken 282624 times.
✓ Branch 5 taken 37997568 times.
 121167360 if (!vs && (pos == vb_pos || pos == vb_pos + 1))
243 1130496 continue;
244
245
3/4
✓ Branch 0 taken 59419296 times.
✓ Branch 1 taken 599136 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 59419296 times.
 120036864 if (pos == (vb_pos - 2) || pos == (vb_pos + 1))
246 1198272 s3 = s2;
247
4/4
✓ Branch 0 taken 59136672 times.
✓ Branch 1 taken 282624 times.
✓ Branch 2 taken 316512 times.
✓ Branch 3 taken 58820160 times.
 118838592 else if (pos == (vb_pos - 1) || pos == vb_pos)
248 1198272 s3 = s2 = s0 = s1;
249
250
251 120036864 sum += filter[0] * (*s0 - *src);
252 120036864 sum += filter[1] * (*(s1 - 1) - *src);
253 120036864 sum += filter[2] * (*(s1 + 1) - *src);
254 120036864 sum += filter[3] * (*(s2 - 1) - *src);
255 120036864 sum += filter[4] * (*s2 - *src);
256 120036864 sum += filter[5] * (*(s2 + 1) - *src);
257 120036864 sum += filter[6] * (*s3 - *src);
258 120036864 sum = av_clip((sum + 64) >> 7, -(1 << (BIT_DEPTH - 1)), (1 << (BIT_DEPTH - 1)) - 1);
259 120036864 sum += *dst;
260 120036864 *dst = av_clip_pixel(sum);
261 }
262 }
263 16704 }
264
265 #define ALF_DIR_VERT 0
266 #define ALF_DIR_HORZ 1
267 #define ALF_DIR_DIGA0 2
268 #define ALF_DIR_DIGA1 3
269
270 32293184 static void FUNC(alf_get_idx)(int *class_idx, int *transpose_idx, const int *sum, const int ac)
271 {
272 static const int arg_var[] = {0, 1, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 4 };
273
274 int hv0, hv1, dir_hv, d0, d1, dir_d, hvd1, hvd0, sum_hv, dir1;
275
276 32293184 dir_hv = sum[ALF_DIR_VERT] <= sum[ALF_DIR_HORZ];
277 32293184 hv1 = FFMAX(sum[ALF_DIR_VERT], sum[ALF_DIR_HORZ]);
278 32293184 hv0 = FFMIN(sum[ALF_DIR_VERT], sum[ALF_DIR_HORZ]);
279
280 32293184 dir_d = sum[ALF_DIR_DIGA0] <= sum[ALF_DIR_DIGA1];
281 32293184 d1 = FFMAX(sum[ALF_DIR_DIGA0], sum[ALF_DIR_DIGA1]);
282 32293184 d0 = FFMIN(sum[ALF_DIR_DIGA0], sum[ALF_DIR_DIGA1]);
283
284 //promote to avoid overflow
285 32293184 dir1 = (uint64_t)d1 * hv0 <= (uint64_t)hv1 * d0;
286
2/2
✓ Branch 0 taken 11930544 times.
✓ Branch 1 taken 4216048 times.
 32293184 hvd1 = dir1 ? hv1 : d1;
287
2/2
✓ Branch 0 taken 11930544 times.
✓ Branch 1 taken 4216048 times.
 32293184 hvd0 = dir1 ? hv0 : d0;
288
289 32293184 sum_hv = sum[ALF_DIR_HORZ] + sum[ALF_DIR_VERT];
290 32293184 *class_idx = arg_var[av_clip_uintp2(sum_hv * ac >> (BIT_DEPTH - 1), 4)];
291
2/2
✓ Branch 0 taken 3185473 times.
✓ Branch 1 taken 12961119 times.
 32293184 if (hvd1 * 2 > 9 * hvd0)
292 6370946 *class_idx += ((dir1 << 1) + 2) * 5;
293
2/2
✓ Branch 0 taken 4293836 times.
✓ Branch 1 taken 8667283 times.
 25922238 else if (hvd1 > 2 * hvd0)
294 8587672 *class_idx += ((dir1 << 1) + 1) * 5;
295
296 32293184 *transpose_idx = dir_d * 2 + dir_hv;
297 32293184 }
298
299 34962 static void FUNC(alf_classify)(int *class_idx, int *transpose_idx,
300 const uint8_t *_src, const ptrdiff_t _src_stride, const int width, const int height,
301 const int vb_pos, int *gradient_tmp)
302 {
303 int *grad;
304
305 34962 const int h = height + ALF_GRADIENT_BORDER * 2;
306 34962 const int w = width + ALF_GRADIENT_BORDER * 2;
307 34962 const int size = (ALF_BLOCK_SIZE + ALF_GRADIENT_BORDER * 2) / ALF_GRADIENT_STEP;
308 34962 const int gstride = (w / ALF_GRADIENT_STEP) * ALF_NUM_DIR;
309
310 34962 const pixel *src = (const pixel *)_src;
311 34962 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
312 34962 src -= (ALF_GRADIENT_BORDER + 1) * src_stride + ALF_GRADIENT_BORDER;
313
314 34962 grad = gradient_tmp;
315
2/2
✓ Branch 0 taken 1076170 times.
✓ Branch 1 taken 17481 times.
 2187302 for (int y = 0; y < h; y += ALF_GRADIENT_STEP) {
316 2152340 const pixel *s0 = src + y * src_stride;
317 2152340 const pixel *s1 = s0 + src_stride;
318 2152340 const pixel *s2 = s1 + src_stride;
319 2152340 const pixel *s3 = s2 + src_stride;
320
321
2/2
✓ Branch 0 taken 14014 times.
✓ Branch 1 taken 1062156 times.
 2152340 if (y == vb_pos) //above
322 28028 s3 = s2;
323
2/2
✓ Branch 0 taken 14014 times.
✓ Branch 1 taken 1048142 times.
 2124312 else if (y == vb_pos + ALF_GRADIENT_BORDER)
324 28028 s0 = s1;
325
326
2/2
✓ Branch 0 taken 68907540 times.
✓ Branch 1 taken 1076170 times.
 139967420 for (int x = 0; x < w; x += ALF_GRADIENT_STEP) {
327 //two points a time
328 137815080 const pixel *a0 = s0 + x;
329 137815080 const pixel *p0 = s1 + x;
330 137815080 const pixel *b0 = s2 + x;
331 137815080 const int val0 = (*p0) << 1;
332
333 137815080 const pixel *a1 = s1 + x + 1;
334 137815080 const pixel *p1 = s2 + x + 1;
335 137815080 const pixel *b1 = s3 + x + 1;
336 137815080 const int val1 = (*p1) << 1;
337
338 137815080 grad[ALF_DIR_VERT] = FFABS(val0 - *a0 - *b0) + FFABS(val1 - *a1 - *b1);
339 137815080 grad[ALF_DIR_HORZ] = FFABS(val0 - *(p0 - 1) - *(p0 + 1)) + FFABS(val1 - *(p1 - 1) - *(p1 + 1));
340 137815080 grad[ALF_DIR_DIGA0] = FFABS(val0 - *(a0 - 1) - *(b0 + 1)) + FFABS(val1 - *(a1 - 1) - *(b1 + 1));
341 137815080 grad[ALF_DIR_DIGA1] = FFABS(val0 - *(a0 + 1) - *(b0 - 1)) + FFABS(val1 - *(a1 + 1) - *(b1 - 1));
342 137815080 grad += ALF_NUM_DIR;
343 }
344 }
345
346
2/2
✓ Branch 0 taken 520604 times.
✓ Branch 1 taken 17481 times.
 1076170 for (int y = 0; y < height ; y += ALF_BLOCK_SIZE ) {
347 1041208 int start = 0;
348 1041208 int end = (ALF_BLOCK_SIZE + ALF_GRADIENT_BORDER * 2) / ALF_GRADIENT_STEP;
349 1041208 int ac = 2;
350
2/2
✓ Branch 0 taken 14014 times.
✓ Branch 1 taken 506590 times.
 1041208 if (y + ALF_BLOCK_SIZE == vb_pos) {
351 28028 end -= ALF_GRADIENT_BORDER / ALF_GRADIENT_STEP;
352 28028 ac = 3;
353
2/2
✓ Branch 0 taken 14014 times.
✓ Branch 1 taken 492576 times.
 1013180 } else if (y == vb_pos) {
354 28028 start += ALF_GRADIENT_BORDER / ALF_GRADIENT_STEP;
355 28028 ac = 3;
356 }
357
2/2
✓ Branch 0 taken 16146592 times.
✓ Branch 1 taken 520604 times.
 33334392 for (int x = 0; x < width; x += ALF_BLOCK_SIZE) {
358 32293184 const int xg = x / ALF_GRADIENT_STEP;
359 32293184 const int yg = y / ALF_GRADIENT_STEP;
360 32293184 int sum[ALF_NUM_DIR] = { 0 };
361
362 32293184 grad = gradient_tmp + (yg + start) * gstride + xg * ALF_NUM_DIR;
363 //todo: optimize this loop
364
2/2
✓ Branch 0 taken 63709184 times.
✓ Branch 1 taken 16146592 times.
 159711552 for (int i = start; i < end; i++) {
365
2/2
✓ Branch 0 taken 254836736 times.
✓ Branch 1 taken 63709184 times.
 637091840 for (int j = 0; j < size; j++) {
366 509673472 sum[ALF_DIR_VERT] += grad[ALF_DIR_VERT];
367 509673472 sum[ALF_DIR_HORZ] += grad[ALF_DIR_HORZ];
368 509673472 sum[ALF_DIR_DIGA0] += grad[ALF_DIR_DIGA0];
369 509673472 sum[ALF_DIR_DIGA1] += grad[ALF_DIR_DIGA1];
370 509673472 grad += ALF_NUM_DIR;
371 }
372 127418368 grad += gstride - size * ALF_NUM_DIR;
373 }
374 32293184 FUNC(alf_get_idx)(class_idx, transpose_idx, sum, ac);
375
376 32293184 class_idx++;
377 32293184 transpose_idx++;
378 }
379 }
380
381 34962 }
382
383 34962 static void FUNC(alf_recon_coeff_and_clip)(int16_t *coeff, int16_t *clip,
384 const int *class_idx, const int *transpose_idx, const int size,
385 const int16_t *coeff_set, const uint8_t *clip_idx_set, const uint8_t *class_to_filt)
386 {
387 const static int index[][ALF_NUM_COEFF_LUMA] = {
388 { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 },
389 { 9, 4, 10, 8, 1, 5, 11, 7, 3, 0, 2, 6 },
390 { 0, 3, 2, 1, 8, 7, 6, 5, 4, 9, 10, 11 },
391 { 9, 8, 10, 4, 3, 7, 11, 5, 1, 0, 2, 6 },
392 };
393
394 34962 const int16_t clip_set[] = {
395 1 << BIT_DEPTH, 1 << (BIT_DEPTH - 3), 1 << (BIT_DEPTH - 5), 1 << (BIT_DEPTH - 7)
396 };
397
398
2/2
✓ Branch 0 taken 16146592 times.
✓ Branch 1 taken 17481 times.
 32328146 for (int i = 0; i < size; i++) {
399 32293184 const int16_t *src_coeff = coeff_set + class_to_filt[class_idx[i]] * ALF_NUM_COEFF_LUMA;
400 32293184 const uint8_t *clip_idx = clip_idx_set + class_idx[i] * ALF_NUM_COEFF_LUMA;
401
402
2/2
✓ Branch 0 taken 193759104 times.
✓ Branch 1 taken 16146592 times.
 419811392 for (int j = 0; j < ALF_NUM_COEFF_LUMA; j++) {
403 387518208 const int idx = index[transpose_idx[i]][j];
404 387518208 *coeff++ = src_coeff[idx];
405 387518208 *clip++ = clip_set[clip_idx[idx]];
406 }
407 }
408 34962 }
409
410 #undef ALF_DIR_HORZ
411 #undef ALF_DIR_VERT
412 #undef ALF_DIR_DIGA0
413 #undef ALF_DIR_DIGA1
414
415 // line zero
416 #define P7 pix[-8 * xstride]
417 #define P6 pix[-7 * xstride]
418 #define P5 pix[-6 * xstride]
419 #define P4 pix[-5 * xstride]
420 #define P3 pix[-4 * xstride]
421 #define P2 pix[-3 * xstride]
422 #define P1 pix[-2 * xstride]
423 #define P0 pix[-1 * xstride]
424 #define Q0 pix[0 * xstride]
425 #define Q1 pix[1 * xstride]
426 #define Q2 pix[2 * xstride]
427 #define Q3 pix[3 * xstride]
428 #define Q4 pix[4 * xstride]
429 #define Q5 pix[5 * xstride]
430 #define Q6 pix[6 * xstride]
431 #define Q7 pix[7 * xstride]
432 #define P(x) pix[(-(x)-1) * xstride]
433 #define Q(x) pix[(x) * xstride]
434
435 // line three. used only for deblocking decision
436 #define TP7 pix[-8 * xstride + 3 * ystride]
437 #define TP6 pix[-7 * xstride + 3 * ystride]
438 #define TP5 pix[-6 * xstride + 3 * ystride]
439 #define TP4 pix[-5 * xstride + 3 * ystride]
440 #define TP3 pix[-4 * xstride + 3 * ystride]
441 #define TP2 pix[-3 * xstride + 3 * ystride]
442 #define TP1 pix[-2 * xstride + 3 * ystride]
443 #define TP0 pix[-1 * xstride + 3 * ystride]
444 #define TQ0 pix[0 * xstride + 3 * ystride]
445 #define TQ1 pix[1 * xstride + 3 * ystride]
446 #define TQ2 pix[2 * xstride + 3 * ystride]
447 #define TQ3 pix[3 * xstride + 3 * ystride]
448 #define TQ4 pix[4 * xstride + 3 * ystride]
449 #define TQ5 pix[5 * xstride + 3 * ystride]
450 #define TQ6 pix[6 * xstride + 3 * ystride]
451 #define TQ7 pix[7 * xstride + 3 * ystride]
452 #define TP(x) pix[(-(x)-1) * xstride + 3 * ystride]
453 #define TQ(x) pix[(x) * xstride + 3 * ystride]
454
455 #define FP3 pix[-4 * xstride + 1 * ystride]
456 #define FP2 pix[-3 * xstride + 1 * ystride]
457 #define FP1 pix[-2 * xstride + 1 * ystride]
458 #define FP0 pix[-1 * xstride + 1 * ystride]
459 #define FQ0 pix[0 * xstride + 1 * ystride]
460 #define FQ1 pix[1 * xstride + 1 * ystride]
461 #define FQ2 pix[2 * xstride + 1 * ystride]
462 #define FQ3 pix[3 * xstride + 1 * ystride]
463
464 #include "libavcodec/h26x/h2656_deblock_template.c"
465
466 1367762 static void FUNC(loop_filter_luma_large)(pixel *pix, const ptrdiff_t xstride, const ptrdiff_t ystride, const int32_t tc,
467 const uint8_t no_p, const uint8_t no_q, const uint8_t max_len_p, const uint8_t max_len_q)
468 {
469
2/2
✓ Branch 0 taken 2735524 times.
✓ Branch 1 taken 683881 times.
 6838810 for (int d = 0; d < 4; d++) {
470 5471048 const int p6 = P6;
471 5471048 const int p5 = P5;
472 5471048 const int p4 = P4;
473 5471048 const int p3 = P3;
474 5471048 const int p2 = P2;
475 5471048 const int p1 = P1;
476 5471048 const int p0 = P0;
477 5471048 const int q0 = Q0;
478 5471048 const int q1 = Q1;
479 5471048 const int q2 = Q2;
480 5471048 const int q3 = Q3;
481 5471048 const int q4 = Q4;
482 5471048 const int q5 = Q5;
483 5471048 const int q6 = Q6;
484 int m;
485
4/4
✓ Branch 0 taken 415440 times.
✓ Branch 1 taken 2320084 times.
✓ Branch 2 taken 235756 times.
✓ Branch 3 taken 179684 times.
 5471048 if (max_len_p == 5 && max_len_q == 5)
486 471512 m = (p4 + p3 + 2 * (p2 + p1 + p0 + q0 + q1 + q2) + q3 + q4 + 8) >> 4;
487
2/2
✓ Branch 0 taken 1133860 times.
✓ Branch 1 taken 1365908 times.
 4999536 else if (max_len_p == max_len_q)
488 2267720 m = (p6 + p5 + p4 + p3 + p2 + p1 + 2 * (p0 + q0) + q1 + q2 + q3 + q4 + q5 + q6 + 8) >> 4;
489
2/2
✓ Branch 0 taken 265300 times.
✓ Branch 1 taken 1100608 times.
 2731816 else if (max_len_p + max_len_q == 12)
490 530600 m = (p5 + p4 + p3 + p2 + 2 * (p1 + p0 + q0 + q1) + q2 + q3 + q4 + q5 + 8) >> 4;
491
2/2
✓ Branch 0 taken 259072 times.
✓ Branch 1 taken 841536 times.
 2201216 else if (max_len_p + max_len_q == 8)
492 518144 m = (p3 + p2 + p1 + p0 + q0 + q1 + q2 + q3 + 4) >> 3;
493
2/2
✓ Branch 0 taken 617740 times.
✓ Branch 1 taken 223796 times.
 1683072 else if (max_len_q == 7)
494 1235480 m = (2 * (p2 + p1 + p0 + q0) + p0 + p1 + q1 + q2 + q3 + q4 + q5 + q6 + 8) >> 4;
495 else
496 447592 m = (p6 + p5 + p4 + p3 + p2 + p1 + 2 * (q2 + q1 + q0 + p0) + q0 + q1 + 8) >> 4;
497
1/2
✓ Branch 0 taken 2735524 times.
✗ Branch 1 not taken.
 5471048 if (!no_p) {
498 5471048 const int refp = (P(max_len_p) + P(max_len_p - 1) + 1) >> 1;
499
2/2
✓ Branch 0 taken 826460 times.
✓ Branch 1 taken 1909064 times.
 5471048 if (max_len_p == 3) {
500 1652920 P0 = p0 + av_clip(((m * 53 + refp * 11 + 32) >> 6) - p0, -(tc * 6 >> 1), (tc * 6 >> 1));
501 1652920 P1 = p1 + av_clip(((m * 32 + refp * 32 + 32) >> 6) - p1, -(tc * 4 >> 1), (tc * 4 >> 1));
502 1652920 P2 = p2 + av_clip(((m * 11 + refp * 53 + 32) >> 6) - p2, -(tc * 2 >> 1), (tc * 2 >> 1));
503
2/2
✓ Branch 0 taken 415440 times.
✓ Branch 1 taken 1493624 times.
 3818128 } else if (max_len_p == 5) {
504 830880 P0 = p0 + av_clip(((m * 58 + refp * 6 + 32) >> 6) - p0, -(tc * 6 >> 1), (tc * 6 >> 1));
505 830880 P1 = p1 + av_clip(((m * 45 + refp * 19 + 32) >> 6) - p1, -(tc * 5 >> 1), (tc * 5 >> 1));
506 830880 P2 = p2 + av_clip(((m * 32 + refp * 32 + 32) >> 6) - p2, -(tc * 4 >> 1), (tc * 4 >> 1));
507 830880 P3 = p3 + av_clip(((m * 19 + refp * 45 + 32) >> 6) - p3, -(tc * 3 >> 1), (tc * 3 >> 1));
508 830880 P4 = p4 + av_clip(((m * 6 + refp * 58 + 32) >> 6) - p4, -(tc * 2 >> 1), (tc * 2 >> 1));
509 } else {
510 2987248 P0 = p0 + av_clip(((m * 59 + refp * 5 + 32) >> 6) - p0, -(tc * 6 >> 1), (tc * 6 >> 1));
511 2987248 P1 = p1 + av_clip(((m * 50 + refp * 14 + 32) >> 6) - p1, -(tc * 5 >> 1), (tc * 5 >> 1));
512 2987248 P2 = p2 + av_clip(((m * 41 + refp * 23 + 32) >> 6) - p2, -(tc * 4 >> 1), (tc * 4 >> 1));
513 2987248 P3 = p3 + av_clip(((m * 32 + refp * 32 + 32) >> 6) - p3, -(tc * 3 >> 1), (tc * 3 >> 1));
514 2987248 P4 = p4 + av_clip(((m * 23 + refp * 41 + 32) >> 6) - p4, -(tc * 2 >> 1), (tc * 2 >> 1));
515 2987248 P5 = p5 + av_clip(((m * 14 + refp * 50 + 32) >> 6) - p5, -(tc * 1 >> 1), (tc * 1 >> 1));
516 2987248 P6 = p6 + av_clip(((m * 5 + refp * 59 + 32) >> 6) - p6, -(tc * 1 >> 1), (tc * 1 >> 1));
517 }
518 }
519
1/2
✓ Branch 0 taken 2735524 times.
✗ Branch 1 not taken.
 5471048 if (!no_q) {
520 5471048 const int refq = (Q(max_len_q) + Q(max_len_q - 1) + 1) >> 1;
521
2/2
✓ Branch 0 taken 274148 times.
✓ Branch 1 taken 2461376 times.
 5471048 if (max_len_q == 3) {
522 548296 Q0 = q0 + av_clip(((m * 53 + refq * 11 + 32) >> 6) - q0, -(tc * 6 >> 1), (tc * 6 >> 1));
523 548296 Q1 = q1 + av_clip(((m * 32 + refq * 32 + 32) >> 6) - q1, -(tc * 4 >> 1), (tc * 4 >> 1));
524 548296 Q2 = q2 + av_clip(((m * 11 + refq * 53 + 32) >> 6) - q2, -(tc * 2 >> 1), (tc * 2 >> 1));
525
2/2
✓ Branch 0 taken 580444 times.
✓ Branch 1 taken 1880932 times.
 4922752 } else if (max_len_q == 5) {
526 1160888 Q0 = q0 + av_clip(((m * 58 + refq * 6 + 32) >> 6) - q0, -(tc * 6 >> 1), (tc * 6 >> 1));
527 1160888 Q1 = q1 + av_clip(((m * 45 + refq * 19 + 32) >> 6) - q1, -(tc * 5 >> 1), (tc * 5 >> 1));
528 1160888 Q2 = q2 + av_clip(((m * 32 + refq * 32 + 32) >> 6) - q2, -(tc * 4 >> 1), (tc * 4 >> 1));
529 1160888 Q3 = q3 + av_clip(((m * 19 + refq * 45 + 32) >> 6) - q3, -(tc * 3 >> 1), (tc * 3 >> 1));
530 1160888 Q4 = q4 + av_clip(((m * 6 + refq * 58 + 32) >> 6) - q4, -(tc * 2 >> 1), (tc * 2 >> 1));
531 } else {
532 3761864 Q0 = q0 + av_clip(((m * 59 + refq * 5 + 32) >> 6) - q0, -(tc * 6 >> 1), (tc * 6 >> 1));
533 3761864 Q1 = q1 + av_clip(((m * 50 + refq * 14 + 32) >> 6) - q1, -(tc * 5 >> 1), (tc * 5 >> 1));
534 3761864 Q2 = q2 + av_clip(((m * 41 + refq * 23 + 32) >> 6) - q2, -(tc * 4 >> 1), (tc * 4 >> 1));
535 3761864 Q3 = q3 + av_clip(((m * 32 + refq * 32 + 32) >> 6) - q3, -(tc * 3 >> 1), (tc * 3 >> 1));
536 3761864 Q4 = q4 + av_clip(((m * 23 + refq * 41 + 32) >> 6) - q4, -(tc * 2 >> 1), (tc * 2 >> 1));
537 3761864 Q5 = q5 + av_clip(((m * 14 + refq * 50 + 32) >> 6) - q5, -(tc * 1 >> 1), (tc * 1 >> 1));
538 3761864 Q6 = q6 + av_clip(((m * 5 + refq * 59 + 32) >> 6) - q6, -(tc * 1 >> 1), (tc * 1 >> 1));
539 }
540
541 }
542 5471048 pix += ystride;
543 }
544 1367762 }
545
546 6221380 static void FUNC(vvc_loop_filter_luma)(uint8_t* _pix, ptrdiff_t _xstride, ptrdiff_t _ystride,
547 const int32_t *_beta, const int32_t *_tc, const uint8_t *_no_p, const uint8_t *_no_q,
548 const uint8_t *_max_len_p, const uint8_t *_max_len_q, const int hor_ctu_edge)
549 {
550 6221380 const ptrdiff_t xstride = _xstride / sizeof(pixel);
551 6221380 const ptrdiff_t ystride = _ystride / sizeof(pixel);
552
553
2/2
✓ Branch 0 taken 6221380 times.
✓ Branch 1 taken 3110690 times.
 18664140 for (int i = 0; i < 2; i++) {
554 #if BIT_DEPTH < 10
555 31856 const int tc = (_tc[i] + (1 << (9 - BIT_DEPTH))) >> (10 - BIT_DEPTH);
556 #else
557 12410904 const int tc = _tc[i] << (BIT_DEPTH - 10);
558 #endif
559
2/2
✓ Branch 0 taken 6097542 times.
✓ Branch 1 taken 123838 times.
 12442760 if (tc) {
560 12195084 pixel* pix = (pixel*)_pix + i * 4 * ystride;
561 12195084 const int dp0 = abs(P2 - 2 * P1 + P0);
562 12195084 const int dq0 = abs(Q2 - 2 * Q1 + Q0);
563 12195084 const int dp3 = abs(TP2 - 2 * TP1 + TP0);
564 12195084 const int dq3 = abs(TQ2 - 2 * TQ1 + TQ0);
565 12195084 const int d0 = dp0 + dq0;
566 12195084 const int d3 = dp3 + dq3;
567 12195084 const int tc25 = ((tc * 5 + 1) >> 1);
568
569 12195084 const int no_p = _no_p[i];
570 12195084 const int no_q = _no_q[i];
571
572 12195084 int max_len_p = _max_len_p[i];
573 12195084 int max_len_q = _max_len_q[i];
574
575
4/4
✓ Branch 0 taken 1877745 times.
✓ Branch 1 taken 4219797 times.
✓ Branch 2 taken 1483605 times.
✓ Branch 3 taken 394140 times.
 12195084 const int large_p = (max_len_p > 3 && !hor_ctu_edge);
576 12195084 const int large_q = max_len_q > 3;
577
578 12195084 const int beta = _beta[i] << BIT_DEPTH - 8;
579 12195084 const int beta_3 = beta >> 3;
580 12195084 const int beta_2 = beta >> 2;
581
582
4/4
✓ Branch 0 taken 4613937 times.
✓ Branch 1 taken 1483605 times.
✓ Branch 2 taken 748257 times.
✓ Branch 3 taken 3865680 times.
 12195084 if (large_p || large_q) {
583
2/2
✓ Branch 0 taken 1483605 times.
✓ Branch 1 taken 748257 times.
 4463724 const int dp0l = large_p ? ((dp0 + abs(P5 - 2 * P4 + P3) + 1) >> 1) : dp0;
584
2/2
✓ Branch 0 taken 1861077 times.
✓ Branch 1 taken 370785 times.
 4463724 const int dq0l = large_q ? ((dq0 + abs(Q5 - 2 * Q4 + Q3) + 1) >> 1) : dq0;
585
2/2
✓ Branch 0 taken 1483605 times.
✓ Branch 1 taken 748257 times.
 4463724 const int dp3l = large_p ? ((dp3 + abs(TP5 - 2 * TP4 + TP3) + 1) >> 1) : dp3;
586
2/2
✓ Branch 0 taken 1861077 times.
✓ Branch 1 taken 370785 times.
 4463724 const int dq3l = large_q ? ((dq3 + abs(TQ5 - 2 * TQ4 + TQ3) + 1) >> 1) : dq3;
587 4463724 const int d0l = dp0l + dq0l;
588 4463724 const int d3l = dp3l + dq3l;
589 4463724 const int beta53 = beta * 3 >> 5;
590 4463724 const int beta_4 = beta >> 4;
591
2/2
✓ Branch 0 taken 1483605 times.
✓ Branch 1 taken 748257 times.
 4463724 max_len_p = large_p ? max_len_p : 3;
592
2/2
✓ Branch 0 taken 1861077 times.
✓ Branch 1 taken 370785 times.
 4463724 max_len_q = large_q ? max_len_q : 3;
593
594
2/2
✓ Branch 0 taken 2088454 times.
✓ Branch 1 taken 143408 times.
 4463724 if (d0l + d3l < beta) {
595
2/2
✓ Branch 0 taken 1129796 times.
✓ Branch 1 taken 958658 times.
 4176908 const int sp0l = abs(P3 - P0) + (max_len_p == 7 ? abs(P7 - P6 - P5 + P4) : 0);
596
2/2
✓ Branch 0 taken 1406765 times.
✓ Branch 1 taken 681689 times.
 4176908 const int sq0l = abs(Q0 - Q3) + (max_len_q == 7 ? abs(Q4 - Q5 - Q6 + Q7) : 0);
597
2/2
✓ Branch 0 taken 1129796 times.
✓ Branch 1 taken 958658 times.
 4176908 const int sp3l = abs(TP3 - TP0) + (max_len_p == 7 ? abs(TP7 - TP6 - TP5 + TP4) : 0);
598
2/2
✓ Branch 0 taken 1406765 times.
✓ Branch 1 taken 681689 times.
 4176908 const int sq3l = abs(TQ0 - TQ3) + (max_len_q == 7 ? abs(TQ4 - TQ5 - TQ6 + TQ7) : 0);
599
2/2
✓ Branch 0 taken 1395445 times.
✓ Branch 1 taken 693009 times.
 4176908 const int sp0 = large_p ? ((sp0l + abs(P3 - P(max_len_p)) + 1) >> 1) : sp0l;
600
2/2
✓ Branch 0 taken 1395445 times.
✓ Branch 1 taken 693009 times.
 4176908 const int sp3 = large_p ? ((sp3l + abs(TP3 - TP(max_len_p)) + 1) >> 1) : sp3l;
601
2/2
✓ Branch 0 taken 1757265 times.
✓ Branch 1 taken 331189 times.
 4176908 const int sq0 = large_q ? ((sq0l + abs(Q3 - Q(max_len_q)) + 1) >> 1) : sq0l;
602
2/2
✓ Branch 0 taken 1757265 times.
✓ Branch 1 taken 331189 times.
 4176908 const int sq3 = large_q ? ((sq3l + abs(TQ3 - TQ(max_len_q)) + 1) >> 1) : sq3l;
603
4/4
✓ Branch 0 taken 848076 times.
✓ Branch 1 taken 1240378 times.
✓ Branch 2 taken 846133 times.
✓ Branch 3 taken 1943 times.
 4176908 if (sp0 + sq0 < beta53 && abs(P0 - Q0) < tc25 &&
604
4/4
✓ Branch 0 taken 747660 times.
✓ Branch 1 taken 98473 times.
✓ Branch 2 taken 747073 times.
✓ Branch 3 taken 587 times.
 1692266 sp3 + sq3 < beta53 && abs(TP0 - TQ0) < tc25 &&
605
4/4
✓ Branch 0 taken 705522 times.
✓ Branch 1 taken 41551 times.
✓ Branch 2 taken 683881 times.
✓ Branch 3 taken 21641 times.
 1494146 (d0l << 1) < beta_4 && (d3l << 1) < beta_4) {
606 1367762 FUNC(loop_filter_luma_large)(pix, xstride, ystride, tc, no_p, no_q, max_len_p, max_len_q);
607 1367762 continue;
608 }
609 }
610 }
611
2/2
✓ Branch 0 taken 3993591 times.
✓ Branch 1 taken 1420070 times.
 10827322 if (d0 + d3 < beta) {
612
3/4
✓ Branch 0 taken 2975490 times.
✓ Branch 1 taken 1018101 times.
✓ Branch 2 taken 2975490 times.
✗ Branch 3 not taken.
 7987182 if (max_len_p > 2 && max_len_q > 2 &&
613
4/4
✓ Branch 0 taken 994195 times.
✓ Branch 1 taken 1981295 times.
✓ Branch 2 taken 984230 times.
✓ Branch 3 taken 9965 times.
 5950980 abs(P3 - P0) + abs(Q3 - Q0) < beta_3 && abs(P0 - Q0) < tc25 &&
614
4/4
✓ Branch 0 taken 746133 times.
✓ Branch 1 taken 238097 times.
✓ Branch 2 taken 743701 times.
✓ Branch 3 taken 2432 times.
 1968460 abs(TP3 - TP0) + abs(TQ3 - TQ0) < beta_3 && abs(TP0 - TQ0) < tc25 &&
615
4/4
✓ Branch 0 taken 732769 times.
✓ Branch 1 taken 10932 times.
✓ Branch 2 taken 728143 times.
✓ Branch 3 taken 4626 times.
 1487402 (d0 << 1) < beta_2 && (d3 << 1) < beta_2) {
616 1456286 FUNC(loop_filter_luma_strong)(pix, xstride, ystride, tc, tc << 1, tc * 3, no_p, no_q);
617 } else {
618 6530896 int nd_p = 1;
619 6530896 int nd_q = 1;
620
3/4
✓ Branch 0 taken 2406823 times.
✓ Branch 1 taken 858625 times.
✓ Branch 2 taken 2406823 times.
✗ Branch 3 not taken.
 6530896 if (max_len_p > 1 && max_len_q > 1) {
621
2/2
✓ Branch 0 taken 1946790 times.
✓ Branch 1 taken 460033 times.
 4813646 if (dp0 + dp3 < ((beta + (beta >> 1)) >> 3))
622 3893580 nd_p = 2;
623
2/2
✓ Branch 0 taken 1874648 times.
✓ Branch 1 taken 532175 times.
 4813646 if (dq0 + dq3 < ((beta + (beta >> 1)) >> 3))
624 3749296 nd_q = 2;
625 }
626 6530896 FUNC(loop_filter_luma_weak)(pix, xstride, ystride, tc, beta, no_p, no_q, nd_p, nd_q);
627 }
628 }
629 }
630 }
631 6221380 }
632
633 2580734 static void FUNC(loop_filter_chroma_strong)(pixel *pix, const ptrdiff_t xstride, const ptrdiff_t ystride,
634 const int size, const int32_t tc, const uint8_t no_p, const uint8_t no_q)
635 {
636
2/2
✓ Branch 0 taken 2874144 times.
✓ Branch 1 taken 1290367 times.
 8329022 for (int d = 0; d < size; d++) {
637 5748288 const int p3 = P3;
638 5748288 const int p2 = P2;
639 5748288 const int p1 = P1;
640 5748288 const int p0 = P0;
641 5748288 const int q0 = Q0;
642 5748288 const int q1 = Q1;
643 5748288 const int q2 = Q2;
644 5748288 const int q3 = Q3;
645
1/2
✓ Branch 0 taken 2874144 times.
✗ Branch 1 not taken.
 5748288 if (!no_p) {
646 5748288 P0 = av_clip((p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2 + 4) >> 3, p0 - tc, p0 + tc);
647 5748288 P1 = av_clip((2 * p3 + p2 + 2 * p1 + p0 + q0 + q1 + 4) >> 3, p1 - tc, p1 + tc);
648 5748288 P2 = av_clip((3 * p3 + 2 * p2 + p1 + p0 + q0 + 4) >> 3, p2 - tc, p2 + tc );
649 }
650
1/2
✓ Branch 0 taken 2874144 times.
✗ Branch 1 not taken.
 5748288 if (!no_q) {
651 5748288 Q0 = av_clip((p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3 + 4) >> 3, q0 - tc, q0 + tc);
652 5748288 Q1 = av_clip((p1 + p0 + q0 + 2 * q1 + q2 + 2 * q3 + 4) >> 3, q1 - tc, q1 + tc);
653 5748288 Q2 = av_clip((p0 + q0 + q1 + 2 * q2 + 3 * q3 + 4) >> 3, q2 - tc, q2 + tc);
654 }
655 5748288 pix += ystride;
656 }
657 2580734 }
658
659 467506 static void FUNC(loop_filter_chroma_strong_one_side)(pixel *pix, const ptrdiff_t xstride, const ptrdiff_t ystride,
660 const int size, const int32_t tc, const uint8_t no_p, const uint8_t no_q)
661 {
662
2/2
✓ Branch 0 taken 511328 times.
✓ Branch 1 taken 233753 times.
 1490162 for (int d = 0; d < size; d++) {
663 1022656 const int p1 = P1;
664 1022656 const int p0 = P0;
665 1022656 const int q0 = Q0;
666 1022656 const int q1 = Q1;
667 1022656 const int q2 = Q2;
668 1022656 const int q3 = Q3;
669
1/2
✓ Branch 0 taken 511328 times.
✗ Branch 1 not taken.
 1022656 if (!no_p) {
670 1022656 P0 = av_clip((3 * p1 + 2 * p0 + q0 + q1 + q2 + 4) >> 3, p0 - tc, p0 + tc);
671 }
672
1/2
✓ Branch 0 taken 511328 times.
✗ Branch 1 not taken.
 1022656 if (!no_q) {
673 1022656 Q0 = av_clip((2 * p1 + p0 + 2 * q0 + q1 + q2 + q3 + 4) >> 3, q0 - tc, q0 + tc);
674 1022656 Q1 = av_clip((p1 + p0 + q0 + 2 * q1 + q2 + 2 * q3 + 4) >> 3, q1 - tc, q1 + tc);
675 1022656 Q2 = av_clip((p0 + q0 + q1 + 2 * q2 + 3 * q3 + 4) >> 3, q2 - tc, q2 + tc);
676 }
677 1022656 pix += ystride;
678 }
679 467506 }
680
681 1903456 static void FUNC(vvc_loop_filter_chroma)(uint8_t *_pix, const ptrdiff_t _xstride, const ptrdiff_t _ystride,
682 const int32_t *_beta, const int32_t *_tc, const uint8_t *_no_p, const uint8_t *_no_q,
683 const uint8_t *_max_len_p, const uint8_t *_max_len_q, const int shift)
684 {
685 1903456 const ptrdiff_t xstride = _xstride / sizeof(pixel);
686 1903456 const ptrdiff_t ystride = _ystride / sizeof(pixel);
687
2/2
✓ Branch 0 taken 690843 times.
✓ Branch 1 taken 260885 times.
 1903456 const int size = shift ? 2 : 4;
688 1903456 const int end = 8 / size; // 8 samples a loop
689
690
2/2
✓ Branch 0 taken 3285142 times.
✓ Branch 1 taken 951728 times.
 8473740 for (int i = 0; i < end; i++) {
691 #if BIT_DEPTH < 10
692 23504 const int tc = (_tc[i] + (1 << (9 - BIT_DEPTH))) >> (10 - BIT_DEPTH);
693 #else
694 6546780 const int tc = _tc[i] << (BIT_DEPTH - 10);
695 #endif
696
2/2
✓ Branch 0 taken 3141554 times.
✓ Branch 1 taken 143588 times.
 6570284 if (tc) {
697 6283108 pixel *pix = (pixel *)_pix + i * size * ystride;
698 6283108 const uint8_t no_p = _no_p[i];
699 6283108 const uint8_t no_q = _no_q[i];
700
701 6283108 const int beta = _beta[i] << (BIT_DEPTH - 8);
702 6283108 const int beta_3 = beta >> 3;
703 6283108 const int beta_2 = beta >> 2;
704
705 6283108 const int tc25 = ((tc * 5 + 1) >> 1);
706
707 6283108 uint8_t max_len_p = _max_len_p[i];
708 6283108 uint8_t max_len_q = _max_len_q[i];
709
710
3/4
✓ Branch 0 taken 3067201 times.
✓ Branch 1 taken 74353 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 3067201 times.
 6283108 if (!max_len_p || !max_len_q)
711 148706 continue;
712
713
2/2
✓ Branch 0 taken 2248349 times.
✓ Branch 1 taken 818852 times.
 6134402 if (max_len_q == 3){
714
2/2
✓ Branch 0 taken 1833784 times.
✓ Branch 1 taken 414565 times.
 4496698 const int p1n = shift ? FP1 : TP1;
715
4/4
✓ Branch 0 taken 1966234 times.
✓ Branch 1 taken 282115 times.
✓ Branch 2 taken 1590352 times.
✓ Branch 3 taken 375882 times.
 4496698 const int p2n = max_len_p == 1 ? p1n : (shift ? FP2 : TP2);
716
2/2
✓ Branch 0 taken 1833784 times.
✓ Branch 1 taken 414565 times.
 4496698 const int p0n = shift ? FP0 : TP0;
717
2/2
✓ Branch 0 taken 1833784 times.
✓ Branch 1 taken 414565 times.
 4496698 const int q0n = shift ? FQ0 : TQ0;
718
2/2
✓ Branch 0 taken 1833784 times.
✓ Branch 1 taken 414565 times.
 4496698 const int q1n = shift ? FQ1 : TQ1;
719
2/2
✓ Branch 0 taken 1833784 times.
✓ Branch 1 taken 414565 times.
 4496698 const int q2n = shift ? FQ2 : TQ2;
720
2/2
✓ Branch 0 taken 282115 times.
✓ Branch 1 taken 1966234 times.
 4496698 const int p3 = max_len_p == 1 ? P1 : P3;
721
2/2
✓ Branch 0 taken 282115 times.
✓ Branch 1 taken 1966234 times.
 4496698 const int p2 = max_len_p == 1 ? P1 : P2;
722 4496698 const int p1 = P1;
723 4496698 const int p0 = P0;
724 4496698 const int dp0 = abs(p2 - 2 * p1 + p0);
725 4496698 const int dq0 = abs(Q2 - 2 * Q1 + Q0);
726
727 4496698 const int dp1 = abs(p2n - 2 * p1n + p0n);
728 4496698 const int dq1 = abs(q2n - 2 * q1n + q0n);
729 4496698 const int d0 = dp0 + dq0;
730 4496698 const int d1 = dp1 + dq1;
731
732
2/2
✓ Branch 0 taken 2118593 times.
✓ Branch 1 taken 129756 times.
 4496698 if (d0 + d1 < beta) {
733
4/4
✓ Branch 0 taken 1842904 times.
✓ Branch 1 taken 275689 times.
✓ Branch 2 taken 1558136 times.
✓ Branch 3 taken 284768 times.
 4237186 const int p3n = max_len_p == 1 ? p1n : (shift ? FP3 : TP3);
734
2/2
✓ Branch 0 taken 1799256 times.
✓ Branch 1 taken 319337 times.
 4237186 const int q3n = shift ? FQ3 : TQ3;
735
4/4
✓ Branch 0 taken 1898875 times.
✓ Branch 1 taken 219718 times.
✓ Branch 2 taken 1587412 times.
✓ Branch 3 taken 311463 times.
 7412010 const int dsam0 = (d0 << 1) < beta_2 && (abs(p3 - p0) + abs(Q0 - Q3) < beta_3) &&
736
2/2
✓ Branch 0 taken 1582197 times.
✓ Branch 1 taken 5215 times.
 3174824 abs(p0 - Q0) < tc25;
737
4/4
✓ Branch 0 taken 1900887 times.
✓ Branch 1 taken 217706 times.
✓ Branch 2 taken 1589916 times.
✓ Branch 3 taken 310971 times.
 7417018 const int dsam1 = (d1 << 1) < beta_2 && (abs(p3n - p0n) + abs(q0n - q3n) < beta_3) &&
738
2/2
✓ Branch 0 taken 1584437 times.
✓ Branch 1 taken 5479 times.
 3179832 abs(p0n - q0n) < tc25;
739
4/4
✓ Branch 0 taken 1582197 times.
✓ Branch 1 taken 536396 times.
✓ Branch 2 taken 58077 times.
✓ Branch 3 taken 1524120 times.
 4237186 if (!dsam0 || !dsam1)
740 1188946 max_len_p = max_len_q = 1;
741 } else {
742 259512 max_len_p = max_len_q = 1;
743 }
744 }
745
746
3/4
✓ Branch 0 taken 1290367 times.
✓ Branch 1 taken 1776834 times.
✓ Branch 2 taken 1290367 times.
✗ Branch 3 not taken.
 6134402 if (max_len_p == 3 && max_len_q == 3)
747 2580734 FUNC(loop_filter_chroma_strong)(pix, xstride, ystride, size, tc, no_p, no_q);
748
2/2
✓ Branch 0 taken 233753 times.
✓ Branch 1 taken 1543081 times.
 3553668 else if (max_len_q == 3)
749 467506 FUNC(loop_filter_chroma_strong_one_side)(pix, xstride, ystride, size, tc, no_p, no_q);
750 else
751 3086162 FUNC(loop_filter_chroma_weak)(pix, xstride, ystride, size, tc, no_p, no_q);
752 }
753 }
754 1903456 }
755
756 975946 static void FUNC(vvc_h_loop_filter_chroma)(uint8_t *pix, ptrdiff_t stride,
757 const int32_t *beta, const int32_t *tc, const uint8_t *no_p, const uint8_t *no_q,
758 const uint8_t *max_len_p, const uint8_t *max_len_q, int shift)
759 {
760 975946 FUNC(vvc_loop_filter_chroma)(pix, stride, sizeof(pixel), beta, tc,
761 no_p, no_q, max_len_p, max_len_q, shift);
762 975946 }
763
764 927510 static void FUNC(vvc_v_loop_filter_chroma)(uint8_t *pix, ptrdiff_t stride,
765 const int32_t *beta, const int32_t *tc, const uint8_t *no_p, const uint8_t *no_q,
766 const uint8_t *max_len_p, const uint8_t *max_len_q, int shift)
767 {
768 927510 FUNC(vvc_loop_filter_chroma)(pix, sizeof(pixel), stride, beta, tc,
769 no_p, no_q, max_len_p, max_len_q, shift);
770 927510 }
771
772 3194932 static void FUNC(vvc_h_loop_filter_luma)(uint8_t *pix, ptrdiff_t stride,
773 const int32_t *beta, const int32_t *tc, const uint8_t *no_p, const uint8_t *no_q,
774 const uint8_t *max_len_p, const uint8_t *max_len_q, const int hor_ctu_edge)
775 {
776 3194932 FUNC(vvc_loop_filter_luma)(pix, stride, sizeof(pixel), beta, tc,
777 no_p, no_q, max_len_p, max_len_q, hor_ctu_edge);
778 3194932 }
779
780 3026448 static void FUNC(vvc_v_loop_filter_luma)(uint8_t *pix, ptrdiff_t stride,
781 const int32_t *beta, const int32_t *tc, const uint8_t *no_p, const uint8_t *no_q,
782 const uint8_t *max_len_p, const uint8_t *max_len_q, const int hor_ctu_edge)
783 {
784 3026448 FUNC(vvc_loop_filter_luma)(pix, sizeof(pixel), stride, beta, tc,
785 no_p, no_q, max_len_p, max_len_q, hor_ctu_edge);
786 3026448 }
787
788 ✗ static int FUNC(vvc_loop_ladf_level)(const uint8_t *_pix, const ptrdiff_t _xstride, const ptrdiff_t _ystride)
789 {
790 ✗ const pixel *pix = (pixel *)_pix;
791 ✗ const ptrdiff_t xstride = _xstride / sizeof(pixel);
792 ✗ const ptrdiff_t ystride = _ystride / sizeof(pixel);
793 ✗ return (P0 + TP0 + Q0 + TQ0) >> 2;
794 }
795
796 ✗ static int FUNC(vvc_h_loop_ladf_level)(const uint8_t *pix, ptrdiff_t stride)
797 {
798 ✗ return FUNC(vvc_loop_ladf_level)(pix, stride, sizeof(pixel));
799 }
800
801 ✗ static int FUNC(vvc_v_loop_ladf_level)(const uint8_t *pix, ptrdiff_t stride)
802 {
803 ✗ return FUNC(vvc_loop_ladf_level)(pix, sizeof(pixel), stride);
804 }
805
806 #undef P7
807 #undef P6
808 #undef P5
809 #undef P4
810 #undef P3
811 #undef P2
812 #undef P1
813 #undef P0
814 #undef Q0
815 #undef Q1
816 #undef Q2
817 #undef Q3
818 #undef Q4
819 #undef Q5
820 #undef Q6
821 #undef Q7
822
823 #undef TP7
824 #undef TP6
825 #undef TP5
826 #undef TP4
827 #undef TP3
828 #undef TP2
829 #undef TP1
830 #undef TP0
831 #undef TQ0
832 #undef TQ1
833 #undef TQ2
834 #undef TQ3
835 #undef TQ4
836 #undef TQ5
837 #undef TQ6
838 #undef TQ7
839
840 2134 static void FUNC(ff_vvc_lmcs_dsp_init)(VVCLMCSDSPContext *const lmcs)
841 {
842 2134 lmcs->filter = FUNC(lmcs_filter_luma);
843 2134 }
844
845 2134 static void FUNC(ff_vvc_lf_dsp_init)(VVCLFDSPContext *const lf)
846 {
847 2134 lf->ladf_level[0] = FUNC(vvc_h_loop_ladf_level);
848 2134 lf->ladf_level[1] = FUNC(vvc_v_loop_ladf_level);
849 2134 lf->filter_luma[0] = FUNC(vvc_h_loop_filter_luma);
850 2134 lf->filter_luma[1] = FUNC(vvc_v_loop_filter_luma);
851 2134 lf->filter_chroma[0] = FUNC(vvc_h_loop_filter_chroma);
852 2134 lf->filter_chroma[1] = FUNC(vvc_v_loop_filter_chroma);
853 2134 }
854
855 2134 static void FUNC(ff_vvc_sao_dsp_init)(VVCSAODSPContext *const sao)
856 {
857
2/2
✓ Branch 0 taken 9603 times.
✓ Branch 1 taken 1067 times.
 21340 for (int i = 0; i < FF_ARRAY_ELEMS(sao->band_filter); i++)
858 19206 sao->band_filter[i] = FUNC(sao_band_filter);
859
2/2
✓ Branch 0 taken 9603 times.
✓ Branch 1 taken 1067 times.
 21340 for (int i = 0; i < FF_ARRAY_ELEMS(sao->edge_filter); i++)
860 19206 sao->edge_filter[i] = FUNC(sao_edge_filter);
861 2134 sao->edge_restore[0] = FUNC(sao_edge_restore_0);
862 2134 sao->edge_restore[1] = FUNC(sao_edge_restore_1);
863 2134 }
864
865 2134 static void FUNC(ff_vvc_alf_dsp_init)(VVCALFDSPContext *const alf)
866 {
867 2134 alf->filter[LUMA] = FUNC(alf_filter_luma);
868 2134 alf->filter[CHROMA] = FUNC(alf_filter_chroma);
869 2134 alf->filter_cc = FUNC(alf_filter_cc);
870 2134 alf->classify = FUNC(alf_classify);
871 2134 alf->recon_coeff_and_clip = FUNC(alf_recon_coeff_and_clip);
872 2134 }
873