FFmpeg coverage

Directory: ../../../ffmpeg/
File: src/libavcodec/h26x/h2656_inter_template.c
Date: 2025-01-14 03:02:21
Exec Total Coverage
Lines: 325 325 100.0%
Functions: 49 63 77.8%
Branches: 106 106 100.0%
Line Branch Exec Source
1 /*
2 * inter prediction template for HEVC/VVC
3 *
4 * Copyright (C) 2022 Nuo Mi
5 * Copyright (C) 2024 Wu Jianhua
6 *
7 * This file is part of FFmpeg.
8 *
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
13 *
14 * FFmpeg is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
18 *
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 */
23
24 #define CHROMA_EXTRA_BEFORE 1
25 #define CHROMA_EXTRA 3
26 #define LUMA_EXTRA_BEFORE 3
27 #define LUMA_EXTRA 7
28
29 8597510 static void FUNC(put_pixels)(int16_t *dst,
30 const uint8_t *_src, const ptrdiff_t _src_stride,
31 const int height, const int8_t *hf, const int8_t *vf, const int width)
32 {
33 8597510 const pixel *src = (const pixel *)_src;
34 8597510 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
35
36
2/2
✓ Branch 0 taken 48506554 times.
✓ Branch 1 taken 4298755 times.
 105610618 for (int y = 0; y < height; y++) {
37
2/2
✓ Branch 0 taken 788889220 times.
✓ Branch 1 taken 48506554 times.
 1674791548 for (int x = 0; x < width; x++)
38 1577778440 dst[x] = src[x] << (14 - BIT_DEPTH);
39 97013108 src += src_stride;
40 97013108 dst += MAX_PB_SIZE;
41 }
42 8597510 }
43
44 2527502 static void FUNC(put_uni_pixels)(uint8_t *_dst, const ptrdiff_t _dst_stride,
45 const uint8_t *_src, const ptrdiff_t _src_stride, const int height,
46 const int8_t *hf, const int8_t *vf, const int width)
47 {
48 2527502 const pixel *src = (const pixel *)_src;
49 2527502 pixel *dst = (pixel *)_dst;
50 2527502 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
51 2527502 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
52
53
2/2
✓ Branch 0 taken 8548784 times.
✓ Branch 1 taken 1263751 times.
 19625070 for (int y = 0; y < height; y++) {
54 17097568 memcpy(dst, src, width * sizeof(pixel));
55 17097568 src += src_stride;
56 17097568 dst += dst_stride;
57 }
58 2527502 }
59
60 256604 static void FUNC(put_uni_w_pixels)(uint8_t *_dst, const ptrdiff_t _dst_stride,
61 const uint8_t *_src, const ptrdiff_t _src_stride, const int height,
62 const int denom, const int wx, const int _ox, const int8_t *hf, const int8_t *vf,
63 const int width)
64 {
65 256604 const pixel *src = (const pixel *)_src;
66 256604 pixel *dst = (pixel *)_dst;
67 256604 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
68 256604 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
69 256604 const int shift = denom + 14 - BIT_DEPTH;
70 #if BIT_DEPTH < 14
71 256604 const int offset = 1 << (shift - 1);
72 #else
73 const int offset = 0;
74 #endif
75 256604 const int ox = _ox * (1 << (BIT_DEPTH - 8));
76
77
2/2
✓ Branch 0 taken 637352 times.
✓ Branch 1 taken 128302 times.
 1531308 for (int y = 0; y < height; y++) {
78
2/2
✓ Branch 0 taken 5833840 times.
✓ Branch 1 taken 637352 times.
 12942384 for (int x = 0; x < width; x++) {
79 11667680 const int v = (src[x] << (14 - BIT_DEPTH));
80 11667680 dst[x] = av_clip_pixel(((v * wx + offset) >> shift) + ox);
81 }
82 1274704 src += src_stride;
83 1274704 dst += dst_stride;
84 }
85 256604 }
86
87 #define LUMA_FILTER(src, stride) \
88 (filter[0] * src[x - 3 * stride] + \
89 filter[1] * src[x - 2 * stride] + \
90 filter[2] * src[x - stride] + \
91 filter[3] * src[x ] + \
92 filter[4] * src[x + stride] + \
93 filter[5] * src[x + 2 * stride] + \
94 filter[6] * src[x + 3 * stride] + \
95 filter[7] * src[x + 4 * stride])
96
97 1112090 static void FUNC(put_luma_h)(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride,
98 const int height, const int8_t *hf, const int8_t *vf, const int width)
99 {
100 1112090 const pixel *src = (const pixel*)_src;
101 1112090 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
102 1112090 const int8_t *filter = hf;
103
104
2/2
✓ Branch 0 taken 4960232 times.
✓ Branch 1 taken 556045 times.
 11032554 for (int y = 0; y < height; y++) {
105
2/2
✓ Branch 0 taken 76802160 times.
✓ Branch 1 taken 4960232 times.
 163524784 for (int x = 0; x < width; x++)
106 153604320 dst[x] = LUMA_FILTER(src, 1) >> (BIT_DEPTH - 8);
107 9920464 src += src_stride;
108 9920464 dst += MAX_PB_SIZE;
109 }
110 1112090 }
111
112 1393438 static void FUNC(put_luma_v)(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride,
113 const int height, const int8_t *hf, const int8_t *vf, const int width)
114 {
115 1393438 const pixel *src = (pixel*)_src;
116 1393438 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
117 1393438 const int8_t *filter = vf;
118
119
2/2
✓ Branch 0 taken 6372900 times.
✓ Branch 1 taken 696719 times.
 14139238 for (int y = 0; y < height; y++) {
120
2/2
✓ Branch 0 taken 96518144 times.
✓ Branch 1 taken 6372900 times.
 205782088 for (int x = 0; x < width; x++)
121 193036288 dst[x] = LUMA_FILTER(src, src_stride) >> (BIT_DEPTH - 8);
122 12745800 src += src_stride;
123 12745800 dst += MAX_PB_SIZE;
124 }
125 1393438 }
126
127 8301924 static void FUNC(put_luma_hv)(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride,
128 const int height, const int8_t *hf, const int8_t *vf, const int width)
129 {
130 int16_t tmp_array[(MAX_PB_SIZE + LUMA_EXTRA) * MAX_PB_SIZE];
131 8301924 int16_t *tmp = tmp_array;
132 8301924 const pixel *src = (const pixel*)_src;
133 8301924 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
134 8301924 const int8_t *filter = hf;
135
136 8301924 src -= LUMA_EXTRA_BEFORE * src_stride;
137
2/2
✓ Branch 0 taken 57385190 times.
✓ Branch 1 taken 4150962 times.
 123072304 for (int y = 0; y < height + LUMA_EXTRA; y++) {
138
2/2
✓ Branch 0 taken 512223568 times.
✓ Branch 1 taken 57385190 times.
 1139217516 for (int x = 0; x < width; x++)
139 1024447136 tmp[x] = LUMA_FILTER(src, 1) >> (BIT_DEPTH - 8);
140 114770380 src += src_stride;
141 114770380 tmp += MAX_PB_SIZE;
142 }
143
144 8301924 tmp = tmp_array + LUMA_EXTRA_BEFORE * MAX_PB_SIZE;
145 8301924 filter = vf;
146
2/2
✓ Branch 0 taken 28328456 times.
✓ Branch 1 taken 4150962 times.
 64958836 for (int y = 0; y < height; y++) {
147
2/2
✓ Branch 0 taken 314863216 times.
✓ Branch 1 taken 28328456 times.
 686383344 for (int x = 0; x < width; x++)
148 629726432 dst[x] = LUMA_FILTER(tmp, MAX_PB_SIZE) >> 6;
149 56656912 tmp += MAX_PB_SIZE;
150 56656912 dst += MAX_PB_SIZE;
151 }
152 8301924 }
153
154 97956 static void FUNC(put_uni_luma_h)(uint8_t *_dst, const ptrdiff_t _dst_stride,
155 const uint8_t *_src, const ptrdiff_t _src_stride,
156 const int height, const int8_t *hf, const int8_t *vf, const int width)
157 {
158 97956 const pixel *src = (const pixel*)_src;
159 97956 pixel *dst = (pixel *)_dst;
160 97956 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
161 97956 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
162 97956 const int8_t *filter = hf;
163 97956 const int shift = 14 - BIT_DEPTH;
164 #if BIT_DEPTH < 14
165 97956 const int offset = 1 << (shift - 1);
166 #else
167 const int offset = 0;
168 #endif
169
170
2/2
✓ Branch 0 taken 566512 times.
✓ Branch 1 taken 48978 times.
 1230980 for (int y = 0; y < height; y++) {
171
2/2
✓ Branch 0 taken 12226768 times.
✓ Branch 1 taken 566512 times.
 25586560 for (int x = 0; x < width; x++) {
172 24453536 const int val = LUMA_FILTER(src, 1) >> (BIT_DEPTH - 8);
173 24453536 dst[x] = av_clip_pixel((val + offset) >> shift);
174 }
175 1133024 src += src_stride;
176 1133024 dst += dst_stride;
177 }
178 97956 }
179
180 118844 static void FUNC(put_uni_luma_v)(uint8_t *_dst, const ptrdiff_t _dst_stride,
181 const uint8_t *_src, const ptrdiff_t _src_stride,
182 const int height, const int8_t *hf, const int8_t *vf, const int width)
183 {
184
185 118844 const pixel *src = (const pixel*)_src;
186 118844 pixel *dst = (pixel *)_dst;
187 118844 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
188 118844 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
189 118844 const int8_t *filter = vf;
190 118844 const int shift = 14 - BIT_DEPTH;
191 #if BIT_DEPTH < 14
192 118844 const int offset = 1 << (shift - 1);
193 #else
194 const int offset = 0;
195 #endif
196
197
2/2
✓ Branch 0 taken 656348 times.
✓ Branch 1 taken 59422 times.
 1431540 for (int y = 0; y < height; y++) {
198
2/2
✓ Branch 0 taken 13769984 times.
✓ Branch 1 taken 656348 times.
 28852664 for (int x = 0; x < width; x++) {
199 27539968 const int val = LUMA_FILTER(src, src_stride) >> (BIT_DEPTH - 8);
200 27539968 dst[x] = av_clip_pixel((val + offset) >> shift);
201 }
202 1312696 src += src_stride;
203 1312696 dst += dst_stride;
204 }
205 118844 }
206
207 369286 static void FUNC(put_uni_luma_hv)(uint8_t *_dst, const ptrdiff_t _dst_stride,
208 const uint8_t *_src, const ptrdiff_t _src_stride,
209 const int height, const int8_t *hf, const int8_t *vf, const int width)
210 {
211 int16_t tmp_array[(MAX_PB_SIZE + LUMA_EXTRA) * MAX_PB_SIZE];
212 369286 int16_t *tmp = tmp_array;
213 369286 const pixel *src = (const pixel*)_src;
214 369286 pixel *dst = (pixel *)_dst;
215 369286 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
216 369286 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
217 369286 const int8_t *filter = hf;
218 369286 const int shift = 14 - BIT_DEPTH;
219 #if BIT_DEPTH < 14
220 369286 const int offset = 1 << (shift - 1);
221 #else
222 const int offset = 0;
223 #endif
224
225 369286 src -= LUMA_EXTRA_BEFORE * src_stride;
226
2/2
✓ Branch 0 taken 3316897 times.
✓ Branch 1 taken 184643 times.
 7003080 for (int y = 0; y < height + LUMA_EXTRA; y++) {
227
2/2
✓ Branch 0 taken 49113924 times.
✓ Branch 1 taken 3316897 times.
 104861642 for (int x = 0; x < width; x++)
228 98227848 tmp[x] = LUMA_FILTER(src, 1) >> (BIT_DEPTH - 8);
229 6633794 src += src_stride;
230 6633794 tmp += MAX_PB_SIZE;
231 }
232
233 369286 tmp = tmp_array + LUMA_EXTRA_BEFORE * MAX_PB_SIZE;
234 369286 filter = vf;
235
236
2/2
✓ Branch 0 taken 2024396 times.
✓ Branch 1 taken 184643 times.
 4418078 for (int y = 0; y < height; y++) {
237
2/2
✓ Branch 0 taken 35564304 times.
✓ Branch 1 taken 2024396 times.
 75177400 for (int x = 0; x < width; x++) {
238 71128608 const int val = LUMA_FILTER(tmp, MAX_PB_SIZE) >> 6;
239 71128608 dst[x] = av_clip_pixel((val + offset) >> shift);
240 }
241 4048792 tmp += MAX_PB_SIZE;
242 4048792 dst += dst_stride;
243 }
244
245 369286 }
246
247 22844 static void FUNC(put_uni_luma_w_h)(uint8_t *_dst, const ptrdiff_t _dst_stride,
248 const uint8_t *_src, const ptrdiff_t _src_stride, int height,
249 const int denom, const int wx, const int _ox, const int8_t *hf, const int8_t *vf,
250 const int width)
251 {
252 22844 const pixel *src = (const pixel*)_src;
253 22844 pixel *dst = (pixel *)_dst;
254 22844 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
255 22844 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
256 22844 const int8_t *filter = hf;
257 22844 const int ox = _ox * (1 << (BIT_DEPTH - 8));
258 22844 const int shift = denom + 14 - BIT_DEPTH;
259 #if BIT_DEPTH < 14
260 22844 const int offset = 1 << (shift - 1);
261 #else
262 const int offset = 0;
263 #endif
264
265
2/2
✓ Branch 0 taken 58128 times.
✓ Branch 1 taken 11422 times.
 139100 for (int y = 0; y < height; y++) {
266
2/2
✓ Branch 0 taken 527680 times.
✓ Branch 1 taken 58128 times.
 1171616 for (int x = 0; x < width; x++)
267 1055360 dst[x] = av_clip_pixel((((LUMA_FILTER(src, 1) >> (BIT_DEPTH - 8)) * wx + offset) >> shift) + ox);
268 116256 src += src_stride;
269 116256 dst += dst_stride;
270 }
271 22844 }
272
273 22448 static void FUNC(put_uni_luma_w_v)(uint8_t *_dst, const ptrdiff_t _dst_stride,
274 const uint8_t *_src, const ptrdiff_t _src_stride, const int height,
275 const int denom, const int wx, const int _ox, const int8_t *hf, const int8_t *vf,
276 const int width)
277 {
278 22448 const pixel *src = (const pixel*)_src;
279 22448 pixel *dst = (pixel *)_dst;
280 22448 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
281 22448 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
282 22448 const int8_t *filter = vf;
283 22448 const int ox = _ox * (1 << (BIT_DEPTH - 8));
284 22448 const int shift = denom + 14 - BIT_DEPTH;
285 #if BIT_DEPTH < 14
286 22448 const int offset = 1 << (shift - 1);
287 #else
288 const int offset = 0;
289 #endif
290
291
2/2
✓ Branch 0 taken 54840 times.
✓ Branch 1 taken 11224 times.
 132128 for (int y = 0; y < height; y++) {
292
2/2
✓ Branch 0 taken 440416 times.
✓ Branch 1 taken 54840 times.
 990512 for (int x = 0; x < width; x++)
293 880832 dst[x] = av_clip_pixel((((LUMA_FILTER(src, src_stride) >> (BIT_DEPTH - 8)) * wx + offset) >> shift) + ox);
294 109680 src += src_stride;
295 109680 dst += dst_stride;
296 }
297 22448 }
298
299 76152 static void FUNC(put_uni_luma_w_hv)(uint8_t *_dst, const ptrdiff_t _dst_stride,
300 const uint8_t *_src, const ptrdiff_t _src_stride, const int height, const int denom,
301 const int wx, const int _ox, const int8_t *hf, const int8_t *vf, const int width)
302 {
303 int16_t tmp_array[(MAX_PB_SIZE + LUMA_EXTRA) * MAX_PB_SIZE];
304 76152 int16_t *tmp = tmp_array;
305 76152 const pixel *src = (const pixel*)_src;
306 76152 pixel *dst = (pixel *)_dst;
307 76152 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
308 76152 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
309 76152 const int8_t *filter = hf;
310 76152 const int ox = _ox * (1 << (BIT_DEPTH - 8));
311 76152 const int shift = denom + 14 - BIT_DEPTH;
312 #if BIT_DEPTH < 14
313 76152 const int offset = 1 << (shift - 1);
314 #else
315 const int offset = 0;
316 #endif
317
318 76152 src -= LUMA_EXTRA_BEFORE * src_stride;
319
2/2
✓ Branch 0 taken 456544 times.
✓ Branch 1 taken 38076 times.
 989240 for (int y = 0; y < height + LUMA_EXTRA; y++) {
320
2/2
✓ Branch 0 taken 3176504 times.
✓ Branch 1 taken 456544 times.
 7266096 for (int x = 0; x < width; x++)
321 6353008 tmp[x] = LUMA_FILTER(src, 1) >> (BIT_DEPTH - 8);
322 913088 src += src_stride;
323 913088 tmp += MAX_PB_SIZE;
324 }
325
326 76152 tmp = tmp_array + LUMA_EXTRA_BEFORE * MAX_PB_SIZE;
327 76152 filter = vf;
328
2/2
✓ Branch 0 taken 190012 times.
✓ Branch 1 taken 38076 times.
 456176 for (int y = 0; y < height; y++) {
329
2/2
✓ Branch 0 taken 1834912 times.
✓ Branch 1 taken 190012 times.
 4049848 for (int x = 0; x < width; x++)
330 3669824 dst[x] = av_clip_pixel((((LUMA_FILTER(tmp, MAX_PB_SIZE) >> 6) * wx + offset) >> shift) + ox);
331 380024 tmp += MAX_PB_SIZE;
332 380024 dst += dst_stride;
333 }
334 76152 }
335
336 #define CHROMA_FILTER(src, stride) \
337 (filter[0] * src[x - stride] + \
338 filter[1] * src[x] + \
339 filter[2] * src[x + stride] + \
340 filter[3] * src[x + 2 * stride])
341
342 1144510 static void FUNC(put_chroma_h)(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride,
343 const int height, const int8_t *hf, const int8_t *vf, const int width)
344 {
345 1144510 const pixel *src = (const pixel *)_src;
346 1144510 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
347 1144510 const int8_t *filter = hf;
348
349
2/2
✓ Branch 0 taken 4821498 times.
✓ Branch 1 taken 572255 times.
 10787506 for (int y = 0; y < height; y++) {
350
2/2
✓ Branch 0 taken 79161476 times.
✓ Branch 1 taken 4821498 times.
 167965948 for (int x = 0; x < width; x++)
351 158322952 dst[x] = CHROMA_FILTER(src, 1) >> (BIT_DEPTH - 8);
352 9642996 src += src_stride;
353 9642996 dst += MAX_PB_SIZE;
354 }
355 1144510 }
356
357 1712798 static void FUNC(put_chroma_v)(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride,
358 const int height, const int8_t *hf, const int8_t *vf, const int width)
359 {
360 1712798 const pixel *src = (const pixel *)_src;
361 1712798 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
362 1712798 const int8_t *filter = vf;
363
364
2/2
✓ Branch 0 taken 6863842 times.
✓ Branch 1 taken 856399 times.
 15440482 for (int y = 0; y < height; y++) {
365
2/2
✓ Branch 0 taken 103285956 times.
✓ Branch 1 taken 6863842 times.
 220299596 for (int x = 0; x < width; x++)
366 206571912 dst[x] = CHROMA_FILTER(src, src_stride) >> (BIT_DEPTH - 8);
367 13727684 src += src_stride;
368 13727684 dst += MAX_PB_SIZE;
369 }
370 1712798 }
371
372 7323794 static void FUNC(put_chroma_hv)(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride,
373 const int height, const int8_t *hf, const int8_t *vf, const int width)
374 {
375 int16_t tmp_array[(MAX_PB_SIZE + CHROMA_EXTRA) * MAX_PB_SIZE];
376 7323794 int16_t *tmp = tmp_array;
377 7323794 const pixel *src = (const pixel *)_src;
378 7323794 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
379 7323794 const int8_t *filter = hf;
380
381 7323794 src -= CHROMA_EXTRA_BEFORE * src_stride;
382
383
2/2
✓ Branch 0 taken 36614905 times.
✓ Branch 1 taken 3661897 times.
 80553604 for (int y = 0; y < height + CHROMA_EXTRA; y++) {
384
2/2
✓ Branch 0 taken 344250534 times.
✓ Branch 1 taken 36614905 times.
 761730878 for (int x = 0; x < width; x++)
385 688501068 tmp[x] = CHROMA_FILTER(src, 1) >> (BIT_DEPTH - 8);
386 73229810 src += src_stride;
387 73229810 tmp += MAX_PB_SIZE;
388 }
389
390 7323794 tmp = tmp_array + CHROMA_EXTRA_BEFORE * MAX_PB_SIZE;
391 7323794 filter = vf;
392
393
2/2
✓ Branch 0 taken 25629214 times.
✓ Branch 1 taken 3661897 times.
 58582222 for (int y = 0; y < height; y++) {
394
2/2
✓ Branch 0 taken 267738660 times.
✓ Branch 1 taken 25629214 times.
 586735748 for (int x = 0; x < width; x++)
395 535477320 dst[x] = CHROMA_FILTER(tmp, MAX_PB_SIZE) >> 6;
396 51258428 tmp += MAX_PB_SIZE;
397 51258428 dst += MAX_PB_SIZE;
398 }
399 7323794 }
400
401 411272 static void FUNC(put_uni_chroma_h)(uint8_t *_dst, const ptrdiff_t _dst_stride,
402 const uint8_t *_src, const ptrdiff_t _src_stride,
403 const int height, const int8_t *hf, const int8_t *vf, const int width)
404 {
405 411272 const pixel *src = (const pixel *)_src;
406 411272 pixel *dst = (pixel *)_dst;
407 411272 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
408 411272 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
409 411272 const int8_t *filter = hf;
410 411272 const int shift = 14 - BIT_DEPTH;
411 #if BIT_DEPTH < 14
412 411272 const int offset = 1 << (shift - 1);
413 #else
414 const int offset = 0;
415 #endif
416
417
2/2
✓ Branch 0 taken 1106040 times.
✓ Branch 1 taken 205636 times.
 2623352 for (int y = 0; y < height; y++) {
418
2/2
✓ Branch 0 taken 13607200 times.
✓ Branch 1 taken 1106040 times.
 29426480 for (int x = 0; x < width; x++)
419 27214400 dst[x] = av_clip_pixel(((CHROMA_FILTER(src, 1) >> (BIT_DEPTH - 8)) + offset) >> shift);
420 2212080 src += src_stride;
421 2212080 dst += dst_stride;
422 }
423 411272 }
424
425 409020 static void FUNC(put_uni_chroma_v)(uint8_t *_dst, const ptrdiff_t _dst_stride,
426 const uint8_t *_src, const ptrdiff_t _src_stride,
427 const int height, const int8_t *hf, const int8_t *vf, const int width)
428 {
429 409020 const pixel *src = (const pixel *)_src;
430 409020 pixel *dst = (pixel *)_dst;
431 409020 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
432 409020 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
433 409020 const int8_t *filter = vf;
434 409020 const int shift = 14 - BIT_DEPTH;
435 #if BIT_DEPTH < 14
436 409020 const int offset = 1 << (shift - 1);
437 #else
438 const int offset = 0;
439 #endif
440
441
2/2
✓ Branch 0 taken 1111240 times.
✓ Branch 1 taken 204510 times.
 2631500 for (int y = 0; y < height; y++) {
442
2/2
✓ Branch 0 taken 13987024 times.
✓ Branch 1 taken 1111240 times.
 30196528 for (int x = 0; x < width; x++)
443 27974048 dst[x] = av_clip_pixel(((CHROMA_FILTER(src, src_stride) >> (BIT_DEPTH - 8)) + offset) >> shift);
444 2222480 src += src_stride;
445 2222480 dst += dst_stride;
446 }
447 409020 }
448
449 3039880 static void FUNC(put_uni_chroma_hv)(uint8_t *_dst, const ptrdiff_t _dst_stride,
450 const uint8_t *_src, const ptrdiff_t _src_stride,
451 const int height, const int8_t *hf, const int8_t *vf, const int width)
452 {
453 int16_t tmp_array[(MAX_PB_SIZE + CHROMA_EXTRA) * MAX_PB_SIZE];
454 3039880 int16_t *tmp = tmp_array;
455 3039880 const pixel *src = (const pixel *)_src;
456 3039880 pixel *dst = (pixel *)_dst;
457 3039880 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
458 3039880 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
459 3039880 const int8_t *filter = hf;
460 3039880 const int shift = 14 - BIT_DEPTH;
461 #if BIT_DEPTH < 14
462 3039880 const int offset = 1 << (shift - 1);
463 #else
464 const int offset = 0;
465 #endif
466
467 3039880 src -= CHROMA_EXTRA_BEFORE * src_stride;
468
469
2/2
✓ Branch 0 taken 11558756 times.
✓ Branch 1 taken 1519940 times.
 26157392 for (int y = 0; y < height + CHROMA_EXTRA; y++) {
470
2/2
✓ Branch 0 taken 71100140 times.
✓ Branch 1 taken 11558756 times.
 165317792 for (int x = 0; x < width; x++)
471 142200280 tmp[x] = CHROMA_FILTER(src, 1) >> (BIT_DEPTH - 8);
472 23117512 src += src_stride;
473 23117512 tmp += MAX_PB_SIZE;
474 }
475
476 3039880 tmp = tmp_array + CHROMA_EXTRA_BEFORE * MAX_PB_SIZE;
477 3039880 filter = vf;
478
479
2/2
✓ Branch 0 taken 6998936 times.
✓ Branch 1 taken 1519940 times.
 17037752 for (int y = 0; y < height; y++) {
480
2/2
✓ Branch 0 taken 50368448 times.
✓ Branch 1 taken 6998936 times.
 114734768 for (int x = 0; x < width; x++)
481 100736896 dst[x] = av_clip_pixel(((CHROMA_FILTER(tmp, MAX_PB_SIZE) >> 6) + offset) >> shift);
482 13997872 tmp += MAX_PB_SIZE;
483 13997872 dst += dst_stride;
484 }
485 3039880 }
486
487 19480 static void FUNC(put_uni_chroma_w_h)(uint8_t *_dst, ptrdiff_t _dst_stride,
488 const uint8_t *_src, ptrdiff_t _src_stride, int height, int denom, int wx, int ox,
489 const int8_t *hf, const int8_t *vf, int width)
490 {
491 19480 const pixel *src = (const pixel *)_src;
492 19480 pixel *dst = (pixel *)_dst;
493 19480 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
494 19480 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
495 19480 const int8_t *filter = hf;
496 19480 const int shift = denom + 14 - BIT_DEPTH;
497 #if BIT_DEPTH < 14
498 19480 const int offset = 1 << (shift - 1);
499 #else
500 const int offset = 0;
501 #endif
502
503 19480 ox = ox * (1 << (BIT_DEPTH - 8));
504
2/2
✓ Branch 0 taken 46728 times.
✓ Branch 1 taken 9740 times.
 112936 for (int y = 0; y < height; y++) {
505
2/2
✓ Branch 0 taken 292128 times.
✓ Branch 1 taken 46728 times.
 677712 for (int x = 0; x < width; x++) {
506 584256 dst[x] = av_clip_pixel((((CHROMA_FILTER(src, 1) >> (BIT_DEPTH - 8)) * wx + offset) >> shift) + ox);
507 }
508 93456 dst += dst_stride;
509 93456 src += src_stride;
510 }
511 19480 }
512
513 19836 static void FUNC(put_uni_chroma_w_v)(uint8_t *_dst, const ptrdiff_t _dst_stride,
514 const uint8_t *_src, const ptrdiff_t _src_stride, const int height,
515 const int denom, const int wx, const int _ox, const int8_t *hf, const int8_t *vf,
516 const int width)
517 {
518 19836 const pixel *src = (const pixel *)_src;
519 19836 pixel *dst = (pixel *)_dst;
520 19836 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
521 19836 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
522 19836 const int8_t *filter = vf;
523 19836 const int shift = denom + 14 - BIT_DEPTH;
524 19836 const int ox = _ox * (1 << (BIT_DEPTH - 8));
525 #if BIT_DEPTH < 14
526 19836 int offset = 1 << (shift - 1);
527 #else
528 int offset = 0;
529 #endif
530
531
2/2
✓ Branch 0 taken 46288 times.
✓ Branch 1 taken 9918 times.
 112412 for (int y = 0; y < height; y++) {
532
2/2
✓ Branch 0 taken 301520 times.
✓ Branch 1 taken 46288 times.
 695616 for (int x = 0; x < width; x++) {
533 603040 dst[x] = av_clip_pixel((((CHROMA_FILTER(src, src_stride) >> (BIT_DEPTH - 8)) * wx + offset) >> shift) + ox);
534 }
535 92576 dst += dst_stride;
536 92576 src += src_stride;
537 }
538 19836 }
539
540 110112 static void FUNC(put_uni_chroma_w_hv)(uint8_t *_dst, ptrdiff_t _dst_stride,
541 const uint8_t *_src, ptrdiff_t _src_stride, int height, int denom, int wx, int ox,
542 const int8_t *hf, const int8_t *vf, int width)
543 {
544 int16_t tmp_array[(MAX_PB_SIZE + CHROMA_EXTRA) * MAX_PB_SIZE];
545 110112 int16_t *tmp = tmp_array;
546 110112 const pixel *src = (const pixel *)_src;
547 110112 pixel *dst = (pixel *)_dst;
548 110112 const ptrdiff_t src_stride = _src_stride / sizeof(pixel);
549 110112 const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel);
550 110112 const int8_t *filter = hf;
551 110112 const int shift = denom + 14 - BIT_DEPTH;
552 #if BIT_DEPTH < 14
553 110112 const int offset = 1 << (shift - 1);
554 #else
555 const int offset = 0;
556 #endif
557
558 110112 src -= CHROMA_EXTRA_BEFORE * src_stride;
559
560
2/2
✓ Branch 0 taken 417916 times.
✓ Branch 1 taken 55056 times.
 945944 for (int y = 0; y < height + CHROMA_EXTRA; y++) {
561
2/2
✓ Branch 0 taken 2399396 times.
✓ Branch 1 taken 417916 times.
 5634624 for (int x = 0; x < width; x++)
562 4798792 tmp[x] = CHROMA_FILTER(src, 1) >> (BIT_DEPTH - 8);
563 835832 src += src_stride;
564 835832 tmp += MAX_PB_SIZE;
565 }
566
567 110112 tmp = tmp_array + CHROMA_EXTRA_BEFORE * MAX_PB_SIZE;
568 110112 filter = vf;
569
570 110112 ox = ox * (1 << (BIT_DEPTH - 8));
571
2/2
✓ Branch 0 taken 252748 times.
✓ Branch 1 taken 55056 times.
 615608 for (int y = 0; y < height; y++) {
572
2/2
✓ Branch 0 taken 1626656 times.
✓ Branch 1 taken 252748 times.
 3758808 for (int x = 0; x < width; x++)
573 3253312 dst[x] = av_clip_pixel((((CHROMA_FILTER(tmp, MAX_PB_SIZE) >> 6) * wx + offset) >> shift) + ox);
574 505496 tmp += MAX_PB_SIZE;
575 505496 dst += dst_stride;
576 }
577 110112 }
578