FFmpeg coverage

Directory: ../../../ffmpeg/
File: src/libavcodec/x86/hevc/dsp_init.c
Date: 2026-08-15 14:54:27
Exec Total Coverage
Lines: 635 641 99.1%
Functions: 1015 1015 100.0%
Branches: 408 412 99.0%
Line Branch Exec Source
1 /*
2 * Copyright (c) 2013 Seppo Tomperi
3 * Copyright (c) 2013-2014 Pierre-Edouard Lepere
4 * Copyright (c) 2023-2024 Wu Jianhua
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 "config.h"
24
25 #include "libavutil/cpu.h"
26 #include "libavutil/mem_internal.h"
27 #include "libavutil/x86/cpu.h"
28 #include "libavcodec/hevc/dsp.h"
29 #include "libavcodec/x86/hevc/dsp.h"
30 #include "libavcodec/x86/h26x/h2656dsp.h"
31
32 void ff_hevc_dequant_8_ssse3(int16_t *coeffs, int16_t log2_size);
33
34 #define LFC_FUNC(DIR, DEPTH, OPT) \
35 void ff_hevc_ ## DIR ## _loop_filter_chroma_ ## DEPTH ## _ ## OPT(uint8_t *pix, ptrdiff_t stride, const int *tc, const uint8_t *no_p, const uint8_t *no_q);
36
37 #define LFL_FUNC(DIR, DEPTH, OPT) \
38 void ff_hevc_ ## DIR ## _loop_filter_luma_ ## DEPTH ## _ ## OPT(uint8_t *pix, ptrdiff_t stride, int beta, const int *tc, const uint8_t *no_p, const uint8_t *no_q);
39
40 #define LFC_FUNCS(type, depth, opt) \
41 LFC_FUNC(h, depth, opt) \
42 LFC_FUNC(v, depth, opt)
43
44 #define LFL_FUNCS(type, depth, opt) \
45 LFL_FUNC(h, depth, opt) \
46 LFL_FUNC(v, depth, opt)
47
48 LFC_FUNCS(uint8_t, 8, sse2)
49 LFC_FUNCS(uint8_t, 10, sse2)
50 LFC_FUNCS(uint8_t, 12, sse2)
51 LFC_FUNCS(uint8_t, 8, avx)
52 LFC_FUNCS(uint8_t, 10, avx)
53 LFC_FUNCS(uint8_t, 12, avx)
54 LFL_FUNCS(uint8_t, 8, sse2)
55 LFL_FUNCS(uint8_t, 10, sse2)
56 LFL_FUNCS(uint8_t, 12, sse2)
57 LFL_FUNCS(uint8_t, 8, ssse3)
58 LFL_FUNCS(uint8_t, 10, ssse3)
59 LFL_FUNCS(uint8_t, 12, ssse3)
60 LFL_FUNCS(uint8_t, 8, avx)
61 LFL_FUNCS(uint8_t, 10, avx)
62 LFL_FUNCS(uint8_t, 12, avx)
63
64 #define IDCT_DC_FUNCS(W, opt) \
65 void ff_hevc_idct_ ## W ## _dc_8_ ## opt(int16_t *coeffs); \
66 void ff_hevc_idct_ ## W ## _dc_10_ ## opt(int16_t *coeffs); \
67 void ff_hevc_idct_ ## W ## _dc_12_ ## opt(int16_t *coeffs)
68
69 IDCT_DC_FUNCS(4x4, sse2);
70 IDCT_DC_FUNCS(8x8, sse2);
71 IDCT_DC_FUNCS(16x16, sse2);
72 IDCT_DC_FUNCS(32x32, sse2);
73 IDCT_DC_FUNCS(16x16, avx2);
74 IDCT_DC_FUNCS(32x32, avx2);
75
76 #define IDCT_FUNCS(opt) \
77 void ff_hevc_idct_4x4_8_ ## opt(int16_t *coeffs, int col_limit); \
78 void ff_hevc_idct_4x4_10_ ## opt(int16_t *coeffs, int col_limit); \
79 void ff_hevc_idct_8x8_8_ ## opt(int16_t *coeffs, int col_limit); \
80 void ff_hevc_idct_8x8_10_ ## opt(int16_t *coeffs, int col_limit); \
81 void ff_hevc_idct_16x16_8_ ## opt(int16_t *coeffs, int col_limit); \
82 void ff_hevc_idct_16x16_10_ ## opt(int16_t *coeffs, int col_limit); \
83 void ff_hevc_idct_32x32_8_ ## opt(int16_t *coeffs, int col_limit); \
84 void ff_hevc_idct_32x32_10_ ## opt(int16_t *coeffs, int col_limit);
85
86 IDCT_FUNCS(sse2)
87 IDCT_FUNCS(avx)
88
89 #define TRANSFORM_LUMA_FUNCS(opt) \
90 void ff_hevc_transform_4x4_luma_8_ ## opt(int16_t *coeffs); \
91 void ff_hevc_transform_4x4_luma_10_ ## opt(int16_t *coeffs)
92
93 TRANSFORM_LUMA_FUNCS(sse2);
94 TRANSFORM_LUMA_FUNCS(avx);
95
96 #define ff_hevc_pel_filters ff_hevc_qpel_filters
97 #define DECL_HV_FILTER(f) \
98 const uint8_t *hf = ff_hevc_ ## f ## _filters[mx]; \
99 const uint8_t *vf = ff_hevc_ ## f ## _filters[my];
100
101 #define FW_PUT(p, a, b, depth, opt) \
102 static void hevc_put_ ## a ## _ ## depth ## _##opt(int16_t *dst, const uint8_t *src, ptrdiff_t srcstride, \
103 int height, intptr_t mx, intptr_t my,int width) \
104 { \
105 DECL_HV_FILTER(p) \
106 ff_h2656_put_ ## b ## _ ## depth ## _##opt(dst, 2 * MAX_PB_SIZE, src, srcstride, height, hf, vf, width); \
107 }
108
109 #define FW_PUT_UNI(p, a, b, depth, opt) \
110 static void hevc_put_uni_ ## a ## _ ## depth ## _##opt(uint8_t *dst, ptrdiff_t dststride, \
111 const uint8_t *src, ptrdiff_t srcstride, \
112 int height, intptr_t mx, intptr_t my, int width) \
113 { \
114 DECL_HV_FILTER(p) \
115 ff_h2656_put_uni_ ## b ## _ ## depth ## _##opt(dst, dststride, src, srcstride, height, hf, vf, width); \
116 }
117
118 #if ARCH_X86_64 && HAVE_SSE4_EXTERNAL
119
120 #define FW_PUT_FUNCS(p, a, b, depth, opt) \
121 FW_PUT(p, a, b, depth, opt) \
122 FW_PUT_UNI(p, a, b, depth, opt)
123
124 #define FW_PEL(w, depth, opt) FW_PUT_FUNCS(pel, pel_pixels##w, pixels##w, depth, opt)
125
126 #define FW_DIR(npel, n, w, depth, opt) \
127 FW_PUT_FUNCS(npel, npel ## _h##w, n ## tap_h##w, depth, opt) \
128 FW_PUT_FUNCS(npel, npel ## _v##w, n ## tap_v##w, depth, opt)
129
130 #define FW_DIR_HV(npel, n, w, depth, opt) \
131 FW_PUT_FUNCS(npel, npel ## _hv##w, n ## tap_hv##w, depth, opt)
132
133 21598 FW_PEL(4, 8, sse4)
134 104 FW_PEL(6, 8, sse4)
135 31412 FW_PEL(8, 8, sse4)
136 104 FW_PEL(12, 8, sse4)
137 156094 FW_PEL(16, 8, sse4)
138 2512 FW_PEL(4, 10, sse4)
139 104 FW_PEL(6, 10, sse4)
140 34298 FW_PEL(8, 10, sse4)
141 416 FW_PEL(4, 12, sse4)
142 104 FW_PEL(6, 12, sse4)
143 2496 FW_PEL(8, 12, sse4)
144
145 #define FW_EPEL(w, depth, opt) FW_DIR(epel, 4, w, depth, opt)
146 #define FW_EPEL_HV(w, depth, opt) FW_DIR_HV(epel, 4, w, depth, opt)
147 #define FW_EPEL_FUNCS(w, depth, opt) \
148 FW_EPEL(w, depth, opt) \
149 FW_EPEL_HV(w, depth, opt)
150
151 208 FW_EPEL(12, 8, sse4)
152
153 19552 FW_EPEL_FUNCS(4, 8, sse4)
154 312 FW_EPEL_FUNCS(6, 8, sse4)
155 17604 FW_EPEL_FUNCS(8, 8, sse4)
156 23120 FW_EPEL_FUNCS(16, 8, sse4)
157 7220 FW_EPEL_FUNCS(4, 10, sse4)
158 312 FW_EPEL_FUNCS(6, 10, sse4)
159 33092 FW_EPEL_FUNCS(8, 10, sse4)
160 1248 FW_EPEL_FUNCS(4, 12, sse4)
161 312 FW_EPEL_FUNCS(6, 12, sse4)
162 7488 FW_EPEL_FUNCS(8, 12, sse4)
163
164 #define FW_QPEL(w, depth, opt) FW_DIR(qpel, 8, w, depth, opt)
165 #define FW_QPEL_HV(w, depth, opt) FW_DIR_HV(qpel, 8, w, depth, opt)
166 #define FW_QPEL_FUNCS(w, depth, opt) \
167 FW_QPEL(w, depth, opt) \
168 FW_QPEL_HV(w, depth, opt)
169
170 208 FW_QPEL(12, 8, sse4)
171 3706 FW_QPEL(16, 8, sse4)
172
173 1414 FW_QPEL_FUNCS(4, 8, sse4)
174 12234 FW_QPEL_FUNCS(8, 8, sse4)
175 1248 FW_QPEL_FUNCS(4, 10, sse4)
176 16296 FW_QPEL_FUNCS(8, 10, sse4)
177 1248 FW_QPEL_FUNCS(4, 12, sse4)
178 7488 FW_QPEL_FUNCS(8, 12, sse4)
179
180 #if HAVE_AVX2_EXTERNAL
181
182 29180 FW_PEL(32, 8, avx2)
183 3039 FW_PUT(pel, pel_pixels16, pixels16, 10, avx2)
184
185 1428 FW_EPEL(32, 8, avx2)
186 5200 FW_EPEL(16, 10, avx2)
187
188 768 FW_EPEL_HV(32, 8, avx2)
189 6776 FW_EPEL_HV(16, 10, avx2)
190
191 340 FW_QPEL(32, 8, avx2)
192 9048 FW_QPEL(16, 10, avx2)
193
194 11398 FW_QPEL_HV(16, 10, avx2)
195
196 #endif
197 #endif
198
199 #define mc_rep_func(name, bitd, step, W, opt) \
200 static void hevc_put_##name##W##_##bitd##_##opt(int16_t *_dst, \
201 const uint8_t *_src, ptrdiff_t _srcstride, int height, \
202 intptr_t mx, intptr_t my, int width) \
203 { \
204 int i; \
205 int16_t *dst; \
206 for (i = 0; i < W; i += step) { \
207 const uint8_t *src = _src + (i * ((bitd + 7) / 8)); \
208 dst = _dst + i; \
209 hevc_put_##name##step##_##bitd##_##opt(dst, src, _srcstride, height, mx, my, width); \
210 } \
211 }
212 #define mc_rep_uni_func(name, bitd, step, W, opt) \
213 static void hevc_put_uni_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, \
214 const uint8_t *_src, ptrdiff_t _srcstride, int height, \
215 intptr_t mx, intptr_t my, int width) \
216 { \
217 int i; \
218 uint8_t *dst; \
219 for (i = 0; i < W; i += step) { \
220 const uint8_t *src = _src + (i * ((bitd + 7) / 8)); \
221 dst = _dst + (i * ((bitd + 7) / 8)); \
222 hevc_put_uni_##name##step##_##bitd##_##opt(dst, dststride, src, _srcstride, \
223 height, mx, my, width); \
224 } \
225 }
226 #define mc_rep_bi_func(name, bitd, step, W, opt) \
227 static void ff_hevc_put_bi_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, const uint8_t *_src, \
228 ptrdiff_t _srcstride, const int16_t *_src2, \
229 int height, intptr_t mx, intptr_t my, int width) \
230 { \
231 int i; \
232 uint8_t *dst; \
233 for (i = 0; i < W ; i += step) { \
234 const uint8_t *src = _src + (i * ((bitd + 7) / 8)); \
235 const int16_t *src2 = _src2 + i; \
236 dst = _dst + (i * ((bitd + 7) / 8)); \
237 ff_hevc_put_bi_##name##step##_##bitd##_##opt(dst, dststride, src, _srcstride, src2, \
238 height, mx, my, width); \
239 } \
240 }
241
242 #define mc_rep_funcs(name, bitd, step, W, opt) \
243 mc_rep_func(name, bitd, step, W, opt) \
244 mc_rep_uni_func(name, bitd, step, W, opt) \
245 mc_rep_bi_func(name, bitd, step, W, opt)
246
247 #define mc_rep_func2(name, bitd, step1, step2, W, opt) \
248 static void hevc_put_##name##W##_##bitd##_##opt(int16_t *dst, \
249 const uint8_t *src, ptrdiff_t _srcstride, int height, \
250 intptr_t mx, intptr_t my, int width) \
251 { \
252 hevc_put_##name##step1##_##bitd##_##opt(dst, src, _srcstride, height, mx, my, width); \
253 hevc_put_##name##step2##_##bitd##_##opt(dst + step1, src + (step1 * ((bitd + 7) / 8)), \
254 _srcstride, height, mx, my, width); \
255 }
256 #define mc_rep_uni_func2(name, bitd, step1, step2, W, opt) \
257 static void hevc_put_uni_##name##W##_##bitd##_##opt(uint8_t *dst, ptrdiff_t dststride, \
258 const uint8_t *src, ptrdiff_t _srcstride, int height, \
259 intptr_t mx, intptr_t my, int width) \
260 { \
261 hevc_put_uni_##name##step1##_##bitd##_##opt(dst, dststride, src, _srcstride, height, mx, my, width); \
262 hevc_put_uni_##name##step2##_##bitd##_##opt(dst + (step1 * ((bitd + 7) / 8)), dststride, \
263 src + (step1 * ((bitd + 7) / 8)), _srcstride, \
264 height, mx, my, width); \
265 }
266 #define mc_rep_bi_func2(name, bitd, step1, step2, W, opt) \
267 static void ff_hevc_put_bi_##name##W##_##bitd##_##opt(uint8_t *dst, ptrdiff_t dststride, const uint8_t *src, \
268 ptrdiff_t _srcstride, const int16_t *src2, \
269 int height, intptr_t mx, intptr_t my, int width) \
270 { \
271 ff_hevc_put_bi_##name##step1##_##bitd##_##opt(dst, dststride, src, _srcstride, src2, height, mx, my, width);\
272 ff_hevc_put_bi_##name##step2##_##bitd##_##opt(dst + (step1 * ((bitd + 7) / 8)), dststride, \
273 src + (step1 * ((bitd + 7) / 8)), _srcstride, \
274 src2 + step1, height, mx, my, width); \
275 }
276
277 #define mc_rep_funcs2(name, bitd, step1, step2, W, opt) \
278 mc_rep_func2(name, bitd, step1, step2, W, opt) \
279 mc_rep_uni_func2(name, bitd, step1, step2, W, opt) \
280 mc_rep_bi_func2(name, bitd, step1, step2, W, opt)
281
282 #if ARCH_X86_64 && HAVE_SSE4_EXTERNAL
283
284 #define mc_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \
285 static void hevc_put_##name##width1##_10_##opt1(int16_t *dst, const uint8_t *src, ptrdiff_t _srcstride, \
286 int height, intptr_t mx, intptr_t my, int width) \
287 \
288 { \
289 hevc_put_##name##width2##_10_##opt1(dst, src, _srcstride, height, mx, my, width); \
290 hevc_put_##name##width3##_10_##opt2(dst+ width2, src+ width4, _srcstride, height, mx, my, width); \
291 }
292
293 #define mc_bi_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \
294 static void ff_hevc_put_bi_##name##width1##_10_##opt1(uint8_t *dst, ptrdiff_t dststride, const uint8_t *src, \
295 ptrdiff_t _srcstride, const int16_t *src2, \
296 int height, intptr_t mx, intptr_t my, int width) \
297 { \
298 ff_hevc_put_bi_##name##width2##_10_##opt1(dst, dststride, src, _srcstride, src2, \
299 height, mx, my, width); \
300 ff_hevc_put_bi_##name##width3##_10_##opt2(dst+width4, dststride, src+width4, _srcstride, src2+width2, \
301 height, mx, my, width); \
302 }
303
304 #define mc_uni_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \
305 static void hevc_put_uni_##name##width1##_10_##opt1(uint8_t *dst, ptrdiff_t dststride, \
306 const uint8_t *src, ptrdiff_t _srcstride, int height, \
307 intptr_t mx, intptr_t my, int width) \
308 { \
309 hevc_put_uni_##name##width2##_10_##opt1(dst, dststride, src, _srcstride, \
310 height, mx, my, width); \
311 hevc_put_uni_##name##width3##_10_##opt2(dst+width4, dststride, src+width4, _srcstride, \
312 height, mx, my, width); \
313 }
314
315 #define mc_rep_mixs_10(name, width1, width2, width3, opt1, opt2, width4) \
316 mc_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \
317 mc_bi_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \
318 mc_uni_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4)
319
320 #define mc_rep_mix_8(name, width1, width2, width3, opt1, opt2) \
321 static void hevc_put_##name##width1##_8_##opt1(int16_t *dst, const uint8_t *src, ptrdiff_t _srcstride, \
322 int height, intptr_t mx, intptr_t my, int width) \
323 \
324 { \
325 hevc_put_##name##width2##_8_##opt1(dst, src, _srcstride, height, mx, my, width); \
326 hevc_put_##name##width3##_8_##opt2(dst+ width2, src+ width2, _srcstride, height, mx, my, width); \
327 }
328
329 #define mc_bi_rep_mix_8(name, width1, width2, width3, opt1, opt2) \
330 static void ff_hevc_put_bi_##name##width1##_8_##opt1(uint8_t *dst, ptrdiff_t dststride, const uint8_t *src, \
331 ptrdiff_t _srcstride, const int16_t *src2, \
332 int height, intptr_t mx, intptr_t my, int width) \
333 { \
334 ff_hevc_put_bi_##name##width2##_8_##opt1(dst, dststride, src, _srcstride, \
335 src2, height, mx, my, width); \
336 ff_hevc_put_bi_##name##width3##_8_##opt2(dst+width2, dststride, src+width2, _srcstride, \
337 src2+width2, height, mx, my, width); \
338 }
339
340 #define mc_uni_rep_mix_8(name, width1, width2, width3, opt1, opt2) \
341 static void hevc_put_uni_##name##width1##_8_##opt1(uint8_t *dst, ptrdiff_t dststride, \
342 const uint8_t *src, ptrdiff_t _srcstride, int height, \
343 intptr_t mx, intptr_t my, int width) \
344 { \
345 hevc_put_uni_##name##width2##_8_##opt1(dst, dststride, src, _srcstride, \
346 height, mx, my, width); \
347 hevc_put_uni_##name##width3##_8_##opt2(dst+width2, dststride, src+width2, _srcstride, \
348 height, mx, my, width); \
349 }
350
351 #define mc_rep_mixs_8(name, width1, width2, width3, opt1, opt2) \
352 mc_rep_mix_8(name, width1, width2, width3, opt1, opt2) \
353 mc_bi_rep_mix_8(name, width1, width2, width3, opt1, opt2) \
354 mc_uni_rep_mix_8(name, width1, width2, width3, opt1, opt2)
355
356 #if HAVE_AVX2_EXTERNAL
357
358 16 mc_rep_mixs_8(pel_pixels, 48, 32, 16, avx2, sse4)
359 12 mc_rep_mixs_8(epel_hv, 48, 32, 16, avx2, sse4)
360 12 mc_rep_mixs_8(epel_h , 48, 32, 16, avx2, sse4)
361 12 mc_rep_mixs_8(epel_v , 48, 32, 16, avx2, sse4)
362
363 6 mc_rep_mix_10(pel_pixels, 24, 16, 8, avx2, sse4, 32)
364 2 mc_bi_rep_mix_10(pel_pixels,24, 16, 8, avx2, sse4, 32)
365 12 mc_rep_mixs_10(epel_hv, 24, 16, 8, avx2, sse4, 32)
366 12 mc_rep_mixs_10(epel_h , 24, 16, 8, avx2, sse4, 32)
367 12 mc_rep_mixs_10(epel_v , 24, 16, 8, avx2, sse4, 32)
368
369
370 12 mc_rep_mixs_10(qpel_h , 24, 16, 8, avx2, sse4, 32)
371 12 mc_rep_mixs_10(qpel_v , 24, 16, 8, avx2, sse4, 32)
372 12 mc_rep_mixs_10(qpel_hv, 24, 16, 8, avx2, sse4, 32)
373
374
375
2/2
✓ Branch 1 taken 6734 times.
✓ Branch 2 taken 3367 times.
 20202 mc_rep_funcs(pel_pixels, 8, 32, 64, avx2)
376
377
2/2
✓ Branch 1 taken 158 times.
✓ Branch 2 taken 79 times.
 237 mc_rep_uni_func(pel_pixels, 8, 64, 128, avx2)//used for 10bit
378
2/2
✓ Branch 1 taken 6 times.
✓ Branch 2 taken 2 times.
 8 mc_rep_uni_func(pel_pixels, 8, 32, 96, avx2) //used for 10bit
379
380
2/2
✓ Branch 1 taken 2032 times.
✓ Branch 2 taken 1016 times.
 3048 mc_rep_func(pel_pixels, 10, 16, 32, avx2)
381
2/2
✓ Branch 1 taken 6 times.
✓ Branch 2 taken 2 times.
 8 mc_rep_func(pel_pixels, 10, 16, 48, avx2)
382
2/2
✓ Branch 1 taken 306 times.
✓ Branch 2 taken 153 times.
 459 mc_rep_func(pel_pixels, 10, 32, 64, avx2)
383
384
2/2
✓ Branch 1 taken 182 times.
✓ Branch 2 taken 91 times.
 273 mc_rep_bi_func(pel_pixels, 10, 16, 32, avx2)
385
2/2
✓ Branch 1 taken 6 times.
✓ Branch 2 taken 2 times.
 8 mc_rep_bi_func(pel_pixels, 10, 16, 48, avx2)
386
2/2
✓ Branch 1 taken 44 times.
✓ Branch 2 taken 22 times.
 66 mc_rep_bi_func(pel_pixels, 10, 32, 64, avx2)
387
388
2/2
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 6 times.
 36 mc_rep_funcs(epel_h, 8, 32, 64, avx2)
389
390
2/2
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 6 times.
 36 mc_rep_funcs(epel_v, 8, 32, 64, avx2)
391
392
2/2
✓ Branch 1 taken 1308 times.
✓ Branch 2 taken 654 times.
 3924 mc_rep_funcs(epel_h, 10, 16, 32, avx2)
393
2/2
✓ Branch 1 taken 18 times.
✓ Branch 2 taken 6 times.
 48 mc_rep_funcs(epel_h, 10, 16, 48, avx2)
394
2/2
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 6 times.
 36 mc_rep_funcs(epel_h, 10, 32, 64, avx2)
395
396
2/2
✓ Branch 1 taken 172 times.
✓ Branch 2 taken 86 times.
 516 mc_rep_funcs(epel_v, 10, 16, 32, avx2)
397
2/2
✓ Branch 1 taken 18 times.
✓ Branch 2 taken 6 times.
 48 mc_rep_funcs(epel_v, 10, 16, 48, avx2)
398
2/2
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 6 times.
 36 mc_rep_funcs(epel_v, 10, 32, 64, avx2)
399
400
401
2/2
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 6 times.
 36 mc_rep_funcs(epel_hv, 8, 32, 64, avx2)
402
403
2/2
✓ Branch 1 taken 1616 times.
✓ Branch 2 taken 808 times.
 4848 mc_rep_funcs(epel_hv, 10, 16, 32, avx2)
404
2/2
✓ Branch 1 taken 18 times.
✓ Branch 2 taken 6 times.
 48 mc_rep_funcs(epel_hv, 10, 16, 48, avx2)
405
2/2
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 6 times.
 36 mc_rep_funcs(epel_hv, 10, 32, 64, avx2)
406
407
2/2
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 6 times.
 36 mc_rep_funcs(qpel_h, 8, 32, 64, avx2)
408 12 mc_rep_mixs_8(qpel_h , 48, 32, 16, avx2, sse4)
409
410
2/2
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 6 times.
 36 mc_rep_funcs(qpel_v, 8, 32, 64, avx2)
411 12 mc_rep_mixs_8(qpel_v, 48, 32, 16, avx2, sse4)
412
413
2/2
✓ Branch 1 taken 2000 times.
✓ Branch 2 taken 1000 times.
 6000 mc_rep_funcs(qpel_h, 10, 16, 32, avx2)
414
2/2
✓ Branch 1 taken 18 times.
✓ Branch 2 taken 6 times.
 48 mc_rep_funcs(qpel_h, 10, 16, 48, avx2)
415
2/2
✓ Branch 1 taken 518 times.
✓ Branch 2 taken 259 times.
 1554 mc_rep_funcs(qpel_h, 10, 32, 64, avx2)
416
417
2/2
✓ Branch 1 taken 928 times.
✓ Branch 2 taken 464 times.
 2784 mc_rep_funcs(qpel_v, 10, 16, 32, avx2)
418
2/2
✓ Branch 1 taken 18 times.
✓ Branch 2 taken 6 times.
 48 mc_rep_funcs(qpel_v, 10, 16, 48, avx2)
419
2/2
✓ Branch 1 taken 208 times.
✓ Branch 2 taken 104 times.
 624 mc_rep_funcs(qpel_v, 10, 32, 64, avx2)
420
421
2/2
✓ Branch 1 taken 3134 times.
✓ Branch 2 taken 1567 times.
 9402 mc_rep_funcs(qpel_hv, 10, 16, 32, avx2)
422
2/2
✓ Branch 1 taken 18 times.
✓ Branch 2 taken 6 times.
 48 mc_rep_funcs(qpel_hv, 10, 16, 48, avx2)
423
2/2
✓ Branch 1 taken 674 times.
✓ Branch 2 taken 337 times.
 2022 mc_rep_funcs(qpel_hv, 10, 32, 64, avx2)
424
425 #endif //AVX2
426
427
2/2
✓ Branch 1 taken 30984 times.
✓ Branch 2 taken 7746 times.
 77460 mc_rep_funcs(pel_pixels, 8, 16, 64, sse4)
428
2/2
✓ Branch 1 taken 171 times.
✓ Branch 2 taken 57 times.
 456 mc_rep_funcs(pel_pixels, 8, 16, 48, sse4)
429
2/2
✓ Branch 1 taken 34012 times.
✓ Branch 2 taken 17006 times.
 102036 mc_rep_funcs(pel_pixels, 8, 16, 32, sse4)
430
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(pel_pixels, 8, 8, 24, sse4)
431
2/2
✓ Branch 1 taken 3080 times.
✓ Branch 2 taken 385 times.
 6930 mc_rep_funcs(pel_pixels,10, 8, 64, sse4)
432
2/2
✓ Branch 1 taken 342 times.
✓ Branch 2 taken 57 times.
 798 mc_rep_funcs(pel_pixels,10, 8, 48, sse4)
433
2/2
✓ Branch 1 taken 6268 times.
✓ Branch 2 taken 1567 times.
 15670 mc_rep_funcs(pel_pixels,10, 8, 32, sse4)
434
2/2
✓ Branch 1 taken 189 times.
✓ Branch 2 taken 63 times.
 504 mc_rep_funcs(pel_pixels,10, 8, 24, sse4)
435
2/2
✓ Branch 1 taken 4050 times.
✓ Branch 2 taken 2025 times.
 12150 mc_rep_funcs(pel_pixels,10, 8, 16, sse4)
436
2/2
✓ Branch 1 taken 222 times.
✓ Branch 2 taken 74 times.
 592 mc_rep_funcs(pel_pixels,10, 4, 12, sse4)
437
2/2
✓ Branch 1 taken 432 times.
✓ Branch 2 taken 54 times.
 972 mc_rep_funcs(pel_pixels,12, 8, 64, sse4)
438
2/2
✓ Branch 1 taken 324 times.
✓ Branch 2 taken 54 times.
 756 mc_rep_funcs(pel_pixels,12, 8, 48, sse4)
439
2/2
✓ Branch 1 taken 216 times.
✓ Branch 2 taken 54 times.
 540 mc_rep_funcs(pel_pixels,12, 8, 32, sse4)
440
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(pel_pixels,12, 8, 24, sse4)
441
2/2
✓ Branch 1 taken 108 times.
✓ Branch 2 taken 54 times.
 324 mc_rep_funcs(pel_pixels,12, 8, 16, sse4)
442
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(pel_pixels,12, 4, 12, sse4)
443
444
2/2
✓ Branch 1 taken 228 times.
✓ Branch 2 taken 57 times.
 570 mc_rep_funcs(epel_h, 8, 16, 64, sse4)
445
2/2
✓ Branch 1 taken 171 times.
✓ Branch 2 taken 57 times.
 456 mc_rep_funcs(epel_h, 8, 16, 48, sse4)
446
2/2
✓ Branch 1 taken 1106 times.
✓ Branch 2 taken 553 times.
 3318 mc_rep_funcs(epel_h, 8, 16, 32, sse4)
447
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(epel_h, 8, 8, 24, sse4)
448
2/2
✓ Branch 1 taken 456 times.
✓ Branch 2 taken 57 times.
 1026 mc_rep_funcs(epel_h,10, 8, 64, sse4)
449
2/2
✓ Branch 1 taken 342 times.
✓ Branch 2 taken 57 times.
 798 mc_rep_funcs(epel_h,10, 8, 48, sse4)
450
2/2
✓ Branch 1 taken 1076 times.
✓ Branch 2 taken 269 times.
 2690 mc_rep_funcs(epel_h,10, 8, 32, sse4)
451
2/2
✓ Branch 1 taken 171 times.
✓ Branch 2 taken 57 times.
 456 mc_rep_funcs(epel_h,10, 8, 24, sse4)
452
2/2
✓ Branch 1 taken 1170 times.
✓ Branch 2 taken 585 times.
 3510 mc_rep_funcs(epel_h,10, 8, 16, sse4)
453
2/2
✓ Branch 1 taken 174 times.
✓ Branch 2 taken 58 times.
 464 mc_rep_funcs(epel_h,10, 4, 12, sse4)
454
2/2
✓ Branch 1 taken 432 times.
✓ Branch 2 taken 54 times.
 972 mc_rep_funcs(epel_h,12, 8, 64, sse4)
455
2/2
✓ Branch 1 taken 324 times.
✓ Branch 2 taken 54 times.
 756 mc_rep_funcs(epel_h,12, 8, 48, sse4)
456
2/2
✓ Branch 1 taken 216 times.
✓ Branch 2 taken 54 times.
 540 mc_rep_funcs(epel_h,12, 8, 32, sse4)
457
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(epel_h,12, 8, 24, sse4)
458
2/2
✓ Branch 1 taken 108 times.
✓ Branch 2 taken 54 times.
 324 mc_rep_funcs(epel_h,12, 8, 16, sse4)
459
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(epel_h,12, 4, 12, sse4)
460
2/2
✓ Branch 1 taken 228 times.
✓ Branch 2 taken 57 times.
 570 mc_rep_funcs(epel_v, 8, 16, 64, sse4)
461
2/2
✓ Branch 1 taken 171 times.
✓ Branch 2 taken 57 times.
 456 mc_rep_funcs(epel_v, 8, 16, 48, sse4)
462
2/2
✓ Branch 1 taken 2086 times.
✓ Branch 2 taken 1043 times.
 6258 mc_rep_funcs(epel_v, 8, 16, 32, sse4)
463
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(epel_v, 8, 8, 24, sse4)
464
2/2
✓ Branch 1 taken 456 times.
✓ Branch 2 taken 57 times.
 1026 mc_rep_funcs(epel_v,10, 8, 64, sse4)
465
2/2
✓ Branch 1 taken 342 times.
✓ Branch 2 taken 57 times.
 798 mc_rep_funcs(epel_v,10, 8, 48, sse4)
466
2/2
✓ Branch 1 taken 356 times.
✓ Branch 2 taken 89 times.
 890 mc_rep_funcs(epel_v,10, 8, 32, sse4)
467
2/2
✓ Branch 1 taken 171 times.
✓ Branch 2 taken 57 times.
 456 mc_rep_funcs(epel_v,10, 8, 24, sse4)
468
2/2
✓ Branch 1 taken 234 times.
✓ Branch 2 taken 117 times.
 702 mc_rep_funcs(epel_v,10, 8, 16, sse4)
469
2/2
✓ Branch 1 taken 186 times.
✓ Branch 2 taken 62 times.
 496 mc_rep_funcs(epel_v,10, 4, 12, sse4)
470
2/2
✓ Branch 1 taken 432 times.
✓ Branch 2 taken 54 times.
 972 mc_rep_funcs(epel_v,12, 8, 64, sse4)
471
2/2
✓ Branch 1 taken 324 times.
✓ Branch 2 taken 54 times.
 756 mc_rep_funcs(epel_v,12, 8, 48, sse4)
472
2/2
✓ Branch 1 taken 216 times.
✓ Branch 2 taken 54 times.
 540 mc_rep_funcs(epel_v,12, 8, 32, sse4)
473
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(epel_v,12, 8, 24, sse4)
474
2/2
✓ Branch 1 taken 108 times.
✓ Branch 2 taken 54 times.
 324 mc_rep_funcs(epel_v,12, 8, 16, sse4)
475
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(epel_v,12, 4, 12, sse4)
476
2/2
✓ Branch 1 taken 228 times.
✓ Branch 2 taken 57 times.
 570 mc_rep_funcs(epel_hv, 8, 16, 64, sse4)
477
2/2
✓ Branch 1 taken 171 times.
✓ Branch 2 taken 57 times.
 456 mc_rep_funcs(epel_hv, 8, 16, 48, sse4)
478
2/2
✓ Branch 1 taken 1890 times.
✓ Branch 2 taken 945 times.
 5670 mc_rep_funcs(epel_hv, 8, 16, 32, sse4)
479
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(epel_hv, 8, 8, 24, sse4)
480 108 mc_rep_funcs2(epel_hv,8, 8, 4, 12, sse4)
481
2/2
✓ Branch 1 taken 456 times.
✓ Branch 2 taken 57 times.
 1026 mc_rep_funcs(epel_hv,10, 8, 64, sse4)
482
2/2
✓ Branch 1 taken 342 times.
✓ Branch 2 taken 57 times.
 798 mc_rep_funcs(epel_hv,10, 8, 48, sse4)
483
2/2
✓ Branch 1 taken 372 times.
✓ Branch 2 taken 93 times.
 930 mc_rep_funcs(epel_hv,10, 8, 32, sse4)
484
2/2
✓ Branch 1 taken 171 times.
✓ Branch 2 taken 57 times.
 456 mc_rep_funcs(epel_hv,10, 8, 24, sse4)
485
2/2
✓ Branch 1 taken 266 times.
✓ Branch 2 taken 133 times.
 798 mc_rep_funcs(epel_hv,10, 8, 16, sse4)
486
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(epel_hv,10, 4, 12, sse4)
487
2/2
✓ Branch 1 taken 432 times.
✓ Branch 2 taken 54 times.
 972 mc_rep_funcs(epel_hv,12, 8, 64, sse4)
488
2/2
✓ Branch 1 taken 324 times.
✓ Branch 2 taken 54 times.
 756 mc_rep_funcs(epel_hv,12, 8, 48, sse4)
489
2/2
✓ Branch 1 taken 216 times.
✓ Branch 2 taken 54 times.
 540 mc_rep_funcs(epel_hv,12, 8, 32, sse4)
490
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(epel_hv,12, 8, 24, sse4)
491
2/2
✓ Branch 1 taken 108 times.
✓ Branch 2 taken 54 times.
 324 mc_rep_funcs(epel_hv,12, 8, 16, sse4)
492
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(epel_hv,12, 4, 12, sse4)
493
494
2/2
✓ Branch 1 taken 228 times.
✓ Branch 2 taken 57 times.
 570 mc_rep_funcs(qpel_h, 8, 16, 64, sse4)
495
2/2
✓ Branch 1 taken 171 times.
✓ Branch 2 taken 57 times.
 456 mc_rep_funcs(qpel_h, 8, 16, 48, sse4)
496
2/2
✓ Branch 1 taken 188 times.
✓ Branch 2 taken 94 times.
 564 mc_rep_funcs(qpel_h, 8, 16, 32, sse4)
497
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(qpel_h, 8, 8, 24, sse4)
498
2/2
✓ Branch 1 taken 1304 times.
✓ Branch 2 taken 163 times.
 2934 mc_rep_funcs(qpel_h,10, 8, 64, sse4)
499
2/2
✓ Branch 1 taken 342 times.
✓ Branch 2 taken 57 times.
 798 mc_rep_funcs(qpel_h,10, 8, 48, sse4)
500
2/2
✓ Branch 1 taken 1308 times.
✓ Branch 2 taken 327 times.
 3270 mc_rep_funcs(qpel_h,10, 8, 32, sse4)
501
2/2
✓ Branch 1 taken 177 times.
✓ Branch 2 taken 59 times.
 472 mc_rep_funcs(qpel_h,10, 8, 24, sse4)
502
2/2
✓ Branch 1 taken 454 times.
✓ Branch 2 taken 227 times.
 1362 mc_rep_funcs(qpel_h,10, 8, 16, sse4)
503
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(qpel_h,10, 4, 12, sse4)
504
2/2
✓ Branch 1 taken 432 times.
✓ Branch 2 taken 54 times.
 972 mc_rep_funcs(qpel_h,12, 8, 64, sse4)
505
2/2
✓ Branch 1 taken 324 times.
✓ Branch 2 taken 54 times.
 756 mc_rep_funcs(qpel_h,12, 8, 48, sse4)
506
2/2
✓ Branch 1 taken 216 times.
✓ Branch 2 taken 54 times.
 540 mc_rep_funcs(qpel_h,12, 8, 32, sse4)
507
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(qpel_h,12, 8, 24, sse4)
508
2/2
✓ Branch 1 taken 108 times.
✓ Branch 2 taken 54 times.
 324 mc_rep_funcs(qpel_h,12, 8, 16, sse4)
509
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(qpel_h,12, 4, 12, sse4)
510
2/2
✓ Branch 1 taken 236 times.
✓ Branch 2 taken 59 times.
 590 mc_rep_funcs(qpel_v, 8, 16, 64, sse4)
511
2/2
✓ Branch 1 taken 171 times.
✓ Branch 2 taken 57 times.
 456 mc_rep_funcs(qpel_v, 8, 16, 48, sse4)
512
2/2
✓ Branch 1 taken 164 times.
✓ Branch 2 taken 82 times.
 492 mc_rep_funcs(qpel_v, 8, 16, 32, sse4)
513
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(qpel_v, 8, 8, 24, sse4)
514
2/2
✓ Branch 1 taken 584 times.
✓ Branch 2 taken 73 times.
 1314 mc_rep_funcs(qpel_v,10, 8, 64, sse4)
515
2/2
✓ Branch 1 taken 342 times.
✓ Branch 2 taken 57 times.
 798 mc_rep_funcs(qpel_v,10, 8, 48, sse4)
516
2/2
✓ Branch 1 taken 364 times.
✓ Branch 2 taken 91 times.
 910 mc_rep_funcs(qpel_v,10, 8, 32, sse4)
517
2/2
✓ Branch 1 taken 183 times.
✓ Branch 2 taken 61 times.
 488 mc_rep_funcs(qpel_v,10, 8, 24, sse4)
518
2/2
✓ Branch 1 taken 126 times.
✓ Branch 2 taken 63 times.
 378 mc_rep_funcs(qpel_v,10, 8, 16, sse4)
519
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(qpel_v,10, 4, 12, sse4)
520
2/2
✓ Branch 1 taken 432 times.
✓ Branch 2 taken 54 times.
 972 mc_rep_funcs(qpel_v,12, 8, 64, sse4)
521
2/2
✓ Branch 1 taken 324 times.
✓ Branch 2 taken 54 times.
 756 mc_rep_funcs(qpel_v,12, 8, 48, sse4)
522
2/2
✓ Branch 1 taken 216 times.
✓ Branch 2 taken 54 times.
 540 mc_rep_funcs(qpel_v,12, 8, 32, sse4)
523
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(qpel_v,12, 8, 24, sse4)
524
2/2
✓ Branch 1 taken 108 times.
✓ Branch 2 taken 54 times.
 324 mc_rep_funcs(qpel_v,12, 8, 16, sse4)
525
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(qpel_v,12, 4, 12, sse4)
526
2/2
✓ Branch 1 taken 1048 times.
✓ Branch 2 taken 131 times.
 2358 mc_rep_funcs(qpel_hv, 8, 8, 64, sse4)
527
2/2
✓ Branch 1 taken 324 times.
✓ Branch 2 taken 54 times.
 756 mc_rep_funcs(qpel_hv, 8, 8, 48, sse4)
528
2/2
✓ Branch 1 taken 764 times.
✓ Branch 2 taken 191 times.
 1910 mc_rep_funcs(qpel_hv, 8, 8, 32, sse4)
529
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(qpel_hv, 8, 8, 24, sse4)
530
2/2
✓ Branch 1 taken 1252 times.
✓ Branch 2 taken 626 times.
 3756 mc_rep_funcs(qpel_hv, 8, 8, 16, sse4)
531 108 mc_rep_funcs2(qpel_hv,8, 8, 4, 12, sse4)
532
2/2
✓ Branch 1 taken 600 times.
✓ Branch 2 taken 75 times.
 1350 mc_rep_funcs(qpel_hv,10, 8, 64, sse4)
533
2/2
✓ Branch 1 taken 342 times.
✓ Branch 2 taken 57 times.
 798 mc_rep_funcs(qpel_hv,10, 8, 48, sse4)
534
2/2
✓ Branch 1 taken 332 times.
✓ Branch 2 taken 83 times.
 830 mc_rep_funcs(qpel_hv,10, 8, 32, sse4)
535
2/2
✓ Branch 1 taken 171 times.
✓ Branch 2 taken 57 times.
 456 mc_rep_funcs(qpel_hv,10, 8, 24, sse4)
536
2/2
✓ Branch 1 taken 130 times.
✓ Branch 2 taken 65 times.
 390 mc_rep_funcs(qpel_hv,10, 8, 16, sse4)
537
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(qpel_hv,10, 4, 12, sse4)
538
2/2
✓ Branch 1 taken 432 times.
✓ Branch 2 taken 54 times.
 972 mc_rep_funcs(qpel_hv,12, 8, 64, sse4)
539
2/2
✓ Branch 1 taken 324 times.
✓ Branch 2 taken 54 times.
 756 mc_rep_funcs(qpel_hv,12, 8, 48, sse4)
540
2/2
✓ Branch 1 taken 216 times.
✓ Branch 2 taken 54 times.
 540 mc_rep_funcs(qpel_hv,12, 8, 32, sse4)
541
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(qpel_hv,12, 8, 24, sse4)
542
2/2
✓ Branch 1 taken 108 times.
✓ Branch 2 taken 54 times.
 324 mc_rep_funcs(qpel_hv,12, 8, 16, sse4)
543
2/2
✓ Branch 1 taken 162 times.
✓ Branch 2 taken 54 times.
 432 mc_rep_funcs(qpel_hv,12, 4, 12, sse4)
544
545 #define mc_rep_uni_w(bitd, step, W, opt) \
546 void ff_hevc_put_uni_w##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, const int16_t *_src, \
547 int height, int denom, int _wx, int _ox) \
548 { \
549 int i; \
550 uint8_t *dst; \
551 for (i = 0; i < W; i += step) { \
552 const int16_t *src = _src + i; \
553 dst= _dst + (i * ((bitd + 7) / 8)); \
554 ff_hevc_put_uni_w##step##_##bitd##_##opt(dst, dststride, src, \
555 height, denom, _wx, _ox); \
556 } \
557 }
558
559
2/2
✓ Branch 1 taken 336 times.
✓ Branch 2 taken 168 times.
 504 mc_rep_uni_w(8, 6, 12, sse4)
560
2/2
✓ Branch 1 taken 23652 times.
✓ Branch 2 taken 11826 times.
 35478 mc_rep_uni_w(8, 8, 16, sse4)
561
2/2
✓ Branch 1 taken 504 times.
✓ Branch 2 taken 168 times.
 672 mc_rep_uni_w(8, 8, 24, sse4)
562
2/2
✓ Branch 1 taken 55472 times.
✓ Branch 2 taken 13868 times.
 69340 mc_rep_uni_w(8, 8, 32, sse4)
563
2/2
✓ Branch 1 taken 1008 times.
✓ Branch 2 taken 168 times.
 1176 mc_rep_uni_w(8, 8, 48, sse4)
564
2/2
✓ Branch 1 taken 43352 times.
✓ Branch 2 taken 5419 times.
 48771 mc_rep_uni_w(8, 8, 64, sse4)
565
566
2/2
✓ Branch 1 taken 368 times.
✓ Branch 2 taken 184 times.
 552 mc_rep_uni_w(10, 6, 12, sse4)
567
2/2
✓ Branch 1 taken 3304 times.
✓ Branch 2 taken 1652 times.
 4956 mc_rep_uni_w(10, 8, 16, sse4)
568
2/2
✓ Branch 1 taken 528 times.
✓ Branch 2 taken 176 times.
 704 mc_rep_uni_w(10, 8, 24, sse4)
569
2/2
✓ Branch 1 taken 4896 times.
✓ Branch 2 taken 1224 times.
 6120 mc_rep_uni_w(10, 8, 32, sse4)
570
2/2
✓ Branch 1 taken 1008 times.
✓ Branch 2 taken 168 times.
 1176 mc_rep_uni_w(10, 8, 48, sse4)
571
2/2
✓ Branch 1 taken 3168 times.
✓ Branch 2 taken 396 times.
 3564 mc_rep_uni_w(10, 8, 64, sse4)
572
573
2/2
✓ Branch 1 taken 336 times.
✓ Branch 2 taken 168 times.
 504 mc_rep_uni_w(12, 6, 12, sse4)
574
2/2
✓ Branch 1 taken 336 times.
✓ Branch 2 taken 168 times.
 504 mc_rep_uni_w(12, 8, 16, sse4)
575
2/2
✓ Branch 1 taken 504 times.
✓ Branch 2 taken 168 times.
 672 mc_rep_uni_w(12, 8, 24, sse4)
576
2/2
✓ Branch 1 taken 672 times.
✓ Branch 2 taken 168 times.
 840 mc_rep_uni_w(12, 8, 32, sse4)
577
2/2
✓ Branch 1 taken 1008 times.
✓ Branch 2 taken 168 times.
 1176 mc_rep_uni_w(12, 8, 48, sse4)
578
2/2
✓ Branch 1 taken 1344 times.
✓ Branch 2 taken 168 times.
 1512 mc_rep_uni_w(12, 8, 64, sse4)
579
580 #define mc_rep_bi_w(bitd, step, W, opt) \
581 void ff_hevc_put_bi_w##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, const int16_t *_src, \
582 const int16_t *_src2, int height, \
583 int denom, int wx0, int wx1, int ox) \
584 { \
585 int i; \
586 uint8_t *dst; \
587 for (i = 0; i < W; i += step) { \
588 const int16_t *src = _src + i; \
589 const int16_t *src2 = _src2 + i; \
590 dst = _dst + (i * ((bitd + 7) / 8)); \
591 ff_hevc_put_bi_w##step##_##bitd##_##opt(dst, dststride, src, src2, \
592 height, denom, wx0, wx1, ox); \
593 } \
594 }
595
596
2/2
✓ Branch 1 taken 336 times.
✓ Branch 2 taken 168 times.
 504 mc_rep_bi_w(8, 6, 12, sse4)
597
2/2
✓ Branch 1 taken 4980 times.
✓ Branch 2 taken 2490 times.
 7470 mc_rep_bi_w(8, 8, 16, sse4)
598
2/2
✓ Branch 1 taken 504 times.
✓ Branch 2 taken 168 times.
 672 mc_rep_bi_w(8, 8, 24, sse4)
599
2/2
✓ Branch 1 taken 23720 times.
✓ Branch 2 taken 5930 times.
 29650 mc_rep_bi_w(8, 8, 32, sse4)
600
2/2
✓ Branch 1 taken 1008 times.
✓ Branch 2 taken 168 times.
 1176 mc_rep_bi_w(8, 8, 48, sse4)
601
2/2
✓ Branch 1 taken 21480 times.
✓ Branch 2 taken 2685 times.
 24165 mc_rep_bi_w(8, 8, 64, sse4)
602
603
2/2
✓ Branch 1 taken 352 times.
✓ Branch 2 taken 176 times.
 528 mc_rep_bi_w(10, 6, 12, sse4)
604
2/2
✓ Branch 1 taken 3000 times.
✓ Branch 2 taken 1500 times.
 4500 mc_rep_bi_w(10, 8, 16, sse4)
605
2/2
✓ Branch 1 taken 516 times.
✓ Branch 2 taken 172 times.
 688 mc_rep_bi_w(10, 8, 24, sse4)
606
2/2
✓ Branch 1 taken 4928 times.
✓ Branch 2 taken 1232 times.
 6160 mc_rep_bi_w(10, 8, 32, sse4)
607
2/2
✓ Branch 1 taken 1008 times.
✓ Branch 2 taken 168 times.
 1176 mc_rep_bi_w(10, 8, 48, sse4)
608
2/2
✓ Branch 1 taken 3264 times.
✓ Branch 2 taken 408 times.
 3672 mc_rep_bi_w(10, 8, 64, sse4)
609
610
2/2
✓ Branch 1 taken 336 times.
✓ Branch 2 taken 168 times.
 504 mc_rep_bi_w(12, 6, 12, sse4)
611
2/2
✓ Branch 1 taken 336 times.
✓ Branch 2 taken 168 times.
 504 mc_rep_bi_w(12, 8, 16, sse4)
612
2/2
✓ Branch 1 taken 504 times.
✓ Branch 2 taken 168 times.
 672 mc_rep_bi_w(12, 8, 24, sse4)
613
2/2
✓ Branch 1 taken 672 times.
✓ Branch 2 taken 168 times.
 840 mc_rep_bi_w(12, 8, 32, sse4)
614
2/2
✓ Branch 1 taken 1008 times.
✓ Branch 2 taken 168 times.
 1176 mc_rep_bi_w(12, 8, 48, sse4)
615
2/2
✓ Branch 1 taken 1344 times.
✓ Branch 2 taken 168 times.
 1512 mc_rep_bi_w(12, 8, 64, sse4)
616
617 #define mc_uni_w_func(name, bitd, W, opt) \
618 static void hevc_put_uni_w_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t _dststride, \
619 const uint8_t *_src, ptrdiff_t _srcstride, \
620 int height, int denom, \
621 int _wx, int _ox, \
622 intptr_t mx, intptr_t my, int width) \
623 { \
624 LOCAL_ALIGNED_16(int16_t, temp, [71 * MAX_PB_SIZE]); \
625 hevc_put_##name##W##_##bitd##_##opt(temp, _src, _srcstride, height, mx, my, width); \
626 ff_hevc_put_uni_w##W##_##bitd##_##opt(_dst, _dststride, temp, height, denom, _wx, _ox); \
627 }
628
629 #define mc_uni_w_funcs(name, bitd, opt) \
630 mc_uni_w_func(name, bitd, 4, opt) \
631 mc_uni_w_func(name, bitd, 8, opt) \
632 mc_uni_w_func(name, bitd, 12, opt) \
633 mc_uni_w_func(name, bitd, 16, opt) \
634 mc_uni_w_func(name, bitd, 24, opt) \
635 mc_uni_w_func(name, bitd, 32, opt) \
636 mc_uni_w_func(name, bitd, 48, opt) \
637 mc_uni_w_func(name, bitd, 64, opt)
638
639 86198 mc_uni_w_funcs(pel_pixels, 8, sse4)
640 24 mc_uni_w_func(pel_pixels, 8, 6, sse4)
641 10360 mc_uni_w_funcs(epel_h, 8, sse4)
642 24 mc_uni_w_func(epel_h, 8, 6, sse4)
643 9276 mc_uni_w_funcs(epel_v, 8, sse4)
644 24 mc_uni_w_func(epel_v, 8, 6, sse4)
645 13300 mc_uni_w_funcs(epel_hv, 8, sse4)
646 24 mc_uni_w_func(epel_hv, 8, 6, sse4)
647 1952 mc_uni_w_funcs(qpel_h, 8, sse4)
648 1508 mc_uni_w_funcs(qpel_v, 8, sse4)
649 3960 mc_uni_w_funcs(qpel_hv, 8, sse4)
650
651 5076 mc_uni_w_funcs(pel_pixels, 10, sse4)
652 24 mc_uni_w_func(pel_pixels, 10, 6, sse4)
653 1448 mc_uni_w_funcs(epel_h, 10, sse4)
654 24 mc_uni_w_func(epel_h, 10, 6, sse4)
655 456 mc_uni_w_funcs(epel_v, 10, sse4)
656 24 mc_uni_w_func(epel_v, 10, 6, sse4)
657 632 mc_uni_w_funcs(epel_hv, 10, sse4)
658 24 mc_uni_w_func(epel_hv, 10, 6, sse4)
659 940 mc_uni_w_funcs(qpel_h, 10, sse4)
660 428 mc_uni_w_funcs(qpel_v, 10, sse4)
661 476 mc_uni_w_funcs(qpel_hv, 10, sse4)
662
663 384 mc_uni_w_funcs(pel_pixels, 12, sse4)
664 24 mc_uni_w_func(pel_pixels, 12, 6, sse4)
665 384 mc_uni_w_funcs(epel_h, 12, sse4)
666 24 mc_uni_w_func(epel_h, 12, 6, sse4)
667 384 mc_uni_w_funcs(epel_v, 12, sse4)
668 24 mc_uni_w_func(epel_v, 12, 6, sse4)
669 384 mc_uni_w_funcs(epel_hv, 12, sse4)
670 24 mc_uni_w_func(epel_hv, 12, 6, sse4)
671 384 mc_uni_w_funcs(qpel_h, 12, sse4)
672 384 mc_uni_w_funcs(qpel_v, 12, sse4)
673 384 mc_uni_w_funcs(qpel_hv, 12, sse4)
674
675 #define mc_bi_w_func(name, bitd, W, opt) \
676 static void hevc_put_bi_w_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t _dststride, \
677 const uint8_t *_src, ptrdiff_t _srcstride, \
678 const int16_t *_src2, \
679 int height, int denom, \
680 int wx0, int wx1, int ox, \
681 intptr_t mx, intptr_t my, int width) \
682 { \
683 LOCAL_ALIGNED_16(int16_t, temp, [71 * MAX_PB_SIZE]); \
684 hevc_put_##name##W##_##bitd##_##opt(temp, _src, _srcstride, height, mx, my, width); \
685 ff_hevc_put_bi_w##W##_##bitd##_##opt(_dst, _dststride, temp, _src2, \
686 height, denom, wx0, wx1, ox); \
687 }
688
689 #define mc_bi_w_funcs(name, bitd, opt) \
690 mc_bi_w_func(name, bitd, 4, opt) \
691 mc_bi_w_func(name, bitd, 8, opt) \
692 mc_bi_w_func(name, bitd, 12, opt) \
693 mc_bi_w_func(name, bitd, 16, opt) \
694 mc_bi_w_func(name, bitd, 24, opt) \
695 mc_bi_w_func(name, bitd, 32, opt) \
696 mc_bi_w_func(name, bitd, 48, opt) \
697 mc_bi_w_func(name, bitd, 64, opt)
698
699 24022 mc_bi_w_funcs(pel_pixels, 8, sse4)
700 24 mc_bi_w_func(pel_pixels, 8, 6, sse4)
701 1336 mc_bi_w_funcs(epel_h, 8, sse4)
702 24 mc_bi_w_func(epel_h, 8, 6, sse4)
703 4956 mc_bi_w_funcs(epel_v, 8, sse4)
704 24 mc_bi_w_func(epel_v, 8, 6, sse4)
705 2512 mc_bi_w_funcs(epel_hv, 8, sse4)
706 24 mc_bi_w_func(epel_hv, 8, 6, sse4)
707 384 mc_bi_w_funcs(qpel_h, 8, sse4)
708 384 mc_bi_w_funcs(qpel_v, 8, sse4)
709 384 mc_bi_w_funcs(qpel_hv, 8, sse4)
710
711 4432 mc_bi_w_funcs(pel_pixels, 10, sse4)
712 24 mc_bi_w_func(pel_pixels, 10, 6, sse4)
713 1488 mc_bi_w_funcs(epel_h, 10, sse4)
714 24 mc_bi_w_func(epel_h, 10, 6, sse4)
715 544 mc_bi_w_funcs(epel_v, 10, sse4)
716 24 mc_bi_w_func(epel_v, 10, 6, sse4)
717 408 mc_bi_w_funcs(epel_hv, 10, sse4)
718 24 mc_bi_w_func(epel_hv, 10, 6, sse4)
719 932 mc_bi_w_funcs(qpel_h, 10, sse4)
720 464 mc_bi_w_funcs(qpel_v, 10, sse4)
721 396 mc_bi_w_funcs(qpel_hv, 10, sse4)
722
723 384 mc_bi_w_funcs(pel_pixels, 12, sse4)
724 24 mc_bi_w_func(pel_pixels, 12, 6, sse4)
725 384 mc_bi_w_funcs(epel_h, 12, sse4)
726 24 mc_bi_w_func(epel_h, 12, 6, sse4)
727 384 mc_bi_w_funcs(epel_v, 12, sse4)
728 24 mc_bi_w_func(epel_v, 12, 6, sse4)
729 384 mc_bi_w_funcs(epel_hv, 12, sse4)
730 24 mc_bi_w_func(epel_hv, 12, 6, sse4)
731 384 mc_bi_w_funcs(qpel_h, 12, sse4)
732 384 mc_bi_w_funcs(qpel_v, 12, sse4)
733 384 mc_bi_w_funcs(qpel_hv, 12, sse4)
734 #endif //ARCH_X86_64 && HAVE_SSE4_EXTERNAL
735
736 #define SAO_BAND_FILTER_FUNCS(bitd, opt) \
737 void ff_hevc_sao_band_filter_8_##bitd##_##opt(uint8_t *_dst, const uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
738 const int16_t *sao_offset_val, int sao_left_class, int width, int height); \
739 void ff_hevc_sao_band_filter_16_##bitd##_##opt(uint8_t *_dst, const uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
740 const int16_t *sao_offset_val, int sao_left_class, int width, int height); \
741 void ff_hevc_sao_band_filter_32_##bitd##_##opt(uint8_t *_dst, const uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
742 const int16_t *sao_offset_val, int sao_left_class, int width, int height); \
743 void ff_hevc_sao_band_filter_48_##bitd##_##opt(uint8_t *_dst, const uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
744 const int16_t *sao_offset_val, int sao_left_class, int width, int height); \
745 void ff_hevc_sao_band_filter_64_##bitd##_##opt(uint8_t *_dst, const uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
746 const int16_t *sao_offset_val, int sao_left_class, int width, int height);
747
748 SAO_BAND_FILTER_FUNCS(8, sse2)
749 SAO_BAND_FILTER_FUNCS(10, sse2)
750 SAO_BAND_FILTER_FUNCS(12, sse2)
751 SAO_BAND_FILTER_FUNCS(8, avx)
752 SAO_BAND_FILTER_FUNCS(10, avx)
753 SAO_BAND_FILTER_FUNCS(12, avx)
754 SAO_BAND_FILTER_FUNCS(8, avx2)
755 SAO_BAND_FILTER_FUNCS(10, avx2)
756 SAO_BAND_FILTER_FUNCS(12, avx2)
757
758 #define SAO_BAND_INIT(bitd, opt) do { \
759 c->sao_band_filter[0] = ff_hevc_sao_band_filter_8_##bitd##_##opt; \
760 c->sao_band_filter[1] = ff_hevc_sao_band_filter_16_##bitd##_##opt; \
761 c->sao_band_filter[2] = ff_hevc_sao_band_filter_32_##bitd##_##opt; \
762 c->sao_band_filter[3] = ff_hevc_sao_band_filter_48_##bitd##_##opt; \
763 c->sao_band_filter[4] = ff_hevc_sao_band_filter_64_##bitd##_##opt; \
764 } while (0)
765
766 #define SAO_EDGE_FILTER_FUNCS(bitd, opt) \
767 void ff_hevc_sao_edge_filter_8_##bitd##_##opt(uint8_t *_dst, const uint8_t *_src, ptrdiff_t stride_dst, \
768 const int16_t *sao_offset_val, int eo, int width, int height); \
769 void ff_hevc_sao_edge_filter_16_##bitd##_##opt(uint8_t *_dst, const uint8_t *_src, ptrdiff_t stride_dst, \
770 const int16_t *sao_offset_val, int eo, int width, int height); \
771 void ff_hevc_sao_edge_filter_32_##bitd##_##opt(uint8_t *_dst, const uint8_t *_src, ptrdiff_t stride_dst, \
772 const int16_t *sao_offset_val, int eo, int width, int height); \
773 void ff_hevc_sao_edge_filter_48_##bitd##_##opt(uint8_t *_dst, const uint8_t *_src, ptrdiff_t stride_dst, \
774 const int16_t *sao_offset_val, int eo, int width, int height); \
775 void ff_hevc_sao_edge_filter_64_##bitd##_##opt(uint8_t *_dst, const uint8_t *_src, ptrdiff_t stride_dst, \
776 const int16_t *sao_offset_val, int eo, int width, int height); \
777
778 SAO_EDGE_FILTER_FUNCS(8, ssse3)
779 SAO_EDGE_FILTER_FUNCS(8, avx2)
780 SAO_EDGE_FILTER_FUNCS(10, sse2)
781 SAO_EDGE_FILTER_FUNCS(10, avx2)
782 SAO_EDGE_FILTER_FUNCS(12, sse2)
783 SAO_EDGE_FILTER_FUNCS(12, avx2)
784
785 #define SAO_EDGE_INIT(bitd, opt) do { \
786 c->sao_edge_filter[0] = ff_hevc_sao_edge_filter_8_##bitd##_##opt; \
787 c->sao_edge_filter[1] = ff_hevc_sao_edge_filter_16_##bitd##_##opt; \
788 c->sao_edge_filter[2] = ff_hevc_sao_edge_filter_32_##bitd##_##opt; \
789 c->sao_edge_filter[3] = ff_hevc_sao_edge_filter_48_##bitd##_##opt; \
790 c->sao_edge_filter[4] = ff_hevc_sao_edge_filter_64_##bitd##_##opt; \
791 } while (0)
792
793 #define PEL_LINK(dst, idx1, idx2, idx3, name, D, opt) \
794 dst [idx1][idx2][idx3] = hevc_put_ ## name ## _ ## D ## _##opt; \
795 dst ## _bi [idx1][idx2][idx3] = ff_hevc_put_bi_ ## name ## _ ## D ## _##opt; \
796 dst ## _uni [idx1][idx2][idx3] = hevc_put_uni_ ## name ## _ ## D ## _##opt; \
797 dst ## _uni_w[idx1][idx2][idx3] = hevc_put_uni_w_ ## name ## _ ## D ## _##opt; \
798 dst ## _bi_w [idx1][idx2][idx3] = hevc_put_bi_w_ ## name ## _ ## D ## _##opt
799
800 #define EPEL_LINKS(pointer, my, mx, fname, bitd, opt ) \
801 PEL_LINK(pointer, 1, my , mx , fname##4 , bitd, opt ); \
802 PEL_LINK(pointer, 2, my , mx , fname##6 , bitd, opt ); \
803 PEL_LINK(pointer, 3, my , mx , fname##8 , bitd, opt ); \
804 PEL_LINK(pointer, 4, my , mx , fname##12, bitd, opt ); \
805 PEL_LINK(pointer, 5, my , mx , fname##16, bitd, opt ); \
806 PEL_LINK(pointer, 6, my , mx , fname##24, bitd, opt ); \
807 PEL_LINK(pointer, 7, my , mx , fname##32, bitd, opt ); \
808 PEL_LINK(pointer, 8, my , mx , fname##48, bitd, opt ); \
809 PEL_LINK(pointer, 9, my , mx , fname##64, bitd, opt )
810 #define QPEL_LINKS(pointer, my, mx, fname, bitd, opt) \
811 PEL_LINK(pointer, 1, my , mx , fname##4 , bitd, opt ); \
812 PEL_LINK(pointer, 3, my , mx , fname##8 , bitd, opt ); \
813 PEL_LINK(pointer, 4, my , mx , fname##12, bitd, opt ); \
814 PEL_LINK(pointer, 5, my , mx , fname##16, bitd, opt ); \
815 PEL_LINK(pointer, 6, my , mx , fname##24, bitd, opt ); \
816 PEL_LINK(pointer, 7, my , mx , fname##32, bitd, opt ); \
817 PEL_LINK(pointer, 8, my , mx , fname##48, bitd, opt ); \
818 PEL_LINK(pointer, 9, my , mx , fname##64, bitd, opt )
819
820 1651 void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth)
821 {
822 1651 int cpu_flags = av_get_cpu_flags();
823
824
2/2
✓ Branch 0 taken 685 times.
✓ Branch 1 taken 966 times.
 1651 if (bit_depth == 8) {
825
2/2
✓ Branch 0 taken 219 times.
✓ Branch 1 taken 466 times.
 685 if (EXTERNAL_SSE2(cpu_flags)) {
826 219 c->add_residual[0] = ff_hevc_add_residual_4_8_sse2;
827
828 219 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_sse2;
829 219 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_sse2;
830 #if ARCH_X86_64
831 219 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_sse2;
832 219 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_sse2;
833
834 219 c->idct[2] = ff_hevc_idct_16x16_8_sse2;
835 219 c->idct[3] = ff_hevc_idct_32x32_8_sse2;
836 #endif
837 219 SAO_BAND_INIT(8, sse2);
838
839 219 c->idct_dc[0] = ff_hevc_idct_4x4_dc_8_sse2;
840 219 c->idct_dc[1] = ff_hevc_idct_8x8_dc_8_sse2;
841 219 c->idct_dc[2] = ff_hevc_idct_16x16_dc_8_sse2;
842 219 c->idct_dc[3] = ff_hevc_idct_32x32_dc_8_sse2;
843
844 219 c->idct[0] = ff_hevc_idct_4x4_8_sse2;
845 219 c->idct[1] = ff_hevc_idct_8x8_8_sse2;
846 219 c->transform_4x4_luma = ff_hevc_transform_4x4_luma_8_sse2;
847
848 219 c->add_residual[1] = ff_hevc_add_residual_8_8_sse2;
849 219 c->add_residual[2] = ff_hevc_add_residual_16_8_sse2;
850 219 c->add_residual[3] = ff_hevc_add_residual_32_8_sse2;
851 }
852
2/2
✓ Branch 0 taken 177 times.
✓ Branch 1 taken 508 times.
 685 if (EXTERNAL_SSSE3(cpu_flags)) {
853 #if ARCH_X86_64
854 177 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_ssse3;
855 177 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_ssse3;
856 #endif
857 177 c->dequant = ff_hevc_dequant_8_ssse3;
858 177 SAO_EDGE_INIT(8, ssse3);
859 }
860 #if HAVE_SSE4_EXTERNAL && ARCH_X86_64
861
2/2
✓ Branch 0 taken 156 times.
✓ Branch 1 taken 529 times.
 685 if (EXTERNAL_SSE4(cpu_flags)) {
862
863 156 EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 8, sse4);
864 156 EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 8, sse4);
865 156 EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 8, sse4);
866 156 EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 8, sse4);
867
868 156 QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 8, sse4);
869 156 QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 8, sse4);
870 156 QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 8, sse4);
871 156 QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 8, sse4);
872 }
873 #endif
874
2/2
✓ Branch 0 taken 72 times.
✓ Branch 1 taken 613 times.
 685 if (EXTERNAL_AVX(cpu_flags)) {
875 72 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_avx;
876 72 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_avx;
877 #if ARCH_X86_64
878 72 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_avx;
879 72 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_avx;
880
881 72 c->idct[2] = ff_hevc_idct_16x16_8_avx;
882 72 c->idct[3] = ff_hevc_idct_32x32_8_avx;
883 #endif
884 72 SAO_BAND_INIT(8, avx);
885
886 72 c->idct[0] = ff_hevc_idct_4x4_8_avx;
887 72 c->idct[1] = ff_hevc_idct_8x8_8_avx;
888 72 c->transform_4x4_luma = ff_hevc_transform_4x4_luma_8_avx;
889 }
890
2/2
✓ Branch 0 taken 30 times.
✓ Branch 1 taken 655 times.
 685 if (EXTERNAL_AVX2(cpu_flags)) {
891 30 c->sao_band_filter[0] = ff_hevc_sao_band_filter_8_8_avx2;
892 30 c->sao_band_filter[1] = ff_hevc_sao_band_filter_16_8_avx2;
893 }
894 #if HAVE_AVX2_EXTERNAL
895
3/4
✓ Branch 0 taken 30 times.
✓ Branch 1 taken 655 times.
✓ Branch 2 taken 30 times.
✗ Branch 3 not taken.
 685 if (EXTERNAL_AVX2_FAST(cpu_flags)) {
896 30 c->idct_dc[2] = ff_hevc_idct_16x16_dc_8_avx2;
897 30 c->idct_dc[3] = ff_hevc_idct_32x32_dc_8_avx2;
898
899 #if ARCH_X86_64
900 30 c->put_hevc_epel[7][0][0] = hevc_put_pel_pixels32_8_avx2;
901 30 c->put_hevc_epel[8][0][0] = hevc_put_pel_pixels48_8_avx2;
902 30 c->put_hevc_epel[9][0][0] = hevc_put_pel_pixels64_8_avx2;
903
904 30 c->put_hevc_qpel[7][0][0] = hevc_put_pel_pixels32_8_avx2;
905 30 c->put_hevc_qpel[8][0][0] = hevc_put_pel_pixels48_8_avx2;
906 30 c->put_hevc_qpel[9][0][0] = hevc_put_pel_pixels64_8_avx2;
907
908 30 c->put_hevc_epel_uni[7][0][0] = hevc_put_uni_pel_pixels32_8_avx2;
909 30 c->put_hevc_epel_uni[8][0][0] = hevc_put_uni_pel_pixels48_8_avx2;
910 30 c->put_hevc_epel_uni[9][0][0] = hevc_put_uni_pel_pixels64_8_avx2;
911
912 30 c->put_hevc_qpel_uni[7][0][0] = hevc_put_uni_pel_pixels32_8_avx2;
913 30 c->put_hevc_qpel_uni[8][0][0] = hevc_put_uni_pel_pixels48_8_avx2;
914 30 c->put_hevc_qpel_uni[9][0][0] = hevc_put_uni_pel_pixels64_8_avx2;
915
916 30 c->put_hevc_qpel_bi[7][0][0] = ff_hevc_put_bi_pel_pixels32_8_avx2;
917 30 c->put_hevc_qpel_bi[8][0][0] = ff_hevc_put_bi_pel_pixels48_8_avx2;
918 30 c->put_hevc_qpel_bi[9][0][0] = ff_hevc_put_bi_pel_pixels64_8_avx2;
919
920 30 c->put_hevc_epel_bi[7][0][0] = ff_hevc_put_bi_pel_pixels32_8_avx2;
921 30 c->put_hevc_epel_bi[8][0][0] = ff_hevc_put_bi_pel_pixels48_8_avx2;
922 30 c->put_hevc_epel_bi[9][0][0] = ff_hevc_put_bi_pel_pixels64_8_avx2;
923
924 30 c->put_hevc_epel[7][0][1] = hevc_put_epel_h32_8_avx2;
925 30 c->put_hevc_epel[8][0][1] = hevc_put_epel_h48_8_avx2;
926 30 c->put_hevc_epel[9][0][1] = hevc_put_epel_h64_8_avx2;
927
928 30 c->put_hevc_epel_uni[7][0][1] = hevc_put_uni_epel_h32_8_avx2;
929 30 c->put_hevc_epel_uni[8][0][1] = hevc_put_uni_epel_h48_8_avx2;
930 30 c->put_hevc_epel_uni[9][0][1] = hevc_put_uni_epel_h64_8_avx2;
931
932 30 c->put_hevc_epel_bi[7][0][1] = ff_hevc_put_bi_epel_h32_8_avx2;
933 30 c->put_hevc_epel_bi[8][0][1] = ff_hevc_put_bi_epel_h48_8_avx2;
934 30 c->put_hevc_epel_bi[9][0][1] = ff_hevc_put_bi_epel_h64_8_avx2;
935
936 30 c->put_hevc_epel[7][1][0] = hevc_put_epel_v32_8_avx2;
937 30 c->put_hevc_epel[8][1][0] = hevc_put_epel_v48_8_avx2;
938 30 c->put_hevc_epel[9][1][0] = hevc_put_epel_v64_8_avx2;
939
940 30 c->put_hevc_epel_uni[7][1][0] = hevc_put_uni_epel_v32_8_avx2;
941 30 c->put_hevc_epel_uni[8][1][0] = hevc_put_uni_epel_v48_8_avx2;
942 30 c->put_hevc_epel_uni[9][1][0] = hevc_put_uni_epel_v64_8_avx2;
943
944 30 c->put_hevc_epel_bi[7][1][0] = ff_hevc_put_bi_epel_v32_8_avx2;
945 30 c->put_hevc_epel_bi[8][1][0] = ff_hevc_put_bi_epel_v48_8_avx2;
946 30 c->put_hevc_epel_bi[9][1][0] = ff_hevc_put_bi_epel_v64_8_avx2;
947
948 30 c->put_hevc_epel[7][1][1] = hevc_put_epel_hv32_8_avx2;
949 30 c->put_hevc_epel[8][1][1] = hevc_put_epel_hv48_8_avx2;
950 30 c->put_hevc_epel[9][1][1] = hevc_put_epel_hv64_8_avx2;
951
952 30 c->put_hevc_epel_uni[7][1][1] = hevc_put_uni_epel_hv32_8_avx2;
953 30 c->put_hevc_epel_uni[8][1][1] = hevc_put_uni_epel_hv48_8_avx2;
954 30 c->put_hevc_epel_uni[9][1][1] = hevc_put_uni_epel_hv64_8_avx2;
955
956 30 c->put_hevc_epel_bi[7][1][1] = ff_hevc_put_bi_epel_hv32_8_avx2;
957 30 c->put_hevc_epel_bi[8][1][1] = ff_hevc_put_bi_epel_hv48_8_avx2;
958 30 c->put_hevc_epel_bi[9][1][1] = ff_hevc_put_bi_epel_hv64_8_avx2;
959
960 30 c->put_hevc_qpel[7][0][1] = hevc_put_qpel_h32_8_avx2;
961 30 c->put_hevc_qpel[8][0][1] = hevc_put_qpel_h48_8_avx2;
962 30 c->put_hevc_qpel[9][0][1] = hevc_put_qpel_h64_8_avx2;
963
964 30 c->put_hevc_qpel[7][1][0] = hevc_put_qpel_v32_8_avx2;
965 30 c->put_hevc_qpel[8][1][0] = hevc_put_qpel_v48_8_avx2;
966 30 c->put_hevc_qpel[9][1][0] = hevc_put_qpel_v64_8_avx2;
967
968 30 c->put_hevc_qpel_uni[7][0][1] = hevc_put_uni_qpel_h32_8_avx2;
969 30 c->put_hevc_qpel_uni[8][0][1] = hevc_put_uni_qpel_h48_8_avx2;
970 30 c->put_hevc_qpel_uni[9][0][1] = hevc_put_uni_qpel_h64_8_avx2;
971
972 30 c->put_hevc_qpel_uni[7][1][0] = hevc_put_uni_qpel_v32_8_avx2;
973 30 c->put_hevc_qpel_uni[8][1][0] = hevc_put_uni_qpel_v48_8_avx2;
974 30 c->put_hevc_qpel_uni[9][1][0] = hevc_put_uni_qpel_v64_8_avx2;
975
976 30 c->put_hevc_qpel_bi[7][0][1] = ff_hevc_put_bi_qpel_h32_8_avx2;
977 30 c->put_hevc_qpel_bi[8][0][1] = ff_hevc_put_bi_qpel_h48_8_avx2;
978 30 c->put_hevc_qpel_bi[9][0][1] = ff_hevc_put_bi_qpel_h64_8_avx2;
979
980 30 c->put_hevc_qpel_bi[7][1][0] = ff_hevc_put_bi_qpel_v32_8_avx2;
981 30 c->put_hevc_qpel_bi[8][1][0] = ff_hevc_put_bi_qpel_v48_8_avx2;
982 30 c->put_hevc_qpel_bi[9][1][0] = ff_hevc_put_bi_qpel_v64_8_avx2;
983 #endif /* ARCH_X86_64 */
984
985 30 SAO_BAND_INIT(8, avx2);
986
987 30 c->sao_edge_filter[2] = ff_hevc_sao_edge_filter_32_8_avx2;
988 30 c->sao_edge_filter[3] = ff_hevc_sao_edge_filter_48_8_avx2;
989 30 c->sao_edge_filter[4] = ff_hevc_sao_edge_filter_64_8_avx2;
990
991 30 c->add_residual[3] = ff_hevc_add_residual_32_8_avx2;
992 }
993 #endif /* HAVE_AVX2_EXTERNAL */
994 #if ARCH_X86_64
995
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 685 times.
 685 if (EXTERNAL_AVX512ICL(cpu_flags)) {
996 ✗ c->put_hevc_qpel[1][0][1] = ff_hevc_put_qpel_h4_8_avx512icl;
997 ✗ c->put_hevc_qpel[3][0][1] = ff_hevc_put_qpel_h8_8_avx512icl;
998 ✗ c->put_hevc_qpel[5][0][1] = ff_hevc_put_qpel_h16_8_avx512icl;
999 ✗ c->put_hevc_qpel[7][0][1] = ff_hevc_put_qpel_h32_8_avx512icl;
1000 ✗ c->put_hevc_qpel[9][0][1] = ff_hevc_put_qpel_h64_8_avx512icl;
1001 ✗ c->put_hevc_qpel[3][1][1] = ff_hevc_put_qpel_hv8_8_avx512icl;
1002 }
1003 #endif
1004
2/2
✓ Branch 0 taken 354 times.
✓ Branch 1 taken 612 times.
 966 } else if (bit_depth == 10) {
1005
2/2
✓ Branch 0 taken 222 times.
✓ Branch 1 taken 132 times.
 354 if (EXTERNAL_SSE2(cpu_flags)) {
1006 222 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_sse2;
1007 222 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_10_sse2;
1008 #if ARCH_X86_64
1009 222 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_sse2;
1010 222 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_sse2;
1011
1012 222 c->idct[2] = ff_hevc_idct_16x16_10_sse2;
1013 222 c->idct[3] = ff_hevc_idct_32x32_10_sse2;
1014 #endif
1015 222 SAO_BAND_INIT(10, sse2);
1016 222 SAO_EDGE_INIT(10, sse2);
1017
1018 222 c->idct_dc[0] = ff_hevc_idct_4x4_dc_10_sse2;
1019 222 c->idct_dc[1] = ff_hevc_idct_8x8_dc_10_sse2;
1020 222 c->idct_dc[2] = ff_hevc_idct_16x16_dc_10_sse2;
1021 222 c->idct_dc[3] = ff_hevc_idct_32x32_dc_10_sse2;
1022
1023 222 c->idct[0] = ff_hevc_idct_4x4_10_sse2;
1024 222 c->idct[1] = ff_hevc_idct_8x8_10_sse2;
1025 222 c->transform_4x4_luma = ff_hevc_transform_4x4_luma_10_sse2;
1026
1027 222 c->add_residual[0] = ff_hevc_add_residual_4_10_sse2;
1028 222 c->add_residual[1] = ff_hevc_add_residual_8_10_sse2;
1029 222 c->add_residual[2] = ff_hevc_add_residual_16_10_sse2;
1030 222 c->add_residual[3] = ff_hevc_add_residual_32_10_sse2;
1031 }
1032 #if ARCH_X86_64
1033
2/2
✓ Branch 0 taken 180 times.
✓ Branch 1 taken 174 times.
 354 if (EXTERNAL_SSSE3(cpu_flags)) {
1034 180 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_ssse3;
1035 180 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_ssse3;
1036 }
1037 #endif
1038 #if HAVE_SSE4_EXTERNAL && ARCH_X86_64
1039
2/2
✓ Branch 0 taken 159 times.
✓ Branch 1 taken 195 times.
 354 if (EXTERNAL_SSE4(cpu_flags)) {
1040 159 EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 10, sse4);
1041 159 EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 10, sse4);
1042 159 EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 10, sse4);
1043 159 EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 10, sse4);
1044
1045 159 QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 10, sse4);
1046 159 QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 10, sse4);
1047 159 QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 10, sse4);
1048 159 QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 10, sse4);
1049 }
1050 #endif
1051
2/2
✓ Branch 0 taken 75 times.
✓ Branch 1 taken 279 times.
 354 if (EXTERNAL_AVX(cpu_flags)) {
1052 75 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_avx;
1053 75 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_10_avx;
1054 #if ARCH_X86_64
1055 75 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_avx;
1056 75 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_avx;
1057
1058 75 c->idct[2] = ff_hevc_idct_16x16_10_avx;
1059 75 c->idct[3] = ff_hevc_idct_32x32_10_avx;
1060 #endif
1061
1062 75 c->idct[0] = ff_hevc_idct_4x4_10_avx;
1063 75 c->idct[1] = ff_hevc_idct_8x8_10_avx;
1064 75 c->transform_4x4_luma = ff_hevc_transform_4x4_luma_10_avx;
1065
1066 75 SAO_BAND_INIT(10, avx);
1067 }
1068
2/2
✓ Branch 0 taken 33 times.
✓ Branch 1 taken 321 times.
 354 if (EXTERNAL_AVX2(cpu_flags)) {
1069 33 c->sao_band_filter[0] = ff_hevc_sao_band_filter_8_10_avx2;
1070 }
1071 #if HAVE_AVX2_EXTERNAL
1072
3/4
✓ Branch 0 taken 33 times.
✓ Branch 1 taken 321 times.
✓ Branch 2 taken 33 times.
✗ Branch 3 not taken.
 354 if (EXTERNAL_AVX2_FAST(cpu_flags)) {
1073 33 c->idct_dc[2] = ff_hevc_idct_16x16_dc_10_avx2;
1074 33 c->idct_dc[3] = ff_hevc_idct_32x32_dc_10_avx2;
1075
1076 #if ARCH_X86_64
1077 33 c->put_hevc_epel[5][0][0] = hevc_put_pel_pixels16_10_avx2;
1078 33 c->put_hevc_epel[6][0][0] = hevc_put_pel_pixels24_10_avx2;
1079 33 c->put_hevc_epel[7][0][0] = hevc_put_pel_pixels32_10_avx2;
1080 33 c->put_hevc_epel[8][0][0] = hevc_put_pel_pixels48_10_avx2;
1081 33 c->put_hevc_epel[9][0][0] = hevc_put_pel_pixels64_10_avx2;
1082
1083 33 c->put_hevc_qpel[5][0][0] = hevc_put_pel_pixels16_10_avx2;
1084 33 c->put_hevc_qpel[6][0][0] = hevc_put_pel_pixels24_10_avx2;
1085 33 c->put_hevc_qpel[7][0][0] = hevc_put_pel_pixels32_10_avx2;
1086 33 c->put_hevc_qpel[8][0][0] = hevc_put_pel_pixels48_10_avx2;
1087 33 c->put_hevc_qpel[9][0][0] = hevc_put_pel_pixels64_10_avx2;
1088
1089 33 c->put_hevc_epel_uni[5][0][0] = hevc_put_uni_pel_pixels32_8_avx2;
1090 33 c->put_hevc_epel_uni[6][0][0] = hevc_put_uni_pel_pixels48_8_avx2;
1091 33 c->put_hevc_epel_uni[7][0][0] = hevc_put_uni_pel_pixels64_8_avx2;
1092 33 c->put_hevc_epel_uni[8][0][0] = hevc_put_uni_pel_pixels96_8_avx2;
1093 33 c->put_hevc_epel_uni[9][0][0] = hevc_put_uni_pel_pixels128_8_avx2;
1094
1095 33 c->put_hevc_qpel_uni[5][0][0] = hevc_put_uni_pel_pixels32_8_avx2;
1096 33 c->put_hevc_qpel_uni[6][0][0] = hevc_put_uni_pel_pixels48_8_avx2;
1097 33 c->put_hevc_qpel_uni[7][0][0] = hevc_put_uni_pel_pixels64_8_avx2;
1098 33 c->put_hevc_qpel_uni[8][0][0] = hevc_put_uni_pel_pixels96_8_avx2;
1099 33 c->put_hevc_qpel_uni[9][0][0] = hevc_put_uni_pel_pixels128_8_avx2;
1100
1101 33 c->put_hevc_epel_bi[5][0][0] = ff_hevc_put_bi_pel_pixels16_10_avx2;
1102 33 c->put_hevc_epel_bi[6][0][0] = ff_hevc_put_bi_pel_pixels24_10_avx2;
1103 33 c->put_hevc_epel_bi[7][0][0] = ff_hevc_put_bi_pel_pixels32_10_avx2;
1104 33 c->put_hevc_epel_bi[8][0][0] = ff_hevc_put_bi_pel_pixels48_10_avx2;
1105 33 c->put_hevc_epel_bi[9][0][0] = ff_hevc_put_bi_pel_pixels64_10_avx2;
1106 33 c->put_hevc_qpel_bi[5][0][0] = ff_hevc_put_bi_pel_pixels16_10_avx2;
1107 33 c->put_hevc_qpel_bi[6][0][0] = ff_hevc_put_bi_pel_pixels24_10_avx2;
1108 33 c->put_hevc_qpel_bi[7][0][0] = ff_hevc_put_bi_pel_pixels32_10_avx2;
1109 33 c->put_hevc_qpel_bi[8][0][0] = ff_hevc_put_bi_pel_pixels48_10_avx2;
1110 33 c->put_hevc_qpel_bi[9][0][0] = ff_hevc_put_bi_pel_pixels64_10_avx2;
1111
1112 33 c->put_hevc_epel[5][0][1] = hevc_put_epel_h16_10_avx2;
1113 33 c->put_hevc_epel[6][0][1] = hevc_put_epel_h24_10_avx2;
1114 33 c->put_hevc_epel[7][0][1] = hevc_put_epel_h32_10_avx2;
1115 33 c->put_hevc_epel[8][0][1] = hevc_put_epel_h48_10_avx2;
1116 33 c->put_hevc_epel[9][0][1] = hevc_put_epel_h64_10_avx2;
1117
1118 33 c->put_hevc_epel_uni[5][0][1] = hevc_put_uni_epel_h16_10_avx2;
1119 33 c->put_hevc_epel_uni[6][0][1] = hevc_put_uni_epel_h24_10_avx2;
1120 33 c->put_hevc_epel_uni[7][0][1] = hevc_put_uni_epel_h32_10_avx2;
1121 33 c->put_hevc_epel_uni[8][0][1] = hevc_put_uni_epel_h48_10_avx2;
1122 33 c->put_hevc_epel_uni[9][0][1] = hevc_put_uni_epel_h64_10_avx2;
1123
1124 33 c->put_hevc_epel_bi[5][0][1] = ff_hevc_put_bi_epel_h16_10_avx2;
1125 33 c->put_hevc_epel_bi[6][0][1] = ff_hevc_put_bi_epel_h24_10_avx2;
1126 33 c->put_hevc_epel_bi[7][0][1] = ff_hevc_put_bi_epel_h32_10_avx2;
1127 33 c->put_hevc_epel_bi[8][0][1] = ff_hevc_put_bi_epel_h48_10_avx2;
1128 33 c->put_hevc_epel_bi[9][0][1] = ff_hevc_put_bi_epel_h64_10_avx2;
1129
1130 33 c->put_hevc_epel[5][1][0] = hevc_put_epel_v16_10_avx2;
1131 33 c->put_hevc_epel[6][1][0] = hevc_put_epel_v24_10_avx2;
1132 33 c->put_hevc_epel[7][1][0] = hevc_put_epel_v32_10_avx2;
1133 33 c->put_hevc_epel[8][1][0] = hevc_put_epel_v48_10_avx2;
1134 33 c->put_hevc_epel[9][1][0] = hevc_put_epel_v64_10_avx2;
1135
1136 33 c->put_hevc_epel_uni[5][1][0] = hevc_put_uni_epel_v16_10_avx2;
1137 33 c->put_hevc_epel_uni[6][1][0] = hevc_put_uni_epel_v24_10_avx2;
1138 33 c->put_hevc_epel_uni[7][1][0] = hevc_put_uni_epel_v32_10_avx2;
1139 33 c->put_hevc_epel_uni[8][1][0] = hevc_put_uni_epel_v48_10_avx2;
1140 33 c->put_hevc_epel_uni[9][1][0] = hevc_put_uni_epel_v64_10_avx2;
1141
1142 33 c->put_hevc_epel_bi[5][1][0] = ff_hevc_put_bi_epel_v16_10_avx2;
1143 33 c->put_hevc_epel_bi[6][1][0] = ff_hevc_put_bi_epel_v24_10_avx2;
1144 33 c->put_hevc_epel_bi[7][1][0] = ff_hevc_put_bi_epel_v32_10_avx2;
1145 33 c->put_hevc_epel_bi[8][1][0] = ff_hevc_put_bi_epel_v48_10_avx2;
1146 33 c->put_hevc_epel_bi[9][1][0] = ff_hevc_put_bi_epel_v64_10_avx2;
1147
1148 33 c->put_hevc_epel[5][1][1] = hevc_put_epel_hv16_10_avx2;
1149 33 c->put_hevc_epel[6][1][1] = hevc_put_epel_hv24_10_avx2;
1150 33 c->put_hevc_epel[7][1][1] = hevc_put_epel_hv32_10_avx2;
1151 33 c->put_hevc_epel[8][1][1] = hevc_put_epel_hv48_10_avx2;
1152 33 c->put_hevc_epel[9][1][1] = hevc_put_epel_hv64_10_avx2;
1153
1154 33 c->put_hevc_epel_uni[5][1][1] = hevc_put_uni_epel_hv16_10_avx2;
1155 33 c->put_hevc_epel_uni[6][1][1] = hevc_put_uni_epel_hv24_10_avx2;
1156 33 c->put_hevc_epel_uni[7][1][1] = hevc_put_uni_epel_hv32_10_avx2;
1157 33 c->put_hevc_epel_uni[8][1][1] = hevc_put_uni_epel_hv48_10_avx2;
1158 33 c->put_hevc_epel_uni[9][1][1] = hevc_put_uni_epel_hv64_10_avx2;
1159
1160 33 c->put_hevc_epel_bi[5][1][1] = ff_hevc_put_bi_epel_hv16_10_avx2;
1161 33 c->put_hevc_epel_bi[6][1][1] = ff_hevc_put_bi_epel_hv24_10_avx2;
1162 33 c->put_hevc_epel_bi[7][1][1] = ff_hevc_put_bi_epel_hv32_10_avx2;
1163 33 c->put_hevc_epel_bi[8][1][1] = ff_hevc_put_bi_epel_hv48_10_avx2;
1164 33 c->put_hevc_epel_bi[9][1][1] = ff_hevc_put_bi_epel_hv64_10_avx2;
1165
1166 33 c->put_hevc_qpel[5][0][1] = hevc_put_qpel_h16_10_avx2;
1167 33 c->put_hevc_qpel[6][0][1] = hevc_put_qpel_h24_10_avx2;
1168 33 c->put_hevc_qpel[7][0][1] = hevc_put_qpel_h32_10_avx2;
1169 33 c->put_hevc_qpel[8][0][1] = hevc_put_qpel_h48_10_avx2;
1170 33 c->put_hevc_qpel[9][0][1] = hevc_put_qpel_h64_10_avx2;
1171
1172 33 c->put_hevc_qpel_uni[5][0][1] = hevc_put_uni_qpel_h16_10_avx2;
1173 33 c->put_hevc_qpel_uni[6][0][1] = hevc_put_uni_qpel_h24_10_avx2;
1174 33 c->put_hevc_qpel_uni[7][0][1] = hevc_put_uni_qpel_h32_10_avx2;
1175 33 c->put_hevc_qpel_uni[8][0][1] = hevc_put_uni_qpel_h48_10_avx2;
1176 33 c->put_hevc_qpel_uni[9][0][1] = hevc_put_uni_qpel_h64_10_avx2;
1177
1178 33 c->put_hevc_qpel_bi[5][0][1] = ff_hevc_put_bi_qpel_h16_10_avx2;
1179 33 c->put_hevc_qpel_bi[6][0][1] = ff_hevc_put_bi_qpel_h24_10_avx2;
1180 33 c->put_hevc_qpel_bi[7][0][1] = ff_hevc_put_bi_qpel_h32_10_avx2;
1181 33 c->put_hevc_qpel_bi[8][0][1] = ff_hevc_put_bi_qpel_h48_10_avx2;
1182 33 c->put_hevc_qpel_bi[9][0][1] = ff_hevc_put_bi_qpel_h64_10_avx2;
1183
1184 33 c->put_hevc_qpel[5][1][0] = hevc_put_qpel_v16_10_avx2;
1185 33 c->put_hevc_qpel[6][1][0] = hevc_put_qpel_v24_10_avx2;
1186 33 c->put_hevc_qpel[7][1][0] = hevc_put_qpel_v32_10_avx2;
1187 33 c->put_hevc_qpel[8][1][0] = hevc_put_qpel_v48_10_avx2;
1188 33 c->put_hevc_qpel[9][1][0] = hevc_put_qpel_v64_10_avx2;
1189
1190 33 c->put_hevc_qpel_uni[5][1][0] = hevc_put_uni_qpel_v16_10_avx2;
1191 33 c->put_hevc_qpel_uni[6][1][0] = hevc_put_uni_qpel_v24_10_avx2;
1192 33 c->put_hevc_qpel_uni[7][1][0] = hevc_put_uni_qpel_v32_10_avx2;
1193 33 c->put_hevc_qpel_uni[8][1][0] = hevc_put_uni_qpel_v48_10_avx2;
1194 33 c->put_hevc_qpel_uni[9][1][0] = hevc_put_uni_qpel_v64_10_avx2;
1195
1196 33 c->put_hevc_qpel_bi[5][1][0] = ff_hevc_put_bi_qpel_v16_10_avx2;
1197 33 c->put_hevc_qpel_bi[6][1][0] = ff_hevc_put_bi_qpel_v24_10_avx2;
1198 33 c->put_hevc_qpel_bi[7][1][0] = ff_hevc_put_bi_qpel_v32_10_avx2;
1199 33 c->put_hevc_qpel_bi[8][1][0] = ff_hevc_put_bi_qpel_v48_10_avx2;
1200 33 c->put_hevc_qpel_bi[9][1][0] = ff_hevc_put_bi_qpel_v64_10_avx2;
1201
1202 33 c->put_hevc_qpel[5][1][1] = hevc_put_qpel_hv16_10_avx2;
1203 33 c->put_hevc_qpel[6][1][1] = hevc_put_qpel_hv24_10_avx2;
1204 33 c->put_hevc_qpel[7][1][1] = hevc_put_qpel_hv32_10_avx2;
1205 33 c->put_hevc_qpel[8][1][1] = hevc_put_qpel_hv48_10_avx2;
1206 33 c->put_hevc_qpel[9][1][1] = hevc_put_qpel_hv64_10_avx2;
1207
1208 33 c->put_hevc_qpel_uni[5][1][1] = hevc_put_uni_qpel_hv16_10_avx2;
1209 33 c->put_hevc_qpel_uni[6][1][1] = hevc_put_uni_qpel_hv24_10_avx2;
1210 33 c->put_hevc_qpel_uni[7][1][1] = hevc_put_uni_qpel_hv32_10_avx2;
1211 33 c->put_hevc_qpel_uni[8][1][1] = hevc_put_uni_qpel_hv48_10_avx2;
1212 33 c->put_hevc_qpel_uni[9][1][1] = hevc_put_uni_qpel_hv64_10_avx2;
1213
1214 33 c->put_hevc_qpel_bi[5][1][1] = ff_hevc_put_bi_qpel_hv16_10_avx2;
1215 33 c->put_hevc_qpel_bi[6][1][1] = ff_hevc_put_bi_qpel_hv24_10_avx2;
1216 33 c->put_hevc_qpel_bi[7][1][1] = ff_hevc_put_bi_qpel_hv32_10_avx2;
1217 33 c->put_hevc_qpel_bi[8][1][1] = ff_hevc_put_bi_qpel_hv48_10_avx2;
1218 33 c->put_hevc_qpel_bi[9][1][1] = ff_hevc_put_bi_qpel_hv64_10_avx2;
1219 #endif /* ARCH_X86_64 */
1220
1221 33 SAO_BAND_INIT(10, avx2);
1222 33 SAO_EDGE_INIT(10, avx2);
1223
1224 33 c->add_residual[2] = ff_hevc_add_residual_16_10_avx2;
1225 33 c->add_residual[3] = ff_hevc_add_residual_32_10_avx2;
1226 }
1227 #endif /* HAVE_AVX2_EXTERNAL */
1228
2/2
✓ Branch 0 taken 304 times.
✓ Branch 1 taken 308 times.
 612 } else if (bit_depth == 12) {
1229
2/2
✓ Branch 0 taken 210 times.
✓ Branch 1 taken 94 times.
 304 if (EXTERNAL_SSE2(cpu_flags)) {
1230 210 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_12_sse2;
1231 210 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_12_sse2;
1232 #if ARCH_X86_64
1233 210 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_sse2;
1234 210 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_sse2;
1235 #endif
1236 210 SAO_BAND_INIT(12, sse2);
1237 210 SAO_EDGE_INIT(12, sse2);
1238
1239 210 c->idct_dc[0] = ff_hevc_idct_4x4_dc_12_sse2;
1240 210 c->idct_dc[1] = ff_hevc_idct_8x8_dc_12_sse2;
1241 210 c->idct_dc[2] = ff_hevc_idct_16x16_dc_12_sse2;
1242 210 c->idct_dc[3] = ff_hevc_idct_32x32_dc_12_sse2;
1243 }
1244 #if ARCH_X86_64
1245
2/2
✓ Branch 0 taken 168 times.
✓ Branch 1 taken 136 times.
 304 if (EXTERNAL_SSSE3(cpu_flags)) {
1246 168 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_ssse3;
1247 168 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_ssse3;
1248 }
1249 #endif
1250 #if HAVE_SSE4_EXTERNAL && ARCH_X86_64
1251
2/2
✓ Branch 0 taken 147 times.
✓ Branch 1 taken 157 times.
 304 if (EXTERNAL_SSE4(cpu_flags)) {
1252 147 EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 12, sse4);
1253 147 EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 12, sse4);
1254 147 EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 12, sse4);
1255 147 EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 12, sse4);
1256
1257 147 QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 12, sse4);
1258 147 QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 12, sse4);
1259 147 QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 12, sse4);
1260 147 QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 12, sse4);
1261 }
1262 #endif
1263
2/2
✓ Branch 0 taken 63 times.
✓ Branch 1 taken 241 times.
 304 if (EXTERNAL_AVX(cpu_flags)) {
1264 63 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_12_avx;
1265 63 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_12_avx;
1266 #if ARCH_X86_64
1267 63 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_avx;
1268 63 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_avx;
1269 #endif
1270 63 SAO_BAND_INIT(12, avx);
1271 }
1272
2/2
✓ Branch 0 taken 21 times.
✓ Branch 1 taken 283 times.
 304 if (EXTERNAL_AVX2(cpu_flags)) {
1273 21 c->sao_band_filter[0] = ff_hevc_sao_band_filter_8_12_avx2;
1274 }
1275
3/4
✓ Branch 0 taken 21 times.
✓ Branch 1 taken 283 times.
✓ Branch 2 taken 21 times.
✗ Branch 3 not taken.
 304 if (EXTERNAL_AVX2_FAST(cpu_flags)) {
1276 21 c->idct_dc[2] = ff_hevc_idct_16x16_dc_12_avx2;
1277 21 c->idct_dc[3] = ff_hevc_idct_32x32_dc_12_avx2;
1278
1279 21 SAO_BAND_INIT(12, avx2);
1280 21 SAO_EDGE_INIT(12, avx2);
1281 }
1282 }
1283 1651 }
1284