1 /*
2 * Copyright (c) 2013 Seppo Tomperi
3 * Copyright (c) 2013 - 2014 Pierre-Edouard Lepere
4 *
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22 #include "config.h"
23
24 #include "libavutil/cpu.h"
25 #include "libavutil/mem_internal.h"
26 #include "libavutil/x86/asm.h"
27 #include "libavutil/x86/cpu.h"
28 #include "libavcodec/get_bits.h" /* required for hevcdsp.h GetBitContext */
29 #include "libavcodec/hevcdsp.h"
30 #include "libavcodec/x86/hevcdsp.h"
31
32 #define LFC_FUNC(DIR, DEPTH, OPT) \
33 void ff_hevc_ ## DIR ## _loop_filter_chroma_ ## DEPTH ## _ ## OPT(uint8_t *pix, ptrdiff_t stride, int *tc, uint8_t *no_p, uint8_t *no_q);
34
35 #define LFL_FUNC(DIR, DEPTH, OPT) \
36 void ff_hevc_ ## DIR ## _loop_filter_luma_ ## DEPTH ## _ ## OPT(uint8_t *pix, ptrdiff_t stride, int beta, int *tc, uint8_t *no_p, uint8_t *no_q);
37
38 #define LFC_FUNCS(type, depth, opt) \
39 LFC_FUNC(h, depth, opt) \
40 LFC_FUNC(v, depth, opt)
41
42 #define LFL_FUNCS(type, depth, opt) \
43 LFL_FUNC(h, depth, opt) \
44 LFL_FUNC(v, depth, opt)
45
46 LFC_FUNCS(uint8_t, 8, sse2)
47 LFC_FUNCS(uint8_t, 10, sse2)
48 LFC_FUNCS(uint8_t, 12, sse2)
49 LFC_FUNCS(uint8_t, 8, avx)
50 LFC_FUNCS(uint8_t, 10, avx)
51 LFC_FUNCS(uint8_t, 12, avx)
52 LFL_FUNCS(uint8_t, 8, sse2)
53 LFL_FUNCS(uint8_t, 10, sse2)
54 LFL_FUNCS(uint8_t, 12, sse2)
55 LFL_FUNCS(uint8_t, 8, ssse3)
56 LFL_FUNCS(uint8_t, 10, ssse3)
57 LFL_FUNCS(uint8_t, 12, ssse3)
58 LFL_FUNCS(uint8_t, 8, avx)
59 LFL_FUNCS(uint8_t, 10, avx)
60 LFL_FUNCS(uint8_t, 12, avx)
61
62 #define IDCT_DC_FUNCS(W, opt) \
63 void ff_hevc_idct_ ## W ## _dc_8_ ## opt(int16_t *coeffs); \
64 void ff_hevc_idct_ ## W ## _dc_10_ ## opt(int16_t *coeffs); \
65 void ff_hevc_idct_ ## W ## _dc_12_ ## opt(int16_t *coeffs)
66
67 IDCT_DC_FUNCS(4x4, mmxext);
68 IDCT_DC_FUNCS(8x8, mmxext);
69 IDCT_DC_FUNCS(8x8, sse2);
70 IDCT_DC_FUNCS(16x16, sse2);
71 IDCT_DC_FUNCS(32x32, sse2);
72 IDCT_DC_FUNCS(16x16, avx2);
73 IDCT_DC_FUNCS(32x32, avx2);
74
75 #define IDCT_FUNCS(opt) \
76 void ff_hevc_idct_4x4_8_ ## opt(int16_t *coeffs, int col_limit); \
77 void ff_hevc_idct_4x4_10_ ## opt(int16_t *coeffs, int col_limit); \
78 void ff_hevc_idct_8x8_8_ ## opt(int16_t *coeffs, int col_limit); \
79 void ff_hevc_idct_8x8_10_ ## opt(int16_t *coeffs, int col_limit); \
80 void ff_hevc_idct_16x16_8_ ## opt(int16_t *coeffs, int col_limit); \
81 void ff_hevc_idct_16x16_10_ ## opt(int16_t *coeffs, int col_limit); \
82 void ff_hevc_idct_32x32_8_ ## opt(int16_t *coeffs, int col_limit); \
83 void ff_hevc_idct_32x32_10_ ## opt(int16_t *coeffs, int col_limit);
84
85 IDCT_FUNCS(sse2)
IDCT_FUNCS(avx)86 IDCT_FUNCS(avx)
87
88 #define mc_rep_func(name, bitd, step, W, opt) \
89 void ff_hevc_put_hevc_##name##W##_##bitd##_##opt(int16_t *_dst, \
90 uint8_t *_src, ptrdiff_t _srcstride, int height, \
91 intptr_t mx, intptr_t my, int width) \
92 { \
93 int i; \
94 uint8_t *src; \
95 int16_t *dst; \
96 for (i = 0; i < W; i += step) { \
97 src = _src + (i * ((bitd + 7) / 8)); \
98 dst = _dst + i; \
99 ff_hevc_put_hevc_##name##step##_##bitd##_##opt(dst, src, _srcstride, height, mx, my, width); \
100 } \
101 }
102 #define mc_rep_uni_func(name, bitd, step, W, opt) \
103 void ff_hevc_put_hevc_uni_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, \
104 uint8_t *_src, ptrdiff_t _srcstride, int height, \
105 intptr_t mx, intptr_t my, int width) \
106 { \
107 int i; \
108 uint8_t *src; \
109 uint8_t *dst; \
110 for (i = 0; i < W; i += step) { \
111 src = _src + (i * ((bitd + 7) / 8)); \
112 dst = _dst + (i * ((bitd + 7) / 8)); \
113 ff_hevc_put_hevc_uni_##name##step##_##bitd##_##opt(dst, dststride, src, _srcstride, \
114 height, mx, my, width); \
115 } \
116 }
117 #define mc_rep_bi_func(name, bitd, step, W, opt) \
118 void ff_hevc_put_hevc_bi_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, uint8_t *_src, \
119 ptrdiff_t _srcstride, int16_t* _src2, \
120 int height, intptr_t mx, intptr_t my, int width) \
121 { \
122 int i; \
123 uint8_t *src; \
124 uint8_t *dst; \
125 int16_t *src2; \
126 for (i = 0; i < W ; i += step) { \
127 src = _src + (i * ((bitd + 7) / 8)); \
128 dst = _dst + (i * ((bitd + 7) / 8)); \
129 src2 = _src2 + i; \
130 ff_hevc_put_hevc_bi_##name##step##_##bitd##_##opt(dst, dststride, src, _srcstride, src2, \
131 height, mx, my, width); \
132 } \
133 }
134
135 #define mc_rep_funcs(name, bitd, step, W, opt) \
136 mc_rep_func(name, bitd, step, W, opt) \
137 mc_rep_uni_func(name, bitd, step, W, opt) \
138 mc_rep_bi_func(name, bitd, step, W, opt)
139
140 #define mc_rep_func2(name, bitd, step1, step2, W, opt) \
141 void ff_hevc_put_hevc_##name##W##_##bitd##_##opt(int16_t *dst, \
142 uint8_t *src, ptrdiff_t _srcstride, int height, \
143 intptr_t mx, intptr_t my, int width) \
144 { \
145 ff_hevc_put_hevc_##name##step1##_##bitd##_##opt(dst, src, _srcstride, height, mx, my, width); \
146 ff_hevc_put_hevc_##name##step2##_##bitd##_##opt(dst + step1, src + (step1 * ((bitd + 7) / 8)), \
147 _srcstride, height, mx, my, width); \
148 }
149 #define mc_rep_uni_func2(name, bitd, step1, step2, W, opt) \
150 void ff_hevc_put_hevc_uni_##name##W##_##bitd##_##opt(uint8_t *dst, ptrdiff_t dststride, \
151 uint8_t *src, ptrdiff_t _srcstride, int height, \
152 intptr_t mx, intptr_t my, int width) \
153 { \
154 ff_hevc_put_hevc_uni_##name##step1##_##bitd##_##opt(dst, dststride, src, _srcstride, height, mx, my, width);\
155 ff_hevc_put_hevc_uni_##name##step2##_##bitd##_##opt(dst + (step1 * ((bitd + 7) / 8)), dststride, \
156 src + (step1 * ((bitd + 7) / 8)), _srcstride, \
157 height, mx, my, width); \
158 }
159 #define mc_rep_bi_func2(name, bitd, step1, step2, W, opt) \
160 void ff_hevc_put_hevc_bi_##name##W##_##bitd##_##opt(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, \
161 ptrdiff_t _srcstride, int16_t* src2, \
162 int height, intptr_t mx, intptr_t my, int width) \
163 { \
164 ff_hevc_put_hevc_bi_##name##step1##_##bitd##_##opt(dst, dststride, src, _srcstride, src2, height, mx, my, width);\
165 ff_hevc_put_hevc_bi_##name##step2##_##bitd##_##opt(dst + (step1 * ((bitd + 7) / 8)), dststride, \
166 src + (step1 * ((bitd + 7) / 8)), _srcstride, \
167 src2 + step1, height, mx, my, width); \
168 }
169
170 #define mc_rep_funcs2(name, bitd, step1, step2, W, opt) \
171 mc_rep_func2(name, bitd, step1, step2, W, opt) \
172 mc_rep_uni_func2(name, bitd, step1, step2, W, opt) \
173 mc_rep_bi_func2(name, bitd, step1, step2, W, opt)
174
175 #if ARCH_X86_64 && HAVE_SSE4_EXTERNAL
176
177 #define mc_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \
178 void ff_hevc_put_hevc_##name##width1##_10_##opt1(int16_t *dst, uint8_t *src, ptrdiff_t _srcstride, \
179 int height, intptr_t mx, intptr_t my, int width) \
180 \
181 { \
182 ff_hevc_put_hevc_##name##width2##_10_##opt1(dst, src, _srcstride, height, mx, my, width); \
183 ff_hevc_put_hevc_##name##width3##_10_##opt2(dst+ width2, src+ width4, _srcstride, height, mx, my, width); \
184 }
185
186 #define mc_bi_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \
187 void ff_hevc_put_hevc_bi_##name##width1##_10_##opt1(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, \
188 ptrdiff_t _srcstride, int16_t *src2, \
189 int height, intptr_t mx, intptr_t my, int width) \
190 { \
191 ff_hevc_put_hevc_bi_##name##width2##_10_##opt1(dst, dststride, src, _srcstride, src2, \
192 height, mx, my, width); \
193 ff_hevc_put_hevc_bi_##name##width3##_10_##opt2(dst+width4, dststride, src+width4, _srcstride, src2+width2,\
194 height, mx, my, width); \
195 }
196
197 #define mc_uni_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \
198 void ff_hevc_put_hevc_uni_##name##width1##_10_##opt1(uint8_t *dst, ptrdiff_t dststride, \
199 uint8_t *src, ptrdiff_t _srcstride, int height, \
200 intptr_t mx, intptr_t my, int width) \
201 { \
202 ff_hevc_put_hevc_uni_##name##width2##_10_##opt1(dst, dststride, src, _srcstride, \
203 height, mx, my, width); \
204 ff_hevc_put_hevc_uni_##name##width3##_10_##opt2(dst+width4, dststride, src+width4, _srcstride, \
205 height, mx, my, width); \
206 }
207
208 #define mc_rep_mixs_10(name, width1, width2, width3, opt1, opt2, width4) \
209 mc_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \
210 mc_bi_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \
211 mc_uni_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4)
212
213 #define mc_rep_mix_8(name, width1, width2, width3, opt1, opt2) \
214 void ff_hevc_put_hevc_##name##width1##_8_##opt1(int16_t *dst, uint8_t *src, ptrdiff_t _srcstride, \
215 int height, intptr_t mx, intptr_t my, int width) \
216 \
217 { \
218 ff_hevc_put_hevc_##name##width2##_8_##opt1(dst, src, _srcstride, height, mx, my, width); \
219 ff_hevc_put_hevc_##name##width3##_8_##opt2(dst+ width2, src+ width2, _srcstride, height, mx, my, width); \
220 }
221
222 #define mc_bi_rep_mix_8(name, width1, width2, width3, opt1, opt2) \
223 void ff_hevc_put_hevc_bi_##name##width1##_8_##opt1(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, \
224 ptrdiff_t _srcstride, int16_t* src2, \
225 int height, intptr_t mx, intptr_t my, int width) \
226 { \
227 ff_hevc_put_hevc_bi_##name##width2##_8_##opt1(dst, dststride, src, _srcstride, \
228 src2, height, mx, my, width); \
229 ff_hevc_put_hevc_bi_##name##width3##_8_##opt2(dst+width2, dststride, src+width2, _srcstride, \
230 src2+width2, height, mx, my, width); \
231 }
232
233 #define mc_uni_rep_mix_8(name, width1, width2, width3, opt1, opt2) \
234 void ff_hevc_put_hevc_uni_##name##width1##_8_##opt1(uint8_t *dst, ptrdiff_t dststride, \
235 uint8_t *src, ptrdiff_t _srcstride, int height, \
236 intptr_t mx, intptr_t my, int width) \
237 { \
238 ff_hevc_put_hevc_uni_##name##width2##_8_##opt1(dst, dststride, src, _srcstride, \
239 height, mx, my, width); \
240 ff_hevc_put_hevc_uni_##name##width3##_8_##opt2(dst+width2, dststride, src+width2, _srcstride, \
241 height, mx, my, width); \
242 }
243
244 #define mc_rep_mixs_8(name, width1, width2, width3, opt1, opt2) \
245 mc_rep_mix_8(name, width1, width2, width3, opt1, opt2) \
246 mc_bi_rep_mix_8(name, width1, width2, width3, opt1, opt2) \
247 mc_uni_rep_mix_8(name, width1, width2, width3, opt1, opt2)
248
249 #if HAVE_AVX2_EXTERNAL
250
251 mc_rep_mixs_8(pel_pixels, 48, 32, 16, avx2, sse4)
252 mc_rep_mixs_8(epel_hv, 48, 32, 16, avx2, sse4)
253 mc_rep_mixs_8(epel_h , 48, 32, 16, avx2, sse4)
254 mc_rep_mixs_8(epel_v , 48, 32, 16, avx2, sse4)
255
256 mc_rep_mix_10(pel_pixels, 24, 16, 8, avx2, sse4, 32)
257 mc_bi_rep_mix_10(pel_pixels,24, 16, 8, avx2, sse4, 32)
258 mc_rep_mixs_10(epel_hv, 24, 16, 8, avx2, sse4, 32)
259 mc_rep_mixs_10(epel_h , 24, 16, 8, avx2, sse4, 32)
260 mc_rep_mixs_10(epel_v , 24, 16, 8, avx2, sse4, 32)
261
262
263 mc_rep_mixs_10(qpel_h , 24, 16, 8, avx2, sse4, 32)
264 mc_rep_mixs_10(qpel_v , 24, 16, 8, avx2, sse4, 32)
265 mc_rep_mixs_10(qpel_hv, 24, 16, 8, avx2, sse4, 32)
266
267
268 mc_rep_uni_func(pel_pixels, 8, 64, 128, avx2)//used for 10bit
269 mc_rep_uni_func(pel_pixels, 8, 32, 96, avx2) //used for 10bit
270
271 mc_rep_funcs(pel_pixels, 8, 32, 64, avx2)
272
273 mc_rep_func(pel_pixels, 10, 16, 32, avx2)
274 mc_rep_func(pel_pixels, 10, 16, 48, avx2)
275 mc_rep_func(pel_pixels, 10, 32, 64, avx2)
276
277 mc_rep_bi_func(pel_pixels, 10, 16, 32, avx2)
278 mc_rep_bi_func(pel_pixels, 10, 16, 48, avx2)
279 mc_rep_bi_func(pel_pixels, 10, 32, 64, avx2)
280
281 mc_rep_funcs(epel_h, 8, 32, 64, avx2)
282
283 mc_rep_funcs(epel_v, 8, 32, 64, avx2)
284
285 mc_rep_funcs(epel_h, 10, 16, 32, avx2)
286 mc_rep_funcs(epel_h, 10, 16, 48, avx2)
287 mc_rep_funcs(epel_h, 10, 32, 64, avx2)
288
289 mc_rep_funcs(epel_v, 10, 16, 32, avx2)
290 mc_rep_funcs(epel_v, 10, 16, 48, avx2)
291 mc_rep_funcs(epel_v, 10, 32, 64, avx2)
292
293
294 mc_rep_funcs(epel_hv, 8, 32, 64, avx2)
295
296 mc_rep_funcs(epel_hv, 10, 16, 32, avx2)
297 mc_rep_funcs(epel_hv, 10, 16, 48, avx2)
298 mc_rep_funcs(epel_hv, 10, 32, 64, avx2)
299
300 mc_rep_funcs(qpel_h, 8, 32, 64, avx2)
301 mc_rep_mixs_8(qpel_h , 48, 32, 16, avx2, sse4)
302
303 mc_rep_funcs(qpel_v, 8, 32, 64, avx2)
304 mc_rep_mixs_8(qpel_v, 48, 32, 16, avx2, sse4)
305
306 mc_rep_funcs(qpel_h, 10, 16, 32, avx2)
307 mc_rep_funcs(qpel_h, 10, 16, 48, avx2)
308 mc_rep_funcs(qpel_h, 10, 32, 64, avx2)
309
310 mc_rep_funcs(qpel_v, 10, 16, 32, avx2)
311 mc_rep_funcs(qpel_v, 10, 16, 48, avx2)
312 mc_rep_funcs(qpel_v, 10, 32, 64, avx2)
313
314 mc_rep_funcs(qpel_hv, 10, 16, 32, avx2)
315 mc_rep_funcs(qpel_hv, 10, 16, 48, avx2)
316 mc_rep_funcs(qpel_hv, 10, 32, 64, avx2)
317
318 #endif //AVX2
319
320 mc_rep_funcs(pel_pixels, 8, 16, 64, sse4)
321 mc_rep_funcs(pel_pixels, 8, 16, 48, sse4)
322 mc_rep_funcs(pel_pixels, 8, 16, 32, sse4)
323 mc_rep_funcs(pel_pixels, 8, 8, 24, sse4)
324 mc_rep_funcs(pel_pixels,10, 8, 64, sse4)
325 mc_rep_funcs(pel_pixels,10, 8, 48, sse4)
326 mc_rep_funcs(pel_pixels,10, 8, 32, sse4)
327 mc_rep_funcs(pel_pixels,10, 8, 24, sse4)
328 mc_rep_funcs(pel_pixels,10, 8, 16, sse4)
329 mc_rep_funcs(pel_pixels,10, 4, 12, sse4)
330 mc_rep_funcs(pel_pixels,12, 8, 64, sse4)
331 mc_rep_funcs(pel_pixels,12, 8, 48, sse4)
332 mc_rep_funcs(pel_pixels,12, 8, 32, sse4)
333 mc_rep_funcs(pel_pixels,12, 8, 24, sse4)
334 mc_rep_funcs(pel_pixels,12, 8, 16, sse4)
335 mc_rep_funcs(pel_pixels,12, 4, 12, sse4)
336
337 mc_rep_funcs(epel_h, 8, 16, 64, sse4)
338 mc_rep_funcs(epel_h, 8, 16, 48, sse4)
339 mc_rep_funcs(epel_h, 8, 16, 32, sse4)
340 mc_rep_funcs(epel_h, 8, 8, 24, sse4)
341 mc_rep_funcs(epel_h,10, 8, 64, sse4)
342 mc_rep_funcs(epel_h,10, 8, 48, sse4)
343 mc_rep_funcs(epel_h,10, 8, 32, sse4)
344 mc_rep_funcs(epel_h,10, 8, 24, sse4)
345 mc_rep_funcs(epel_h,10, 8, 16, sse4)
346 mc_rep_funcs(epel_h,10, 4, 12, sse4)
347 mc_rep_funcs(epel_h,12, 8, 64, sse4)
348 mc_rep_funcs(epel_h,12, 8, 48, sse4)
349 mc_rep_funcs(epel_h,12, 8, 32, sse4)
350 mc_rep_funcs(epel_h,12, 8, 24, sse4)
351 mc_rep_funcs(epel_h,12, 8, 16, sse4)
352 mc_rep_funcs(epel_h,12, 4, 12, sse4)
353 mc_rep_funcs(epel_v, 8, 16, 64, sse4)
354 mc_rep_funcs(epel_v, 8, 16, 48, sse4)
355 mc_rep_funcs(epel_v, 8, 16, 32, sse4)
356 mc_rep_funcs(epel_v, 8, 8, 24, sse4)
357 mc_rep_funcs(epel_v,10, 8, 64, sse4)
358 mc_rep_funcs(epel_v,10, 8, 48, sse4)
359 mc_rep_funcs(epel_v,10, 8, 32, sse4)
360 mc_rep_funcs(epel_v,10, 8, 24, sse4)
361 mc_rep_funcs(epel_v,10, 8, 16, sse4)
362 mc_rep_funcs(epel_v,10, 4, 12, sse4)
363 mc_rep_funcs(epel_v,12, 8, 64, sse4)
364 mc_rep_funcs(epel_v,12, 8, 48, sse4)
365 mc_rep_funcs(epel_v,12, 8, 32, sse4)
366 mc_rep_funcs(epel_v,12, 8, 24, sse4)
367 mc_rep_funcs(epel_v,12, 8, 16, sse4)
368 mc_rep_funcs(epel_v,12, 4, 12, sse4)
369 mc_rep_funcs(epel_hv, 8, 16, 64, sse4)
370 mc_rep_funcs(epel_hv, 8, 16, 48, sse4)
371 mc_rep_funcs(epel_hv, 8, 16, 32, sse4)
372 mc_rep_funcs(epel_hv, 8, 8, 24, sse4)
373 mc_rep_funcs2(epel_hv,8, 8, 4, 12, sse4)
374 mc_rep_funcs(epel_hv,10, 8, 64, sse4)
375 mc_rep_funcs(epel_hv,10, 8, 48, sse4)
376 mc_rep_funcs(epel_hv,10, 8, 32, sse4)
377 mc_rep_funcs(epel_hv,10, 8, 24, sse4)
378 mc_rep_funcs(epel_hv,10, 8, 16, sse4)
379 mc_rep_funcs(epel_hv,10, 4, 12, sse4)
380 mc_rep_funcs(epel_hv,12, 8, 64, sse4)
381 mc_rep_funcs(epel_hv,12, 8, 48, sse4)
382 mc_rep_funcs(epel_hv,12, 8, 32, sse4)
383 mc_rep_funcs(epel_hv,12, 8, 24, sse4)
384 mc_rep_funcs(epel_hv,12, 8, 16, sse4)
385 mc_rep_funcs(epel_hv,12, 4, 12, sse4)
386
387 mc_rep_funcs(qpel_h, 8, 16, 64, sse4)
388 mc_rep_funcs(qpel_h, 8, 16, 48, sse4)
389 mc_rep_funcs(qpel_h, 8, 16, 32, sse4)
390 mc_rep_funcs(qpel_h, 8, 8, 24, sse4)
391 mc_rep_funcs(qpel_h,10, 8, 64, sse4)
392 mc_rep_funcs(qpel_h,10, 8, 48, sse4)
393 mc_rep_funcs(qpel_h,10, 8, 32, sse4)
394 mc_rep_funcs(qpel_h,10, 8, 24, sse4)
395 mc_rep_funcs(qpel_h,10, 8, 16, sse4)
396 mc_rep_funcs(qpel_h,10, 4, 12, sse4)
397 mc_rep_funcs(qpel_h,12, 8, 64, sse4)
398 mc_rep_funcs(qpel_h,12, 8, 48, sse4)
399 mc_rep_funcs(qpel_h,12, 8, 32, sse4)
400 mc_rep_funcs(qpel_h,12, 8, 24, sse4)
401 mc_rep_funcs(qpel_h,12, 8, 16, sse4)
402 mc_rep_funcs(qpel_h,12, 4, 12, sse4)
403 mc_rep_funcs(qpel_v, 8, 16, 64, sse4)
404 mc_rep_funcs(qpel_v, 8, 16, 48, sse4)
405 mc_rep_funcs(qpel_v, 8, 16, 32, sse4)
406 mc_rep_funcs(qpel_v, 8, 8, 24, sse4)
407 mc_rep_funcs(qpel_v,10, 8, 64, sse4)
408 mc_rep_funcs(qpel_v,10, 8, 48, sse4)
409 mc_rep_funcs(qpel_v,10, 8, 32, sse4)
410 mc_rep_funcs(qpel_v,10, 8, 24, sse4)
411 mc_rep_funcs(qpel_v,10, 8, 16, sse4)
412 mc_rep_funcs(qpel_v,10, 4, 12, sse4)
413 mc_rep_funcs(qpel_v,12, 8, 64, sse4)
414 mc_rep_funcs(qpel_v,12, 8, 48, sse4)
415 mc_rep_funcs(qpel_v,12, 8, 32, sse4)
416 mc_rep_funcs(qpel_v,12, 8, 24, sse4)
417 mc_rep_funcs(qpel_v,12, 8, 16, sse4)
418 mc_rep_funcs(qpel_v,12, 4, 12, sse4)
419 mc_rep_funcs(qpel_hv, 8, 8, 64, sse4)
420 mc_rep_funcs(qpel_hv, 8, 8, 48, sse4)
421 mc_rep_funcs(qpel_hv, 8, 8, 32, sse4)
422 mc_rep_funcs(qpel_hv, 8, 8, 24, sse4)
423 mc_rep_funcs(qpel_hv, 8, 8, 16, sse4)
424 mc_rep_funcs2(qpel_hv,8, 8, 4, 12, sse4)
425 mc_rep_funcs(qpel_hv,10, 8, 64, sse4)
426 mc_rep_funcs(qpel_hv,10, 8, 48, sse4)
427 mc_rep_funcs(qpel_hv,10, 8, 32, sse4)
428 mc_rep_funcs(qpel_hv,10, 8, 24, sse4)
429 mc_rep_funcs(qpel_hv,10, 8, 16, sse4)
430 mc_rep_funcs(qpel_hv,10, 4, 12, sse4)
431 mc_rep_funcs(qpel_hv,12, 8, 64, sse4)
432 mc_rep_funcs(qpel_hv,12, 8, 48, sse4)
433 mc_rep_funcs(qpel_hv,12, 8, 32, sse4)
434 mc_rep_funcs(qpel_hv,12, 8, 24, sse4)
435 mc_rep_funcs(qpel_hv,12, 8, 16, sse4)
436 mc_rep_funcs(qpel_hv,12, 4, 12, sse4)
437
438 #define mc_rep_uni_w(bitd, step, W, opt) \
439 void ff_hevc_put_hevc_uni_w##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, int16_t *_src, \
440 int height, int denom, int _wx, int _ox) \
441 { \
442 int i; \
443 int16_t *src; \
444 uint8_t *dst; \
445 for (i = 0; i < W; i += step) { \
446 src= _src + i; \
447 dst= _dst + (i * ((bitd + 7) / 8)); \
448 ff_hevc_put_hevc_uni_w##step##_##bitd##_##opt(dst, dststride, src, \
449 height, denom, _wx, _ox); \
450 } \
451 }
452
453 mc_rep_uni_w(8, 6, 12, sse4)
454 mc_rep_uni_w(8, 8, 16, sse4)
455 mc_rep_uni_w(8, 8, 24, sse4)
456 mc_rep_uni_w(8, 8, 32, sse4)
457 mc_rep_uni_w(8, 8, 48, sse4)
458 mc_rep_uni_w(8, 8, 64, sse4)
459
460 mc_rep_uni_w(10, 6, 12, sse4)
461 mc_rep_uni_w(10, 8, 16, sse4)
462 mc_rep_uni_w(10, 8, 24, sse4)
463 mc_rep_uni_w(10, 8, 32, sse4)
464 mc_rep_uni_w(10, 8, 48, sse4)
465 mc_rep_uni_w(10, 8, 64, sse4)
466
467 mc_rep_uni_w(12, 6, 12, sse4)
468 mc_rep_uni_w(12, 8, 16, sse4)
469 mc_rep_uni_w(12, 8, 24, sse4)
470 mc_rep_uni_w(12, 8, 32, sse4)
471 mc_rep_uni_w(12, 8, 48, sse4)
472 mc_rep_uni_w(12, 8, 64, sse4)
473
474 #define mc_rep_bi_w(bitd, step, W, opt) \
475 void ff_hevc_put_hevc_bi_w##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, int16_t *_src, \
476 int16_t *_src2, int height, \
477 int denom, int _wx0, int _wx1, int _ox0, int _ox1) \
478 { \
479 int i; \
480 int16_t *src; \
481 int16_t *src2; \
482 uint8_t *dst; \
483 for (i = 0; i < W; i += step) { \
484 src = _src + i; \
485 src2 = _src2 + i; \
486 dst = _dst + (i * ((bitd + 7) / 8)); \
487 ff_hevc_put_hevc_bi_w##step##_##bitd##_##opt(dst, dststride, src, src2, \
488 height, denom, _wx0, _wx1, _ox0, _ox1); \
489 } \
490 }
491
492 mc_rep_bi_w(8, 6, 12, sse4)
493 mc_rep_bi_w(8, 8, 16, sse4)
494 mc_rep_bi_w(8, 8, 24, sse4)
495 mc_rep_bi_w(8, 8, 32, sse4)
496 mc_rep_bi_w(8, 8, 48, sse4)
497 mc_rep_bi_w(8, 8, 64, sse4)
498
499 mc_rep_bi_w(10, 6, 12, sse4)
500 mc_rep_bi_w(10, 8, 16, sse4)
501 mc_rep_bi_w(10, 8, 24, sse4)
502 mc_rep_bi_w(10, 8, 32, sse4)
503 mc_rep_bi_w(10, 8, 48, sse4)
504 mc_rep_bi_w(10, 8, 64, sse4)
505
506 mc_rep_bi_w(12, 6, 12, sse4)
507 mc_rep_bi_w(12, 8, 16, sse4)
508 mc_rep_bi_w(12, 8, 24, sse4)
509 mc_rep_bi_w(12, 8, 32, sse4)
510 mc_rep_bi_w(12, 8, 48, sse4)
511 mc_rep_bi_w(12, 8, 64, sse4)
512
513 #define mc_uni_w_func(name, bitd, W, opt) \
514 void ff_hevc_put_hevc_uni_w_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t _dststride, \
515 uint8_t *_src, ptrdiff_t _srcstride, \
516 int height, int denom, \
517 int _wx, int _ox, \
518 intptr_t mx, intptr_t my, int width) \
519 { \
520 LOCAL_ALIGNED_16(int16_t, temp, [71 * MAX_PB_SIZE]); \
521 ff_hevc_put_hevc_##name##W##_##bitd##_##opt(temp, _src, _srcstride, height, mx, my, width); \
522 ff_hevc_put_hevc_uni_w##W##_##bitd##_##opt(_dst, _dststride, temp, height, denom, _wx, _ox);\
523 }
524
525 #define mc_uni_w_funcs(name, bitd, opt) \
526 mc_uni_w_func(name, bitd, 4, opt) \
527 mc_uni_w_func(name, bitd, 8, opt) \
528 mc_uni_w_func(name, bitd, 12, opt) \
529 mc_uni_w_func(name, bitd, 16, opt) \
530 mc_uni_w_func(name, bitd, 24, opt) \
531 mc_uni_w_func(name, bitd, 32, opt) \
532 mc_uni_w_func(name, bitd, 48, opt) \
533 mc_uni_w_func(name, bitd, 64, opt)
534
535 mc_uni_w_funcs(pel_pixels, 8, sse4)
536 mc_uni_w_func(pel_pixels, 8, 6, sse4)
537 mc_uni_w_funcs(epel_h, 8, sse4)
538 mc_uni_w_func(epel_h, 8, 6, sse4)
539 mc_uni_w_funcs(epel_v, 8, sse4)
540 mc_uni_w_func(epel_v, 8, 6, sse4)
541 mc_uni_w_funcs(epel_hv, 8, sse4)
542 mc_uni_w_func(epel_hv, 8, 6, sse4)
543 mc_uni_w_funcs(qpel_h, 8, sse4)
544 mc_uni_w_funcs(qpel_v, 8, sse4)
545 mc_uni_w_funcs(qpel_hv, 8, sse4)
546
547 mc_uni_w_funcs(pel_pixels, 10, sse4)
548 mc_uni_w_func(pel_pixels, 10, 6, sse4)
549 mc_uni_w_funcs(epel_h, 10, sse4)
550 mc_uni_w_func(epel_h, 10, 6, sse4)
551 mc_uni_w_funcs(epel_v, 10, sse4)
552 mc_uni_w_func(epel_v, 10, 6, sse4)
553 mc_uni_w_funcs(epel_hv, 10, sse4)
554 mc_uni_w_func(epel_hv, 10, 6, sse4)
555 mc_uni_w_funcs(qpel_h, 10, sse4)
556 mc_uni_w_funcs(qpel_v, 10, sse4)
557 mc_uni_w_funcs(qpel_hv, 10, sse4)
558
559 mc_uni_w_funcs(pel_pixels, 12, sse4)
560 mc_uni_w_func(pel_pixels, 12, 6, sse4)
561 mc_uni_w_funcs(epel_h, 12, sse4)
562 mc_uni_w_func(epel_h, 12, 6, sse4)
563 mc_uni_w_funcs(epel_v, 12, sse4)
564 mc_uni_w_func(epel_v, 12, 6, sse4)
565 mc_uni_w_funcs(epel_hv, 12, sse4)
566 mc_uni_w_func(epel_hv, 12, 6, sse4)
567 mc_uni_w_funcs(qpel_h, 12, sse4)
568 mc_uni_w_funcs(qpel_v, 12, sse4)
569 mc_uni_w_funcs(qpel_hv, 12, sse4)
570
571 #define mc_bi_w_func(name, bitd, W, opt) \
572 void ff_hevc_put_hevc_bi_w_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t _dststride, \
573 uint8_t *_src, ptrdiff_t _srcstride, \
574 int16_t *_src2, \
575 int height, int denom, \
576 int _wx0, int _wx1, int _ox0, int _ox1, \
577 intptr_t mx, intptr_t my, int width) \
578 { \
579 LOCAL_ALIGNED_16(int16_t, temp, [71 * MAX_PB_SIZE]); \
580 ff_hevc_put_hevc_##name##W##_##bitd##_##opt(temp, _src, _srcstride, height, mx, my, width); \
581 ff_hevc_put_hevc_bi_w##W##_##bitd##_##opt(_dst, _dststride, temp, _src2, \
582 height, denom, _wx0, _wx1, _ox0, _ox1); \
583 }
584
585 #define mc_bi_w_funcs(name, bitd, opt) \
586 mc_bi_w_func(name, bitd, 4, opt) \
587 mc_bi_w_func(name, bitd, 8, opt) \
588 mc_bi_w_func(name, bitd, 12, opt) \
589 mc_bi_w_func(name, bitd, 16, opt) \
590 mc_bi_w_func(name, bitd, 24, opt) \
591 mc_bi_w_func(name, bitd, 32, opt) \
592 mc_bi_w_func(name, bitd, 48, opt) \
593 mc_bi_w_func(name, bitd, 64, opt)
594
595 mc_bi_w_funcs(pel_pixels, 8, sse4)
596 mc_bi_w_func(pel_pixels, 8, 6, sse4)
597 mc_bi_w_funcs(epel_h, 8, sse4)
598 mc_bi_w_func(epel_h, 8, 6, sse4)
599 mc_bi_w_funcs(epel_v, 8, sse4)
600 mc_bi_w_func(epel_v, 8, 6, sse4)
601 mc_bi_w_funcs(epel_hv, 8, sse4)
602 mc_bi_w_func(epel_hv, 8, 6, sse4)
603 mc_bi_w_funcs(qpel_h, 8, sse4)
604 mc_bi_w_funcs(qpel_v, 8, sse4)
605 mc_bi_w_funcs(qpel_hv, 8, sse4)
606
607 mc_bi_w_funcs(pel_pixels, 10, sse4)
608 mc_bi_w_func(pel_pixels, 10, 6, sse4)
609 mc_bi_w_funcs(epel_h, 10, sse4)
610 mc_bi_w_func(epel_h, 10, 6, sse4)
611 mc_bi_w_funcs(epel_v, 10, sse4)
612 mc_bi_w_func(epel_v, 10, 6, sse4)
613 mc_bi_w_funcs(epel_hv, 10, sse4)
614 mc_bi_w_func(epel_hv, 10, 6, sse4)
615 mc_bi_w_funcs(qpel_h, 10, sse4)
616 mc_bi_w_funcs(qpel_v, 10, sse4)
617 mc_bi_w_funcs(qpel_hv, 10, sse4)
618
619 mc_bi_w_funcs(pel_pixels, 12, sse4)
620 mc_bi_w_func(pel_pixels, 12, 6, sse4)
621 mc_bi_w_funcs(epel_h, 12, sse4)
622 mc_bi_w_func(epel_h, 12, 6, sse4)
623 mc_bi_w_funcs(epel_v, 12, sse4)
624 mc_bi_w_func(epel_v, 12, 6, sse4)
625 mc_bi_w_funcs(epel_hv, 12, sse4)
626 mc_bi_w_func(epel_hv, 12, 6, sse4)
627 mc_bi_w_funcs(qpel_h, 12, sse4)
628 mc_bi_w_funcs(qpel_v, 12, sse4)
629 mc_bi_w_funcs(qpel_hv, 12, sse4)
630 #endif //ARCH_X86_64 && HAVE_SSE4_EXTERNAL
631
632 #define SAO_BAND_FILTER_FUNCS(bitd, opt) \
633 void ff_hevc_sao_band_filter_8_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
634 int16_t *sao_offset_val, int sao_left_class, int width, int height); \
635 void ff_hevc_sao_band_filter_16_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
636 int16_t *sao_offset_val, int sao_left_class, int width, int height); \
637 void ff_hevc_sao_band_filter_32_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
638 int16_t *sao_offset_val, int sao_left_class, int width, int height); \
639 void ff_hevc_sao_band_filter_48_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
640 int16_t *sao_offset_val, int sao_left_class, int width, int height); \
641 void ff_hevc_sao_band_filter_64_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
642 int16_t *sao_offset_val, int sao_left_class, int width, int height);
643
644 SAO_BAND_FILTER_FUNCS(8, sse2)
645 SAO_BAND_FILTER_FUNCS(10, sse2)
646 SAO_BAND_FILTER_FUNCS(12, sse2)
647 SAO_BAND_FILTER_FUNCS(8, avx)
648 SAO_BAND_FILTER_FUNCS(10, avx)
649 SAO_BAND_FILTER_FUNCS(12, avx)
650 SAO_BAND_FILTER_FUNCS(8, avx2)
651 SAO_BAND_FILTER_FUNCS(10, avx2)
652 SAO_BAND_FILTER_FUNCS(12, avx2)
653
654 #define SAO_BAND_INIT(bitd, opt) do { \
655 c->sao_band_filter[0] = ff_hevc_sao_band_filter_8_##bitd##_##opt; \
656 c->sao_band_filter[1] = ff_hevc_sao_band_filter_16_##bitd##_##opt; \
657 c->sao_band_filter[2] = ff_hevc_sao_band_filter_32_##bitd##_##opt; \
658 c->sao_band_filter[3] = ff_hevc_sao_band_filter_48_##bitd##_##opt; \
659 c->sao_band_filter[4] = ff_hevc_sao_band_filter_64_##bitd##_##opt; \
660 } while (0)
661
662 #define SAO_EDGE_FILTER_FUNCS(bitd, opt) \
663 void ff_hevc_sao_edge_filter_8_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \
664 int eo, int width, int height); \
665 void ff_hevc_sao_edge_filter_16_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \
666 int eo, int width, int height); \
667 void ff_hevc_sao_edge_filter_32_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \
668 int eo, int width, int height); \
669 void ff_hevc_sao_edge_filter_48_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \
670 int eo, int width, int height); \
671 void ff_hevc_sao_edge_filter_64_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \
672 int eo, int width, int height); \
673
674 SAO_EDGE_FILTER_FUNCS(8, ssse3)
675 SAO_EDGE_FILTER_FUNCS(8, avx2)
676 SAO_EDGE_FILTER_FUNCS(10, sse2)
677 SAO_EDGE_FILTER_FUNCS(10, avx2)
678 SAO_EDGE_FILTER_FUNCS(12, sse2)
679 SAO_EDGE_FILTER_FUNCS(12, avx2)
680
681 #define SAO_EDGE_INIT(bitd, opt) do { \
682 c->sao_edge_filter[0] = ff_hevc_sao_edge_filter_8_##bitd##_##opt; \
683 c->sao_edge_filter[1] = ff_hevc_sao_edge_filter_16_##bitd##_##opt; \
684 c->sao_edge_filter[2] = ff_hevc_sao_edge_filter_32_##bitd##_##opt; \
685 c->sao_edge_filter[3] = ff_hevc_sao_edge_filter_48_##bitd##_##opt; \
686 c->sao_edge_filter[4] = ff_hevc_sao_edge_filter_64_##bitd##_##opt; \
687 } while (0)
688
689 #define EPEL_LINKS(pointer, my, mx, fname, bitd, opt ) \
690 PEL_LINK(pointer, 1, my , mx , fname##4 , bitd, opt ); \
691 PEL_LINK(pointer, 2, my , mx , fname##6 , bitd, opt ); \
692 PEL_LINK(pointer, 3, my , mx , fname##8 , bitd, opt ); \
693 PEL_LINK(pointer, 4, my , mx , fname##12, bitd, opt ); \
694 PEL_LINK(pointer, 5, my , mx , fname##16, bitd, opt ); \
695 PEL_LINK(pointer, 6, my , mx , fname##24, bitd, opt ); \
696 PEL_LINK(pointer, 7, my , mx , fname##32, bitd, opt ); \
697 PEL_LINK(pointer, 8, my , mx , fname##48, bitd, opt ); \
698 PEL_LINK(pointer, 9, my , mx , fname##64, bitd, opt )
699 #define QPEL_LINKS(pointer, my, mx, fname, bitd, opt) \
700 PEL_LINK(pointer, 1, my , mx , fname##4 , bitd, opt ); \
701 PEL_LINK(pointer, 3, my , mx , fname##8 , bitd, opt ); \
702 PEL_LINK(pointer, 4, my , mx , fname##12, bitd, opt ); \
703 PEL_LINK(pointer, 5, my , mx , fname##16, bitd, opt ); \
704 PEL_LINK(pointer, 6, my , mx , fname##24, bitd, opt ); \
705 PEL_LINK(pointer, 7, my , mx , fname##32, bitd, opt ); \
706 PEL_LINK(pointer, 8, my , mx , fname##48, bitd, opt ); \
707 PEL_LINK(pointer, 9, my , mx , fname##64, bitd, opt )
708
709 void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth)
710 {
711 int cpu_flags = av_get_cpu_flags();
712
713 if (bit_depth == 8) {
714 if (EXTERNAL_MMXEXT(cpu_flags)) {
715 c->idct_dc[0] = ff_hevc_idct_4x4_dc_8_mmxext;
716 c->idct_dc[1] = ff_hevc_idct_8x8_dc_8_mmxext;
717
718 c->add_residual[0] = ff_hevc_add_residual_4_8_mmxext;
719 }
720 if (EXTERNAL_SSE2(cpu_flags)) {
721 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_sse2;
722 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_sse2;
723 if (ARCH_X86_64) {
724 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_sse2;
725 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_sse2;
726
727 c->idct[2] = ff_hevc_idct_16x16_8_sse2;
728 c->idct[3] = ff_hevc_idct_32x32_8_sse2;
729 }
730 SAO_BAND_INIT(8, sse2);
731
732 c->idct_dc[1] = ff_hevc_idct_8x8_dc_8_sse2;
733 c->idct_dc[2] = ff_hevc_idct_16x16_dc_8_sse2;
734 c->idct_dc[3] = ff_hevc_idct_32x32_dc_8_sse2;
735
736 c->idct[0] = ff_hevc_idct_4x4_8_sse2;
737 c->idct[1] = ff_hevc_idct_8x8_8_sse2;
738
739 c->add_residual[1] = ff_hevc_add_residual_8_8_sse2;
740 c->add_residual[2] = ff_hevc_add_residual_16_8_sse2;
741 c->add_residual[3] = ff_hevc_add_residual_32_8_sse2;
742 }
743 if (EXTERNAL_SSSE3(cpu_flags)) {
744 if(ARCH_X86_64) {
745 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_ssse3;
746 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_ssse3;
747 }
748 SAO_EDGE_INIT(8, ssse3);
749 }
750 if (EXTERNAL_SSE4(cpu_flags) && ARCH_X86_64) {
751
752 EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 8, sse4);
753 EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 8, sse4);
754 EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 8, sse4);
755 EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 8, sse4);
756
757 QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 8, sse4);
758 QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 8, sse4);
759 QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 8, sse4);
760 QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 8, sse4);
761 }
762 if (EXTERNAL_AVX(cpu_flags)) {
763 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_avx;
764 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_avx;
765 if (ARCH_X86_64) {
766 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_avx;
767 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_avx;
768
769 c->idct[2] = ff_hevc_idct_16x16_8_avx;
770 c->idct[3] = ff_hevc_idct_32x32_8_avx;
771 }
772 SAO_BAND_INIT(8, avx);
773
774 c->idct[0] = ff_hevc_idct_4x4_8_avx;
775 c->idct[1] = ff_hevc_idct_8x8_8_avx;
776
777 c->add_residual[1] = ff_hevc_add_residual_8_8_avx;
778 c->add_residual[2] = ff_hevc_add_residual_16_8_avx;
779 c->add_residual[3] = ff_hevc_add_residual_32_8_avx;
780 }
781 if (EXTERNAL_AVX2(cpu_flags)) {
782 c->sao_band_filter[0] = ff_hevc_sao_band_filter_8_8_avx2;
783 c->sao_band_filter[1] = ff_hevc_sao_band_filter_16_8_avx2;
784 }
785 if (EXTERNAL_AVX2_FAST(cpu_flags)) {
786 c->idct_dc[2] = ff_hevc_idct_16x16_dc_8_avx2;
787 c->idct_dc[3] = ff_hevc_idct_32x32_dc_8_avx2;
788 if (ARCH_X86_64) {
789 c->put_hevc_epel[7][0][0] = ff_hevc_put_hevc_pel_pixels32_8_avx2;
790 c->put_hevc_epel[8][0][0] = ff_hevc_put_hevc_pel_pixels48_8_avx2;
791 c->put_hevc_epel[9][0][0] = ff_hevc_put_hevc_pel_pixels64_8_avx2;
792
793 c->put_hevc_qpel[7][0][0] = ff_hevc_put_hevc_pel_pixels32_8_avx2;
794 c->put_hevc_qpel[8][0][0] = ff_hevc_put_hevc_pel_pixels48_8_avx2;
795 c->put_hevc_qpel[9][0][0] = ff_hevc_put_hevc_pel_pixels64_8_avx2;
796
797 c->put_hevc_epel_uni[7][0][0] = ff_hevc_put_hevc_uni_pel_pixels32_8_avx2;
798 c->put_hevc_epel_uni[8][0][0] = ff_hevc_put_hevc_uni_pel_pixels48_8_avx2;
799 c->put_hevc_epel_uni[9][0][0] = ff_hevc_put_hevc_uni_pel_pixels64_8_avx2;
800
801 c->put_hevc_qpel_uni[7][0][0] = ff_hevc_put_hevc_uni_pel_pixels32_8_avx2;
802 c->put_hevc_qpel_uni[8][0][0] = ff_hevc_put_hevc_uni_pel_pixels48_8_avx2;
803 c->put_hevc_qpel_uni[9][0][0] = ff_hevc_put_hevc_uni_pel_pixels64_8_avx2;
804
805 c->put_hevc_qpel_bi[7][0][0] = ff_hevc_put_hevc_bi_pel_pixels32_8_avx2;
806 c->put_hevc_qpel_bi[8][0][0] = ff_hevc_put_hevc_bi_pel_pixels48_8_avx2;
807 c->put_hevc_qpel_bi[9][0][0] = ff_hevc_put_hevc_bi_pel_pixels64_8_avx2;
808
809 c->put_hevc_epel_bi[7][0][0] = ff_hevc_put_hevc_bi_pel_pixels32_8_avx2;
810 c->put_hevc_epel_bi[8][0][0] = ff_hevc_put_hevc_bi_pel_pixels48_8_avx2;
811 c->put_hevc_epel_bi[9][0][0] = ff_hevc_put_hevc_bi_pel_pixels64_8_avx2;
812
813 c->put_hevc_epel[7][0][1] = ff_hevc_put_hevc_epel_h32_8_avx2;
814 c->put_hevc_epel[8][0][1] = ff_hevc_put_hevc_epel_h48_8_avx2;
815 c->put_hevc_epel[9][0][1] = ff_hevc_put_hevc_epel_h64_8_avx2;
816
817 c->put_hevc_epel_uni[7][0][1] = ff_hevc_put_hevc_uni_epel_h32_8_avx2;
818 c->put_hevc_epel_uni[8][0][1] = ff_hevc_put_hevc_uni_epel_h48_8_avx2;
819 c->put_hevc_epel_uni[9][0][1] = ff_hevc_put_hevc_uni_epel_h64_8_avx2;
820
821 c->put_hevc_epel_bi[7][0][1] = ff_hevc_put_hevc_bi_epel_h32_8_avx2;
822 c->put_hevc_epel_bi[8][0][1] = ff_hevc_put_hevc_bi_epel_h48_8_avx2;
823 c->put_hevc_epel_bi[9][0][1] = ff_hevc_put_hevc_bi_epel_h64_8_avx2;
824
825 c->put_hevc_epel[7][1][0] = ff_hevc_put_hevc_epel_v32_8_avx2;
826 c->put_hevc_epel[8][1][0] = ff_hevc_put_hevc_epel_v48_8_avx2;
827 c->put_hevc_epel[9][1][0] = ff_hevc_put_hevc_epel_v64_8_avx2;
828
829 c->put_hevc_epel_uni[7][1][0] = ff_hevc_put_hevc_uni_epel_v32_8_avx2;
830 c->put_hevc_epel_uni[8][1][0] = ff_hevc_put_hevc_uni_epel_v48_8_avx2;
831 c->put_hevc_epel_uni[9][1][0] = ff_hevc_put_hevc_uni_epel_v64_8_avx2;
832
833 c->put_hevc_epel_bi[7][1][0] = ff_hevc_put_hevc_bi_epel_v32_8_avx2;
834 c->put_hevc_epel_bi[8][1][0] = ff_hevc_put_hevc_bi_epel_v48_8_avx2;
835 c->put_hevc_epel_bi[9][1][0] = ff_hevc_put_hevc_bi_epel_v64_8_avx2;
836
837 c->put_hevc_epel[7][1][1] = ff_hevc_put_hevc_epel_hv32_8_avx2;
838 c->put_hevc_epel[8][1][1] = ff_hevc_put_hevc_epel_hv48_8_avx2;
839 c->put_hevc_epel[9][1][1] = ff_hevc_put_hevc_epel_hv64_8_avx2;
840
841 c->put_hevc_epel_uni[7][1][1] = ff_hevc_put_hevc_uni_epel_hv32_8_avx2;
842 c->put_hevc_epel_uni[8][1][1] = ff_hevc_put_hevc_uni_epel_hv48_8_avx2;
843 c->put_hevc_epel_uni[9][1][1] = ff_hevc_put_hevc_uni_epel_hv64_8_avx2;
844
845 c->put_hevc_epel_bi[7][1][1] = ff_hevc_put_hevc_bi_epel_hv32_8_avx2;
846 c->put_hevc_epel_bi[8][1][1] = ff_hevc_put_hevc_bi_epel_hv48_8_avx2;
847 c->put_hevc_epel_bi[9][1][1] = ff_hevc_put_hevc_bi_epel_hv64_8_avx2;
848
849 c->put_hevc_qpel[7][0][1] = ff_hevc_put_hevc_qpel_h32_8_avx2;
850 c->put_hevc_qpel[8][0][1] = ff_hevc_put_hevc_qpel_h48_8_avx2;
851 c->put_hevc_qpel[9][0][1] = ff_hevc_put_hevc_qpel_h64_8_avx2;
852
853 c->put_hevc_qpel[7][1][0] = ff_hevc_put_hevc_qpel_v32_8_avx2;
854 c->put_hevc_qpel[8][1][0] = ff_hevc_put_hevc_qpel_v48_8_avx2;
855 c->put_hevc_qpel[9][1][0] = ff_hevc_put_hevc_qpel_v64_8_avx2;
856
857 c->put_hevc_qpel_uni[7][0][1] = ff_hevc_put_hevc_uni_qpel_h32_8_avx2;
858 c->put_hevc_qpel_uni[8][0][1] = ff_hevc_put_hevc_uni_qpel_h48_8_avx2;
859 c->put_hevc_qpel_uni[9][0][1] = ff_hevc_put_hevc_uni_qpel_h64_8_avx2;
860
861 c->put_hevc_qpel_uni[7][1][0] = ff_hevc_put_hevc_uni_qpel_v32_8_avx2;
862 c->put_hevc_qpel_uni[8][1][0] = ff_hevc_put_hevc_uni_qpel_v48_8_avx2;
863 c->put_hevc_qpel_uni[9][1][0] = ff_hevc_put_hevc_uni_qpel_v64_8_avx2;
864
865 c->put_hevc_qpel_bi[7][0][1] = ff_hevc_put_hevc_bi_qpel_h32_8_avx2;
866 c->put_hevc_qpel_bi[8][0][1] = ff_hevc_put_hevc_bi_qpel_h48_8_avx2;
867 c->put_hevc_qpel_bi[9][0][1] = ff_hevc_put_hevc_bi_qpel_h64_8_avx2;
868
869 c->put_hevc_qpel_bi[7][1][0] = ff_hevc_put_hevc_bi_qpel_v32_8_avx2;
870 c->put_hevc_qpel_bi[8][1][0] = ff_hevc_put_hevc_bi_qpel_v48_8_avx2;
871 c->put_hevc_qpel_bi[9][1][0] = ff_hevc_put_hevc_bi_qpel_v64_8_avx2;
872 }
873 SAO_BAND_INIT(8, avx2);
874
875 c->sao_edge_filter[2] = ff_hevc_sao_edge_filter_32_8_avx2;
876 c->sao_edge_filter[3] = ff_hevc_sao_edge_filter_48_8_avx2;
877 c->sao_edge_filter[4] = ff_hevc_sao_edge_filter_64_8_avx2;
878
879 c->add_residual[3] = ff_hevc_add_residual_32_8_avx2;
880 }
881 } else if (bit_depth == 10) {
882 if (EXTERNAL_MMXEXT(cpu_flags)) {
883 c->add_residual[0] = ff_hevc_add_residual_4_10_mmxext;
884 c->idct_dc[0] = ff_hevc_idct_4x4_dc_10_mmxext;
885 c->idct_dc[1] = ff_hevc_idct_8x8_dc_10_mmxext;
886 }
887 if (EXTERNAL_SSE2(cpu_flags)) {
888 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_sse2;
889 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_10_sse2;
890 if (ARCH_X86_64) {
891 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_sse2;
892 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_sse2;
893
894 c->idct[2] = ff_hevc_idct_16x16_10_sse2;
895 c->idct[3] = ff_hevc_idct_32x32_10_sse2;
896 }
897 SAO_BAND_INIT(10, sse2);
898 SAO_EDGE_INIT(10, sse2);
899
900 c->idct_dc[1] = ff_hevc_idct_8x8_dc_10_sse2;
901 c->idct_dc[2] = ff_hevc_idct_16x16_dc_10_sse2;
902 c->idct_dc[3] = ff_hevc_idct_32x32_dc_10_sse2;
903
904 c->idct[0] = ff_hevc_idct_4x4_10_sse2;
905 c->idct[1] = ff_hevc_idct_8x8_10_sse2;
906
907 c->add_residual[1] = ff_hevc_add_residual_8_10_sse2;
908 c->add_residual[2] = ff_hevc_add_residual_16_10_sse2;
909 c->add_residual[3] = ff_hevc_add_residual_32_10_sse2;
910 }
911 if (EXTERNAL_SSSE3(cpu_flags) && ARCH_X86_64) {
912 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_ssse3;
913 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_ssse3;
914 }
915 if (EXTERNAL_SSE4(cpu_flags) && ARCH_X86_64) {
916 EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 10, sse4);
917 EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 10, sse4);
918 EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 10, sse4);
919 EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 10, sse4);
920
921 QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 10, sse4);
922 QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 10, sse4);
923 QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 10, sse4);
924 QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 10, sse4);
925 }
926 if (EXTERNAL_AVX(cpu_flags)) {
927 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_avx;
928 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_10_avx;
929 if (ARCH_X86_64) {
930 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_avx;
931 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_avx;
932
933 c->idct[2] = ff_hevc_idct_16x16_10_avx;
934 c->idct[3] = ff_hevc_idct_32x32_10_avx;
935 }
936
937 c->idct[0] = ff_hevc_idct_4x4_10_avx;
938 c->idct[1] = ff_hevc_idct_8x8_10_avx;
939
940 SAO_BAND_INIT(10, avx);
941 }
942 if (EXTERNAL_AVX2(cpu_flags)) {
943 c->sao_band_filter[0] = ff_hevc_sao_band_filter_8_10_avx2;
944 }
945 if (EXTERNAL_AVX2_FAST(cpu_flags)) {
946 c->idct_dc[2] = ff_hevc_idct_16x16_dc_10_avx2;
947 c->idct_dc[3] = ff_hevc_idct_32x32_dc_10_avx2;
948 if (ARCH_X86_64) {
949 c->put_hevc_epel[5][0][0] = ff_hevc_put_hevc_pel_pixels16_10_avx2;
950 c->put_hevc_epel[6][0][0] = ff_hevc_put_hevc_pel_pixels24_10_avx2;
951 c->put_hevc_epel[7][0][0] = ff_hevc_put_hevc_pel_pixels32_10_avx2;
952 c->put_hevc_epel[8][0][0] = ff_hevc_put_hevc_pel_pixels48_10_avx2;
953 c->put_hevc_epel[9][0][0] = ff_hevc_put_hevc_pel_pixels64_10_avx2;
954
955 c->put_hevc_qpel[5][0][0] = ff_hevc_put_hevc_pel_pixels16_10_avx2;
956 c->put_hevc_qpel[6][0][0] = ff_hevc_put_hevc_pel_pixels24_10_avx2;
957 c->put_hevc_qpel[7][0][0] = ff_hevc_put_hevc_pel_pixels32_10_avx2;
958 c->put_hevc_qpel[8][0][0] = ff_hevc_put_hevc_pel_pixels48_10_avx2;
959 c->put_hevc_qpel[9][0][0] = ff_hevc_put_hevc_pel_pixels64_10_avx2;
960
961 c->put_hevc_epel_uni[5][0][0] = ff_hevc_put_hevc_uni_pel_pixels32_8_avx2;
962 c->put_hevc_epel_uni[6][0][0] = ff_hevc_put_hevc_uni_pel_pixels48_8_avx2;
963 c->put_hevc_epel_uni[7][0][0] = ff_hevc_put_hevc_uni_pel_pixels64_8_avx2;
964 c->put_hevc_epel_uni[8][0][0] = ff_hevc_put_hevc_uni_pel_pixels96_8_avx2;
965 c->put_hevc_epel_uni[9][0][0] = ff_hevc_put_hevc_uni_pel_pixels128_8_avx2;
966
967 c->put_hevc_qpel_uni[5][0][0] = ff_hevc_put_hevc_uni_pel_pixels32_8_avx2;
968 c->put_hevc_qpel_uni[6][0][0] = ff_hevc_put_hevc_uni_pel_pixels48_8_avx2;
969 c->put_hevc_qpel_uni[7][0][0] = ff_hevc_put_hevc_uni_pel_pixels64_8_avx2;
970 c->put_hevc_qpel_uni[8][0][0] = ff_hevc_put_hevc_uni_pel_pixels96_8_avx2;
971 c->put_hevc_qpel_uni[9][0][0] = ff_hevc_put_hevc_uni_pel_pixels128_8_avx2;
972
973 c->put_hevc_epel_bi[5][0][0] = ff_hevc_put_hevc_bi_pel_pixels16_10_avx2;
974 c->put_hevc_epel_bi[6][0][0] = ff_hevc_put_hevc_bi_pel_pixels24_10_avx2;
975 c->put_hevc_epel_bi[7][0][0] = ff_hevc_put_hevc_bi_pel_pixels32_10_avx2;
976 c->put_hevc_epel_bi[8][0][0] = ff_hevc_put_hevc_bi_pel_pixels48_10_avx2;
977 c->put_hevc_epel_bi[9][0][0] = ff_hevc_put_hevc_bi_pel_pixels64_10_avx2;
978 c->put_hevc_qpel_bi[5][0][0] = ff_hevc_put_hevc_bi_pel_pixels16_10_avx2;
979 c->put_hevc_qpel_bi[6][0][0] = ff_hevc_put_hevc_bi_pel_pixels24_10_avx2;
980 c->put_hevc_qpel_bi[7][0][0] = ff_hevc_put_hevc_bi_pel_pixels32_10_avx2;
981 c->put_hevc_qpel_bi[8][0][0] = ff_hevc_put_hevc_bi_pel_pixels48_10_avx2;
982 c->put_hevc_qpel_bi[9][0][0] = ff_hevc_put_hevc_bi_pel_pixels64_10_avx2;
983
984 c->put_hevc_epel[5][0][1] = ff_hevc_put_hevc_epel_h16_10_avx2;
985 c->put_hevc_epel[6][0][1] = ff_hevc_put_hevc_epel_h24_10_avx2;
986 c->put_hevc_epel[7][0][1] = ff_hevc_put_hevc_epel_h32_10_avx2;
987 c->put_hevc_epel[8][0][1] = ff_hevc_put_hevc_epel_h48_10_avx2;
988 c->put_hevc_epel[9][0][1] = ff_hevc_put_hevc_epel_h64_10_avx2;
989
990 c->put_hevc_epel_uni[5][0][1] = ff_hevc_put_hevc_uni_epel_h16_10_avx2;
991 c->put_hevc_epel_uni[6][0][1] = ff_hevc_put_hevc_uni_epel_h24_10_avx2;
992 c->put_hevc_epel_uni[7][0][1] = ff_hevc_put_hevc_uni_epel_h32_10_avx2;
993 c->put_hevc_epel_uni[8][0][1] = ff_hevc_put_hevc_uni_epel_h48_10_avx2;
994 c->put_hevc_epel_uni[9][0][1] = ff_hevc_put_hevc_uni_epel_h64_10_avx2;
995
996 c->put_hevc_epel_bi[5][0][1] = ff_hevc_put_hevc_bi_epel_h16_10_avx2;
997 c->put_hevc_epel_bi[6][0][1] = ff_hevc_put_hevc_bi_epel_h24_10_avx2;
998 c->put_hevc_epel_bi[7][0][1] = ff_hevc_put_hevc_bi_epel_h32_10_avx2;
999 c->put_hevc_epel_bi[8][0][1] = ff_hevc_put_hevc_bi_epel_h48_10_avx2;
1000 c->put_hevc_epel_bi[9][0][1] = ff_hevc_put_hevc_bi_epel_h64_10_avx2;
1001
1002 c->put_hevc_epel[5][1][0] = ff_hevc_put_hevc_epel_v16_10_avx2;
1003 c->put_hevc_epel[6][1][0] = ff_hevc_put_hevc_epel_v24_10_avx2;
1004 c->put_hevc_epel[7][1][0] = ff_hevc_put_hevc_epel_v32_10_avx2;
1005 c->put_hevc_epel[8][1][0] = ff_hevc_put_hevc_epel_v48_10_avx2;
1006 c->put_hevc_epel[9][1][0] = ff_hevc_put_hevc_epel_v64_10_avx2;
1007
1008 c->put_hevc_epel_uni[5][1][0] = ff_hevc_put_hevc_uni_epel_v16_10_avx2;
1009 c->put_hevc_epel_uni[6][1][0] = ff_hevc_put_hevc_uni_epel_v24_10_avx2;
1010 c->put_hevc_epel_uni[7][1][0] = ff_hevc_put_hevc_uni_epel_v32_10_avx2;
1011 c->put_hevc_epel_uni[8][1][0] = ff_hevc_put_hevc_uni_epel_v48_10_avx2;
1012 c->put_hevc_epel_uni[9][1][0] = ff_hevc_put_hevc_uni_epel_v64_10_avx2;
1013
1014 c->put_hevc_epel_bi[5][1][0] = ff_hevc_put_hevc_bi_epel_v16_10_avx2;
1015 c->put_hevc_epel_bi[6][1][0] = ff_hevc_put_hevc_bi_epel_v24_10_avx2;
1016 c->put_hevc_epel_bi[7][1][0] = ff_hevc_put_hevc_bi_epel_v32_10_avx2;
1017 c->put_hevc_epel_bi[8][1][0] = ff_hevc_put_hevc_bi_epel_v48_10_avx2;
1018 c->put_hevc_epel_bi[9][1][0] = ff_hevc_put_hevc_bi_epel_v64_10_avx2;
1019
1020 c->put_hevc_epel[5][1][1] = ff_hevc_put_hevc_epel_hv16_10_avx2;
1021 c->put_hevc_epel[6][1][1] = ff_hevc_put_hevc_epel_hv24_10_avx2;
1022 c->put_hevc_epel[7][1][1] = ff_hevc_put_hevc_epel_hv32_10_avx2;
1023 c->put_hevc_epel[8][1][1] = ff_hevc_put_hevc_epel_hv48_10_avx2;
1024 c->put_hevc_epel[9][1][1] = ff_hevc_put_hevc_epel_hv64_10_avx2;
1025
1026 c->put_hevc_epel_uni[5][1][1] = ff_hevc_put_hevc_uni_epel_hv16_10_avx2;
1027 c->put_hevc_epel_uni[6][1][1] = ff_hevc_put_hevc_uni_epel_hv24_10_avx2;
1028 c->put_hevc_epel_uni[7][1][1] = ff_hevc_put_hevc_uni_epel_hv32_10_avx2;
1029 c->put_hevc_epel_uni[8][1][1] = ff_hevc_put_hevc_uni_epel_hv48_10_avx2;
1030 c->put_hevc_epel_uni[9][1][1] = ff_hevc_put_hevc_uni_epel_hv64_10_avx2;
1031
1032 c->put_hevc_epel_bi[5][1][1] = ff_hevc_put_hevc_bi_epel_hv16_10_avx2;
1033 c->put_hevc_epel_bi[6][1][1] = ff_hevc_put_hevc_bi_epel_hv24_10_avx2;
1034 c->put_hevc_epel_bi[7][1][1] = ff_hevc_put_hevc_bi_epel_hv32_10_avx2;
1035 c->put_hevc_epel_bi[8][1][1] = ff_hevc_put_hevc_bi_epel_hv48_10_avx2;
1036 c->put_hevc_epel_bi[9][1][1] = ff_hevc_put_hevc_bi_epel_hv64_10_avx2;
1037
1038 c->put_hevc_qpel[5][0][1] = ff_hevc_put_hevc_qpel_h16_10_avx2;
1039 c->put_hevc_qpel[6][0][1] = ff_hevc_put_hevc_qpel_h24_10_avx2;
1040 c->put_hevc_qpel[7][0][1] = ff_hevc_put_hevc_qpel_h32_10_avx2;
1041 c->put_hevc_qpel[8][0][1] = ff_hevc_put_hevc_qpel_h48_10_avx2;
1042 c->put_hevc_qpel[9][0][1] = ff_hevc_put_hevc_qpel_h64_10_avx2;
1043
1044 c->put_hevc_qpel_uni[5][0][1] = ff_hevc_put_hevc_uni_qpel_h16_10_avx2;
1045 c->put_hevc_qpel_uni[6][0][1] = ff_hevc_put_hevc_uni_qpel_h24_10_avx2;
1046 c->put_hevc_qpel_uni[7][0][1] = ff_hevc_put_hevc_uni_qpel_h32_10_avx2;
1047 c->put_hevc_qpel_uni[8][0][1] = ff_hevc_put_hevc_uni_qpel_h48_10_avx2;
1048 c->put_hevc_qpel_uni[9][0][1] = ff_hevc_put_hevc_uni_qpel_h64_10_avx2;
1049
1050 c->put_hevc_qpel_bi[5][0][1] = ff_hevc_put_hevc_bi_qpel_h16_10_avx2;
1051 c->put_hevc_qpel_bi[6][0][1] = ff_hevc_put_hevc_bi_qpel_h24_10_avx2;
1052 c->put_hevc_qpel_bi[7][0][1] = ff_hevc_put_hevc_bi_qpel_h32_10_avx2;
1053 c->put_hevc_qpel_bi[8][0][1] = ff_hevc_put_hevc_bi_qpel_h48_10_avx2;
1054 c->put_hevc_qpel_bi[9][0][1] = ff_hevc_put_hevc_bi_qpel_h64_10_avx2;
1055
1056 c->put_hevc_qpel[5][1][0] = ff_hevc_put_hevc_qpel_v16_10_avx2;
1057 c->put_hevc_qpel[6][1][0] = ff_hevc_put_hevc_qpel_v24_10_avx2;
1058 c->put_hevc_qpel[7][1][0] = ff_hevc_put_hevc_qpel_v32_10_avx2;
1059 c->put_hevc_qpel[8][1][0] = ff_hevc_put_hevc_qpel_v48_10_avx2;
1060 c->put_hevc_qpel[9][1][0] = ff_hevc_put_hevc_qpel_v64_10_avx2;
1061
1062 c->put_hevc_qpel_uni[5][1][0] = ff_hevc_put_hevc_uni_qpel_v16_10_avx2;
1063 c->put_hevc_qpel_uni[6][1][0] = ff_hevc_put_hevc_uni_qpel_v24_10_avx2;
1064 c->put_hevc_qpel_uni[7][1][0] = ff_hevc_put_hevc_uni_qpel_v32_10_avx2;
1065 c->put_hevc_qpel_uni[8][1][0] = ff_hevc_put_hevc_uni_qpel_v48_10_avx2;
1066 c->put_hevc_qpel_uni[9][1][0] = ff_hevc_put_hevc_uni_qpel_v64_10_avx2;
1067
1068 c->put_hevc_qpel_bi[5][1][0] = ff_hevc_put_hevc_bi_qpel_v16_10_avx2;
1069 c->put_hevc_qpel_bi[6][1][0] = ff_hevc_put_hevc_bi_qpel_v24_10_avx2;
1070 c->put_hevc_qpel_bi[7][1][0] = ff_hevc_put_hevc_bi_qpel_v32_10_avx2;
1071 c->put_hevc_qpel_bi[8][1][0] = ff_hevc_put_hevc_bi_qpel_v48_10_avx2;
1072 c->put_hevc_qpel_bi[9][1][0] = ff_hevc_put_hevc_bi_qpel_v64_10_avx2;
1073
1074 c->put_hevc_qpel[5][1][1] = ff_hevc_put_hevc_qpel_hv16_10_avx2;
1075 c->put_hevc_qpel[6][1][1] = ff_hevc_put_hevc_qpel_hv24_10_avx2;
1076 c->put_hevc_qpel[7][1][1] = ff_hevc_put_hevc_qpel_hv32_10_avx2;
1077 c->put_hevc_qpel[8][1][1] = ff_hevc_put_hevc_qpel_hv48_10_avx2;
1078 c->put_hevc_qpel[9][1][1] = ff_hevc_put_hevc_qpel_hv64_10_avx2;
1079
1080 c->put_hevc_qpel_uni[5][1][1] = ff_hevc_put_hevc_uni_qpel_hv16_10_avx2;
1081 c->put_hevc_qpel_uni[6][1][1] = ff_hevc_put_hevc_uni_qpel_hv24_10_avx2;
1082 c->put_hevc_qpel_uni[7][1][1] = ff_hevc_put_hevc_uni_qpel_hv32_10_avx2;
1083 c->put_hevc_qpel_uni[8][1][1] = ff_hevc_put_hevc_uni_qpel_hv48_10_avx2;
1084 c->put_hevc_qpel_uni[9][1][1] = ff_hevc_put_hevc_uni_qpel_hv64_10_avx2;
1085
1086 c->put_hevc_qpel_bi[5][1][1] = ff_hevc_put_hevc_bi_qpel_hv16_10_avx2;
1087 c->put_hevc_qpel_bi[6][1][1] = ff_hevc_put_hevc_bi_qpel_hv24_10_avx2;
1088 c->put_hevc_qpel_bi[7][1][1] = ff_hevc_put_hevc_bi_qpel_hv32_10_avx2;
1089 c->put_hevc_qpel_bi[8][1][1] = ff_hevc_put_hevc_bi_qpel_hv48_10_avx2;
1090 c->put_hevc_qpel_bi[9][1][1] = ff_hevc_put_hevc_bi_qpel_hv64_10_avx2;
1091 }
1092 SAO_BAND_INIT(10, avx2);
1093 SAO_EDGE_INIT(10, avx2);
1094
1095 c->add_residual[2] = ff_hevc_add_residual_16_10_avx2;
1096 c->add_residual[3] = ff_hevc_add_residual_32_10_avx2;
1097 }
1098 } else if (bit_depth == 12) {
1099 if (EXTERNAL_MMXEXT(cpu_flags)) {
1100 c->idct_dc[0] = ff_hevc_idct_4x4_dc_12_mmxext;
1101 c->idct_dc[1] = ff_hevc_idct_8x8_dc_12_mmxext;
1102 }
1103 if (EXTERNAL_SSE2(cpu_flags)) {
1104 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_12_sse2;
1105 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_12_sse2;
1106 if (ARCH_X86_64) {
1107 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_sse2;
1108 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_sse2;
1109 }
1110 SAO_BAND_INIT(12, sse2);
1111 SAO_EDGE_INIT(12, sse2);
1112
1113 c->idct_dc[1] = ff_hevc_idct_8x8_dc_12_sse2;
1114 c->idct_dc[2] = ff_hevc_idct_16x16_dc_12_sse2;
1115 c->idct_dc[3] = ff_hevc_idct_32x32_dc_12_sse2;
1116 }
1117 if (EXTERNAL_SSSE3(cpu_flags) && ARCH_X86_64) {
1118 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_ssse3;
1119 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_ssse3;
1120 }
1121 if (EXTERNAL_SSE4(cpu_flags) && ARCH_X86_64) {
1122 EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 12, sse4);
1123 EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 12, sse4);
1124 EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 12, sse4);
1125 EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 12, sse4);
1126
1127 QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 12, sse4);
1128 QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 12, sse4);
1129 QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 12, sse4);
1130 QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 12, sse4);
1131 }
1132 if (EXTERNAL_AVX(cpu_flags)) {
1133 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_12_avx;
1134 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_12_avx;
1135 if (ARCH_X86_64) {
1136 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_avx;
1137 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_avx;
1138 }
1139 SAO_BAND_INIT(12, avx);
1140 }
1141 if (EXTERNAL_AVX2(cpu_flags)) {
1142 c->sao_band_filter[0] = ff_hevc_sao_band_filter_8_12_avx2;
1143 }
1144 if (EXTERNAL_AVX2_FAST(cpu_flags)) {
1145 c->idct_dc[2] = ff_hevc_idct_16x16_dc_12_avx2;
1146 c->idct_dc[3] = ff_hevc_idct_32x32_dc_12_avx2;
1147
1148 SAO_BAND_INIT(12, avx2);
1149 SAO_EDGE_INIT(12, avx2);
1150 }
1151 }
1152 }
1153