2 * Copyright (c) 2013 Seppo Tomperi
3 * Copyright (c) 2013 - 2014 Pierre-Edouard Lepere
6 * This file is part of FFmpeg.
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.
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.
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
24 #include "libavutil/cpu.h"
25 #include "libavutil/x86/asm.h"
26 #include "libavutil/x86/cpu.h"
27 #include "libavcodec/get_bits.h" /* required for hevcdsp.h GetBitContext */
28 #include "libavcodec/hevcdsp.h"
29 #include "libavcodec/x86/hevcdsp.h"
31 #define LFC_FUNC(DIR, DEPTH, OPT) \
32 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 #define LFL_FUNC(DIR, DEPTH, OPT) \
35 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 #define LFC_FUNCS(type, depth, opt) \
38 LFC_FUNC(h, depth, opt) \
39 LFC_FUNC(v, depth, opt)
41 #define LFL_FUNCS(type, depth, opt) \
42 LFL_FUNC(h, depth, opt) \
43 LFL_FUNC(v, depth, opt)
45 LFC_FUNCS(uint8_t, 8, sse2)
46 LFC_FUNCS(uint8_t, 10, sse2)
47 LFC_FUNCS(uint8_t, 12, sse2)
48 LFC_FUNCS(uint8_t, 8, avx)
49 LFC_FUNCS(uint8_t, 10, avx)
50 LFC_FUNCS(uint8_t, 12, avx)
51 LFL_FUNCS(uint8_t, 8, sse2)
52 LFL_FUNCS(uint8_t, 10, sse2)
53 LFL_FUNCS(uint8_t, 12, sse2)
54 LFL_FUNCS(uint8_t, 8, ssse3)
55 LFL_FUNCS(uint8_t, 10, ssse3)
56 LFL_FUNCS(uint8_t, 12, ssse3)
57 LFL_FUNCS(uint8_t, 8, avx)
58 LFL_FUNCS(uint8_t, 10, avx)
59 LFL_FUNCS(uint8_t, 12, avx)
61 #define IDCT_FUNCS(W, opt) \
62 void ff_hevc_idct##W##_dc_8_##opt(int16_t *coeffs); \
63 void ff_hevc_idct##W##_dc_10_##opt(int16_t *coeffs); \
64 void ff_hevc_idct##W##_dc_12_##opt(int16_t *coeffs)
66 IDCT_FUNCS(4x4, mmxext);
67 IDCT_FUNCS(8x8, mmxext);
68 IDCT_FUNCS(8x8, sse2);
69 IDCT_FUNCS(16x16, sse2);
70 IDCT_FUNCS(32x32, sse2);
71 IDCT_FUNCS(16x16, avx2);
72 IDCT_FUNCS(32x32, avx2);
74 #define mc_rep_func(name, bitd, step, W, opt) \
75 void ff_hevc_put_hevc_##name##W##_##bitd##_##opt(int16_t *_dst, \
76 uint8_t *_src, ptrdiff_t _srcstride, int height, \
77 intptr_t mx, intptr_t my, int width) \
82 for (i = 0; i < W; i += step) { \
83 src = _src + (i * ((bitd + 7) / 8)); \
85 ff_hevc_put_hevc_##name##step##_##bitd##_##opt(dst, src, _srcstride, height, mx, my, width); \
88 #define mc_rep_uni_func(name, bitd, step, W, opt) \
89 void ff_hevc_put_hevc_uni_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, \
90 uint8_t *_src, ptrdiff_t _srcstride, int height, \
91 intptr_t mx, intptr_t my, int width) \
96 for (i = 0; i < W; i += step) { \
97 src = _src + (i * ((bitd + 7) / 8)); \
98 dst = _dst + (i * ((bitd + 7) / 8)); \
99 ff_hevc_put_hevc_uni_##name##step##_##bitd##_##opt(dst, dststride, src, _srcstride, \
100 height, mx, my, width); \
103 #define mc_rep_bi_func(name, bitd, step, W, opt) \
104 void ff_hevc_put_hevc_bi_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, uint8_t *_src, \
105 ptrdiff_t _srcstride, int16_t* _src2, \
106 int height, intptr_t mx, intptr_t my, int width) \
112 for (i = 0; i < W ; i += step) { \
113 src = _src + (i * ((bitd + 7) / 8)); \
114 dst = _dst + (i * ((bitd + 7) / 8)); \
116 ff_hevc_put_hevc_bi_##name##step##_##bitd##_##opt(dst, dststride, src, _srcstride, src2, \
117 height, mx, my, width); \
121 #define mc_rep_funcs(name, bitd, step, W, opt) \
122 mc_rep_func(name, bitd, step, W, opt); \
123 mc_rep_uni_func(name, bitd, step, W, opt); \
124 mc_rep_bi_func(name, bitd, step, W, opt)
126 #define mc_rep_func2(name, bitd, step1, step2, W, opt) \
127 void ff_hevc_put_hevc_##name##W##_##bitd##_##opt(int16_t *dst, \
128 uint8_t *src, ptrdiff_t _srcstride, int height, \
129 intptr_t mx, intptr_t my, int width) \
131 ff_hevc_put_hevc_##name##step1##_##bitd##_##opt(dst, src, _srcstride, height, mx, my, width); \
132 ff_hevc_put_hevc_##name##step2##_##bitd##_##opt(dst + step1, src + (step1 * ((bitd + 7) / 8)), \
133 _srcstride, height, mx, my, width); \
135 #define mc_rep_uni_func2(name, bitd, step1, step2, W, opt) \
136 void ff_hevc_put_hevc_uni_##name##W##_##bitd##_##opt(uint8_t *dst, ptrdiff_t dststride, \
137 uint8_t *src, ptrdiff_t _srcstride, int height, \
138 intptr_t mx, intptr_t my, int width) \
140 ff_hevc_put_hevc_uni_##name##step1##_##bitd##_##opt(dst, dststride, src, _srcstride, height, mx, my, width);\
141 ff_hevc_put_hevc_uni_##name##step2##_##bitd##_##opt(dst + (step1 * ((bitd + 7) / 8)), dststride, \
142 src + (step1 * ((bitd + 7) / 8)), _srcstride, \
143 height, mx, my, width); \
145 #define mc_rep_bi_func2(name, bitd, step1, step2, W, opt) \
146 void ff_hevc_put_hevc_bi_##name##W##_##bitd##_##opt(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, \
147 ptrdiff_t _srcstride, int16_t* src2, \
148 int height, intptr_t mx, intptr_t my, int width) \
150 ff_hevc_put_hevc_bi_##name##step1##_##bitd##_##opt(dst, dststride, src, _srcstride, src2, height, mx, my, width);\
151 ff_hevc_put_hevc_bi_##name##step2##_##bitd##_##opt(dst + (step1 * ((bitd + 7) / 8)), dststride, \
152 src + (step1 * ((bitd + 7) / 8)), _srcstride, \
153 src2 + step1, height, mx, my, width); \
156 #define mc_rep_funcs(name, bitd, step, W, opt) \
157 mc_rep_func(name, bitd, step, W, opt); \
158 mc_rep_uni_func(name, bitd, step, W, opt); \
159 mc_rep_bi_func(name, bitd, step, W, opt)
161 #define mc_rep_funcs2(name, bitd, step1, step2, W, opt) \
162 mc_rep_func2(name, bitd, step1, step2, W, opt); \
163 mc_rep_uni_func2(name, bitd, step1, step2, W, opt); \
164 mc_rep_bi_func2(name, bitd, step1, step2, W, opt)
166 #if ARCH_X86_64 && HAVE_SSE4_EXTERNAL
168 #define mc_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \
169 void ff_hevc_put_hevc_##name##width1##_10_##opt1(int16_t *dst, uint8_t *src, ptrdiff_t _srcstride, \
170 int height, intptr_t mx, intptr_t my, int width) \
173 ff_hevc_put_hevc_##name##width2##_10_##opt1(dst, src, _srcstride, height, mx, my, width); \
174 ff_hevc_put_hevc_##name##width3##_10_##opt2(dst+ width2, src+ width4, _srcstride, height, mx, my, width); \
177 #define mc_bi_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \
178 void ff_hevc_put_hevc_bi_##name##width1##_10_##opt1(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, \
179 ptrdiff_t _srcstride, int16_t *src2, \
180 int height, intptr_t mx, intptr_t my, int width) \
182 ff_hevc_put_hevc_bi_##name##width2##_10_##opt1(dst, dststride, src, _srcstride, src2, \
183 height, mx, my, width); \
184 ff_hevc_put_hevc_bi_##name##width3##_10_##opt2(dst+width4, dststride, src+width4, _srcstride, src2+width2,\
185 height, mx, my, width); \
188 #define mc_uni_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \
189 void ff_hevc_put_hevc_uni_##name##width1##_10_##opt1(uint8_t *dst, ptrdiff_t dststride, \
190 uint8_t *src, ptrdiff_t _srcstride, int height, \
191 intptr_t mx, intptr_t my, int width) \
193 ff_hevc_put_hevc_uni_##name##width2##_10_##opt1(dst, dststride, src, _srcstride, \
194 height, mx, my, width); \
195 ff_hevc_put_hevc_uni_##name##width3##_10_##opt2(dst+width4, dststride, src+width4, _srcstride, \
196 height, mx, my, width); \
199 #define mc_rep_mixs_10(name, width1, width2, width3, opt1, opt2, width4) \
200 mc_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4); \
201 mc_bi_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4); \
202 mc_uni_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4)
204 #define mc_rep_mix_8(name, width1, width2, width3, opt1, opt2) \
205 void ff_hevc_put_hevc_##name##width1##_8_##opt1(int16_t *dst, uint8_t *src, ptrdiff_t _srcstride, \
206 int height, intptr_t mx, intptr_t my, int width) \
209 ff_hevc_put_hevc_##name##width2##_8_##opt1(dst, src, _srcstride, height, mx, my, width); \
210 ff_hevc_put_hevc_##name##width3##_8_##opt2(dst+ width2, src+ width2, _srcstride, height, mx, my, width); \
213 #define mc_bi_rep_mix_8(name, width1, width2, width3, opt1, opt2) \
214 void ff_hevc_put_hevc_bi_##name##width1##_8_##opt1(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, \
215 ptrdiff_t _srcstride, int16_t* src2, \
216 int height, intptr_t mx, intptr_t my, int width) \
218 ff_hevc_put_hevc_bi_##name##width2##_8_##opt1(dst, dststride, src, _srcstride, \
219 src2, height, mx, my, width); \
220 ff_hevc_put_hevc_bi_##name##width3##_8_##opt2(dst+width2, dststride, src+width2, _srcstride, \
221 src2+width2, height, mx, my, width); \
224 #define mc_uni_rep_mix_8(name, width1, width2, width3, opt1, opt2) \
225 void ff_hevc_put_hevc_uni_##name##width1##_8_##opt1(uint8_t *dst, ptrdiff_t dststride, \
226 uint8_t *src, ptrdiff_t _srcstride, int height, \
227 intptr_t mx, intptr_t my, int width) \
229 ff_hevc_put_hevc_uni_##name##width2##_8_##opt1(dst, dststride, src, _srcstride, \
230 height, mx, my, width); \
231 ff_hevc_put_hevc_uni_##name##width3##_8_##opt2(dst+width2, dststride, src+width2, _srcstride, \
232 height, mx, my, width); \
235 #define mc_rep_mixs_8(name, width1, width2, width3, opt1, opt2) \
236 mc_rep_mix_8(name, width1, width2, width3, opt1, opt2); \
237 mc_bi_rep_mix_8(name, width1, width2, width3, opt1, opt2); \
238 mc_uni_rep_mix_8(name, width1, width2, width3, opt1, opt2)
240 #if HAVE_AVX2_EXTERNAL
242 mc_rep_mixs_8(pel_pixels, 48, 32, 16, avx2, sse4);
243 mc_rep_mixs_8(epel_hv, 48, 32, 16, avx2, sse4);
244 mc_rep_mixs_8(epel_h , 48, 32, 16, avx2, sse4);
245 mc_rep_mixs_8(epel_v , 48, 32, 16, avx2, sse4);
247 mc_rep_mix_10(pel_pixels, 24, 16, 8, avx2, sse4, 32);
248 mc_bi_rep_mix_10(pel_pixels,24, 16, 8, avx2, sse4, 32);
249 mc_rep_mixs_10(epel_hv, 24, 16, 8, avx2, sse4, 32);
250 mc_rep_mixs_10(epel_h , 24, 16, 8, avx2, sse4, 32);
251 mc_rep_mixs_10(epel_v , 24, 16, 8, avx2, sse4, 32);
254 mc_rep_mixs_10(qpel_h , 24, 16, 8, avx2, sse4, 32);
255 mc_rep_mixs_10(qpel_v , 24, 16, 8, avx2, sse4, 32);
256 mc_rep_mixs_10(qpel_hv, 24, 16, 8, avx2, sse4, 32);
259 mc_rep_uni_func(pel_pixels, 8, 64, 128, avx2);//used for 10bit
260 mc_rep_uni_func(pel_pixels, 8, 32, 96, avx2); //used for 10bit
262 mc_rep_funcs(pel_pixels, 8, 32, 64, avx2);
264 mc_rep_func(pel_pixels, 10, 16, 32, avx2);
265 mc_rep_func(pel_pixels, 10, 16, 48, avx2);
266 mc_rep_func(pel_pixels, 10, 32, 64, avx2);
268 mc_rep_bi_func(pel_pixels, 10, 16, 32, avx2);
269 mc_rep_bi_func(pel_pixels, 10, 16, 48, avx2);
270 mc_rep_bi_func(pel_pixels, 10, 32, 64, avx2);
272 mc_rep_funcs(epel_h, 8, 32, 64, avx2);
274 mc_rep_funcs(epel_v, 8, 32, 64, avx2);
276 mc_rep_funcs(epel_h, 10, 16, 32, avx2);
277 mc_rep_funcs(epel_h, 10, 16, 48, avx2);
278 mc_rep_funcs(epel_h, 10, 32, 64, avx2);
280 mc_rep_funcs(epel_v, 10, 16, 32, avx2);
281 mc_rep_funcs(epel_v, 10, 16, 48, avx2);
282 mc_rep_funcs(epel_v, 10, 32, 64, avx2);
285 mc_rep_funcs(epel_hv, 8, 32, 64, avx2);
287 mc_rep_funcs(epel_hv, 10, 16, 32, avx2);
288 mc_rep_funcs(epel_hv, 10, 16, 48, avx2);
289 mc_rep_funcs(epel_hv, 10, 32, 64, avx2);
291 mc_rep_funcs(qpel_h, 8, 32, 64, avx2);
292 mc_rep_mixs_8(qpel_h , 48, 32, 16, avx2, sse4);
294 mc_rep_funcs(qpel_v, 8, 32, 64, avx2);
295 mc_rep_mixs_8(qpel_v, 48, 32, 16, avx2, sse4);
297 mc_rep_funcs(qpel_h, 10, 16, 32, avx2);
298 mc_rep_funcs(qpel_h, 10, 16, 48, avx2);
299 mc_rep_funcs(qpel_h, 10, 32, 64, avx2);
301 mc_rep_funcs(qpel_v, 10, 16, 32, avx2);
302 mc_rep_funcs(qpel_v, 10, 16, 48, avx2);
303 mc_rep_funcs(qpel_v, 10, 32, 64, avx2);
305 mc_rep_funcs(qpel_hv, 10, 16, 32, avx2);
306 mc_rep_funcs(qpel_hv, 10, 16, 48, avx2);
307 mc_rep_funcs(qpel_hv, 10, 32, 64, avx2);
311 mc_rep_funcs(pel_pixels, 8, 16, 64, sse4);
312 mc_rep_funcs(pel_pixels, 8, 16, 48, sse4);
313 mc_rep_funcs(pel_pixels, 8, 16, 32, sse4);
314 mc_rep_funcs(pel_pixels, 8, 8, 24, sse4);
315 mc_rep_funcs(pel_pixels,10, 8, 64, sse4);
316 mc_rep_funcs(pel_pixels,10, 8, 48, sse4);
317 mc_rep_funcs(pel_pixels,10, 8, 32, sse4);
318 mc_rep_funcs(pel_pixels,10, 8, 24, sse4);
319 mc_rep_funcs(pel_pixels,10, 8, 16, sse4);
320 mc_rep_funcs(pel_pixels,10, 4, 12, sse4);
321 mc_rep_funcs(pel_pixels,12, 8, 64, sse4);
322 mc_rep_funcs(pel_pixels,12, 8, 48, sse4);
323 mc_rep_funcs(pel_pixels,12, 8, 32, sse4);
324 mc_rep_funcs(pel_pixels,12, 8, 24, sse4);
325 mc_rep_funcs(pel_pixels,12, 8, 16, sse4);
326 mc_rep_funcs(pel_pixels,12, 4, 12, sse4);
328 mc_rep_funcs(epel_h, 8, 16, 64, sse4);
329 mc_rep_funcs(epel_h, 8, 16, 48, sse4);
330 mc_rep_funcs(epel_h, 8, 16, 32, sse4);
331 mc_rep_funcs(epel_h, 8, 8, 24, sse4);
332 mc_rep_funcs(epel_h,10, 8, 64, sse4);
333 mc_rep_funcs(epel_h,10, 8, 48, sse4);
334 mc_rep_funcs(epel_h,10, 8, 32, sse4);
335 mc_rep_funcs(epel_h,10, 8, 24, sse4);
336 mc_rep_funcs(epel_h,10, 8, 16, sse4);
337 mc_rep_funcs(epel_h,10, 4, 12, sse4);
338 mc_rep_funcs(epel_h,12, 8, 64, sse4);
339 mc_rep_funcs(epel_h,12, 8, 48, sse4);
340 mc_rep_funcs(epel_h,12, 8, 32, sse4);
341 mc_rep_funcs(epel_h,12, 8, 24, sse4);
342 mc_rep_funcs(epel_h,12, 8, 16, sse4);
343 mc_rep_funcs(epel_h,12, 4, 12, sse4);
344 mc_rep_funcs(epel_v, 8, 16, 64, sse4);
345 mc_rep_funcs(epel_v, 8, 16, 48, sse4);
346 mc_rep_funcs(epel_v, 8, 16, 32, sse4);
347 mc_rep_funcs(epel_v, 8, 8, 24, sse4);
348 mc_rep_funcs(epel_v,10, 8, 64, sse4);
349 mc_rep_funcs(epel_v,10, 8, 48, sse4);
350 mc_rep_funcs(epel_v,10, 8, 32, sse4);
351 mc_rep_funcs(epel_v,10, 8, 24, sse4);
352 mc_rep_funcs(epel_v,10, 8, 16, sse4);
353 mc_rep_funcs(epel_v,10, 4, 12, sse4);
354 mc_rep_funcs(epel_v,12, 8, 64, sse4);
355 mc_rep_funcs(epel_v,12, 8, 48, sse4);
356 mc_rep_funcs(epel_v,12, 8, 32, sse4);
357 mc_rep_funcs(epel_v,12, 8, 24, sse4);
358 mc_rep_funcs(epel_v,12, 8, 16, sse4);
359 mc_rep_funcs(epel_v,12, 4, 12, sse4);
360 mc_rep_funcs(epel_hv, 8, 16, 64, sse4);
361 mc_rep_funcs(epel_hv, 8, 16, 48, sse4);
362 mc_rep_funcs(epel_hv, 8, 16, 32, sse4);
363 mc_rep_funcs(epel_hv, 8, 8, 24, sse4);
364 mc_rep_funcs2(epel_hv,8, 8, 4, 12, sse4);
365 mc_rep_funcs(epel_hv,10, 8, 64, sse4);
366 mc_rep_funcs(epel_hv,10, 8, 48, sse4);
367 mc_rep_funcs(epel_hv,10, 8, 32, sse4);
368 mc_rep_funcs(epel_hv,10, 8, 24, sse4);
369 mc_rep_funcs(epel_hv,10, 8, 16, sse4);
370 mc_rep_funcs(epel_hv,10, 4, 12, sse4);
371 mc_rep_funcs(epel_hv,12, 8, 64, sse4);
372 mc_rep_funcs(epel_hv,12, 8, 48, sse4);
373 mc_rep_funcs(epel_hv,12, 8, 32, sse4);
374 mc_rep_funcs(epel_hv,12, 8, 24, sse4);
375 mc_rep_funcs(epel_hv,12, 8, 16, sse4);
376 mc_rep_funcs(epel_hv,12, 4, 12, sse4);
378 mc_rep_funcs(qpel_h, 8, 16, 64, sse4);
379 mc_rep_funcs(qpel_h, 8, 16, 48, sse4);
380 mc_rep_funcs(qpel_h, 8, 16, 32, sse4);
381 mc_rep_funcs(qpel_h, 8, 8, 24, sse4);
382 mc_rep_funcs(qpel_h,10, 8, 64, sse4);
383 mc_rep_funcs(qpel_h,10, 8, 48, sse4);
384 mc_rep_funcs(qpel_h,10, 8, 32, sse4);
385 mc_rep_funcs(qpel_h,10, 8, 24, sse4);
386 mc_rep_funcs(qpel_h,10, 8, 16, sse4);
387 mc_rep_funcs(qpel_h,10, 4, 12, sse4);
388 mc_rep_funcs(qpel_h,12, 8, 64, sse4);
389 mc_rep_funcs(qpel_h,12, 8, 48, sse4);
390 mc_rep_funcs(qpel_h,12, 8, 32, sse4);
391 mc_rep_funcs(qpel_h,12, 8, 24, sse4);
392 mc_rep_funcs(qpel_h,12, 8, 16, sse4);
393 mc_rep_funcs(qpel_h,12, 4, 12, sse4);
394 mc_rep_funcs(qpel_v, 8, 16, 64, sse4);
395 mc_rep_funcs(qpel_v, 8, 16, 48, sse4);
396 mc_rep_funcs(qpel_v, 8, 16, 32, sse4);
397 mc_rep_funcs(qpel_v, 8, 8, 24, sse4);
398 mc_rep_funcs(qpel_v,10, 8, 64, sse4);
399 mc_rep_funcs(qpel_v,10, 8, 48, sse4);
400 mc_rep_funcs(qpel_v,10, 8, 32, sse4);
401 mc_rep_funcs(qpel_v,10, 8, 24, sse4);
402 mc_rep_funcs(qpel_v,10, 8, 16, sse4);
403 mc_rep_funcs(qpel_v,10, 4, 12, sse4);
404 mc_rep_funcs(qpel_v,12, 8, 64, sse4);
405 mc_rep_funcs(qpel_v,12, 8, 48, sse4);
406 mc_rep_funcs(qpel_v,12, 8, 32, sse4);
407 mc_rep_funcs(qpel_v,12, 8, 24, sse4);
408 mc_rep_funcs(qpel_v,12, 8, 16, sse4);
409 mc_rep_funcs(qpel_v,12, 4, 12, sse4);
410 mc_rep_funcs(qpel_hv, 8, 8, 64, sse4);
411 mc_rep_funcs(qpel_hv, 8, 8, 48, sse4);
412 mc_rep_funcs(qpel_hv, 8, 8, 32, sse4);
413 mc_rep_funcs(qpel_hv, 8, 8, 24, sse4);
414 mc_rep_funcs(qpel_hv, 8, 8, 16, sse4);
415 mc_rep_funcs2(qpel_hv,8, 8, 4, 12, sse4);
416 mc_rep_funcs(qpel_hv,10, 8, 64, sse4);
417 mc_rep_funcs(qpel_hv,10, 8, 48, sse4);
418 mc_rep_funcs(qpel_hv,10, 8, 32, sse4);
419 mc_rep_funcs(qpel_hv,10, 8, 24, sse4);
420 mc_rep_funcs(qpel_hv,10, 8, 16, sse4);
421 mc_rep_funcs(qpel_hv,10, 4, 12, sse4);
422 mc_rep_funcs(qpel_hv,12, 8, 64, sse4);
423 mc_rep_funcs(qpel_hv,12, 8, 48, sse4);
424 mc_rep_funcs(qpel_hv,12, 8, 32, sse4);
425 mc_rep_funcs(qpel_hv,12, 8, 24, sse4);
426 mc_rep_funcs(qpel_hv,12, 8, 16, sse4);
427 mc_rep_funcs(qpel_hv,12, 4, 12, sse4);
429 #define mc_rep_uni_w(bitd, step, W, opt) \
430 void ff_hevc_put_hevc_uni_w##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, int16_t *_src, \
431 int height, int denom, int _wx, int _ox) \
436 for (i = 0; i < W; i += step) { \
438 dst= _dst + (i * ((bitd + 7) / 8)); \
439 ff_hevc_put_hevc_uni_w##step##_##bitd##_##opt(dst, dststride, src, \
440 height, denom, _wx, _ox); \
444 mc_rep_uni_w(8, 6, 12, sse4);
445 mc_rep_uni_w(8, 8, 16, sse4);
446 mc_rep_uni_w(8, 8, 24, sse4);
447 mc_rep_uni_w(8, 8, 32, sse4);
448 mc_rep_uni_w(8, 8, 48, sse4);
449 mc_rep_uni_w(8, 8, 64, sse4);
451 mc_rep_uni_w(10, 6, 12, sse4);
452 mc_rep_uni_w(10, 8, 16, sse4);
453 mc_rep_uni_w(10, 8, 24, sse4);
454 mc_rep_uni_w(10, 8, 32, sse4);
455 mc_rep_uni_w(10, 8, 48, sse4);
456 mc_rep_uni_w(10, 8, 64, sse4);
458 mc_rep_uni_w(12, 6, 12, sse4);
459 mc_rep_uni_w(12, 8, 16, sse4);
460 mc_rep_uni_w(12, 8, 24, sse4);
461 mc_rep_uni_w(12, 8, 32, sse4);
462 mc_rep_uni_w(12, 8, 48, sse4);
463 mc_rep_uni_w(12, 8, 64, sse4);
465 #define mc_rep_bi_w(bitd, step, W, opt) \
466 void ff_hevc_put_hevc_bi_w##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, int16_t *_src, \
467 int16_t *_src2, int height, \
468 int denom, int _wx0, int _wx1, int _ox0, int _ox1) \
474 for (i = 0; i < W; i += step) { \
477 dst = _dst + (i * ((bitd + 7) / 8)); \
478 ff_hevc_put_hevc_bi_w##step##_##bitd##_##opt(dst, dststride, src, src2, \
479 height, denom, _wx0, _wx1, _ox0, _ox1); \
483 mc_rep_bi_w(8, 6, 12, sse4);
484 mc_rep_bi_w(8, 8, 16, sse4);
485 mc_rep_bi_w(8, 8, 24, sse4);
486 mc_rep_bi_w(8, 8, 32, sse4);
487 mc_rep_bi_w(8, 8, 48, sse4);
488 mc_rep_bi_w(8, 8, 64, sse4);
490 mc_rep_bi_w(10, 6, 12, sse4);
491 mc_rep_bi_w(10, 8, 16, sse4);
492 mc_rep_bi_w(10, 8, 24, sse4);
493 mc_rep_bi_w(10, 8, 32, sse4);
494 mc_rep_bi_w(10, 8, 48, sse4);
495 mc_rep_bi_w(10, 8, 64, sse4);
497 mc_rep_bi_w(12, 6, 12, sse4);
498 mc_rep_bi_w(12, 8, 16, sse4);
499 mc_rep_bi_w(12, 8, 24, sse4);
500 mc_rep_bi_w(12, 8, 32, sse4);
501 mc_rep_bi_w(12, 8, 48, sse4);
502 mc_rep_bi_w(12, 8, 64, sse4);
504 #define mc_uni_w_func(name, bitd, W, opt) \
505 void ff_hevc_put_hevc_uni_w_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t _dststride, \
506 uint8_t *_src, ptrdiff_t _srcstride, \
507 int height, int denom, \
509 intptr_t mx, intptr_t my, int width) \
511 LOCAL_ALIGNED_16(int16_t, temp, [71 * MAX_PB_SIZE]); \
512 ff_hevc_put_hevc_##name##W##_##bitd##_##opt(temp, _src, _srcstride, height, mx, my, width); \
513 ff_hevc_put_hevc_uni_w##W##_##bitd##_##opt(_dst, _dststride, temp, height, denom, _wx, _ox);\
516 #define mc_uni_w_funcs(name, bitd, opt) \
517 mc_uni_w_func(name, bitd, 4, opt); \
518 mc_uni_w_func(name, bitd, 8, opt); \
519 mc_uni_w_func(name, bitd, 12, opt); \
520 mc_uni_w_func(name, bitd, 16, opt); \
521 mc_uni_w_func(name, bitd, 24, opt); \
522 mc_uni_w_func(name, bitd, 32, opt); \
523 mc_uni_w_func(name, bitd, 48, opt); \
524 mc_uni_w_func(name, bitd, 64, opt)
526 mc_uni_w_funcs(pel_pixels, 8, sse4);
527 mc_uni_w_func(pel_pixels, 8, 6, sse4);
528 mc_uni_w_funcs(epel_h, 8, sse4);
529 mc_uni_w_func(epel_h, 8, 6, sse4);
530 mc_uni_w_funcs(epel_v, 8, sse4);
531 mc_uni_w_func(epel_v, 8, 6, sse4);
532 mc_uni_w_funcs(epel_hv, 8, sse4);
533 mc_uni_w_func(epel_hv, 8, 6, sse4);
534 mc_uni_w_funcs(qpel_h, 8, sse4);
535 mc_uni_w_funcs(qpel_v, 8, sse4);
536 mc_uni_w_funcs(qpel_hv, 8, sse4);
538 mc_uni_w_funcs(pel_pixels, 10, sse4);
539 mc_uni_w_func(pel_pixels, 10, 6, sse4);
540 mc_uni_w_funcs(epel_h, 10, sse4);
541 mc_uni_w_func(epel_h, 10, 6, sse4);
542 mc_uni_w_funcs(epel_v, 10, sse4);
543 mc_uni_w_func(epel_v, 10, 6, sse4);
544 mc_uni_w_funcs(epel_hv, 10, sse4);
545 mc_uni_w_func(epel_hv, 10, 6, sse4);
546 mc_uni_w_funcs(qpel_h, 10, sse4);
547 mc_uni_w_funcs(qpel_v, 10, sse4);
548 mc_uni_w_funcs(qpel_hv, 10, sse4);
550 mc_uni_w_funcs(pel_pixels, 12, sse4);
551 mc_uni_w_func(pel_pixels, 12, 6, sse4);
552 mc_uni_w_funcs(epel_h, 12, sse4);
553 mc_uni_w_func(epel_h, 12, 6, sse4);
554 mc_uni_w_funcs(epel_v, 12, sse4);
555 mc_uni_w_func(epel_v, 12, 6, sse4);
556 mc_uni_w_funcs(epel_hv, 12, sse4);
557 mc_uni_w_func(epel_hv, 12, 6, sse4);
558 mc_uni_w_funcs(qpel_h, 12, sse4);
559 mc_uni_w_funcs(qpel_v, 12, sse4);
560 mc_uni_w_funcs(qpel_hv, 12, sse4);
562 #define mc_bi_w_func(name, bitd, W, opt) \
563 void ff_hevc_put_hevc_bi_w_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t _dststride, \
564 uint8_t *_src, ptrdiff_t _srcstride, \
566 int height, int denom, \
567 int _wx0, int _wx1, int _ox0, int _ox1, \
568 intptr_t mx, intptr_t my, int width) \
570 LOCAL_ALIGNED_16(int16_t, temp, [71 * MAX_PB_SIZE]); \
571 ff_hevc_put_hevc_##name##W##_##bitd##_##opt(temp, _src, _srcstride, height, mx, my, width); \
572 ff_hevc_put_hevc_bi_w##W##_##bitd##_##opt(_dst, _dststride, temp, _src2, \
573 height, denom, _wx0, _wx1, _ox0, _ox1); \
576 #define mc_bi_w_funcs(name, bitd, opt) \
577 mc_bi_w_func(name, bitd, 4, opt); \
578 mc_bi_w_func(name, bitd, 8, opt); \
579 mc_bi_w_func(name, bitd, 12, opt); \
580 mc_bi_w_func(name, bitd, 16, opt); \
581 mc_bi_w_func(name, bitd, 24, opt); \
582 mc_bi_w_func(name, bitd, 32, opt); \
583 mc_bi_w_func(name, bitd, 48, opt); \
584 mc_bi_w_func(name, bitd, 64, opt)
586 mc_bi_w_funcs(pel_pixels, 8, sse4);
587 mc_bi_w_func(pel_pixels, 8, 6, sse4);
588 mc_bi_w_funcs(epel_h, 8, sse4);
589 mc_bi_w_func(epel_h, 8, 6, sse4);
590 mc_bi_w_funcs(epel_v, 8, sse4);
591 mc_bi_w_func(epel_v, 8, 6, sse4);
592 mc_bi_w_funcs(epel_hv, 8, sse4);
593 mc_bi_w_func(epel_hv, 8, 6, sse4);
594 mc_bi_w_funcs(qpel_h, 8, sse4);
595 mc_bi_w_funcs(qpel_v, 8, sse4);
596 mc_bi_w_funcs(qpel_hv, 8, sse4);
598 mc_bi_w_funcs(pel_pixels, 10, sse4);
599 mc_bi_w_func(pel_pixels, 10, 6, sse4);
600 mc_bi_w_funcs(epel_h, 10, sse4);
601 mc_bi_w_func(epel_h, 10, 6, sse4);
602 mc_bi_w_funcs(epel_v, 10, sse4);
603 mc_bi_w_func(epel_v, 10, 6, sse4);
604 mc_bi_w_funcs(epel_hv, 10, sse4);
605 mc_bi_w_func(epel_hv, 10, 6, sse4);
606 mc_bi_w_funcs(qpel_h, 10, sse4);
607 mc_bi_w_funcs(qpel_v, 10, sse4);
608 mc_bi_w_funcs(qpel_hv, 10, sse4);
610 mc_bi_w_funcs(pel_pixels, 12, sse4);
611 mc_bi_w_func(pel_pixels, 12, 6, sse4);
612 mc_bi_w_funcs(epel_h, 12, sse4);
613 mc_bi_w_func(epel_h, 12, 6, sse4);
614 mc_bi_w_funcs(epel_v, 12, sse4);
615 mc_bi_w_func(epel_v, 12, 6, sse4);
616 mc_bi_w_funcs(epel_hv, 12, sse4);
617 mc_bi_w_func(epel_hv, 12, 6, sse4);
618 mc_bi_w_funcs(qpel_h, 12, sse4);
619 mc_bi_w_funcs(qpel_v, 12, sse4);
620 mc_bi_w_funcs(qpel_hv, 12, sse4);
621 #endif //ARCH_X86_64 && HAVE_SSE4_EXTERNAL
623 #define SAO_BAND_FILTER_FUNCS(bitd, opt) \
624 void ff_hevc_sao_band_filter_8_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
625 int16_t *sao_offset_val, int sao_left_class, int width, int height); \
626 void ff_hevc_sao_band_filter_16_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
627 int16_t *sao_offset_val, int sao_left_class, int width, int height); \
628 void ff_hevc_sao_band_filter_32_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
629 int16_t *sao_offset_val, int sao_left_class, int width, int height); \
630 void ff_hevc_sao_band_filter_48_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
631 int16_t *sao_offset_val, int sao_left_class, int width, int height); \
632 void ff_hevc_sao_band_filter_64_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \
633 int16_t *sao_offset_val, int sao_left_class, int width, int height)
635 SAO_BAND_FILTER_FUNCS(8, sse2);
636 SAO_BAND_FILTER_FUNCS(10, sse2);
637 SAO_BAND_FILTER_FUNCS(12, sse2);
638 SAO_BAND_FILTER_FUNCS(8, avx);
639 SAO_BAND_FILTER_FUNCS(10, avx);
640 SAO_BAND_FILTER_FUNCS(12, avx);
641 SAO_BAND_FILTER_FUNCS(8, avx2);
642 SAO_BAND_FILTER_FUNCS(10, avx2);
643 SAO_BAND_FILTER_FUNCS(12, avx2);
645 #define SAO_BAND_INIT(bitd, opt) do { \
646 c->sao_band_filter[0] = ff_hevc_sao_band_filter_8_##bitd##_##opt; \
647 c->sao_band_filter[1] = ff_hevc_sao_band_filter_16_##bitd##_##opt; \
648 c->sao_band_filter[2] = ff_hevc_sao_band_filter_32_##bitd##_##opt; \
649 c->sao_band_filter[3] = ff_hevc_sao_band_filter_48_##bitd##_##opt; \
650 c->sao_band_filter[4] = ff_hevc_sao_band_filter_64_##bitd##_##opt; \
653 #define SAO_EDGE_FILTER_FUNCS(bitd, opt) \
654 void ff_hevc_sao_edge_filter_8_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \
655 int eo, int width, int height); \
656 void ff_hevc_sao_edge_filter_16_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \
657 int eo, int width, int height); \
658 void ff_hevc_sao_edge_filter_32_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \
659 int eo, int width, int height); \
660 void ff_hevc_sao_edge_filter_48_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \
661 int eo, int width, int height); \
662 void ff_hevc_sao_edge_filter_64_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \
663 int eo, int width, int height); \
665 SAO_EDGE_FILTER_FUNCS(8, ssse3);
666 SAO_EDGE_FILTER_FUNCS(8, avx2);
667 SAO_EDGE_FILTER_FUNCS(10, sse2);
668 SAO_EDGE_FILTER_FUNCS(10, avx2);
669 SAO_EDGE_FILTER_FUNCS(12, sse2);
670 SAO_EDGE_FILTER_FUNCS(12, avx2);
672 #define SAO_EDGE_INIT(bitd, opt) do { \
673 c->sao_edge_filter[0] = ff_hevc_sao_edge_filter_8_##bitd##_##opt; \
674 c->sao_edge_filter[1] = ff_hevc_sao_edge_filter_16_##bitd##_##opt; \
675 c->sao_edge_filter[2] = ff_hevc_sao_edge_filter_32_##bitd##_##opt; \
676 c->sao_edge_filter[3] = ff_hevc_sao_edge_filter_48_##bitd##_##opt; \
677 c->sao_edge_filter[4] = ff_hevc_sao_edge_filter_64_##bitd##_##opt; \
680 #define EPEL_LINKS(pointer, my, mx, fname, bitd, opt ) \
681 PEL_LINK(pointer, 1, my , mx , fname##4 , bitd, opt ); \
682 PEL_LINK(pointer, 2, my , mx , fname##6 , bitd, opt ); \
683 PEL_LINK(pointer, 3, my , mx , fname##8 , bitd, opt ); \
684 PEL_LINK(pointer, 4, my , mx , fname##12, bitd, opt ); \
685 PEL_LINK(pointer, 5, my , mx , fname##16, bitd, opt ); \
686 PEL_LINK(pointer, 6, my , mx , fname##24, bitd, opt ); \
687 PEL_LINK(pointer, 7, my , mx , fname##32, bitd, opt ); \
688 PEL_LINK(pointer, 8, my , mx , fname##48, bitd, opt ); \
689 PEL_LINK(pointer, 9, my , mx , fname##64, bitd, opt )
690 #define QPEL_LINKS(pointer, my, mx, fname, bitd, opt) \
691 PEL_LINK(pointer, 1, my , mx , fname##4 , 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 )
700 void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth)
702 int cpu_flags = av_get_cpu_flags();
704 if (bit_depth == 8) {
705 if (EXTERNAL_MMXEXT(cpu_flags)) {
706 c->idct_dc[0] = ff_hevc_idct4x4_dc_8_mmxext;
707 c->idct_dc[1] = ff_hevc_idct8x8_dc_8_mmxext;
708 c->transform_add[0] = ff_hevc_transform_add4_8_mmxext;
710 if (EXTERNAL_SSE2(cpu_flags)) {
711 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_sse2;
712 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_sse2;
714 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_sse2;
715 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_sse2;
718 SAO_BAND_INIT(8, sse2);
720 c->idct_dc[1] = ff_hevc_idct8x8_dc_8_sse2;
721 c->idct_dc[2] = ff_hevc_idct16x16_dc_8_sse2;
722 c->idct_dc[3] = ff_hevc_idct32x32_dc_8_sse2;
724 c->transform_add[1] = ff_hevc_transform_add8_8_sse2;
725 c->transform_add[2] = ff_hevc_transform_add16_8_sse2;
726 c->transform_add[3] = ff_hevc_transform_add32_8_sse2;
728 if (EXTERNAL_SSSE3(cpu_flags)) {
730 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_ssse3;
731 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_ssse3;
733 SAO_EDGE_INIT(8, ssse3);
735 if (EXTERNAL_SSE4(cpu_flags) && ARCH_X86_64) {
737 EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 8, sse4);
738 EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 8, sse4);
739 EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 8, sse4);
740 EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 8, sse4);
742 QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 8, sse4);
743 QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 8, sse4);
744 QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 8, sse4);
745 QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 8, sse4);
747 if (EXTERNAL_AVX(cpu_flags)) {
748 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_avx;
749 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_avx;
751 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_avx;
752 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_avx;
754 SAO_BAND_INIT(8, avx);
756 c->transform_add[1] = ff_hevc_transform_add8_8_avx;
757 c->transform_add[2] = ff_hevc_transform_add16_8_avx;
758 c->transform_add[3] = ff_hevc_transform_add32_8_avx;
760 if (EXTERNAL_AVX2(cpu_flags)) {
761 c->idct_dc[2] = ff_hevc_idct16x16_dc_8_avx2;
762 c->idct_dc[3] = ff_hevc_idct32x32_dc_8_avx2;
764 c->put_hevc_epel[7][0][0] = ff_hevc_put_hevc_pel_pixels32_8_avx2;
765 c->put_hevc_epel[8][0][0] = ff_hevc_put_hevc_pel_pixels48_8_avx2;
766 c->put_hevc_epel[9][0][0] = ff_hevc_put_hevc_pel_pixels64_8_avx2;
768 c->put_hevc_qpel[7][0][0] = ff_hevc_put_hevc_pel_pixels32_8_avx2;
769 c->put_hevc_qpel[8][0][0] = ff_hevc_put_hevc_pel_pixels48_8_avx2;
770 c->put_hevc_qpel[9][0][0] = ff_hevc_put_hevc_pel_pixels64_8_avx2;
772 c->put_hevc_epel_uni[7][0][0] = ff_hevc_put_hevc_uni_pel_pixels32_8_avx2;
773 c->put_hevc_epel_uni[8][0][0] = ff_hevc_put_hevc_uni_pel_pixels48_8_avx2;
774 c->put_hevc_epel_uni[9][0][0] = ff_hevc_put_hevc_uni_pel_pixels64_8_avx2;
776 c->put_hevc_qpel_uni[7][0][0] = ff_hevc_put_hevc_uni_pel_pixels32_8_avx2;
777 c->put_hevc_qpel_uni[8][0][0] = ff_hevc_put_hevc_uni_pel_pixels48_8_avx2;
778 c->put_hevc_qpel_uni[9][0][0] = ff_hevc_put_hevc_uni_pel_pixels64_8_avx2;
780 c->put_hevc_qpel_bi[7][0][0] = ff_hevc_put_hevc_bi_pel_pixels32_8_avx2;
781 c->put_hevc_qpel_bi[8][0][0] = ff_hevc_put_hevc_bi_pel_pixels48_8_avx2;
782 c->put_hevc_qpel_bi[9][0][0] = ff_hevc_put_hevc_bi_pel_pixels64_8_avx2;
784 c->put_hevc_epel_bi[7][0][0] = ff_hevc_put_hevc_bi_pel_pixels32_8_avx2;
785 c->put_hevc_epel_bi[8][0][0] = ff_hevc_put_hevc_bi_pel_pixels48_8_avx2;
786 c->put_hevc_epel_bi[9][0][0] = ff_hevc_put_hevc_bi_pel_pixels64_8_avx2;
788 c->put_hevc_epel[7][0][1] = ff_hevc_put_hevc_epel_h32_8_avx2;
789 c->put_hevc_epel[8][0][1] = ff_hevc_put_hevc_epel_h48_8_avx2;
790 c->put_hevc_epel[9][0][1] = ff_hevc_put_hevc_epel_h64_8_avx2;
792 c->put_hevc_epel_uni[7][0][1] = ff_hevc_put_hevc_uni_epel_h32_8_avx2;
793 c->put_hevc_epel_uni[8][0][1] = ff_hevc_put_hevc_uni_epel_h48_8_avx2;
794 c->put_hevc_epel_uni[9][0][1] = ff_hevc_put_hevc_uni_epel_h64_8_avx2;
796 c->put_hevc_epel_bi[7][0][1] = ff_hevc_put_hevc_bi_epel_h32_8_avx2;
797 c->put_hevc_epel_bi[8][0][1] = ff_hevc_put_hevc_bi_epel_h48_8_avx2;
798 c->put_hevc_epel_bi[9][0][1] = ff_hevc_put_hevc_bi_epel_h64_8_avx2;
800 c->put_hevc_epel[7][1][0] = ff_hevc_put_hevc_epel_v32_8_avx2;
801 c->put_hevc_epel[8][1][0] = ff_hevc_put_hevc_epel_v48_8_avx2;
802 c->put_hevc_epel[9][1][0] = ff_hevc_put_hevc_epel_v64_8_avx2;
804 c->put_hevc_epel_uni[7][1][0] = ff_hevc_put_hevc_uni_epel_v32_8_avx2;
805 c->put_hevc_epel_uni[8][1][0] = ff_hevc_put_hevc_uni_epel_v48_8_avx2;
806 c->put_hevc_epel_uni[9][1][0] = ff_hevc_put_hevc_uni_epel_v64_8_avx2;
808 c->put_hevc_epel_bi[7][1][0] = ff_hevc_put_hevc_bi_epel_v32_8_avx2;
809 c->put_hevc_epel_bi[8][1][0] = ff_hevc_put_hevc_bi_epel_v48_8_avx2;
810 c->put_hevc_epel_bi[9][1][0] = ff_hevc_put_hevc_bi_epel_v64_8_avx2;
812 c->put_hevc_epel[7][1][1] = ff_hevc_put_hevc_epel_hv32_8_avx2;
813 c->put_hevc_epel[8][1][1] = ff_hevc_put_hevc_epel_hv48_8_avx2;
814 c->put_hevc_epel[9][1][1] = ff_hevc_put_hevc_epel_hv64_8_avx2;
816 c->put_hevc_epel_uni[7][1][1] = ff_hevc_put_hevc_uni_epel_hv32_8_avx2;
817 c->put_hevc_epel_uni[8][1][1] = ff_hevc_put_hevc_uni_epel_hv48_8_avx2;
818 c->put_hevc_epel_uni[9][1][1] = ff_hevc_put_hevc_uni_epel_hv64_8_avx2;
820 c->put_hevc_epel_bi[7][1][1] = ff_hevc_put_hevc_bi_epel_hv32_8_avx2;
821 c->put_hevc_epel_bi[8][1][1] = ff_hevc_put_hevc_bi_epel_hv48_8_avx2;
822 c->put_hevc_epel_bi[9][1][1] = ff_hevc_put_hevc_bi_epel_hv64_8_avx2;
824 c->put_hevc_qpel[7][0][1] = ff_hevc_put_hevc_qpel_h32_8_avx2;
825 c->put_hevc_qpel[8][0][1] = ff_hevc_put_hevc_qpel_h48_8_avx2;
826 c->put_hevc_qpel[9][0][1] = ff_hevc_put_hevc_qpel_h64_8_avx2;
828 c->put_hevc_qpel[7][1][0] = ff_hevc_put_hevc_qpel_v32_8_avx2;
829 c->put_hevc_qpel[8][1][0] = ff_hevc_put_hevc_qpel_v48_8_avx2;
830 c->put_hevc_qpel[9][1][0] = ff_hevc_put_hevc_qpel_v64_8_avx2;
832 c->put_hevc_qpel_uni[7][0][1] = ff_hevc_put_hevc_uni_qpel_h32_8_avx2;
833 c->put_hevc_qpel_uni[8][0][1] = ff_hevc_put_hevc_uni_qpel_h48_8_avx2;
834 c->put_hevc_qpel_uni[9][0][1] = ff_hevc_put_hevc_uni_qpel_h64_8_avx2;
836 c->put_hevc_qpel_uni[7][1][0] = ff_hevc_put_hevc_uni_qpel_v32_8_avx2;
837 c->put_hevc_qpel_uni[8][1][0] = ff_hevc_put_hevc_uni_qpel_v48_8_avx2;
838 c->put_hevc_qpel_uni[9][1][0] = ff_hevc_put_hevc_uni_qpel_v64_8_avx2;
840 c->put_hevc_qpel_bi[7][0][1] = ff_hevc_put_hevc_bi_qpel_h32_8_avx2;
841 c->put_hevc_qpel_bi[8][0][1] = ff_hevc_put_hevc_bi_qpel_h48_8_avx2;
842 c->put_hevc_qpel_bi[9][0][1] = ff_hevc_put_hevc_bi_qpel_h64_8_avx2;
844 c->put_hevc_qpel_bi[7][1][0] = ff_hevc_put_hevc_bi_qpel_v32_8_avx2;
845 c->put_hevc_qpel_bi[8][1][0] = ff_hevc_put_hevc_bi_qpel_v48_8_avx2;
846 c->put_hevc_qpel_bi[9][1][0] = ff_hevc_put_hevc_bi_qpel_v64_8_avx2;
848 SAO_BAND_INIT(8, avx2);
850 c->sao_edge_filter[2] = ff_hevc_sao_edge_filter_32_8_avx2;
851 c->sao_edge_filter[3] = ff_hevc_sao_edge_filter_48_8_avx2;
852 c->sao_edge_filter[4] = ff_hevc_sao_edge_filter_64_8_avx2;
854 c->transform_add[3] = ff_hevc_transform_add32_8_avx2;
856 } else if (bit_depth == 10) {
857 if (EXTERNAL_MMXEXT(cpu_flags)) {
858 c->transform_add[0] = ff_hevc_transform_add4_10_mmxext;
859 c->idct_dc[0] = ff_hevc_idct4x4_dc_10_mmxext;
860 c->idct_dc[1] = ff_hevc_idct8x8_dc_10_mmxext;
862 if (EXTERNAL_SSE2(cpu_flags)) {
863 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_sse2;
864 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_10_sse2;
866 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_sse2;
867 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_sse2;
869 SAO_BAND_INIT(10, sse2);
870 SAO_EDGE_INIT(10, sse2);
872 c->idct_dc[1] = ff_hevc_idct8x8_dc_10_sse2;
873 c->idct_dc[2] = ff_hevc_idct16x16_dc_10_sse2;
874 c->idct_dc[3] = ff_hevc_idct32x32_dc_10_sse2;
876 c->transform_add[1] = ff_hevc_transform_add8_10_sse2;
877 c->transform_add[2] = ff_hevc_transform_add16_10_sse2;
878 c->transform_add[3] = ff_hevc_transform_add32_10_sse2;
880 if (EXTERNAL_SSSE3(cpu_flags) && ARCH_X86_64) {
881 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_ssse3;
882 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_ssse3;
884 if (EXTERNAL_SSE4(cpu_flags) && ARCH_X86_64) {
885 EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 10, sse4);
886 EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 10, sse4);
887 EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 10, sse4);
888 EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 10, sse4);
890 QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 10, sse4);
891 QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 10, sse4);
892 QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 10, sse4);
893 QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 10, sse4);
895 if (EXTERNAL_AVX(cpu_flags)) {
896 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_avx;
897 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_10_avx;
899 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_avx;
900 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_avx;
902 SAO_BAND_INIT(10, avx);
904 if (EXTERNAL_AVX2(cpu_flags)) {
906 c->idct_dc[2] = ff_hevc_idct16x16_dc_10_avx2;
907 c->idct_dc[3] = ff_hevc_idct32x32_dc_10_avx2;
909 c->put_hevc_epel[5][0][0] = ff_hevc_put_hevc_pel_pixels16_10_avx2;
910 c->put_hevc_epel[6][0][0] = ff_hevc_put_hevc_pel_pixels24_10_avx2;
911 c->put_hevc_epel[7][0][0] = ff_hevc_put_hevc_pel_pixels32_10_avx2;
912 c->put_hevc_epel[8][0][0] = ff_hevc_put_hevc_pel_pixels48_10_avx2;
913 c->put_hevc_epel[9][0][0] = ff_hevc_put_hevc_pel_pixels64_10_avx2;
915 c->put_hevc_qpel[5][0][0] = ff_hevc_put_hevc_pel_pixels16_10_avx2;
916 c->put_hevc_qpel[6][0][0] = ff_hevc_put_hevc_pel_pixels24_10_avx2;
917 c->put_hevc_qpel[7][0][0] = ff_hevc_put_hevc_pel_pixels32_10_avx2;
918 c->put_hevc_qpel[8][0][0] = ff_hevc_put_hevc_pel_pixels48_10_avx2;
919 c->put_hevc_qpel[9][0][0] = ff_hevc_put_hevc_pel_pixels64_10_avx2;
921 c->put_hevc_epel_uni[5][0][0] = ff_hevc_put_hevc_uni_pel_pixels32_8_avx2;
922 c->put_hevc_epel_uni[6][0][0] = ff_hevc_put_hevc_uni_pel_pixels48_8_avx2;
923 c->put_hevc_epel_uni[7][0][0] = ff_hevc_put_hevc_uni_pel_pixels64_8_avx2;
924 c->put_hevc_epel_uni[8][0][0] = ff_hevc_put_hevc_uni_pel_pixels96_8_avx2;
925 c->put_hevc_epel_uni[9][0][0] = ff_hevc_put_hevc_uni_pel_pixels128_8_avx2;
927 c->put_hevc_qpel_uni[5][0][0] = ff_hevc_put_hevc_uni_pel_pixels32_8_avx2;
928 c->put_hevc_qpel_uni[6][0][0] = ff_hevc_put_hevc_uni_pel_pixels48_8_avx2;
929 c->put_hevc_qpel_uni[7][0][0] = ff_hevc_put_hevc_uni_pel_pixels64_8_avx2;
930 c->put_hevc_qpel_uni[8][0][0] = ff_hevc_put_hevc_uni_pel_pixels96_8_avx2;
931 c->put_hevc_qpel_uni[9][0][0] = ff_hevc_put_hevc_uni_pel_pixels128_8_avx2;
933 c->put_hevc_epel_bi[5][0][0] = ff_hevc_put_hevc_bi_pel_pixels16_10_avx2;
934 c->put_hevc_epel_bi[6][0][0] = ff_hevc_put_hevc_bi_pel_pixels24_10_avx2;
935 c->put_hevc_epel_bi[7][0][0] = ff_hevc_put_hevc_bi_pel_pixels32_10_avx2;
936 c->put_hevc_epel_bi[8][0][0] = ff_hevc_put_hevc_bi_pel_pixels48_10_avx2;
937 c->put_hevc_epel_bi[9][0][0] = ff_hevc_put_hevc_bi_pel_pixels64_10_avx2;
938 c->put_hevc_qpel_bi[5][0][0] = ff_hevc_put_hevc_bi_pel_pixels16_10_avx2;
939 c->put_hevc_qpel_bi[6][0][0] = ff_hevc_put_hevc_bi_pel_pixels24_10_avx2;
940 c->put_hevc_qpel_bi[7][0][0] = ff_hevc_put_hevc_bi_pel_pixels32_10_avx2;
941 c->put_hevc_qpel_bi[8][0][0] = ff_hevc_put_hevc_bi_pel_pixels48_10_avx2;
942 c->put_hevc_qpel_bi[9][0][0] = ff_hevc_put_hevc_bi_pel_pixels64_10_avx2;
944 c->put_hevc_epel[5][0][1] = ff_hevc_put_hevc_epel_h16_10_avx2;
945 c->put_hevc_epel[6][0][1] = ff_hevc_put_hevc_epel_h24_10_avx2;
946 c->put_hevc_epel[7][0][1] = ff_hevc_put_hevc_epel_h32_10_avx2;
947 c->put_hevc_epel[8][0][1] = ff_hevc_put_hevc_epel_h48_10_avx2;
948 c->put_hevc_epel[9][0][1] = ff_hevc_put_hevc_epel_h64_10_avx2;
950 c->put_hevc_epel_uni[5][0][1] = ff_hevc_put_hevc_uni_epel_h16_10_avx2;
951 c->put_hevc_epel_uni[6][0][1] = ff_hevc_put_hevc_uni_epel_h24_10_avx2;
952 c->put_hevc_epel_uni[7][0][1] = ff_hevc_put_hevc_uni_epel_h32_10_avx2;
953 c->put_hevc_epel_uni[8][0][1] = ff_hevc_put_hevc_uni_epel_h48_10_avx2;
954 c->put_hevc_epel_uni[9][0][1] = ff_hevc_put_hevc_uni_epel_h64_10_avx2;
956 c->put_hevc_epel_bi[5][0][1] = ff_hevc_put_hevc_bi_epel_h16_10_avx2;
957 c->put_hevc_epel_bi[6][0][1] = ff_hevc_put_hevc_bi_epel_h24_10_avx2;
958 c->put_hevc_epel_bi[7][0][1] = ff_hevc_put_hevc_bi_epel_h32_10_avx2;
959 c->put_hevc_epel_bi[8][0][1] = ff_hevc_put_hevc_bi_epel_h48_10_avx2;
960 c->put_hevc_epel_bi[9][0][1] = ff_hevc_put_hevc_bi_epel_h64_10_avx2;
962 c->put_hevc_epel[5][1][0] = ff_hevc_put_hevc_epel_v16_10_avx2;
963 c->put_hevc_epel[6][1][0] = ff_hevc_put_hevc_epel_v24_10_avx2;
964 c->put_hevc_epel[7][1][0] = ff_hevc_put_hevc_epel_v32_10_avx2;
965 c->put_hevc_epel[8][1][0] = ff_hevc_put_hevc_epel_v48_10_avx2;
966 c->put_hevc_epel[9][1][0] = ff_hevc_put_hevc_epel_v64_10_avx2;
968 c->put_hevc_epel_uni[5][1][0] = ff_hevc_put_hevc_uni_epel_v16_10_avx2;
969 c->put_hevc_epel_uni[6][1][0] = ff_hevc_put_hevc_uni_epel_v24_10_avx2;
970 c->put_hevc_epel_uni[7][1][0] = ff_hevc_put_hevc_uni_epel_v32_10_avx2;
971 c->put_hevc_epel_uni[8][1][0] = ff_hevc_put_hevc_uni_epel_v48_10_avx2;
972 c->put_hevc_epel_uni[9][1][0] = ff_hevc_put_hevc_uni_epel_v64_10_avx2;
974 c->put_hevc_epel_bi[5][1][0] = ff_hevc_put_hevc_bi_epel_v16_10_avx2;
975 c->put_hevc_epel_bi[6][1][0] = ff_hevc_put_hevc_bi_epel_v24_10_avx2;
976 c->put_hevc_epel_bi[7][1][0] = ff_hevc_put_hevc_bi_epel_v32_10_avx2;
977 c->put_hevc_epel_bi[8][1][0] = ff_hevc_put_hevc_bi_epel_v48_10_avx2;
978 c->put_hevc_epel_bi[9][1][0] = ff_hevc_put_hevc_bi_epel_v64_10_avx2;
980 c->put_hevc_epel[5][1][1] = ff_hevc_put_hevc_epel_hv16_10_avx2;
981 c->put_hevc_epel[6][1][1] = ff_hevc_put_hevc_epel_hv24_10_avx2;
982 c->put_hevc_epel[7][1][1] = ff_hevc_put_hevc_epel_hv32_10_avx2;
983 c->put_hevc_epel[8][1][1] = ff_hevc_put_hevc_epel_hv48_10_avx2;
984 c->put_hevc_epel[9][1][1] = ff_hevc_put_hevc_epel_hv64_10_avx2;
986 c->put_hevc_epel_uni[5][1][1] = ff_hevc_put_hevc_uni_epel_hv16_10_avx2;
987 c->put_hevc_epel_uni[6][1][1] = ff_hevc_put_hevc_uni_epel_hv24_10_avx2;
988 c->put_hevc_epel_uni[7][1][1] = ff_hevc_put_hevc_uni_epel_hv32_10_avx2;
989 c->put_hevc_epel_uni[8][1][1] = ff_hevc_put_hevc_uni_epel_hv48_10_avx2;
990 c->put_hevc_epel_uni[9][1][1] = ff_hevc_put_hevc_uni_epel_hv64_10_avx2;
992 c->put_hevc_epel_bi[5][1][1] = ff_hevc_put_hevc_bi_epel_hv16_10_avx2;
993 c->put_hevc_epel_bi[6][1][1] = ff_hevc_put_hevc_bi_epel_hv24_10_avx2;
994 c->put_hevc_epel_bi[7][1][1] = ff_hevc_put_hevc_bi_epel_hv32_10_avx2;
995 c->put_hevc_epel_bi[8][1][1] = ff_hevc_put_hevc_bi_epel_hv48_10_avx2;
996 c->put_hevc_epel_bi[9][1][1] = ff_hevc_put_hevc_bi_epel_hv64_10_avx2;
998 c->put_hevc_qpel[5][0][1] = ff_hevc_put_hevc_qpel_h16_10_avx2;
999 c->put_hevc_qpel[6][0][1] = ff_hevc_put_hevc_qpel_h24_10_avx2;
1000 c->put_hevc_qpel[7][0][1] = ff_hevc_put_hevc_qpel_h32_10_avx2;
1001 c->put_hevc_qpel[8][0][1] = ff_hevc_put_hevc_qpel_h48_10_avx2;
1002 c->put_hevc_qpel[9][0][1] = ff_hevc_put_hevc_qpel_h64_10_avx2;
1004 c->put_hevc_qpel_uni[5][0][1] = ff_hevc_put_hevc_uni_qpel_h16_10_avx2;
1005 c->put_hevc_qpel_uni[6][0][1] = ff_hevc_put_hevc_uni_qpel_h24_10_avx2;
1006 c->put_hevc_qpel_uni[7][0][1] = ff_hevc_put_hevc_uni_qpel_h32_10_avx2;
1007 c->put_hevc_qpel_uni[8][0][1] = ff_hevc_put_hevc_uni_qpel_h48_10_avx2;
1008 c->put_hevc_qpel_uni[9][0][1] = ff_hevc_put_hevc_uni_qpel_h64_10_avx2;
1010 c->put_hevc_qpel_bi[5][0][1] = ff_hevc_put_hevc_bi_qpel_h16_10_avx2;
1011 c->put_hevc_qpel_bi[6][0][1] = ff_hevc_put_hevc_bi_qpel_h24_10_avx2;
1012 c->put_hevc_qpel_bi[7][0][1] = ff_hevc_put_hevc_bi_qpel_h32_10_avx2;
1013 c->put_hevc_qpel_bi[8][0][1] = ff_hevc_put_hevc_bi_qpel_h48_10_avx2;
1014 c->put_hevc_qpel_bi[9][0][1] = ff_hevc_put_hevc_bi_qpel_h64_10_avx2;
1016 c->put_hevc_qpel[5][1][0] = ff_hevc_put_hevc_qpel_v16_10_avx2;
1017 c->put_hevc_qpel[6][1][0] = ff_hevc_put_hevc_qpel_v24_10_avx2;
1018 c->put_hevc_qpel[7][1][0] = ff_hevc_put_hevc_qpel_v32_10_avx2;
1019 c->put_hevc_qpel[8][1][0] = ff_hevc_put_hevc_qpel_v48_10_avx2;
1020 c->put_hevc_qpel[9][1][0] = ff_hevc_put_hevc_qpel_v64_10_avx2;
1022 c->put_hevc_qpel_uni[5][1][0] = ff_hevc_put_hevc_uni_qpel_v16_10_avx2;
1023 c->put_hevc_qpel_uni[6][1][0] = ff_hevc_put_hevc_uni_qpel_v24_10_avx2;
1024 c->put_hevc_qpel_uni[7][1][0] = ff_hevc_put_hevc_uni_qpel_v32_10_avx2;
1025 c->put_hevc_qpel_uni[8][1][0] = ff_hevc_put_hevc_uni_qpel_v48_10_avx2;
1026 c->put_hevc_qpel_uni[9][1][0] = ff_hevc_put_hevc_uni_qpel_v64_10_avx2;
1028 c->put_hevc_qpel_bi[5][1][0] = ff_hevc_put_hevc_bi_qpel_v16_10_avx2;
1029 c->put_hevc_qpel_bi[6][1][0] = ff_hevc_put_hevc_bi_qpel_v24_10_avx2;
1030 c->put_hevc_qpel_bi[7][1][0] = ff_hevc_put_hevc_bi_qpel_v32_10_avx2;
1031 c->put_hevc_qpel_bi[8][1][0] = ff_hevc_put_hevc_bi_qpel_v48_10_avx2;
1032 c->put_hevc_qpel_bi[9][1][0] = ff_hevc_put_hevc_bi_qpel_v64_10_avx2;
1034 c->put_hevc_qpel[5][1][1] = ff_hevc_put_hevc_qpel_hv16_10_avx2;
1035 c->put_hevc_qpel[6][1][1] = ff_hevc_put_hevc_qpel_hv24_10_avx2;
1036 c->put_hevc_qpel[7][1][1] = ff_hevc_put_hevc_qpel_hv32_10_avx2;
1037 c->put_hevc_qpel[8][1][1] = ff_hevc_put_hevc_qpel_hv48_10_avx2;
1038 c->put_hevc_qpel[9][1][1] = ff_hevc_put_hevc_qpel_hv64_10_avx2;
1040 c->put_hevc_qpel_uni[5][1][1] = ff_hevc_put_hevc_uni_qpel_hv16_10_avx2;
1041 c->put_hevc_qpel_uni[6][1][1] = ff_hevc_put_hevc_uni_qpel_hv24_10_avx2;
1042 c->put_hevc_qpel_uni[7][1][1] = ff_hevc_put_hevc_uni_qpel_hv32_10_avx2;
1043 c->put_hevc_qpel_uni[8][1][1] = ff_hevc_put_hevc_uni_qpel_hv48_10_avx2;
1044 c->put_hevc_qpel_uni[9][1][1] = ff_hevc_put_hevc_uni_qpel_hv64_10_avx2;
1046 c->put_hevc_qpel_bi[5][1][1] = ff_hevc_put_hevc_bi_qpel_hv16_10_avx2;
1047 c->put_hevc_qpel_bi[6][1][1] = ff_hevc_put_hevc_bi_qpel_hv24_10_avx2;
1048 c->put_hevc_qpel_bi[7][1][1] = ff_hevc_put_hevc_bi_qpel_hv32_10_avx2;
1049 c->put_hevc_qpel_bi[8][1][1] = ff_hevc_put_hevc_bi_qpel_hv48_10_avx2;
1050 c->put_hevc_qpel_bi[9][1][1] = ff_hevc_put_hevc_bi_qpel_hv64_10_avx2;
1052 SAO_BAND_INIT(10, avx2);
1053 c->sao_edge_filter[2] = ff_hevc_sao_edge_filter_32_10_avx2;
1054 c->sao_edge_filter[3] = ff_hevc_sao_edge_filter_48_10_avx2;
1055 c->sao_edge_filter[4] = ff_hevc_sao_edge_filter_64_10_avx2;
1057 c->transform_add[2] = ff_hevc_transform_add16_10_avx2;
1058 c->transform_add[3] = ff_hevc_transform_add32_10_avx2;
1061 } else if (bit_depth == 12) {
1062 if (EXTERNAL_MMXEXT(cpu_flags)) {
1063 c->idct_dc[0] = ff_hevc_idct4x4_dc_12_mmxext;
1064 c->idct_dc[1] = ff_hevc_idct8x8_dc_12_mmxext;
1066 if (EXTERNAL_SSE2(cpu_flags)) {
1067 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_12_sse2;
1068 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_12_sse2;
1070 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_sse2;
1071 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_sse2;
1073 SAO_BAND_INIT(12, sse2);
1074 SAO_EDGE_INIT(12, sse2);
1076 c->idct_dc[1] = ff_hevc_idct8x8_dc_12_sse2;
1077 c->idct_dc[2] = ff_hevc_idct16x16_dc_12_sse2;
1078 c->idct_dc[3] = ff_hevc_idct32x32_dc_12_sse2;
1080 if (EXTERNAL_SSSE3(cpu_flags) && ARCH_X86_64) {
1081 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_ssse3;
1082 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_ssse3;
1084 if (EXTERNAL_SSE4(cpu_flags) && ARCH_X86_64) {
1085 EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 12, sse4);
1086 EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 12, sse4);
1087 EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 12, sse4);
1088 EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 12, sse4);
1090 QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 12, sse4);
1091 QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 12, sse4);
1092 QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 12, sse4);
1093 QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 12, sse4);
1095 if (EXTERNAL_AVX(cpu_flags)) {
1096 c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_12_avx;
1097 c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_12_avx;
1099 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_avx;
1100 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_avx;
1102 SAO_BAND_INIT(12, avx);
1104 if (EXTERNAL_AVX2(cpu_flags)) {
1105 c->idct_dc[2] = ff_hevc_idct16x16_dc_12_avx2;
1106 c->idct_dc[3] = ff_hevc_idct32x32_dc_12_avx2;
1108 SAO_BAND_INIT(12, avx2);
1109 c->sao_edge_filter[2] = ff_hevc_sao_edge_filter_32_12_avx2;
1110 c->sao_edge_filter[3] = ff_hevc_sao_edge_filter_48_12_avx2;
1111 c->sao_edge_filter[4] = ff_hevc_sao_edge_filter_64_12_avx2;