ffmpeg/libavcodec/x86/hevcdsp_init.c

   1 /*
   2  * Copyright (c) 2013 Seppo Tomperi
   3  * Copyright (c) 2013 - 2014 Pierre-Edouard Lepere
   4  *
   5  *
   6  * This file is part of FFmpeg.
   7  *
   8  * FFmpeg is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * FFmpeg is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with FFmpeg; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 #include "config.h"
  24 #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"
  30
  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);
  33
  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);
  36
  37 #define LFC_FUNCS(type, depth, opt) \
  38     LFC_FUNC(h, depth, opt)  \
  39     LFC_FUNC(v, depth, opt)
  40
  41 #define LFL_FUNCS(type, depth, opt) \
  42     LFL_FUNC(h, depth, opt)  \
  43     LFL_FUNC(v, depth, opt)
  44
  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)
  60
  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)
  65
  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);
  73
  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)                            \
  78 {                                                                                                               \
  79     int i;                                                                                                      \
  80     uint8_t *src;                                                                                               \
  81     int16_t *dst;                                                                                               \
  82     for (i = 0; i < W; i += step) {                                                                             \
  83         src  = _src + (i * ((bitd + 7) / 8));                                                                   \
  84         dst = _dst + i;                                                                                         \
  85         ff_hevc_put_hevc_##name##step##_##bitd##_##opt(dst, src, _srcstride, height, mx, my, width);            \
  86     }                                                                                                           \
  87 }
  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)                        \
  92 {                                                                                                               \
  93     int i;                                                                                                      \
  94     uint8_t *src;                                                                                               \
  95     uint8_t *dst;                                                                                               \
  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);                               \
 101     }                                                                                                           \
 102 }
 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)             \
 107 {                                                                                                               \
 108     int i;                                                                                                      \
 109     uint8_t  *src;                                                                                              \
 110     uint8_t  *dst;                                                                                              \
 111     int16_t  *src2;                                                                                             \
 112     for (i = 0; i < W ; i += step) {                                                                            \
 113         src  = _src + (i * ((bitd + 7) / 8));                                                                   \
 114         dst  = _dst + (i * ((bitd + 7) / 8));                                                                   \
 115         src2 = _src2 + i;                                                                                       \
 116         ff_hevc_put_hevc_bi_##name##step##_##bitd##_##opt(dst, dststride, src, _srcstride, src2,                \
 117                                                           height, mx, my, width);                               \
 118     }                                                                                                           \
 119 }
 120
 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)
 125
 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)                           \
 130 {                                                                                                               \
 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);                         \
 134 }
 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)                       \
 139 {                                                                                                               \
 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);                                 \
 144 }
 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)            \
 149 {                                                                                                               \
 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);                    \
 154 }
 155
 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)
 160
 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)
 165
 166 #if ARCH_X86_64 && HAVE_SSE4_EXTERNAL
 167
 168 mc_rep_funcs(pel_pixels, 8, 16, 64, sse4);
 169 mc_rep_funcs(pel_pixels, 8, 16, 48, sse4);
 170 mc_rep_funcs(pel_pixels, 8, 16, 32, sse4);
 171 mc_rep_funcs(pel_pixels, 8,  8, 24, sse4);
 172 mc_rep_funcs(pel_pixels,10,  8, 64, sse4);
 173 mc_rep_funcs(pel_pixels,10,  8, 48, sse4);
 174 mc_rep_funcs(pel_pixels,10,  8, 32, sse4);
 175 mc_rep_funcs(pel_pixels,10,  8, 24, sse4);
 176 mc_rep_funcs(pel_pixels,10,  8, 16, sse4);
 177 mc_rep_funcs(pel_pixels,10,  4, 12, sse4);
 178 mc_rep_funcs(pel_pixels,12,  8, 64, sse4);
 179 mc_rep_funcs(pel_pixels,12,  8, 48, sse4);
 180 mc_rep_funcs(pel_pixels,12,  8, 32, sse4);
 181 mc_rep_funcs(pel_pixels,12,  8, 24, sse4);
 182 mc_rep_funcs(pel_pixels,12,  8, 16, sse4);
 183 mc_rep_funcs(pel_pixels,12,  4, 12, sse4);
 184
 185 mc_rep_funcs(epel_h, 8, 16, 64, sse4);
 186 mc_rep_funcs(epel_h, 8, 16, 48, sse4);
 187 mc_rep_funcs(epel_h, 8, 16, 32, sse4);
 188 mc_rep_funcs(epel_h, 8,  8, 24, sse4);
 189 mc_rep_funcs(epel_h,10,  8, 64, sse4);
 190 mc_rep_funcs(epel_h,10,  8, 48, sse4);
 191 mc_rep_funcs(epel_h,10,  8, 32, sse4);
 192 mc_rep_funcs(epel_h,10,  8, 24, sse4);
 193 mc_rep_funcs(epel_h,10,  8, 16, sse4);
 194 mc_rep_funcs(epel_h,10,  4, 12, sse4);
 195 mc_rep_funcs(epel_h,12,  8, 64, sse4);
 196 mc_rep_funcs(epel_h,12,  8, 48, sse4);
 197 mc_rep_funcs(epel_h,12,  8, 32, sse4);
 198 mc_rep_funcs(epel_h,12,  8, 24, sse4);
 199 mc_rep_funcs(epel_h,12,  8, 16, sse4);
 200 mc_rep_funcs(epel_h,12,  4, 12, sse4);
 201 mc_rep_funcs(epel_v, 8, 16, 64, sse4);
 202 mc_rep_funcs(epel_v, 8, 16, 48, sse4);
 203 mc_rep_funcs(epel_v, 8, 16, 32, sse4);
 204 mc_rep_funcs(epel_v, 8,  8, 24, sse4);
 205 mc_rep_funcs(epel_v,10,  8, 64, sse4);
 206 mc_rep_funcs(epel_v,10,  8, 48, sse4);
 207 mc_rep_funcs(epel_v,10,  8, 32, sse4);
 208 mc_rep_funcs(epel_v,10,  8, 24, sse4);
 209 mc_rep_funcs(epel_v,10,  8, 16, sse4);
 210 mc_rep_funcs(epel_v,10,  4, 12, sse4);
 211 mc_rep_funcs(epel_v,12,  8, 64, sse4);
 212 mc_rep_funcs(epel_v,12,  8, 48, sse4);
 213 mc_rep_funcs(epel_v,12,  8, 32, sse4);
 214 mc_rep_funcs(epel_v,12,  8, 24, sse4);
 215 mc_rep_funcs(epel_v,12,  8, 16, sse4);
 216 mc_rep_funcs(epel_v,12,  4, 12, sse4);
 217 mc_rep_funcs(epel_hv, 8,  8, 64, sse4);
 218 mc_rep_funcs(epel_hv, 8,  8, 48, sse4);
 219 mc_rep_funcs(epel_hv, 8,  8, 32, sse4);
 220 mc_rep_funcs(epel_hv, 8,  8, 24, sse4);
 221 mc_rep_funcs(epel_hv, 8,  8, 16, sse4);
 222 mc_rep_funcs2(epel_hv,8,  8,  4, 12, sse4);
 223 mc_rep_funcs(epel_hv,10,  8, 64, sse4);
 224 mc_rep_funcs(epel_hv,10,  8, 48, sse4);
 225 mc_rep_funcs(epel_hv,10,  8, 32, sse4);
 226 mc_rep_funcs(epel_hv,10,  8, 24, sse4);
 227 mc_rep_funcs(epel_hv,10,  8, 16, sse4);
 228 mc_rep_funcs(epel_hv,10,  4, 12, sse4);
 229 mc_rep_funcs(epel_hv,12,  8, 64, sse4);
 230 mc_rep_funcs(epel_hv,12,  8, 48, sse4);
 231 mc_rep_funcs(epel_hv,12,  8, 32, sse4);
 232 mc_rep_funcs(epel_hv,12,  8, 24, sse4);
 233 mc_rep_funcs(epel_hv,12,  8, 16, sse4);
 234 mc_rep_funcs(epel_hv,12,  4, 12, sse4);
 235
 236 mc_rep_funcs(qpel_h, 8, 16, 64, sse4);
 237 mc_rep_funcs(qpel_h, 8, 16, 48, sse4);
 238 mc_rep_funcs(qpel_h, 8, 16, 32, sse4);
 239 mc_rep_funcs(qpel_h, 8,  8, 24, sse4);
 240 mc_rep_funcs(qpel_h,10,  8, 64, sse4);
 241 mc_rep_funcs(qpel_h,10,  8, 48, sse4);
 242 mc_rep_funcs(qpel_h,10,  8, 32, sse4);
 243 mc_rep_funcs(qpel_h,10,  8, 24, sse4);
 244 mc_rep_funcs(qpel_h,10,  8, 16, sse4);
 245 mc_rep_funcs(qpel_h,10,  4, 12, sse4);
 246 mc_rep_funcs(qpel_h,12,  8, 64, sse4);
 247 mc_rep_funcs(qpel_h,12,  8, 48, sse4);
 248 mc_rep_funcs(qpel_h,12,  8, 32, sse4);
 249 mc_rep_funcs(qpel_h,12,  8, 24, sse4);
 250 mc_rep_funcs(qpel_h,12,  8, 16, sse4);
 251 mc_rep_funcs(qpel_h,12,  4, 12, sse4);
 252 mc_rep_funcs(qpel_v, 8, 16, 64, sse4);
 253 mc_rep_funcs(qpel_v, 8, 16, 48, sse4);
 254 mc_rep_funcs(qpel_v, 8, 16, 32, sse4);
 255 mc_rep_funcs(qpel_v, 8,  8, 24, sse4);
 256 mc_rep_funcs(qpel_v,10,  8, 64, sse4);
 257 mc_rep_funcs(qpel_v,10,  8, 48, sse4);
 258 mc_rep_funcs(qpel_v,10,  8, 32, sse4);
 259 mc_rep_funcs(qpel_v,10,  8, 24, sse4);
 260 mc_rep_funcs(qpel_v,10,  8, 16, sse4);
 261 mc_rep_funcs(qpel_v,10,  4, 12, sse4);
 262 mc_rep_funcs(qpel_v,12,  8, 64, sse4);
 263 mc_rep_funcs(qpel_v,12,  8, 48, sse4);
 264 mc_rep_funcs(qpel_v,12,  8, 32, sse4);
 265 mc_rep_funcs(qpel_v,12,  8, 24, sse4);
 266 mc_rep_funcs(qpel_v,12,  8, 16, sse4);
 267 mc_rep_funcs(qpel_v,12,  4, 12, sse4);
 268 mc_rep_funcs(qpel_hv, 8,  8, 64, sse4);
 269 mc_rep_funcs(qpel_hv, 8,  8, 48, sse4);
 270 mc_rep_funcs(qpel_hv, 8,  8, 32, sse4);
 271 mc_rep_funcs(qpel_hv, 8,  8, 24, sse4);
 272 mc_rep_funcs(qpel_hv, 8,  8, 16, sse4);
 273 mc_rep_funcs2(qpel_hv,8,  8,  4, 12, sse4);
 274 mc_rep_funcs(qpel_hv,10,  8, 64, sse4);
 275 mc_rep_funcs(qpel_hv,10,  8, 48, sse4);
 276 mc_rep_funcs(qpel_hv,10,  8, 32, sse4);
 277 mc_rep_funcs(qpel_hv,10,  8, 24, sse4);
 278 mc_rep_funcs(qpel_hv,10,  8, 16, sse4);
 279 mc_rep_funcs(qpel_hv,10,  4, 12, sse4);
 280 mc_rep_funcs(qpel_hv,12,  8, 64, sse4);
 281 mc_rep_funcs(qpel_hv,12,  8, 48, sse4);
 282 mc_rep_funcs(qpel_hv,12,  8, 32, sse4);
 283 mc_rep_funcs(qpel_hv,12,  8, 24, sse4);
 284 mc_rep_funcs(qpel_hv,12,  8, 16, sse4);
 285 mc_rep_funcs(qpel_hv,12,  4, 12, sse4);
 286
 287 #define mc_rep_uni_w(bitd, step, W, opt) \
 288 void ff_hevc_put_hevc_uni_w##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, int16_t *_src, ptrdiff_t _srcstride,\
 289                                                int height, int denom,  int _wx, int _ox)                                \
 290 {                                                                                                                       \
 291     int i;                                                                                                              \
 292     int16_t *src;                                                                                                       \
 293     uint8_t *dst;                                                                                                       \
 294     for (i = 0; i < W; i += step) {                                                                                     \
 295         src= _src + i;                                                                                                  \
 296         dst= _dst + (i * ((bitd + 7) / 8));                                                                             \
 297         ff_hevc_put_hevc_uni_w##step##_##bitd##_##opt(dst, dststride, src, _srcstride,                                  \
 298                                                      height, denom, _wx, _ox);                                          \
 299     }                                                                                                                   \
 300 }
 301
 302 mc_rep_uni_w(8, 6, 12, sse4);
 303 mc_rep_uni_w(8, 8, 16, sse4);
 304 mc_rep_uni_w(8, 8, 24, sse4);
 305 mc_rep_uni_w(8, 8, 32, sse4);
 306 mc_rep_uni_w(8, 8, 48, sse4);
 307 mc_rep_uni_w(8, 8, 64, sse4);
 308
 309 mc_rep_uni_w(10, 6, 12, sse4);
 310 mc_rep_uni_w(10, 8, 16, sse4);
 311 mc_rep_uni_w(10, 8, 24, sse4);
 312 mc_rep_uni_w(10, 8, 32, sse4);
 313 mc_rep_uni_w(10, 8, 48, sse4);
 314 mc_rep_uni_w(10, 8, 64, sse4);
 315
 316 mc_rep_uni_w(12, 6, 12, sse4);
 317 mc_rep_uni_w(12, 8, 16, sse4);
 318 mc_rep_uni_w(12, 8, 24, sse4);
 319 mc_rep_uni_w(12, 8, 32, sse4);
 320 mc_rep_uni_w(12, 8, 48, sse4);
 321 mc_rep_uni_w(12, 8, 64, sse4);
 322
 323 #define mc_rep_bi_w(bitd, step, W, opt) \
 324 void ff_hevc_put_hevc_bi_w##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, int16_t *_src, ptrdiff_t _srcstride, \
 325                                               int16_t *_src2, int height,                                               \
 326                                               int denom,  int _wx0,  int _wx1, int _ox0, int _ox1)                      \
 327 {                                                                                                                       \
 328     int i;                                                                                                              \
 329     int16_t *src;                                                                                                       \
 330     int16_t *src2;                                                                                                      \
 331     uint8_t *dst;                                                                                                       \
 332     for (i = 0; i < W; i += step) {                                                                                     \
 333         src  = _src  + i;                                                                                               \
 334         src2 = _src2 + i;                                                                                               \
 335         dst  = _dst  + (i * ((bitd + 7) / 8));                                                                          \
 336         ff_hevc_put_hevc_bi_w##step##_##bitd##_##opt(dst, dststride, src, _srcstride, src2,                             \
 337                                                     height, denom, _wx0, _wx1, _ox0, _ox1);                             \
 338     }                                                                                                                   \
 339 }
 340
 341 mc_rep_bi_w(8, 6, 12, sse4);
 342 mc_rep_bi_w(8, 8, 16, sse4);
 343 mc_rep_bi_w(8, 8, 24, sse4);
 344 mc_rep_bi_w(8, 8, 32, sse4);
 345 mc_rep_bi_w(8, 8, 48, sse4);
 346 mc_rep_bi_w(8, 8, 64, sse4);
 347
 348 mc_rep_bi_w(10, 6, 12, sse4);
 349 mc_rep_bi_w(10, 8, 16, sse4);
 350 mc_rep_bi_w(10, 8, 24, sse4);
 351 mc_rep_bi_w(10, 8, 32, sse4);
 352 mc_rep_bi_w(10, 8, 48, sse4);
 353 mc_rep_bi_w(10, 8, 64, sse4);
 354
 355 mc_rep_bi_w(12, 6, 12, sse4);
 356 mc_rep_bi_w(12, 8, 16, sse4);
 357 mc_rep_bi_w(12, 8, 24, sse4);
 358 mc_rep_bi_w(12, 8, 32, sse4);
 359 mc_rep_bi_w(12, 8, 48, sse4);
 360 mc_rep_bi_w(12, 8, 64, sse4);
 361
 362 #define mc_uni_w_func(name, bitd, W, opt) \
 363 void ff_hevc_put_hevc_uni_w_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t _dststride,         \
 364                                                       uint8_t *_src, ptrdiff_t _srcstride,          \
 365                                                       int height, int denom,                        \
 366                                                       int _wx, int _ox,                             \
 367                                                       intptr_t mx, intptr_t my, int width)          \
 368 {                                                                                                   \
 369     LOCAL_ALIGNED_16(int16_t, temp, [71 * MAX_PB_SIZE]);                                            \
 370     ff_hevc_put_hevc_##name##W##_##bitd##_##opt(temp, _src, _srcstride, height, mx, my, width);     \
 371     ff_hevc_put_hevc_uni_w##W##_##bitd##_##opt(_dst, _dststride, temp, MAX_PB_SIZE, height, denom, _wx, _ox);\
 372 }
 373
 374 #define mc_uni_w_funcs(name, bitd, opt)       \
 375         mc_uni_w_func(name, bitd, 4, opt);    \
 376         mc_uni_w_func(name, bitd, 8, opt);    \
 377         mc_uni_w_func(name, bitd, 12, opt);   \
 378         mc_uni_w_func(name, bitd, 16, opt);   \
 379         mc_uni_w_func(name, bitd, 24, opt);   \
 380         mc_uni_w_func(name, bitd, 32, opt);   \
 381         mc_uni_w_func(name, bitd, 48, opt);   \
 382         mc_uni_w_func(name, bitd, 64, opt)
 383
 384 mc_uni_w_funcs(pel_pixels, 8, sse4);
 385 mc_uni_w_func(pel_pixels, 8, 6, sse4);
 386 mc_uni_w_funcs(epel_h, 8, sse4);
 387 mc_uni_w_func(epel_h, 8, 6, sse4);
 388 mc_uni_w_funcs(epel_v, 8, sse4);
 389 mc_uni_w_func(epel_v, 8, 6, sse4);
 390 mc_uni_w_funcs(epel_hv, 8, sse4);
 391 mc_uni_w_func(epel_hv, 8, 6, sse4);
 392 mc_uni_w_funcs(qpel_h, 8, sse4);
 393 mc_uni_w_funcs(qpel_v, 8, sse4);
 394 mc_uni_w_funcs(qpel_hv, 8, sse4);
 395
 396 mc_uni_w_funcs(pel_pixels, 10, sse4);
 397 mc_uni_w_func(pel_pixels, 10, 6, sse4);
 398 mc_uni_w_funcs(epel_h, 10, sse4);
 399 mc_uni_w_func(epel_h, 10, 6, sse4);
 400 mc_uni_w_funcs(epel_v, 10, sse4);
 401 mc_uni_w_func(epel_v, 10, 6, sse4);
 402 mc_uni_w_funcs(epel_hv, 10, sse4);
 403 mc_uni_w_func(epel_hv, 10, 6, sse4);
 404 mc_uni_w_funcs(qpel_h, 10, sse4);
 405 mc_uni_w_funcs(qpel_v, 10, sse4);
 406 mc_uni_w_funcs(qpel_hv, 10, sse4);
 407
 408 mc_uni_w_funcs(pel_pixels, 12, sse4);
 409 mc_uni_w_func(pel_pixels, 12, 6, sse4);
 410 mc_uni_w_funcs(epel_h, 12, sse4);
 411 mc_uni_w_func(epel_h, 12, 6, sse4);
 412 mc_uni_w_funcs(epel_v, 12, sse4);
 413 mc_uni_w_func(epel_v, 12, 6, sse4);
 414 mc_uni_w_funcs(epel_hv, 12, sse4);
 415 mc_uni_w_func(epel_hv, 12, 6, sse4);
 416 mc_uni_w_funcs(qpel_h, 12, sse4);
 417 mc_uni_w_funcs(qpel_v, 12, sse4);
 418 mc_uni_w_funcs(qpel_hv, 12, sse4);
 419
 420 #define mc_bi_w_func(name, bitd, W, opt) \
 421 void ff_hevc_put_hevc_bi_w_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t _dststride,           \
 422                                                      uint8_t *_src, ptrdiff_t _srcstride,            \
 423                                                      int16_t *_src2,                                 \
 424                                                      int height, int denom,                          \
 425                                                      int _wx0, int _wx1, int _ox0, int _ox1,         \
 426                                                      intptr_t mx, intptr_t my, int width)            \
 427 {                                                                                                    \
 428     LOCAL_ALIGNED_16(int16_t, temp, [71 * MAX_PB_SIZE]);                                             \
 429     ff_hevc_put_hevc_##name##W##_##bitd##_##opt(temp, _src, _srcstride, height, mx, my, width);      \
 430     ff_hevc_put_hevc_bi_w##W##_##bitd##_##opt(_dst, _dststride, temp, MAX_PB_SIZE, _src2,            \
 431                                              height, denom, _wx0, _wx1, _ox0, _ox1);                 \
 432 }
 433
 434 #define mc_bi_w_funcs(name, bitd, opt)       \
 435         mc_bi_w_func(name, bitd, 4, opt);    \
 436         mc_bi_w_func(name, bitd, 8, opt);    \
 437         mc_bi_w_func(name, bitd, 12, opt);   \
 438         mc_bi_w_func(name, bitd, 16, opt);   \
 439         mc_bi_w_func(name, bitd, 24, opt);   \
 440         mc_bi_w_func(name, bitd, 32, opt);   \
 441         mc_bi_w_func(name, bitd, 48, opt);   \
 442         mc_bi_w_func(name, bitd, 64, opt)
 443
 444 mc_bi_w_funcs(pel_pixels, 8, sse4);
 445 mc_bi_w_func(pel_pixels, 8, 6, sse4);
 446 mc_bi_w_funcs(epel_h, 8, sse4);
 447 mc_bi_w_func(epel_h, 8, 6, sse4);
 448 mc_bi_w_funcs(epel_v, 8, sse4);
 449 mc_bi_w_func(epel_v, 8, 6, sse4);
 450 mc_bi_w_funcs(epel_hv, 8, sse4);
 451 mc_bi_w_func(epel_hv, 8, 6, sse4);
 452 mc_bi_w_funcs(qpel_h, 8, sse4);
 453 mc_bi_w_funcs(qpel_v, 8, sse4);
 454 mc_bi_w_funcs(qpel_hv, 8, sse4);
 455
 456 mc_bi_w_funcs(pel_pixels, 10, sse4);
 457 mc_bi_w_func(pel_pixels, 10, 6, sse4);
 458 mc_bi_w_funcs(epel_h, 10, sse4);
 459 mc_bi_w_func(epel_h, 10, 6, sse4);
 460 mc_bi_w_funcs(epel_v, 10, sse4);
 461 mc_bi_w_func(epel_v, 10, 6, sse4);
 462 mc_bi_w_funcs(epel_hv, 10, sse4);
 463 mc_bi_w_func(epel_hv, 10, 6, sse4);
 464 mc_bi_w_funcs(qpel_h, 10, sse4);
 465 mc_bi_w_funcs(qpel_v, 10, sse4);
 466 mc_bi_w_funcs(qpel_hv, 10, sse4);
 467
 468 mc_bi_w_funcs(pel_pixels, 12, sse4);
 469 mc_bi_w_func(pel_pixels, 12, 6, sse4);
 470 mc_bi_w_funcs(epel_h, 12, sse4);
 471 mc_bi_w_func(epel_h, 12, 6, sse4);
 472 mc_bi_w_funcs(epel_v, 12, sse4);
 473 mc_bi_w_func(epel_v, 12, 6, sse4);
 474 mc_bi_w_funcs(epel_hv, 12, sse4);
 475 mc_bi_w_func(epel_hv, 12, 6, sse4);
 476 mc_bi_w_funcs(qpel_h, 12, sse4);
 477 mc_bi_w_funcs(qpel_v, 12, sse4);
 478 mc_bi_w_funcs(qpel_hv, 12, sse4);
 479 #endif //ARCH_X86_64 && HAVE_SSE4_EXTERNAL
 480
 481
 482 #define EPEL_LINKS(pointer, my, mx, fname, bitd, opt )           \
 483         PEL_LINK(pointer, 1, my , mx , fname##4 ,  bitd, opt ); \
 484         PEL_LINK(pointer, 2, my , mx , fname##6 ,  bitd, opt ); \
 485         PEL_LINK(pointer, 3, my , mx , fname##8 ,  bitd, opt ); \
 486         PEL_LINK(pointer, 4, my , mx , fname##12,  bitd, opt ); \
 487         PEL_LINK(pointer, 5, my , mx , fname##16,  bitd, opt ); \
 488         PEL_LINK(pointer, 6, my , mx , fname##24,  bitd, opt ); \
 489         PEL_LINK(pointer, 7, my , mx , fname##32,  bitd, opt ); \
 490         PEL_LINK(pointer, 8, my , mx , fname##48,  bitd, opt ); \
 491         PEL_LINK(pointer, 9, my , mx , fname##64,  bitd, opt )
 492 #define QPEL_LINKS(pointer, my, mx, fname, bitd, opt)           \
 493         PEL_LINK(pointer, 1, my , mx , fname##4 ,  bitd, opt ); \
 494         PEL_LINK(pointer, 3, my , mx , fname##8 ,  bitd, opt ); \
 495         PEL_LINK(pointer, 4, my , mx , fname##12,  bitd, opt ); \
 496         PEL_LINK(pointer, 5, my , mx , fname##16,  bitd, opt ); \
 497         PEL_LINK(pointer, 6, my , mx , fname##24,  bitd, opt ); \
 498         PEL_LINK(pointer, 7, my , mx , fname##32,  bitd, opt ); \
 499         PEL_LINK(pointer, 8, my , mx , fname##48,  bitd, opt ); \
 500         PEL_LINK(pointer, 9, my , mx , fname##64,  bitd, opt )
 501
 502
 503 void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth)
 504 {
 505     int cpu_flags = av_get_cpu_flags();
 506
 507     if (bit_depth == 8) {
 508         if (EXTERNAL_MMXEXT(cpu_flags)) {
 509             c->idct_dc[0] = ff_hevc_idct4x4_dc_8_mmxext;
 510             c->idct_dc[1] = ff_hevc_idct8x8_dc_8_mmxext;
 511             c->transform_add[0]    =  ff_hevc_transform_add4_8_mmxext;
 512         }
 513         if (EXTERNAL_SSE2(cpu_flags)) {
 514             c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_sse2;
 515             c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_sse2;
 516             if (ARCH_X86_64) {
 517                 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_sse2;
 518                 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_sse2;
 519             }
 520             c->idct_dc[1] = ff_hevc_idct8x8_dc_8_sse2;
 521             c->idct_dc[2] = ff_hevc_idct16x16_dc_8_sse2;
 522             c->idct_dc[3] = ff_hevc_idct32x32_dc_8_sse2;
 523
 524             c->transform_add[1]    = ff_hevc_transform_add8_8_sse2;
 525             c->transform_add[2]    = ff_hevc_transform_add16_8_sse2;
 526             c->transform_add[3]    = ff_hevc_transform_add32_8_sse2;
 527         }
 528         if (EXTERNAL_SSSE3(cpu_flags) && ARCH_X86_64) {
 529             c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_ssse3;
 530             c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_ssse3;
 531         }
 532         if (EXTERNAL_SSE4(cpu_flags) && ARCH_X86_64) {
 533
 534             EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels,  8, sse4);
 535             EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h,      8, sse4);
 536             EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v,      8, sse4);
 537             EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv,     8, sse4);
 538
 539             QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 8, sse4);
 540             QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h,     8, sse4);
 541             QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v,     8, sse4);
 542             QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv,    8, sse4);
 543         }
 544         if (EXTERNAL_AVX(cpu_flags)) {
 545             c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_avx;
 546             c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_avx;
 547             if (ARCH_X86_64) {
 548                 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_avx;
 549                 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_avx;
 550             }
 551             c->transform_add[1]    = ff_hevc_transform_add8_8_avx;
 552             c->transform_add[2]    = ff_hevc_transform_add16_8_avx;
 553             c->transform_add[3]    = ff_hevc_transform_add32_8_avx;
 554         }
 555         if (EXTERNAL_AVX2(cpu_flags)) {
 556             c->idct_dc[2] = ff_hevc_idct16x16_dc_8_avx2;
 557             c->idct_dc[3] = ff_hevc_idct32x32_dc_8_avx2;
 558
 559             c->transform_add[3]    = ff_hevc_transform_add32_8_avx2;
 560         }
 561     } else if (bit_depth == 10) {
 562         if (EXTERNAL_MMXEXT(cpu_flags)) {
 563             c->transform_add[0] = ff_hevc_transform_add4_10_mmxext;
 564             c->idct_dc[0] = ff_hevc_idct4x4_dc_10_mmxext;
 565             c->idct_dc[1] = ff_hevc_idct8x8_dc_10_mmxext;
 566         }
 567         if (EXTERNAL_SSE2(cpu_flags)) {
 568             c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_sse2;
 569             c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_10_sse2;
 570             if (ARCH_X86_64) {
 571                 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_sse2;
 572                 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_sse2;
 573             }
 574
 575             c->idct_dc[1] = ff_hevc_idct8x8_dc_10_sse2;
 576             c->idct_dc[2] = ff_hevc_idct16x16_dc_10_sse2;
 577             c->idct_dc[3] = ff_hevc_idct32x32_dc_10_sse2;
 578
 579             c->transform_add[1]    = ff_hevc_transform_add8_10_sse2;
 580             c->transform_add[2]    = ff_hevc_transform_add16_10_sse2;
 581             c->transform_add[3]    = ff_hevc_transform_add32_10_sse2;
 582         }
 583         if (EXTERNAL_SSSE3(cpu_flags) && ARCH_X86_64) {
 584             c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_ssse3;
 585             c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_ssse3;
 586         }
 587         if (EXTERNAL_SSE4(cpu_flags) && ARCH_X86_64) {
 588             EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 10, sse4);
 589             EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h,     10, sse4);
 590             EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v,     10, sse4);
 591             EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv,    10, sse4);
 592
 593             QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 10, sse4);
 594             QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h,     10, sse4);
 595             QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v,     10, sse4);
 596             QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv,    10, sse4);
 597         }
 598         if (EXTERNAL_AVX(cpu_flags)) {
 599             c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_avx;
 600             c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_10_avx;
 601             if (ARCH_X86_64) {
 602                 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_avx;
 603                 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_avx;
 604             }
 605         }
 606         if (EXTERNAL_AVX2(cpu_flags)) {
 607
 608             c->idct_dc[2] = ff_hevc_idct16x16_dc_10_avx2;
 609             c->idct_dc[3] = ff_hevc_idct32x32_dc_10_avx2;
 610
 611             c->transform_add[2] = ff_hevc_transform_add16_10_avx2;
 612             c->transform_add[3] = ff_hevc_transform_add32_10_avx2;
 613
 614         }
 615     } else if (bit_depth == 12) {
 616         if (EXTERNAL_MMXEXT(cpu_flags)) {
 617             c->idct_dc[0] = ff_hevc_idct4x4_dc_12_mmxext;
 618             c->idct_dc[1] = ff_hevc_idct8x8_dc_12_mmxext;
 619         }
 620         if (EXTERNAL_SSE2(cpu_flags)) {
 621             c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_12_sse2;
 622             c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_12_sse2;
 623             if (ARCH_X86_64) {
 624                 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_sse2;
 625                 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_sse2;
 626             }
 627
 628             c->idct_dc[1] = ff_hevc_idct8x8_dc_12_sse2;
 629             c->idct_dc[2] = ff_hevc_idct16x16_dc_12_sse2;
 630             c->idct_dc[3] = ff_hevc_idct32x32_dc_12_sse2;
 631         }
 632         if (EXTERNAL_SSSE3(cpu_flags) && ARCH_X86_64) {
 633             c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_ssse3;
 634             c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_ssse3;
 635         }
 636         if (EXTERNAL_SSE4(cpu_flags) && ARCH_X86_64) {
 637             EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 12, sse4);
 638             EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h,     12, sse4);
 639             EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v,     12, sse4);
 640             EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv,    12, sse4);
 641
 642             QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 12, sse4);
 643             QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h,     12, sse4);
 644             QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v,     12, sse4);
 645             QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv,    12, sse4);
 646         }
 647         if (EXTERNAL_AVX(cpu_flags)) {
 648             c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_12_avx;
 649             c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_12_avx;
 650             if (ARCH_X86_64) {
 651                 c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_avx;
 652                 c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_avx;
 653             }
 654         }
 655         if (EXTERNAL_AVX2(cpu_flags)) {
 656             c->idct_dc[2] = ff_hevc_idct16x16_dc_12_avx2;
 657             c->idct_dc[3] = ff_hevc_idct32x32_dc_12_avx2;
 658         }
 659     }
 660 }