vp8/encoder/x86/variance_x86.h

   1 /*
   2  *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
   3  *
   4  *  Use of this source code is governed by a BSD-style license
   5  *  that can be found in the LICENSE file in the root of the source
   6  *  tree. An additional intellectual property rights grant can be found
   7  *  in the file PATENTS.  All contributing project authors may
   8  *  be found in the AUTHORS file in the root of the source tree.
   9  */
  10
  11
  12 #ifndef VARIANCE_X86_H
  13 #define VARIANCE_X86_H
  14
  15
  16 /* Note:
  17  *
  18  * This platform is commonly built for runtime CPU detection. If you modify
  19  * any of the function mappings present in this file, be sure to also update
  20  * them in the function pointer initialization code
  21  */
  22 #if HAVE_MMX
  23 extern prototype_sad(vp8_sad4x4_mmx);
  24 extern prototype_sad(vp8_sad8x8_mmx);
  25 extern prototype_sad(vp8_sad8x16_mmx);
  26 extern prototype_sad(vp8_sad16x8_mmx);
  27 extern prototype_sad(vp8_sad16x16_mmx);
  28 extern prototype_variance(vp8_variance4x4_mmx);
  29 extern prototype_variance(vp8_variance8x8_mmx);
  30 extern prototype_variance(vp8_variance8x16_mmx);
  31 extern prototype_variance(vp8_variance16x8_mmx);
  32 extern prototype_variance(vp8_variance16x16_mmx);
  33 extern prototype_subpixvariance(vp8_sub_pixel_variance4x4_mmx);
  34 extern prototype_subpixvariance(vp8_sub_pixel_variance8x8_mmx);
  35 extern prototype_subpixvariance(vp8_sub_pixel_variance8x16_mmx);
  36 extern prototype_subpixvariance(vp8_sub_pixel_variance16x8_mmx);
  37 extern prototype_subpixvariance(vp8_sub_pixel_variance16x16_mmx);
  38 extern prototype_variance(vp8_variance_halfpixvar16x16_h_mmx);
  39 extern prototype_variance(vp8_variance_halfpixvar16x16_v_mmx);
  40 extern prototype_variance(vp8_variance_halfpixvar16x16_hv_mmx);
  41 extern prototype_subpixvariance(vp8_sub_pixel_mse16x16_mmx);
  42 extern prototype_getmbss(vp8_get_mb_ss_mmx);
  43 extern prototype_variance(vp8_mse16x16_mmx);
  44 extern prototype_variance2(vp8_get8x8var_mmx);
  45 extern prototype_get16x16prederror(vp8_get4x4sse_cs_mmx);
  46
  47 #if !CONFIG_RUNTIME_CPU_DETECT
  48 #undef  vp8_variance_sad4x4
  49 #define vp8_variance_sad4x4 vp8_sad4x4_mmx
  50
  51 #undef  vp8_variance_sad8x8
  52 #define vp8_variance_sad8x8 vp8_sad8x8_mmx
  53
  54 #undef  vp8_variance_sad8x16
  55 #define vp8_variance_sad8x16 vp8_sad8x16_mmx
  56
  57 #undef  vp8_variance_sad16x8
  58 #define vp8_variance_sad16x8 vp8_sad16x8_mmx
  59
  60 #undef  vp8_variance_sad16x16
  61 #define vp8_variance_sad16x16 vp8_sad16x16_mmx
  62
  63 #undef  vp8_variance_var4x4
  64 #define vp8_variance_var4x4 vp8_variance4x4_mmx
  65
  66 #undef  vp8_variance_var8x8
  67 #define vp8_variance_var8x8 vp8_variance8x8_mmx
  68
  69 #undef  vp8_variance_var8x16
  70 #define vp8_variance_var8x16 vp8_variance8x16_mmx
  71
  72 #undef  vp8_variance_var16x8
  73 #define vp8_variance_var16x8 vp8_variance16x8_mmx
  74
  75 #undef  vp8_variance_var16x16
  76 #define vp8_variance_var16x16 vp8_variance16x16_mmx
  77
  78 #undef  vp8_variance_subpixvar4x4
  79 #define vp8_variance_subpixvar4x4 vp8_sub_pixel_variance4x4_mmx
  80
  81 #undef  vp8_variance_subpixvar8x8
  82 #define vp8_variance_subpixvar8x8 vp8_sub_pixel_variance8x8_mmx
  83
  84 #undef  vp8_variance_subpixvar8x16
  85 #define vp8_variance_subpixvar8x16 vp8_sub_pixel_variance8x16_mmx
  86
  87 #undef  vp8_variance_subpixvar16x8
  88 #define vp8_variance_subpixvar16x8 vp8_sub_pixel_variance16x8_mmx
  89
  90 #undef  vp8_variance_subpixvar16x16
  91 #define vp8_variance_subpixvar16x16 vp8_sub_pixel_variance16x16_mmx
  92
  93 #undef  vp8_variance_halfpixvar16x16_h
  94 #define vp8_variance_halfpixvar16x16_h vp8_variance_halfpixvar16x16_h_mmx
  95
  96 #undef  vp8_variance_halfpixvar16x16_v
  97 #define vp8_variance_halfpixvar16x16_v vp8_variance_halfpixvar16x16_v_mmx
  98
  99 #undef  vp8_variance_halfpixvar16x16_hv
 100 #define vp8_variance_halfpixvar16x16_hv vp8_variance_halfpixvar16x16_hv_mmx
 101
 102 #undef  vp8_variance_subpixmse16x16
 103 #define vp8_variance_subpixmse16x16 vp8_sub_pixel_mse16x16_mmx
 104
 105 #undef  vp8_variance_getmbss
 106 #define vp8_variance_getmbss vp8_get_mb_ss_mmx
 107
 108 #undef  vp8_variance_mse16x16
 109 #define vp8_variance_mse16x16 vp8_mse16x16_mmx
 110
 111 #undef  vp8_variance_get4x4sse_cs
 112 #define vp8_variance_get4x4sse_cs vp8_get4x4sse_cs_mmx
 113
 114 #endif
 115 #endif
 116
 117
 118 #if HAVE_SSE2
 119 extern prototype_sad(vp8_sad4x4_wmt);
 120 extern prototype_sad(vp8_sad8x8_wmt);
 121 extern prototype_sad(vp8_sad8x16_wmt);
 122 extern prototype_sad(vp8_sad16x8_wmt);
 123 extern prototype_sad(vp8_sad16x16_wmt);
 124 extern prototype_sad(vp8_copy32xn_sse2);
 125 extern prototype_variance(vp8_variance4x4_wmt);
 126 extern prototype_variance(vp8_variance8x8_wmt);
 127 extern prototype_variance(vp8_variance8x16_wmt);
 128 extern prototype_variance(vp8_variance16x8_wmt);
 129 extern prototype_variance(vp8_variance16x16_wmt);
 130 extern prototype_subpixvariance(vp8_sub_pixel_variance4x4_wmt);
 131 extern prototype_subpixvariance(vp8_sub_pixel_variance8x8_wmt);
 132 extern prototype_subpixvariance(vp8_sub_pixel_variance8x16_wmt);
 133 extern prototype_subpixvariance(vp8_sub_pixel_variance16x8_wmt);
 134 extern prototype_subpixvariance(vp8_sub_pixel_variance16x16_wmt);
 135 extern prototype_variance(vp8_variance_halfpixvar16x16_h_wmt);
 136 extern prototype_variance(vp8_variance_halfpixvar16x16_v_wmt);
 137 extern prototype_variance(vp8_variance_halfpixvar16x16_hv_wmt);
 138 extern prototype_subpixvariance(vp8_sub_pixel_mse16x16_wmt);
 139 extern prototype_getmbss(vp8_get_mb_ss_sse2);
 140 extern prototype_variance(vp8_mse16x16_wmt);
 141 extern prototype_variance2(vp8_get8x8var_sse2);
 142 extern prototype_variance2(vp8_get16x16var_sse2);
 143 extern prototype_ssimpf(vp8_ssim_parms_8x8_sse2)
 144 extern prototype_ssimpf(vp8_ssim_parms_16x16_sse2)
 145
 146 #if !CONFIG_RUNTIME_CPU_DETECT
 147 #undef  vp8_variance_sad4x4
 148 #define vp8_variance_sad4x4 vp8_sad4x4_wmt
 149
 150 #undef  vp8_variance_sad8x8
 151 #define vp8_variance_sad8x8 vp8_sad8x8_wmt
 152
 153 #undef  vp8_variance_sad8x16
 154 #define vp8_variance_sad8x16 vp8_sad8x16_wmt
 155
 156 #undef  vp8_variance_sad16x8
 157 #define vp8_variance_sad16x8 vp8_sad16x8_wmt
 158
 159 #undef  vp8_variance_sad16x16
 160 #define vp8_variance_sad16x16 vp8_sad16x16_wmt
 161
 162 #undef  vp8_variance_copy32xn
 163 #define vp8_variance_copy32xn vp8_copy32xn_sse2
 164
 165 #undef  vp8_variance_var4x4
 166 #define vp8_variance_var4x4 vp8_variance4x4_wmt
 167
 168 #undef  vp8_variance_var8x8
 169 #define vp8_variance_var8x8 vp8_variance8x8_wmt
 170
 171 #undef  vp8_variance_var8x16
 172 #define vp8_variance_var8x16 vp8_variance8x16_wmt
 173
 174 #undef  vp8_variance_var16x8
 175 #define vp8_variance_var16x8 vp8_variance16x8_wmt
 176
 177 #undef  vp8_variance_var16x16
 178 #define vp8_variance_var16x16 vp8_variance16x16_wmt
 179
 180 #undef  vp8_variance_subpixvar4x4
 181 #define vp8_variance_subpixvar4x4 vp8_sub_pixel_variance4x4_wmt
 182
 183 #undef  vp8_variance_subpixvar8x8
 184 #define vp8_variance_subpixvar8x8 vp8_sub_pixel_variance8x8_wmt
 185
 186 #undef  vp8_variance_subpixvar8x16
 187 #define vp8_variance_subpixvar8x16 vp8_sub_pixel_variance8x16_wmt
 188
 189 #undef  vp8_variance_subpixvar16x8
 190 #define vp8_variance_subpixvar16x8 vp8_sub_pixel_variance16x8_wmt
 191
 192 #undef  vp8_variance_subpixvar16x16
 193 #define vp8_variance_subpixvar16x16 vp8_sub_pixel_variance16x16_wmt
 194
 195 #undef  vp8_variance_halfpixvar16x16_h
 196 #define vp8_variance_halfpixvar16x16_h vp8_variance_halfpixvar16x16_h_wmt
 197
 198 #undef  vp8_variance_halfpixvar16x16_v
 199 #define vp8_variance_halfpixvar16x16_v vp8_variance_halfpixvar16x16_v_wmt
 200
 201 #undef  vp8_variance_halfpixvar16x16_hv
 202 #define vp8_variance_halfpixvar16x16_hv vp8_variance_halfpixvar16x16_hv_wmt
 203
 204 #undef  vp8_variance_subpixmse16x16
 205 #define vp8_variance_subpixmse16x16 vp8_sub_pixel_mse16x16_wmt
 206
 207 #undef  vp8_variance_getmbss
 208 #define vp8_variance_getmbss vp8_get_mb_ss_sse2
 209
 210 #undef  vp8_variance_mse16x16
 211 #define vp8_variance_mse16x16 vp8_mse16x16_wmt
 212
 213 #if ARCH_X86_64
 214 #undef  vp8_ssimpf_8x8
 215 #define vp8_ssimpf_8x8 vp8_ssim_parms_8x8_sse2
 216
 217 #undef  vp8_ssimpf_16x16
 218 #define vp8_ssimpf_16x16 vp8_ssim_parms_16x16_sse2
 219 #endif
 220
 221 #endif
 222 #endif
 223
 224
 225 #if HAVE_SSE3
 226 extern prototype_sad(vp8_sad16x16_sse3);
 227 extern prototype_sad(vp8_sad16x8_sse3);
 228 extern prototype_sad_multi_same_address(vp8_sad16x16x3_sse3);
 229 extern prototype_sad_multi_same_address(vp8_sad16x8x3_sse3);
 230 extern prototype_sad_multi_same_address(vp8_sad8x16x3_sse3);
 231 extern prototype_sad_multi_same_address(vp8_sad8x8x3_sse3);
 232 extern prototype_sad_multi_same_address(vp8_sad4x4x3_sse3);
 233
 234 extern prototype_sad_multi_dif_address(vp8_sad16x16x4d_sse3);
 235 extern prototype_sad_multi_dif_address(vp8_sad16x8x4d_sse3);
 236 extern prototype_sad_multi_dif_address(vp8_sad8x16x4d_sse3);
 237 extern prototype_sad_multi_dif_address(vp8_sad8x8x4d_sse3);
 238 extern prototype_sad_multi_dif_address(vp8_sad4x4x4d_sse3);
 239 extern prototype_sad(vp8_copy32xn_sse3);
 240
 241 #if !CONFIG_RUNTIME_CPU_DETECT
 242
 243 #undef  vp8_variance_sad16x16
 244 #define vp8_variance_sad16x16 vp8_sad16x16_sse3
 245
 246 #undef  vp8_variance_sad16x16x3
 247 #define vp8_variance_sad16x16x3 vp8_sad16x16x3_sse3
 248
 249 #undef  vp8_variance_sad16x8x3
 250 #define vp8_variance_sad16x8x3 vp8_sad16x8x3_sse3
 251
 252 #undef  vp8_variance_sad8x16x3
 253 #define vp8_variance_sad8x16x3 vp8_sad8x16x3_sse3
 254
 255 #undef  vp8_variance_sad8x8x3
 256 #define vp8_variance_sad8x8x3 vp8_sad8x8x3_sse3
 257
 258 #undef  vp8_variance_sad4x4x3
 259 #define vp8_variance_sad4x4x3 vp8_sad4x4x3_sse3
 260
 261 #undef  vp8_variance_sad16x16x4d
 262 #define vp8_variance_sad16x16x4d vp8_sad16x16x4d_sse3
 263
 264 #undef  vp8_variance_sad16x8x4d
 265 #define vp8_variance_sad16x8x4d vp8_sad16x8x4d_sse3
 266
 267 #undef  vp8_variance_sad8x16x4d
 268 #define vp8_variance_sad8x16x4d vp8_sad8x16x4d_sse3
 269
 270 #undef  vp8_variance_sad8x8x4d
 271 #define vp8_variance_sad8x8x4d vp8_sad8x8x4d_sse3
 272
 273 #undef  vp8_variance_sad4x4x4d
 274 #define vp8_variance_sad4x4x4d vp8_sad4x4x4d_sse3
 275
 276 #undef  vp8_variance_copy32xn
 277 #define vp8_variance_copy32xn vp8_copy32xn_sse3
 278
 279 #endif
 280 #endif
 281
 282
 283 #if HAVE_SSSE3
 284 extern prototype_sad_multi_same_address(vp8_sad16x16x3_ssse3);
 285 extern prototype_sad_multi_same_address(vp8_sad16x8x3_ssse3);
 286 extern prototype_subpixvariance(vp8_sub_pixel_variance16x8_ssse3);
 287 extern prototype_subpixvariance(vp8_sub_pixel_variance16x16_ssse3);
 288
 289 #if !CONFIG_RUNTIME_CPU_DETECT
 290 #undef  vp8_variance_sad16x16x3
 291 #define vp8_variance_sad16x16x3 vp8_sad16x16x3_ssse3
 292
 293 #undef  vp8_variance_sad16x8x3
 294 #define vp8_variance_sad16x8x3 vp8_sad16x8x3_ssse3
 295
 296 #undef  vp8_variance_subpixvar16x8
 297 #define vp8_variance_subpixvar16x8 vp8_sub_pixel_variance16x8_ssse3
 298
 299 #undef  vp8_variance_subpixvar16x16
 300 #define vp8_variance_subpixvar16x16 vp8_sub_pixel_variance16x16_ssse3
 301
 302 #endif
 303 #endif
 304
 305
 306 #if HAVE_SSE4_1
 307 extern prototype_sad_multi_same_address_1(vp8_sad16x16x8_sse4);
 308 extern prototype_sad_multi_same_address_1(vp8_sad16x8x8_sse4);
 309 extern prototype_sad_multi_same_address_1(vp8_sad8x16x8_sse4);
 310 extern prototype_sad_multi_same_address_1(vp8_sad8x8x8_sse4);
 311 extern prototype_sad_multi_same_address_1(vp8_sad4x4x8_sse4);
 312
 313 #if !CONFIG_RUNTIME_CPU_DETECT
 314 #undef  vp8_variance_sad16x16x8
 315 #define vp8_variance_sad16x16x8 vp8_sad16x16x8_sse4
 316
 317 #undef  vp8_variance_sad16x8x8
 318 #define vp8_variance_sad16x8x8 vp8_sad16x8x8_sse4
 319
 320 #undef  vp8_variance_sad8x16x8
 321 #define vp8_variance_sad8x16x8 vp8_sad8x16x8_sse4
 322
 323 #undef  vp8_variance_sad8x8x8
 324 #define vp8_variance_sad8x8x8 vp8_sad8x8x8_sse4
 325
 326 #undef  vp8_variance_sad4x4x8
 327 #define vp8_variance_sad4x4x8 vp8_sad4x4x8_sse4
 328
 329 #endif
 330 #endif
 331
 332 #endif