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_sad(vp8_get16x16pred_error_mmx);
  45 extern prototype_variance2(vp8_get8x8var_mmx);
  46 extern prototype_variance2(vp8_get16x16var_mmx);
  47 extern prototype_sad(vp8_get4x4sse_cs_mmx);
  48
  49 #if !CONFIG_RUNTIME_CPU_DETECT
  50 #undef  vp8_variance_sad4x4
  51 #define vp8_variance_sad4x4 vp8_sad4x4_mmx
  52
  53 #undef  vp8_variance_sad8x8
  54 #define vp8_variance_sad8x8 vp8_sad8x8_mmx
  55
  56 #undef  vp8_variance_sad8x16
  57 #define vp8_variance_sad8x16 vp8_sad8x16_mmx
  58
  59 #undef  vp8_variance_sad16x8
  60 #define vp8_variance_sad16x8 vp8_sad16x8_mmx
  61
  62 #undef  vp8_variance_sad16x16
  63 #define vp8_variance_sad16x16 vp8_sad16x16_mmx
  64
  65 #undef  vp8_variance_var4x4
  66 #define vp8_variance_var4x4 vp8_variance4x4_mmx
  67
  68 #undef  vp8_variance_var8x8
  69 #define vp8_variance_var8x8 vp8_variance8x8_mmx
  70
  71 #undef  vp8_variance_var8x16
  72 #define vp8_variance_var8x16 vp8_variance8x16_mmx
  73
  74 #undef  vp8_variance_var16x8
  75 #define vp8_variance_var16x8 vp8_variance16x8_mmx
  76
  77 #undef  vp8_variance_var16x16
  78 #define vp8_variance_var16x16 vp8_variance16x16_mmx
  79
  80 #undef  vp8_variance_subpixvar4x4
  81 #define vp8_variance_subpixvar4x4 vp8_sub_pixel_variance4x4_mmx
  82
  83 #undef  vp8_variance_subpixvar8x8
  84 #define vp8_variance_subpixvar8x8 vp8_sub_pixel_variance8x8_mmx
  85
  86 #undef  vp8_variance_subpixvar8x16
  87 #define vp8_variance_subpixvar8x16 vp8_sub_pixel_variance8x16_mmx
  88
  89 #undef  vp8_variance_subpixvar16x8
  90 #define vp8_variance_subpixvar16x8 vp8_sub_pixel_variance16x8_mmx
  91
  92 #undef  vp8_variance_subpixvar16x16
  93 #define vp8_variance_subpixvar16x16 vp8_sub_pixel_variance16x16_mmx
  94
  95 #undef  vp8_variance_halfpixvar16x16_h
  96 #define vp8_variance_halfpixvar16x16_h vp8_variance_halfpixvar16x16_h_mmx
  97
  98 #undef  vp8_variance_halfpixvar16x16_v
  99 #define vp8_variance_halfpixvar16x16_v vp8_variance_halfpixvar16x16_v_mmx
 100
 101 #undef  vp8_variance_halfpixvar16x16_hv
 102 #define vp8_variance_halfpixvar16x16_hv vp8_variance_halfpixvar16x16_hv_mmx
 103
 104 #undef  vp8_variance_subpixmse16x16
 105 #define vp8_variance_subpixmse16x16 vp8_sub_pixel_mse16x16_mmx
 106
 107 #undef  vp8_variance_getmbss
 108 #define vp8_variance_getmbss vp8_get_mb_ss_mmx
 109
 110 #undef  vp8_variance_mse16x16
 111 #define vp8_variance_mse16x16 vp8_mse16x16_mmx
 112
 113 #undef  vp8_variance_get16x16prederror
 114 #define vp8_variance_get16x16prederror vp8_get16x16pred_error_mmx
 115
 116 #undef  vp8_variance_get8x8var
 117 #define vp8_variance_get8x8var vp8_get8x8var_mmx
 118
 119 #undef  vp8_variance_get16x16var
 120 #define vp8_variance_get16x16var vp8_get16x16var_mmx
 121
 122 #undef  vp8_variance_get4x4sse_cs
 123 #define vp8_variance_get4x4sse_cs vp8_get4x4sse_cs_mmx
 124
 125 #endif
 126 #endif
 127
 128
 129 #if HAVE_SSE2
 130 extern prototype_sad(vp8_sad4x4_wmt);
 131 extern prototype_sad(vp8_sad8x8_wmt);
 132 extern prototype_sad(vp8_sad8x16_wmt);
 133 extern prototype_sad(vp8_sad16x8_wmt);
 134 extern prototype_sad(vp8_sad16x16_wmt);
 135 extern prototype_variance(vp8_variance4x4_wmt);
 136 extern prototype_variance(vp8_variance8x8_wmt);
 137 extern prototype_variance(vp8_variance8x16_wmt);
 138 extern prototype_variance(vp8_variance16x8_wmt);
 139 extern prototype_variance(vp8_variance16x16_wmt);
 140 extern prototype_subpixvariance(vp8_sub_pixel_variance4x4_wmt);
 141 extern prototype_subpixvariance(vp8_sub_pixel_variance8x8_wmt);
 142 extern prototype_subpixvariance(vp8_sub_pixel_variance8x16_wmt);
 143 extern prototype_subpixvariance(vp8_sub_pixel_variance16x8_wmt);
 144 extern prototype_subpixvariance(vp8_sub_pixel_variance16x16_wmt);
 145 extern prototype_variance(vp8_variance_halfpixvar16x16_h_wmt);
 146 extern prototype_variance(vp8_variance_halfpixvar16x16_v_wmt);
 147 extern prototype_variance(vp8_variance_halfpixvar16x16_hv_wmt);
 148 extern prototype_subpixvariance(vp8_sub_pixel_mse16x16_wmt);
 149 extern prototype_getmbss(vp8_get_mb_ss_sse2);
 150 extern prototype_variance(vp8_mse16x16_wmt);
 151 extern prototype_sad(vp8_get16x16pred_error_sse2);
 152 extern prototype_variance2(vp8_get8x8var_sse2);
 153 extern prototype_variance2(vp8_get16x16var_sse2);
 154
 155 #if !CONFIG_RUNTIME_CPU_DETECT
 156 #undef  vp8_variance_sad4x4
 157 #define vp8_variance_sad4x4 vp8_sad4x4_wmt
 158
 159 #undef  vp8_variance_sad8x8
 160 #define vp8_variance_sad8x8 vp8_sad8x8_wmt
 161
 162 #undef  vp8_variance_sad8x16
 163 #define vp8_variance_sad8x16 vp8_sad8x16_wmt
 164
 165 #undef  vp8_variance_sad16x8
 166 #define vp8_variance_sad16x8 vp8_sad16x8_wmt
 167
 168 #undef  vp8_variance_sad16x16
 169 #define vp8_variance_sad16x16 vp8_sad16x16_wmt
 170
 171 #undef  vp8_variance_var4x4
 172 #define vp8_variance_var4x4 vp8_variance4x4_wmt
 173
 174 #undef  vp8_variance_var8x8
 175 #define vp8_variance_var8x8 vp8_variance8x8_wmt
 176
 177 #undef  vp8_variance_var8x16
 178 #define vp8_variance_var8x16 vp8_variance8x16_wmt
 179
 180 #undef  vp8_variance_var16x8
 181 #define vp8_variance_var16x8 vp8_variance16x8_wmt
 182
 183 #undef  vp8_variance_var16x16
 184 #define vp8_variance_var16x16 vp8_variance16x16_wmt
 185
 186 #undef  vp8_variance_subpixvar4x4
 187 #define vp8_variance_subpixvar4x4 vp8_sub_pixel_variance4x4_wmt
 188
 189 #undef  vp8_variance_subpixvar8x8
 190 #define vp8_variance_subpixvar8x8 vp8_sub_pixel_variance8x8_wmt
 191
 192 #undef  vp8_variance_subpixvar8x16
 193 #define vp8_variance_subpixvar8x16 vp8_sub_pixel_variance8x16_wmt
 194
 195 #undef  vp8_variance_subpixvar16x8
 196 #define vp8_variance_subpixvar16x8 vp8_sub_pixel_variance16x8_wmt
 197
 198 #undef  vp8_variance_subpixvar16x16
 199 #define vp8_variance_subpixvar16x16 vp8_sub_pixel_variance16x16_wmt
 200
 201 #undef  vp8_variance_halfpixvar16x16_h
 202 #define vp8_variance_halfpixvar16x16_h vp8_variance_halfpixvar16x16_h_wmt
 203
 204 #undef  vp8_variance_halfpixvar16x16_v
 205 #define vp8_variance_halfpixvar16x16_v vp8_variance_halfpixvar16x16_v_wmt
 206
 207 #undef  vp8_variance_halfpixvar16x16_hv
 208 #define vp8_variance_halfpixvar16x16_hv vp8_variance_halfpixvar16x16_hv_wmt
 209
 210 #undef  vp8_variance_subpixmse16x16
 211 #define vp8_variance_subpixmse16x16 vp8_sub_pixel_mse16x16_wmt
 212
 213 #undef  vp8_variance_getmbss
 214 #define vp8_variance_getmbss vp8_get_mb_ss_sse2
 215
 216 #undef  vp8_variance_mse16x16
 217 #define vp8_variance_mse16x16 vp8_mse16x16_wmt
 218
 219 #undef  vp8_variance_get16x16prederror
 220 #define vp8_variance_get16x16prederror vp8_get16x16pred_error_sse2
 221
 222 #undef  vp8_variance_get8x8var
 223 #define vp8_variance_get8x8var vp8_get8x8var_sse2
 224
 225 #undef  vp8_variance_get16x16var
 226 #define vp8_variance_get16x16var vp8_get16x16var_sse2
 227
 228 #endif
 229 #endif
 230
 231
 232 #if HAVE_SSE3
 233 extern prototype_sad(vp8_sad16x16_sse3);
 234 extern prototype_sad(vp8_sad16x8_sse3);
 235 extern prototype_sad_multi_same_address(vp8_sad16x16x3_sse3);
 236 extern prototype_sad_multi_same_address(vp8_sad16x8x3_sse3);
 237 extern prototype_sad_multi_same_address(vp8_sad8x16x3_sse3);
 238 extern prototype_sad_multi_same_address(vp8_sad8x8x3_sse3);
 239 extern prototype_sad_multi_same_address(vp8_sad4x4x3_sse3);
 240
 241 extern prototype_sad_multi_dif_address(vp8_sad16x16x4d_sse3);
 242 extern prototype_sad_multi_dif_address(vp8_sad16x8x4d_sse3);
 243 extern prototype_sad_multi_dif_address(vp8_sad8x16x4d_sse3);
 244 extern prototype_sad_multi_dif_address(vp8_sad8x8x4d_sse3);
 245 extern prototype_sad_multi_dif_address(vp8_sad4x4x4d_sse3);
 246
 247 #if !CONFIG_RUNTIME_CPU_DETECT
 248
 249 #undef  vp8_variance_sad16x16
 250 #define vp8_variance_sad16x16 vp8_sad16x16_sse3
 251
 252 #undef  vp8_variance_sad16x16x3
 253 #define vp8_variance_sad16x16x3 vp8_sad16x16x3_sse3
 254
 255 #undef  vp8_variance_sad16x8x3
 256 #define vp8_variance_sad16x8x3 vp8_sad16x8x3_sse3
 257
 258 #undef  vp8_variance_sad8x16x3
 259 #define vp8_variance_sad8x16x3 vp8_sad8x16x3_sse3
 260
 261 #undef  vp8_variance_sad8x8x3
 262 #define vp8_variance_sad8x8x3 vp8_sad8x8x3_sse3
 263
 264 #undef  vp8_variance_sad4x4x3
 265 #define vp8_variance_sad4x4x3 vp8_sad4x4x3_sse3
 266
 267 #undef  vp8_variance_sad16x16x4d
 268 #define vp8_variance_sad16x16x4d vp8_sad16x16x4d_sse3
 269
 270 #undef  vp8_variance_sad16x8x4d
 271 #define vp8_variance_sad16x8x4d vp8_sad16x8x4d_sse3
 272
 273 #undef  vp8_variance_sad8x16x4d
 274 #define vp8_variance_sad8x16x4d vp8_sad8x16x4d_sse3
 275
 276 #undef  vp8_variance_sad8x8x4d
 277 #define vp8_variance_sad8x8x4d vp8_sad8x8x4d_sse3
 278
 279 #undef  vp8_variance_sad4x4x4d
 280 #define vp8_variance_sad4x4x4d vp8_sad4x4x4d_sse3
 281
 282 #endif
 283 #endif
 284
 285
 286 #if HAVE_SSSE3
 287 extern prototype_sad_multi_same_address(vp8_sad16x16x3_ssse3);
 288 extern prototype_sad_multi_same_address(vp8_sad16x8x3_ssse3);
 289
 290 #if !CONFIG_RUNTIME_CPU_DETECT
 291 #undef  vp8_variance_sad16x16x3
 292 #define vp8_variance_sad16x16x3 vp8_sad16x16x3_ssse3
 293
 294 #undef  vp8_variance_sad16x8x3
 295 #define vp8_variance_sad16x8x3 vp8_sad16x8x3_ssse3
 296
 297 #endif
 298 #endif
 299
 300
 301 #if HAVE_SSE4_1
 302 extern prototype_sad_multi_same_address_1(vp8_sad16x16x8_sse4);
 303 extern prototype_sad_multi_same_address_1(vp8_sad16x8x8_sse4);
 304 extern prototype_sad_multi_same_address_1(vp8_sad8x16x8_sse4);
 305 extern prototype_sad_multi_same_address_1(vp8_sad8x8x8_sse4);
 306 extern prototype_sad_multi_same_address_1(vp8_sad4x4x8_sse4);
 307
 308 #if !CONFIG_RUNTIME_CPU_DETECT
 309 #undef  vp8_variance_sad16x16x8
 310 #define vp8_variance_sad16x16x8 vp8_sad16x16x8_sse4
 311
 312 #undef  vp8_variance_sad16x8x8
 313 #define vp8_variance_sad16x8x8 vp8_sad16x8x8_sse4
 314
 315 #undef  vp8_variance_sad8x16x8
 316 #define vp8_variance_sad8x16x8 vp8_sad8x16x8_sse4
 317
 318 #undef  vp8_variance_sad8x8x8
 319 #define vp8_variance_sad8x8x8 vp8_sad8x8x8_sse4
 320
 321 #undef  vp8_variance_sad4x4x8
 322 #define vp8_variance_sad4x4x8 vp8_sad4x4x8_sse4
 323
 324 #endif
 325 #endif
 326
 327 #endif