vpx_dsp/vpx_dsp_rtcd_defs.pl

   1 sub vpx_dsp_forward_decls() {
   2 print <<EOF
   3 /*
   4  * DSP
   5  */
   6
   7 #include "vpx/vpx_integer.h"
   8 #include "vpx_dsp/vpx_dsp_common.h"
   9
  10 EOF
  11 }
  12 forward_decls qw/vpx_dsp_forward_decls/;
  13
  14 # x86inc.asm had specific constraints. break it out so it's easy to disable.
  15 # zero all the variables to avoid tricky else conditions.
  16 $mmx_x86inc = $sse_x86inc = $sse2_x86inc = $ssse3_x86inc = $avx_x86inc =
  17   $avx2_x86inc = '';
  18 $mmx_x86_64_x86inc = $sse_x86_64_x86inc = $sse2_x86_64_x86inc =
  19   $ssse3_x86_64_x86inc = $avx_x86_64_x86inc = $avx2_x86_64_x86inc = '';
  20 if (vpx_config("CONFIG_USE_X86INC") eq "yes") {
  21   $mmx_x86inc = 'mmx';
  22   $sse_x86inc = 'sse';
  23   $sse2_x86inc = 'sse2';
  24   $ssse3_x86inc = 'ssse3';
  25   $avx_x86inc = 'avx';
  26   $avx2_x86inc = 'avx2';
  27   if ($opts{arch} eq "x86_64") {
  28     $mmx_x86_64_x86inc = 'mmx';
  29     $sse_x86_64_x86inc = 'sse';
  30     $sse2_x86_64_x86inc = 'sse2';
  31     $ssse3_x86_64_x86inc = 'ssse3';
  32     $avx_x86_64_x86inc = 'avx';
  33     $avx2_x86_64_x86inc = 'avx2';
  34   }
  35 }
  36
  37 # optimizations which depend on multiple features
  38 $avx2_ssse3 = '';
  39 if ((vpx_config("HAVE_AVX2") eq "yes") && (vpx_config("HAVE_SSSE3") eq "yes")) {
  40   $avx2_ssse3 = 'avx2';
  41 }
  42
  43 # functions that are 64 bit only.
  44 $mmx_x86_64 = $sse2_x86_64 = $ssse3_x86_64 = $avx_x86_64 = $avx2_x86_64 = '';
  45 if ($opts{arch} eq "x86_64") {
  46   $mmx_x86_64 = 'mmx';
  47   $sse2_x86_64 = 'sse2';
  48   $ssse3_x86_64 = 'ssse3';
  49   $avx_x86_64 = 'avx';
  50   $avx2_x86_64 = 'avx2';
  51 }
  52
  53 #
  54 # Intra prediction
  55 #
  56
  57 add_proto qw/void vpx_d207_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  58 specialize qw/vpx_d207_predictor_4x4/, "$ssse3_x86inc";
  59
  60 add_proto qw/void vpx_d207e_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  61 specialize qw/vpx_d207e_predictor_4x4/;
  62
  63 add_proto qw/void vpx_d45_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  64 specialize qw/vpx_d45_predictor_4x4 neon/, "$ssse3_x86inc";
  65
  66 add_proto qw/void vpx_d45e_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  67 specialize qw/vpx_d45e_predictor_4x4/;
  68
  69 add_proto qw/void vpx_d63_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  70 specialize qw/vpx_d63_predictor_4x4/, "$ssse3_x86inc";
  71
  72 add_proto qw/void vpx_d63e_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  73 specialize qw/vpx_d63e_predictor_4x4/;
  74
  75 add_proto qw/void vpx_d63f_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  76 specialize qw/vpx_d63f_predictor_4x4/;
  77
  78 add_proto qw/void vpx_h_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  79 specialize qw/vpx_h_predictor_4x4 neon dspr2 msa/, "$sse2_x86inc";
  80
  81 add_proto qw/void vpx_he_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  82 specialize qw/vpx_he_predictor_4x4/;
  83
  84 add_proto qw/void vpx_d117_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  85 specialize qw/vpx_d117_predictor_4x4/;
  86
  87 add_proto qw/void vpx_d135_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  88 specialize qw/vpx_d135_predictor_4x4 neon/;
  89
  90 add_proto qw/void vpx_d153_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  91 specialize qw/vpx_d153_predictor_4x4/, "$ssse3_x86inc";
  92
  93 add_proto qw/void vpx_v_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  94 specialize qw/vpx_v_predictor_4x4 neon msa/, "$sse2_x86inc";
  95
  96 add_proto qw/void vpx_ve_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
  97 specialize qw/vpx_ve_predictor_4x4/;
  98
  99 add_proto qw/void vpx_tm_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 100 specialize qw/vpx_tm_predictor_4x4 neon dspr2 msa/, "$sse2_x86inc";
 101
 102 add_proto qw/void vpx_dc_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 103 specialize qw/vpx_dc_predictor_4x4 dspr2 msa neon/, "$sse2_x86inc";
 104
 105 add_proto qw/void vpx_dc_top_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 106 specialize qw/vpx_dc_top_predictor_4x4 msa neon/, "$sse2_x86inc";
 107
 108 add_proto qw/void vpx_dc_left_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 109 specialize qw/vpx_dc_left_predictor_4x4 msa neon/, "$sse2_x86inc";
 110
 111 add_proto qw/void vpx_dc_128_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 112 specialize qw/vpx_dc_128_predictor_4x4 msa neon/, "$sse2_x86inc";
 113
 114 add_proto qw/void vpx_d207_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 115 specialize qw/vpx_d207_predictor_8x8/, "$ssse3_x86inc";
 116
 117 add_proto qw/void vpx_d207e_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 118 specialize qw/vpx_d207e_predictor_8x8/;
 119
 120 add_proto qw/void vpx_d45_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 121 specialize qw/vpx_d45_predictor_8x8 neon/, "$ssse3_x86inc";
 122
 123 add_proto qw/void vpx_d45e_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 124 specialize qw/vpx_d45e_predictor_8x8/;
 125
 126 add_proto qw/void vpx_d63_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 127 specialize qw/vpx_d63_predictor_8x8/, "$ssse3_x86inc";
 128
 129 add_proto qw/void vpx_d63e_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 130 specialize qw/vpx_d63e_predictor_8x8/;
 131
 132 add_proto qw/void vpx_h_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 133 specialize qw/vpx_h_predictor_8x8 neon dspr2 msa/, "$sse2_x86inc";
 134
 135 add_proto qw/void vpx_d117_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 136 specialize qw/vpx_d117_predictor_8x8/;
 137
 138 add_proto qw/void vpx_d135_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 139 specialize qw/vpx_d135_predictor_8x8/;
 140
 141 add_proto qw/void vpx_d153_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 142 specialize qw/vpx_d153_predictor_8x8/, "$ssse3_x86inc";
 143
 144 add_proto qw/void vpx_v_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 145 specialize qw/vpx_v_predictor_8x8 neon msa/, "$sse2_x86inc";
 146
 147 add_proto qw/void vpx_tm_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 148 specialize qw/vpx_tm_predictor_8x8 neon dspr2 msa/, "$sse2_x86inc";
 149
 150 add_proto qw/void vpx_dc_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 151 specialize qw/vpx_dc_predictor_8x8 dspr2 neon msa/, "$sse2_x86inc";
 152
 153 add_proto qw/void vpx_dc_top_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 154 specialize qw/vpx_dc_top_predictor_8x8 neon msa/, "$sse2_x86inc";
 155
 156 add_proto qw/void vpx_dc_left_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 157 specialize qw/vpx_dc_left_predictor_8x8 neon msa/, "$sse2_x86inc";
 158
 159 add_proto qw/void vpx_dc_128_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 160 specialize qw/vpx_dc_128_predictor_8x8 neon msa/, "$sse2_x86inc";
 161
 162 add_proto qw/void vpx_d207_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 163 specialize qw/vpx_d207_predictor_16x16/, "$ssse3_x86inc";
 164
 165 add_proto qw/void vpx_d207e_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 166 specialize qw/vpx_d207e_predictor_16x16/;
 167
 168 add_proto qw/void vpx_d45_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 169 specialize qw/vpx_d45_predictor_16x16 neon/, "$ssse3_x86inc";
 170
 171 add_proto qw/void vpx_d45e_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 172 specialize qw/vpx_d45e_predictor_16x16/;
 173
 174 add_proto qw/void vpx_d63_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 175 specialize qw/vpx_d63_predictor_16x16/, "$ssse3_x86inc";
 176
 177 add_proto qw/void vpx_d63e_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 178 specialize qw/vpx_d63e_predictor_16x16/;
 179
 180 add_proto qw/void vpx_h_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 181 specialize qw/vpx_h_predictor_16x16 neon dspr2 msa/, "$sse2_x86inc";
 182
 183 add_proto qw/void vpx_d117_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 184 specialize qw/vpx_d117_predictor_16x16/;
 185
 186 add_proto qw/void vpx_d135_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 187 specialize qw/vpx_d135_predictor_16x16/;
 188
 189 add_proto qw/void vpx_d153_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 190 specialize qw/vpx_d153_predictor_16x16/, "$ssse3_x86inc";
 191
 192 add_proto qw/void vpx_v_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 193 specialize qw/vpx_v_predictor_16x16 neon msa/, "$sse2_x86inc";
 194
 195 add_proto qw/void vpx_tm_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 196 specialize qw/vpx_tm_predictor_16x16 neon msa/, "$sse2_x86inc";
 197
 198 add_proto qw/void vpx_dc_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 199 specialize qw/vpx_dc_predictor_16x16 dspr2 neon msa/, "$sse2_x86inc";
 200
 201 add_proto qw/void vpx_dc_top_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 202 specialize qw/vpx_dc_top_predictor_16x16 neon msa/, "$sse2_x86inc";
 203
 204 add_proto qw/void vpx_dc_left_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 205 specialize qw/vpx_dc_left_predictor_16x16 neon msa/, "$sse2_x86inc";
 206
 207 add_proto qw/void vpx_dc_128_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 208 specialize qw/vpx_dc_128_predictor_16x16 neon msa/, "$sse2_x86inc";
 209
 210 add_proto qw/void vpx_d207_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 211 specialize qw/vpx_d207_predictor_32x32/, "$ssse3_x86inc";
 212
 213 add_proto qw/void vpx_d207e_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 214 specialize qw/vpx_d207e_predictor_32x32/;
 215
 216 add_proto qw/void vpx_d45_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 217 specialize qw/vpx_d45_predictor_32x32/, "$ssse3_x86inc";
 218
 219 add_proto qw/void vpx_d45e_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 220 specialize qw/vpx_d45e_predictor_32x32/;
 221
 222 add_proto qw/void vpx_d63_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 223 specialize qw/vpx_d63_predictor_32x32/, "$ssse3_x86inc";
 224
 225 add_proto qw/void vpx_d63e_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 226 specialize qw/vpx_d63e_predictor_32x32/;
 227
 228 add_proto qw/void vpx_h_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 229 specialize qw/vpx_h_predictor_32x32 neon msa/, "$sse2_x86inc";
 230
 231 add_proto qw/void vpx_d117_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 232 specialize qw/vpx_d117_predictor_32x32/;
 233
 234 add_proto qw/void vpx_d135_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 235 specialize qw/vpx_d135_predictor_32x32/;
 236
 237 add_proto qw/void vpx_d153_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 238 specialize qw/vpx_d153_predictor_32x32/, "$ssse3_x86inc";
 239
 240 add_proto qw/void vpx_v_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 241 specialize qw/vpx_v_predictor_32x32 neon msa/, "$sse2_x86inc";
 242
 243 add_proto qw/void vpx_tm_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 244 specialize qw/vpx_tm_predictor_32x32 neon msa/, "$sse2_x86inc";
 245
 246 add_proto qw/void vpx_dc_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 247 specialize qw/vpx_dc_predictor_32x32 msa neon/, "$sse2_x86inc";
 248
 249 add_proto qw/void vpx_dc_top_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 250 specialize qw/vpx_dc_top_predictor_32x32 msa neon/, "$sse2_x86inc";
 251
 252 add_proto qw/void vpx_dc_left_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 253 specialize qw/vpx_dc_left_predictor_32x32 msa neon/, "$sse2_x86inc";
 254
 255 add_proto qw/void vpx_dc_128_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 256 specialize qw/vpx_dc_128_predictor_32x32 msa neon/, "$sse2_x86inc";
 257
 258 # High bitdepth functions
 259 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 260   add_proto qw/void vpx_highbd_d207_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 261   specialize qw/vpx_highbd_d207_predictor_4x4/;
 262
 263   add_proto qw/void vpx_highbd_d207e_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 264   specialize qw/vpx_highbd_d207e_predictor_4x4/;
 265
 266   add_proto qw/void vpx_highbd_d45_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 267   specialize qw/vpx_highbd_d45_predictor_4x4/;
 268
 269   add_proto qw/void vpx_highbd_d45e_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 270   specialize qw/vpx_highbd_d45e_predictor_4x4/;
 271
 272   add_proto qw/void vpx_highbd_d63_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 273   specialize qw/vpx_highbd_d63_predictor_4x4/;
 274
 275   add_proto qw/void vpx_highbd_d63e_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 276   specialize qw/vpx_highbd_d63e_predictor_4x4/;
 277
 278   add_proto qw/void vpx_highbd_h_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 279   specialize qw/vpx_highbd_h_predictor_4x4/;
 280
 281   add_proto qw/void vpx_highbd_d117_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 282   specialize qw/vpx_highbd_d117_predictor_4x4/;
 283
 284   add_proto qw/void vpx_highbd_d135_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 285   specialize qw/vpx_highbd_d135_predictor_4x4/;
 286
 287   add_proto qw/void vpx_highbd_d153_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 288   specialize qw/vpx_highbd_d153_predictor_4x4/;
 289
 290   add_proto qw/void vpx_highbd_v_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 291   specialize qw/vpx_highbd_v_predictor_4x4/, "$sse2_x86inc";
 292
 293   add_proto qw/void vpx_highbd_tm_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 294   specialize qw/vpx_highbd_tm_predictor_4x4/, "$sse2_x86inc";
 295
 296   add_proto qw/void vpx_highbd_dc_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 297   specialize qw/vpx_highbd_dc_predictor_4x4/, "$sse2_x86inc";
 298
 299   add_proto qw/void vpx_highbd_dc_top_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 300   specialize qw/vpx_highbd_dc_top_predictor_4x4/;
 301
 302   add_proto qw/void vpx_highbd_dc_left_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 303   specialize qw/vpx_highbd_dc_left_predictor_4x4/;
 304
 305   add_proto qw/void vpx_highbd_dc_128_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 306   specialize qw/vpx_highbd_dc_128_predictor_4x4/;
 307
 308   add_proto qw/void vpx_highbd_d207_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 309   specialize qw/vpx_highbd_d207_predictor_8x8/;
 310
 311   add_proto qw/void vpx_highbd_d207e_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 312   specialize qw/vpx_highbd_d207e_predictor_8x8/;
 313
 314   add_proto qw/void vpx_highbd_d45_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 315   specialize qw/vpx_highbd_d45_predictor_8x8/;
 316
 317   add_proto qw/void vpx_highbd_d45e_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 318   specialize qw/vpx_highbd_d45e_predictor_8x8/;
 319
 320   add_proto qw/void vpx_highbd_d63_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 321   specialize qw/vpx_highbd_d63_predictor_8x8/;
 322
 323   add_proto qw/void vpx_highbd_d63e_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 324   specialize qw/vpx_highbd_d63e_predictor_8x8/;
 325
 326   add_proto qw/void vpx_highbd_h_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 327   specialize qw/vpx_highbd_h_predictor_8x8/;
 328
 329   add_proto qw/void vpx_highbd_d117_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 330   specialize qw/vpx_highbd_d117_predictor_8x8/;
 331
 332   add_proto qw/void vpx_highbd_d135_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 333   specialize qw/vpx_highbd_d135_predictor_8x8/;
 334
 335   add_proto qw/void vpx_highbd_d153_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 336   specialize qw/vpx_highbd_d153_predictor_8x8/;
 337
 338   add_proto qw/void vpx_highbd_v_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 339   specialize qw/vpx_highbd_v_predictor_8x8/, "$sse2_x86inc";
 340
 341   add_proto qw/void vpx_highbd_tm_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 342   specialize qw/vpx_highbd_tm_predictor_8x8/, "$sse2_x86inc";
 343
 344   add_proto qw/void vpx_highbd_dc_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 345   specialize qw/vpx_highbd_dc_predictor_8x8/, "$sse2_x86inc";;
 346
 347   add_proto qw/void vpx_highbd_dc_top_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 348   specialize qw/vpx_highbd_dc_top_predictor_8x8/;
 349
 350   add_proto qw/void vpx_highbd_dc_left_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 351   specialize qw/vpx_highbd_dc_left_predictor_8x8/;
 352
 353   add_proto qw/void vpx_highbd_dc_128_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 354   specialize qw/vpx_highbd_dc_128_predictor_8x8/;
 355
 356   add_proto qw/void vpx_highbd_d207_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 357   specialize qw/vpx_highbd_d207_predictor_16x16/;
 358
 359   add_proto qw/void vpx_highbd_d207e_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 360   specialize qw/vpx_highbd_d207e_predictor_16x16/;
 361
 362   add_proto qw/void vpx_highbd_d45_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 363   specialize qw/vpx_highbd_d45_predictor_16x16/;
 364
 365   add_proto qw/void vpx_highbd_d45e_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 366   specialize qw/vpx_highbd_d45e_predictor_16x16/;
 367
 368   add_proto qw/void vpx_highbd_d63_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 369   specialize qw/vpx_highbd_d63_predictor_16x16/;
 370
 371   add_proto qw/void vpx_highbd_d63e_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 372   specialize qw/vpx_highbd_d63e_predictor_16x16/;
 373
 374   add_proto qw/void vpx_highbd_h_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 375   specialize qw/vpx_highbd_h_predictor_16x16/;
 376
 377   add_proto qw/void vpx_highbd_d117_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 378   specialize qw/vpx_highbd_d117_predictor_16x16/;
 379
 380   add_proto qw/void vpx_highbd_d135_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 381   specialize qw/vpx_highbd_d135_predictor_16x16/;
 382
 383   add_proto qw/void vpx_highbd_d153_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 384   specialize qw/vpx_highbd_d153_predictor_16x16/;
 385
 386   add_proto qw/void vpx_highbd_v_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 387   specialize qw/vpx_highbd_v_predictor_16x16/, "$sse2_x86inc";
 388
 389   add_proto qw/void vpx_highbd_tm_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 390   specialize qw/vpx_highbd_tm_predictor_16x16/, "$sse2_x86inc";
 391
 392   add_proto qw/void vpx_highbd_dc_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 393   specialize qw/vpx_highbd_dc_predictor_16x16/, "$sse2_x86inc";
 394
 395   add_proto qw/void vpx_highbd_dc_top_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 396   specialize qw/vpx_highbd_dc_top_predictor_16x16/;
 397
 398   add_proto qw/void vpx_highbd_dc_left_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 399   specialize qw/vpx_highbd_dc_left_predictor_16x16/;
 400
 401   add_proto qw/void vpx_highbd_dc_128_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 402   specialize qw/vpx_highbd_dc_128_predictor_16x16/;
 403
 404   add_proto qw/void vpx_highbd_d207_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 405   specialize qw/vpx_highbd_d207_predictor_32x32/;
 406
 407   add_proto qw/void vpx_highbd_d207e_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 408   specialize qw/vpx_highbd_d207e_predictor_32x32/;
 409
 410   add_proto qw/void vpx_highbd_d45_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 411   specialize qw/vpx_highbd_d45_predictor_32x32/;
 412
 413   add_proto qw/void vpx_highbd_d45e_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 414   specialize qw/vpx_highbd_d45e_predictor_32x32/;
 415
 416   add_proto qw/void vpx_highbd_d63_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 417   specialize qw/vpx_highbd_d63_predictor_32x32/;
 418
 419   add_proto qw/void vpx_highbd_d63e_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 420   specialize qw/vpx_highbd_d63e_predictor_32x32/;
 421
 422   add_proto qw/void vpx_highbd_h_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 423   specialize qw/vpx_highbd_h_predictor_32x32/;
 424
 425   add_proto qw/void vpx_highbd_d117_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 426   specialize qw/vpx_highbd_d117_predictor_32x32/;
 427
 428   add_proto qw/void vpx_highbd_d135_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 429   specialize qw/vpx_highbd_d135_predictor_32x32/;
 430
 431   add_proto qw/void vpx_highbd_d153_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 432   specialize qw/vpx_highbd_d153_predictor_32x32/;
 433
 434   add_proto qw/void vpx_highbd_v_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 435   specialize qw/vpx_highbd_v_predictor_32x32/, "$sse2_x86inc";
 436
 437   add_proto qw/void vpx_highbd_tm_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 438   specialize qw/vpx_highbd_tm_predictor_32x32/, "$sse2_x86inc";
 439
 440   add_proto qw/void vpx_highbd_dc_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 441   specialize qw/vpx_highbd_dc_predictor_32x32/, "$sse2_x86inc";
 442
 443   add_proto qw/void vpx_highbd_dc_top_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 444   specialize qw/vpx_highbd_dc_top_predictor_32x32/;
 445
 446   add_proto qw/void vpx_highbd_dc_left_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 447   specialize qw/vpx_highbd_dc_left_predictor_32x32/;
 448
 449   add_proto qw/void vpx_highbd_dc_128_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 450   specialize qw/vpx_highbd_dc_128_predictor_32x32/;
 451 }  # CONFIG_VP9_HIGHBITDEPTH
 452
 453 #
 454 # Sub Pixel Filters
 455 #
 456 add_proto qw/void vpx_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 457 specialize qw/vpx_convolve_copy neon dspr2 msa/, "$sse2_x86inc";
 458
 459 add_proto qw/void vpx_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 460 specialize qw/vpx_convolve_avg neon dspr2 msa/, "$sse2_x86inc";
 461
 462 add_proto qw/void vpx_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 463 specialize qw/vpx_convolve8 sse2 ssse3 neon dspr2 msa/, "$avx2_ssse3";
 464
 465 add_proto qw/void vpx_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 466 specialize qw/vpx_convolve8_horiz sse2 ssse3 neon dspr2 msa/, "$avx2_ssse3";
 467
 468 add_proto qw/void vpx_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 469 specialize qw/vpx_convolve8_vert sse2 ssse3 neon dspr2 msa/, "$avx2_ssse3";
 470
 471 add_proto qw/void vpx_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 472 specialize qw/vpx_convolve8_avg sse2 ssse3 neon dspr2 msa/;
 473
 474 add_proto qw/void vpx_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 475 specialize qw/vpx_convolve8_avg_horiz sse2 ssse3 neon dspr2 msa/;
 476
 477 add_proto qw/void vpx_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 478 specialize qw/vpx_convolve8_avg_vert sse2 ssse3 neon dspr2 msa/;
 479
 480 add_proto qw/void vpx_scaled_2d/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 481 specialize qw/vpx_scaled_2d ssse3/;
 482
 483 add_proto qw/void vpx_scaled_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 484 specialize qw/vpx_scaled_horiz/;
 485
 486 add_proto qw/void vpx_scaled_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 487 specialize qw/vpx_scaled_vert/;
 488
 489 add_proto qw/void vpx_scaled_avg_2d/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 490 specialize qw/vpx_scaled_avg_2d/;
 491
 492 add_proto qw/void vpx_scaled_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 493 specialize qw/vpx_scaled_avg_horiz/;
 494
 495 add_proto qw/void vpx_scaled_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 496 specialize qw/vpx_scaled_avg_vert/;
 497
 498 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 499   #
 500   # Sub Pixel Filters
 501   #
 502   add_proto qw/void vpx_highbd_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
 503   specialize qw/vpx_highbd_convolve_copy/, "$sse2_x86inc";
 504
 505   add_proto qw/void vpx_highbd_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
 506   specialize qw/vpx_highbd_convolve_avg/, "$sse2_x86inc";
 507
 508   add_proto qw/void vpx_highbd_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
 509   specialize qw/vpx_highbd_convolve8/, "$sse2_x86_64";
 510
 511   add_proto qw/void vpx_highbd_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
 512   specialize qw/vpx_highbd_convolve8_horiz/, "$sse2_x86_64";
 513
 514   add_proto qw/void vpx_highbd_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
 515   specialize qw/vpx_highbd_convolve8_vert/, "$sse2_x86_64";
 516
 517   add_proto qw/void vpx_highbd_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
 518   specialize qw/vpx_highbd_convolve8_avg/, "$sse2_x86_64";
 519
 520   add_proto qw/void vpx_highbd_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
 521   specialize qw/vpx_highbd_convolve8_avg_horiz/, "$sse2_x86_64";
 522
 523   add_proto qw/void vpx_highbd_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
 524   specialize qw/vpx_highbd_convolve8_avg_vert/, "$sse2_x86_64";
 525 }  # CONFIG_VP9_HIGHBITDEPTH
 526
 527 #
 528 # Loopfilter
 529 #
 530 add_proto qw/void vpx_lpf_vertical_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
 531 specialize qw/vpx_lpf_vertical_16 sse2 neon_asm dspr2 msa/;
 532 $vpx_lpf_vertical_16_neon_asm=vpx_lpf_vertical_16_neon;
 533
 534 add_proto qw/void vpx_lpf_vertical_16_dual/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
 535 specialize qw/vpx_lpf_vertical_16_dual sse2 neon_asm dspr2 msa/;
 536 $vpx_lpf_vertical_16_dual_neon_asm=vpx_lpf_vertical_16_dual_neon;
 537
 538 add_proto qw/void vpx_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
 539 specialize qw/vpx_lpf_vertical_8 sse2 neon dspr2 msa/;
 540
 541 add_proto qw/void vpx_lpf_vertical_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
 542 specialize qw/vpx_lpf_vertical_8_dual sse2 neon_asm dspr2 msa/;
 543 $vpx_lpf_vertical_8_dual_neon_asm=vpx_lpf_vertical_8_dual_neon;
 544
 545 add_proto qw/void vpx_lpf_vertical_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
 546 specialize qw/vpx_lpf_vertical_4 neon dspr2 msa/, "$mmx_x86inc";
 547
 548 add_proto qw/void vpx_lpf_vertical_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
 549 specialize qw/vpx_lpf_vertical_4_dual sse2 neon dspr2 msa/;
 550
 551 add_proto qw/void vpx_lpf_horizontal_edge_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
 552 specialize qw/vpx_lpf_horizontal_edge_8 sse2 avx2 neon_asm dspr2 msa/;
 553 $vpx_lpf_horizontal_edge_8_neon_asm=vpx_lpf_horizontal_edge_8_neon;
 554
 555 add_proto qw/void vpx_lpf_horizontal_edge_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
 556 specialize qw/vpx_lpf_horizontal_edge_16 sse2 avx2 neon_asm dspr2 msa/;
 557 $vpx_lpf_horizontal_edge_16_neon_asm=vpx_lpf_horizontal_edge_16_neon;
 558
 559 add_proto qw/void vpx_lpf_horizontal_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
 560 specialize qw/vpx_lpf_horizontal_8 sse2 neon dspr2 msa/;
 561
 562 add_proto qw/void vpx_lpf_horizontal_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
 563 specialize qw/vpx_lpf_horizontal_8_dual sse2 neon_asm dspr2 msa/;
 564 $vpx_lpf_horizontal_8_dual_neon_asm=vpx_lpf_horizontal_8_dual_neon;
 565
 566 add_proto qw/void vpx_lpf_horizontal_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
 567 specialize qw/vpx_lpf_horizontal_4 neon dspr2 msa/, "$mmx_x86inc";
 568
 569 add_proto qw/void vpx_lpf_horizontal_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
 570 specialize qw/vpx_lpf_horizontal_4_dual sse2 neon dspr2 msa/;
 571
 572 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 573   add_proto qw/void vpx_highbd_lpf_vertical_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
 574   specialize qw/vpx_highbd_lpf_vertical_16 sse2/;
 575
 576   add_proto qw/void vpx_highbd_lpf_vertical_16_dual/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
 577   specialize qw/vpx_highbd_lpf_vertical_16_dual sse2/;
 578
 579   add_proto qw/void vpx_highbd_lpf_vertical_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
 580   specialize qw/vpx_highbd_lpf_vertical_8 sse2/;
 581
 582   add_proto qw/void vpx_highbd_lpf_vertical_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
 583   specialize qw/vpx_highbd_lpf_vertical_8_dual sse2/;
 584
 585   add_proto qw/void vpx_highbd_lpf_vertical_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
 586   specialize qw/vpx_highbd_lpf_vertical_4 sse2/;
 587
 588   add_proto qw/void vpx_highbd_lpf_vertical_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
 589   specialize qw/vpx_highbd_lpf_vertical_4_dual sse2/;
 590
 591   add_proto qw/void vpx_highbd_lpf_horizontal_edge_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
 592   specialize qw/vpx_highbd_lpf_horizontal_edge_8 sse2/;
 593
 594   add_proto qw/void vpx_highbd_lpf_horizontal_edge_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
 595   specialize qw/vpx_highbd_lpf_horizontal_edge_16 sse2/;
 596
 597   add_proto qw/void vpx_highbd_lpf_horizontal_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
 598   specialize qw/vpx_highbd_lpf_horizontal_8 sse2/;
 599
 600   add_proto qw/void vpx_highbd_lpf_horizontal_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
 601   specialize qw/vpx_highbd_lpf_horizontal_8_dual sse2/;
 602
 603   add_proto qw/void vpx_highbd_lpf_horizontal_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
 604   specialize qw/vpx_highbd_lpf_horizontal_4 sse2/;
 605
 606   add_proto qw/void vpx_highbd_lpf_horizontal_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
 607   specialize qw/vpx_highbd_lpf_horizontal_4_dual sse2/;
 608 }  # CONFIG_VP9_HIGHBITDEPTH
 609
 610 #
 611 # Encoder functions.
 612 #
 613
 614 #
 615 # Forward transform
 616 #
 617 if ((vpx_config("CONFIG_VP9_ENCODER") eq "yes") || (vpx_config("CONFIG_VP10_ENCODER") eq "yes")) {
 618 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 619   add_proto qw/void vpx_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
 620   specialize qw/vpx_fdct4x4 sse2/;
 621
 622   add_proto qw/void vpx_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
 623   specialize qw/vpx_fdct4x4_1 sse2/;
 624
 625   add_proto qw/void vpx_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
 626   specialize qw/vpx_fdct8x8 sse2/;
 627
 628   add_proto qw/void vpx_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
 629   specialize qw/vpx_fdct8x8_1 sse2/;
 630
 631   add_proto qw/void vpx_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
 632   specialize qw/vpx_fdct16x16 sse2/;
 633
 634   add_proto qw/void vpx_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
 635   specialize qw/vpx_fdct16x16_1 sse2/;
 636
 637   add_proto qw/void vpx_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
 638   specialize qw/vpx_fdct32x32 sse2/;
 639
 640   add_proto qw/void vpx_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
 641   specialize qw/vpx_fdct32x32_rd sse2/;
 642
 643   add_proto qw/void vpx_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
 644   specialize qw/vpx_fdct32x32_1 sse2/;
 645
 646   add_proto qw/void vpx_highbd_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
 647   specialize qw/vpx_highbd_fdct4x4 sse2/;
 648
 649   add_proto qw/void vpx_highbd_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
 650   specialize qw/vpx_highbd_fdct8x8 sse2/;
 651
 652   add_proto qw/void vpx_highbd_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
 653   specialize qw/vpx_highbd_fdct8x8_1/;
 654
 655   add_proto qw/void vpx_highbd_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
 656   specialize qw/vpx_highbd_fdct16x16 sse2/;
 657
 658   add_proto qw/void vpx_highbd_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
 659   specialize qw/vpx_highbd_fdct16x16_1/;
 660
 661   add_proto qw/void vpx_highbd_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
 662   specialize qw/vpx_highbd_fdct32x32 sse2/;
 663
 664   add_proto qw/void vpx_highbd_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
 665   specialize qw/vpx_highbd_fdct32x32_rd sse2/;
 666
 667   add_proto qw/void vpx_highbd_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
 668   specialize qw/vpx_highbd_fdct32x32_1/;
 669 } else {
 670   add_proto qw/void vpx_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
 671   specialize qw/vpx_fdct4x4 sse2 msa/;
 672
 673   add_proto qw/void vpx_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
 674   specialize qw/vpx_fdct4x4_1 sse2/;
 675
 676   add_proto qw/void vpx_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
 677   specialize qw/vpx_fdct8x8 sse2 neon msa/, "$ssse3_x86_64_x86inc";
 678
 679   add_proto qw/void vpx_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
 680   specialize qw/vpx_fdct8x8_1 sse2 neon msa/;
 681
 682   add_proto qw/void vpx_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
 683   specialize qw/vpx_fdct16x16 sse2 msa/;
 684
 685   add_proto qw/void vpx_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
 686   specialize qw/vpx_fdct16x16_1 sse2 msa/;
 687
 688   add_proto qw/void vpx_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
 689   specialize qw/vpx_fdct32x32 sse2 avx2 msa/;
 690
 691   add_proto qw/void vpx_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
 692   specialize qw/vpx_fdct32x32_rd sse2 avx2 msa/;
 693
 694   add_proto qw/void vpx_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
 695   specialize qw/vpx_fdct32x32_1 sse2 msa/;
 696 }  # CONFIG_VP9_HIGHBITDEPTH
 697 }  # CONFIG_VP9_ENCODER || CONFIG_VP10_ENCODER
 698
 699 #
 700 # Inverse transform
 701 if ((vpx_config("CONFIG_VP9") eq "yes") || (vpx_config("CONFIG_VP10") eq "yes")) {
 702 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 703   # Note as optimized versions of these functions are added we need to add a check to ensure
 704   # that when CONFIG_EMULATE_HARDWARE is on, it defaults to the C versions only.
 705   add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 706   specialize qw/vpx_iwht4x4_1_add/;
 707
 708   add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 709   specialize qw/vpx_iwht4x4_16_add/, "$sse2_x86inc";
 710
 711   add_proto qw/void vpx_highbd_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 712   specialize qw/vpx_highbd_idct4x4_1_add/;
 713
 714   add_proto qw/void vpx_highbd_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 715   specialize qw/vpx_highbd_idct8x8_1_add/;
 716
 717   add_proto qw/void vpx_highbd_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 718   specialize qw/vpx_highbd_idct16x16_1_add/;
 719
 720   add_proto qw/void vpx_highbd_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 721   specialize qw/vpx_highbd_idct32x32_1024_add/;
 722
 723   add_proto qw/void vpx_highbd_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 724   specialize qw/vpx_highbd_idct32x32_34_add/;
 725
 726   add_proto qw/void vpx_highbd_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 727   specialize qw/vpx_highbd_idct32x32_1_add/;
 728
 729   add_proto qw/void vpx_highbd_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 730   specialize qw/vpx_highbd_iwht4x4_1_add/;
 731
 732   add_proto qw/void vpx_highbd_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 733   specialize qw/vpx_highbd_iwht4x4_16_add/;
 734
 735   # Force C versions if CONFIG_EMULATE_HARDWARE is 1
 736   if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
 737     add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 738     specialize qw/vpx_idct4x4_16_add/;
 739
 740     add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 741     specialize qw/vpx_idct4x4_1_add/;
 742
 743     add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 744     specialize qw/vpx_idct8x8_64_add/;
 745
 746     add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 747     specialize qw/vpx_idct8x8_12_add/;
 748
 749     add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 750     specialize qw/vpx_idct8x8_1_add/;
 751
 752     add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 753     specialize qw/vpx_idct16x16_256_add/;
 754
 755     add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 756     specialize qw/vpx_idct16x16_10_add/;
 757
 758     add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 759     specialize qw/vpx_idct16x16_1_add/;
 760
 761     add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 762     specialize qw/vpx_idct32x32_1024_add/;
 763
 764     add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 765     specialize qw/vpx_idct32x32_135_add/;
 766
 767     add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 768     specialize qw/vpx_idct32x32_34_add/;
 769
 770     add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 771     specialize qw/vpx_idct32x32_1_add/;
 772
 773     add_proto qw/void vpx_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 774     specialize qw/vpx_highbd_idct4x4_16_add/;
 775
 776     add_proto qw/void vpx_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 777     specialize qw/vpx_highbd_idct8x8_64_add/;
 778
 779     add_proto qw/void vpx_highbd_idct8x8_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 780     specialize qw/vpx_highbd_idct8x8_10_add/;
 781
 782     add_proto qw/void vpx_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 783     specialize qw/vpx_highbd_idct16x16_256_add/;
 784
 785     add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 786     specialize qw/vpx_highbd_idct16x16_10_add/;
 787   } else {
 788     add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 789     specialize qw/vpx_idct4x4_16_add sse2/;
 790
 791     add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 792     specialize qw/vpx_idct4x4_1_add sse2/;
 793
 794     add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 795     specialize qw/vpx_idct8x8_64_add sse2/, "$ssse3_x86_64_x86inc";
 796
 797     add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 798     specialize qw/vpx_idct8x8_12_add sse2/, "$ssse3_x86_64_x86inc";
 799
 800     add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 801     specialize qw/vpx_idct8x8_1_add sse2/;
 802
 803     add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 804     specialize qw/vpx_idct16x16_256_add sse2/;
 805
 806     add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 807     specialize qw/vpx_idct16x16_10_add sse2/;
 808
 809     add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 810     specialize qw/vpx_idct16x16_1_add sse2/;
 811
 812     add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 813     specialize qw/vpx_idct32x32_1024_add sse2/, "$ssse3_x86_64_x86inc";
 814
 815     add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 816     specialize qw/vpx_idct32x32_135_add sse2/, "$ssse3_x86_64_x86inc";
 817     # Need to add 135 eob idct32x32 implementations.
 818     $vpx_idct32x32_135_add_sse2=vpx_idct32x32_1024_add_sse2;
 819
 820     add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 821     specialize qw/vpx_idct32x32_34_add sse2/, "$ssse3_x86_64_x86inc";
 822
 823     add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 824     specialize qw/vpx_idct32x32_1_add sse2/;
 825
 826     add_proto qw/void vpx_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 827     specialize qw/vpx_highbd_idct4x4_16_add sse2/;
 828
 829     add_proto qw/void vpx_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 830     specialize qw/vpx_highbd_idct8x8_64_add sse2/;
 831
 832     add_proto qw/void vpx_highbd_idct8x8_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 833     specialize qw/vpx_highbd_idct8x8_10_add sse2/;
 834
 835     add_proto qw/void vpx_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 836     specialize qw/vpx_highbd_idct16x16_256_add sse2/;
 837
 838     add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 839     specialize qw/vpx_highbd_idct16x16_10_add sse2/;
 840   }  # CONFIG_EMULATE_HARDWARE
 841 } else {
 842   # Force C versions if CONFIG_EMULATE_HARDWARE is 1
 843   if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
 844     add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 845     specialize qw/vpx_idct4x4_1_add/;
 846
 847     add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 848     specialize qw/vpx_idct4x4_16_add/;
 849
 850     add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 851     specialize qw/vpx_idct8x8_1_add/;
 852
 853     add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 854     specialize qw/vpx_idct8x8_64_add/;
 855
 856     add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 857     specialize qw/vpx_idct8x8_12_add/;
 858
 859     add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 860     specialize qw/vpx_idct16x16_1_add/;
 861
 862     add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 863     specialize qw/vpx_idct16x16_256_add/;
 864
 865     add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 866     specialize qw/vpx_idct16x16_10_add/;
 867
 868     add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 869     specialize qw/vpx_idct32x32_1024_add/;
 870
 871     add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 872     specialize qw/vpx_idct32x32_135_add/;
 873
 874     add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 875     specialize qw/vpx_idct32x32_34_add/;
 876
 877     add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 878     specialize qw/vpx_idct32x32_1_add/;
 879
 880     add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 881     specialize qw/vpx_iwht4x4_1_add/;
 882
 883     add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 884     specialize qw/vpx_iwht4x4_16_add/;
 885   } else {
 886     add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 887     specialize qw/vpx_idct4x4_1_add sse2 neon dspr2 msa/;
 888
 889     add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 890     specialize qw/vpx_idct4x4_16_add sse2 neon dspr2 msa/;
 891
 892     add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 893     specialize qw/vpx_idct8x8_1_add sse2 neon dspr2 msa/;
 894
 895     add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 896     specialize qw/vpx_idct8x8_64_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
 897
 898     add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 899     specialize qw/vpx_idct8x8_12_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
 900
 901     add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 902     specialize qw/vpx_idct16x16_1_add sse2 neon dspr2 msa/;
 903
 904     add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 905     specialize qw/vpx_idct16x16_256_add sse2 neon dspr2 msa/;
 906
 907     add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 908     specialize qw/vpx_idct16x16_10_add sse2 neon dspr2 msa/;
 909
 910     add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 911     specialize qw/vpx_idct32x32_1024_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
 912
 913     add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 914     specialize qw/vpx_idct32x32_135_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
 915     # Need to add 135 eob idct32x32 implementations.
 916     $vpx_idct32x32_135_add_sse2=vpx_idct32x32_1024_add_sse2;
 917     $vpx_idct32x32_135_add_neon=vpx_idct32x32_1024_add_neon;
 918     $vpx_idct32x32_135_add_dspr2=vpx_idct32x32_1024_add_dspr2;
 919     $vpx_idct32x32_135_add_msa=vpx_idct32x32_1024_add_msa;
 920
 921     add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 922     specialize qw/vpx_idct32x32_34_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
 923     # Need to add 34 eob idct32x32 neon implementation.
 924     $vpx_idct32x32_34_add_neon=vpx_idct32x32_1024_add_neon;
 925
 926     add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 927     specialize qw/vpx_idct32x32_1_add sse2 neon dspr2 msa/;
 928
 929     add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 930     specialize qw/vpx_iwht4x4_1_add msa/;
 931
 932     add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 933     specialize qw/vpx_iwht4x4_16_add msa/, "$sse2_x86inc";
 934   }  # CONFIG_EMULATE_HARDWARE
 935 }  # CONFIG_VP9_HIGHBITDEPTH
 936 }  # CONFIG_VP9 || CONFIG_VP10
 937
 938 #
 939 # Quantization
 940 #
 941 if ((vpx_config("CONFIG_VP9_ENCODER") eq "yes") || (vpx_config("CONFIG_VP10_ENCODER") eq "yes")) {
 942   add_proto qw/void vpx_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
 943   specialize qw/vpx_quantize_b sse2/, "$ssse3_x86_64_x86inc", "$avx_x86_64_x86inc";
 944
 945   add_proto qw/void vpx_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
 946   specialize qw/vpx_quantize_b_32x32/, "$ssse3_x86_64_x86inc", "$avx_x86_64_x86inc";
 947
 948   if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 949     add_proto qw/void vpx_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
 950     specialize qw/vpx_highbd_quantize_b sse2/;
 951
 952     add_proto qw/void vpx_highbd_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
 953     specialize qw/vpx_highbd_quantize_b_32x32 sse2/;
 954   }  # CONFIG_VP9_HIGHBITDEPTH
 955 }  # CONFIG_VP9_ENCODER || CONFIG_VP10_ENCODER
 956
 957 if (vpx_config("CONFIG_ENCODERS") eq "yes") {
 958 #
 959 # Block subtraction
 960 #
 961 add_proto qw/void vpx_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
 962 specialize qw/vpx_subtract_block neon msa/, "$sse2_x86inc";
 963
 964 #
 965 # Single block SAD
 966 #
 967 add_proto qw/unsigned int vpx_sad64x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 968 specialize qw/vpx_sad64x64 avx2 neon msa/, "$sse2_x86inc";
 969
 970 add_proto qw/unsigned int vpx_sad64x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 971 specialize qw/vpx_sad64x32 avx2 msa/, "$sse2_x86inc";
 972
 973 add_proto qw/unsigned int vpx_sad32x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 974 specialize qw/vpx_sad32x64 avx2 msa/, "$sse2_x86inc";
 975
 976 add_proto qw/unsigned int vpx_sad32x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 977 specialize qw/vpx_sad32x32 avx2 neon msa/, "$sse2_x86inc";
 978
 979 add_proto qw/unsigned int vpx_sad32x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 980 specialize qw/vpx_sad32x16 avx2 msa/, "$sse2_x86inc";
 981
 982 add_proto qw/unsigned int vpx_sad16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 983 specialize qw/vpx_sad16x32 msa/, "$sse2_x86inc";
 984
 985 add_proto qw/unsigned int vpx_sad16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 986 specialize qw/vpx_sad16x16 mmx media neon msa/, "$sse2_x86inc";
 987
 988 add_proto qw/unsigned int vpx_sad16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 989 specialize qw/vpx_sad16x8 mmx neon msa/, "$sse2_x86inc";
 990
 991 add_proto qw/unsigned int vpx_sad8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 992 specialize qw/vpx_sad8x16 mmx neon msa/, "$sse2_x86inc";
 993
 994 add_proto qw/unsigned int vpx_sad8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 995 specialize qw/vpx_sad8x8 mmx neon msa/, "$sse2_x86inc";
 996
 997 add_proto qw/unsigned int vpx_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 998 specialize qw/vpx_sad8x4 msa/, "$sse2_x86inc";
 999
1000 add_proto qw/unsigned int vpx_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1001 specialize qw/vpx_sad4x8 msa/, "$sse2_x86inc";
1002
1003 add_proto qw/unsigned int vpx_sad4x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1004 specialize qw/vpx_sad4x4 mmx neon msa/, "$sse2_x86inc";
1005
1006 #
1007 # Avg
1008 #
1009 if ((vpx_config("CONFIG_VP9_ENCODER") eq "yes") || (vpx_config("CONFIG_VP10_ENCODER") eq "yes")) {
1010   add_proto qw/unsigned int vpx_avg_8x8/, "const uint8_t *, int p";
1011   specialize qw/vpx_avg_8x8 sse2 neon msa/;
1012
1013   add_proto qw/unsigned int vpx_avg_4x4/, "const uint8_t *, int p";
1014   specialize qw/vpx_avg_4x4 sse2 neon msa/;
1015
1016   add_proto qw/void vpx_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
1017   specialize qw/vpx_minmax_8x8 sse2 neon/;
1018
1019   add_proto qw/void vpx_hadamard_8x8/, "const int16_t *src_diff, int src_stride, int16_t *coeff";
1020   specialize qw/vpx_hadamard_8x8 sse2/, "$ssse3_x86_64_x86inc";
1021
1022   add_proto qw/void vpx_hadamard_16x16/, "const int16_t *src_diff, int src_stride, int16_t *coeff";
1023   specialize qw/vpx_hadamard_16x16 sse2/;
1024
1025   add_proto qw/int vpx_satd/, "const int16_t *coeff, int length";
1026   specialize qw/vpx_satd sse2 neon/;
1027
1028   add_proto qw/void vpx_int_pro_row/, "int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height";
1029   specialize qw/vpx_int_pro_row sse2 neon/;
1030
1031   add_proto qw/int16_t vpx_int_pro_col/, "const uint8_t *ref, const int width";
1032   specialize qw/vpx_int_pro_col sse2 neon/;
1033
1034   add_proto qw/int vpx_vector_var/, "const int16_t *ref, const int16_t *src, const int bwl";
1035   specialize qw/vpx_vector_var neon sse2/;
1036 }  # CONFIG_VP9_ENCODER || CONFIG_VP10_ENCODER
1037
1038 add_proto qw/unsigned int vpx_sad64x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1039 specialize qw/vpx_sad64x64_avg avx2 msa/, "$sse2_x86inc";
1040
1041 add_proto qw/unsigned int vpx_sad64x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1042 specialize qw/vpx_sad64x32_avg avx2 msa/, "$sse2_x86inc";
1043
1044 add_proto qw/unsigned int vpx_sad32x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1045 specialize qw/vpx_sad32x64_avg avx2 msa/, "$sse2_x86inc";
1046
1047 add_proto qw/unsigned int vpx_sad32x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1048 specialize qw/vpx_sad32x32_avg avx2 msa/, "$sse2_x86inc";
1049
1050 add_proto qw/unsigned int vpx_sad32x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1051 specialize qw/vpx_sad32x16_avg avx2 msa/, "$sse2_x86inc";
1052
1053 add_proto qw/unsigned int vpx_sad16x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1054 specialize qw/vpx_sad16x32_avg msa/, "$sse2_x86inc";
1055
1056 add_proto qw/unsigned int vpx_sad16x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1057 specialize qw/vpx_sad16x16_avg msa/, "$sse2_x86inc";
1058
1059 add_proto qw/unsigned int vpx_sad16x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1060 specialize qw/vpx_sad16x8_avg msa/, "$sse2_x86inc";
1061
1062 add_proto qw/unsigned int vpx_sad8x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1063 specialize qw/vpx_sad8x16_avg msa/, "$sse2_x86inc";
1064
1065 add_proto qw/unsigned int vpx_sad8x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1066 specialize qw/vpx_sad8x8_avg msa/, "$sse2_x86inc";
1067
1068 add_proto qw/unsigned int vpx_sad8x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1069 specialize qw/vpx_sad8x4_avg msa/, "$sse2_x86inc";
1070
1071 add_proto qw/unsigned int vpx_sad4x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1072 specialize qw/vpx_sad4x8_avg msa/, "$sse2_x86inc";
1073
1074 add_proto qw/unsigned int vpx_sad4x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1075 specialize qw/vpx_sad4x4_avg msa/, "$sse2_x86inc";
1076
1077 #
1078 # Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally
1079 #
1080 # Blocks of 3
1081 add_proto qw/void vpx_sad64x64x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1082 specialize qw/vpx_sad64x64x3 msa/;
1083
1084 add_proto qw/void vpx_sad32x32x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1085 specialize qw/vpx_sad32x32x3 msa/;
1086
1087 add_proto qw/void vpx_sad16x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1088 specialize qw/vpx_sad16x16x3 sse3 ssse3 msa/;
1089
1090 add_proto qw/void vpx_sad16x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1091 specialize qw/vpx_sad16x8x3 sse3 ssse3 msa/;
1092
1093 add_proto qw/void vpx_sad8x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1094 specialize qw/vpx_sad8x16x3 sse3 msa/;
1095
1096 add_proto qw/void vpx_sad8x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1097 specialize qw/vpx_sad8x8x3 sse3 msa/;
1098
1099 add_proto qw/void vpx_sad4x4x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1100 specialize qw/vpx_sad4x4x3 sse3 msa/;
1101
1102 # Blocks of 8
1103 add_proto qw/void vpx_sad64x64x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1104 specialize qw/vpx_sad64x64x8 msa/;
1105
1106 add_proto qw/void vpx_sad32x32x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1107 specialize qw/vpx_sad32x32x8 msa/;
1108
1109 add_proto qw/void vpx_sad16x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1110 specialize qw/vpx_sad16x16x8 sse4_1 msa/;
1111
1112 add_proto qw/void vpx_sad16x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1113 specialize qw/vpx_sad16x8x8 sse4_1 msa/;
1114
1115 add_proto qw/void vpx_sad8x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1116 specialize qw/vpx_sad8x16x8 sse4_1 msa/;
1117
1118 add_proto qw/void vpx_sad8x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1119 specialize qw/vpx_sad8x8x8 sse4_1 msa/;
1120
1121 add_proto qw/void vpx_sad8x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1122 specialize qw/vpx_sad8x4x8 msa/;
1123
1124 add_proto qw/void vpx_sad4x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1125 specialize qw/vpx_sad4x8x8 msa/;
1126
1127 add_proto qw/void vpx_sad4x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1128 specialize qw/vpx_sad4x4x8 sse4_1 msa/;
1129
1130 #
1131 # Multi-block SAD, comparing a reference to N independent blocks
1132 #
1133 add_proto qw/void vpx_sad64x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1134 specialize qw/vpx_sad64x64x4d avx2 neon msa/, "$sse2_x86inc";
1135
1136 add_proto qw/void vpx_sad64x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1137 specialize qw/vpx_sad64x32x4d msa/, "$sse2_x86inc";
1138
1139 add_proto qw/void vpx_sad32x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1140 specialize qw/vpx_sad32x64x4d msa/, "$sse2_x86inc";
1141
1142 add_proto qw/void vpx_sad32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1143 specialize qw/vpx_sad32x32x4d avx2 neon msa/, "$sse2_x86inc";
1144
1145 add_proto qw/void vpx_sad32x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1146 specialize qw/vpx_sad32x16x4d msa/, "$sse2_x86inc";
1147
1148 add_proto qw/void vpx_sad16x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1149 specialize qw/vpx_sad16x32x4d msa/, "$sse2_x86inc";
1150
1151 add_proto qw/void vpx_sad16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1152 specialize qw/vpx_sad16x16x4d neon msa/, "$sse2_x86inc";
1153
1154 add_proto qw/void vpx_sad16x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1155 specialize qw/vpx_sad16x8x4d msa/, "$sse2_x86inc";
1156
1157 add_proto qw/void vpx_sad8x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1158 specialize qw/vpx_sad8x16x4d msa/, "$sse2_x86inc";
1159
1160 add_proto qw/void vpx_sad8x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1161 specialize qw/vpx_sad8x8x4d msa/, "$sse2_x86inc";
1162
1163 add_proto qw/void vpx_sad8x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1164 specialize qw/vpx_sad8x4x4d msa/, "$sse2_x86inc";
1165
1166 add_proto qw/void vpx_sad4x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1167 specialize qw/vpx_sad4x8x4d msa/, "$sse2_x86inc";
1168
1169 add_proto qw/void vpx_sad4x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1170 specialize qw/vpx_sad4x4x4d msa/, "$sse2_x86inc";
1171
1172 #
1173 # Structured Similarity (SSIM)
1174 #
1175 if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
1176     add_proto qw/void vpx_ssim_parms_8x8/, "const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
1177     specialize qw/vpx_ssim_parms_8x8/, "$sse2_x86_64";
1178
1179     add_proto qw/void vpx_ssim_parms_16x16/, "const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
1180     specialize qw/vpx_ssim_parms_16x16/, "$sse2_x86_64";
1181 }
1182
1183 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
1184   #
1185   # Block subtraction
1186   #
1187   add_proto qw/void vpx_highbd_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride, int bd";
1188   specialize qw/vpx_highbd_subtract_block/;
1189
1190   #
1191   # Single block SAD
1192   #
1193   add_proto qw/unsigned int vpx_highbd_sad64x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1194   specialize qw/vpx_highbd_sad64x64/, "$sse2_x86inc";
1195
1196   add_proto qw/unsigned int vpx_highbd_sad64x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1197   specialize qw/vpx_highbd_sad64x32/, "$sse2_x86inc";
1198
1199   add_proto qw/unsigned int vpx_highbd_sad32x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1200   specialize qw/vpx_highbd_sad32x64/, "$sse2_x86inc";
1201
1202   add_proto qw/unsigned int vpx_highbd_sad32x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1203   specialize qw/vpx_highbd_sad32x32/, "$sse2_x86inc";
1204
1205   add_proto qw/unsigned int vpx_highbd_sad32x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1206   specialize qw/vpx_highbd_sad32x16/, "$sse2_x86inc";
1207
1208   add_proto qw/unsigned int vpx_highbd_sad16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1209   specialize qw/vpx_highbd_sad16x32/, "$sse2_x86inc";
1210
1211   add_proto qw/unsigned int vpx_highbd_sad16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1212   specialize qw/vpx_highbd_sad16x16/, "$sse2_x86inc";
1213
1214   add_proto qw/unsigned int vpx_highbd_sad16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1215   specialize qw/vpx_highbd_sad16x8/, "$sse2_x86inc";
1216
1217   add_proto qw/unsigned int vpx_highbd_sad8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1218   specialize qw/vpx_highbd_sad8x16/, "$sse2_x86inc";
1219
1220   add_proto qw/unsigned int vpx_highbd_sad8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1221   specialize qw/vpx_highbd_sad8x8/, "$sse2_x86inc";
1222
1223   add_proto qw/unsigned int vpx_highbd_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1224   specialize qw/vpx_highbd_sad8x4/, "$sse2_x86inc";
1225
1226   add_proto qw/unsigned int vpx_highbd_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1227   specialize qw/vpx_highbd_sad4x8/;
1228
1229   add_proto qw/unsigned int vpx_highbd_sad4x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1230   specialize qw/vpx_highbd_sad4x4/;
1231
1232   #
1233   # Avg
1234   #
1235   add_proto qw/unsigned int vpx_highbd_avg_8x8/, "const uint8_t *, int p";
1236   specialize qw/vpx_highbd_avg_8x8/;
1237   add_proto qw/unsigned int vpx_highbd_avg_4x4/, "const uint8_t *, int p";
1238   specialize qw/vpx_highbd_avg_4x4/;
1239   add_proto qw/void vpx_highbd_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
1240   specialize qw/vpx_highbd_minmax_8x8/;
1241
1242   add_proto qw/unsigned int vpx_highbd_sad64x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1243   specialize qw/vpx_highbd_sad64x64_avg/, "$sse2_x86inc";
1244
1245   add_proto qw/unsigned int vpx_highbd_sad64x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1246   specialize qw/vpx_highbd_sad64x32_avg/, "$sse2_x86inc";
1247
1248   add_proto qw/unsigned int vpx_highbd_sad32x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1249   specialize qw/vpx_highbd_sad32x64_avg/, "$sse2_x86inc";
1250
1251   add_proto qw/unsigned int vpx_highbd_sad32x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1252   specialize qw/vpx_highbd_sad32x32_avg/, "$sse2_x86inc";
1253
1254   add_proto qw/unsigned int vpx_highbd_sad32x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1255   specialize qw/vpx_highbd_sad32x16_avg/, "$sse2_x86inc";
1256
1257   add_proto qw/unsigned int vpx_highbd_sad16x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1258   specialize qw/vpx_highbd_sad16x32_avg/, "$sse2_x86inc";
1259
1260   add_proto qw/unsigned int vpx_highbd_sad16x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1261   specialize qw/vpx_highbd_sad16x16_avg/, "$sse2_x86inc";
1262
1263   add_proto qw/unsigned int vpx_highbd_sad16x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1264   specialize qw/vpx_highbd_sad16x8_avg/, "$sse2_x86inc";
1265
1266   add_proto qw/unsigned int vpx_highbd_sad8x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1267   specialize qw/vpx_highbd_sad8x16_avg/, "$sse2_x86inc";
1268
1269   add_proto qw/unsigned int vpx_highbd_sad8x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1270   specialize qw/vpx_highbd_sad8x8_avg/, "$sse2_x86inc";
1271
1272   add_proto qw/unsigned int vpx_highbd_sad8x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1273   specialize qw/vpx_highbd_sad8x4_avg/, "$sse2_x86inc";
1274
1275   add_proto qw/unsigned int vpx_highbd_sad4x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1276   specialize qw/vpx_highbd_sad4x8_avg/;
1277
1278   add_proto qw/unsigned int vpx_highbd_sad4x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1279   specialize qw/vpx_highbd_sad4x4_avg/;
1280
1281   #
1282   # Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally
1283   #
1284   # Blocks of 3
1285   add_proto qw/void vpx_highbd_sad64x64x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1286   specialize qw/vpx_highbd_sad64x64x3/;
1287
1288   add_proto qw/void vpx_highbd_sad32x32x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1289   specialize qw/vpx_highbd_sad32x32x3/;
1290
1291   add_proto qw/void vpx_highbd_sad16x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1292   specialize qw/vpx_highbd_sad16x16x3/;
1293
1294   add_proto qw/void vpx_highbd_sad16x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1295   specialize qw/vpx_highbd_sad16x8x3/;
1296
1297   add_proto qw/void vpx_highbd_sad8x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1298   specialize qw/vpx_highbd_sad8x16x3/;
1299
1300   add_proto qw/void vpx_highbd_sad8x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1301   specialize qw/vpx_highbd_sad8x8x3/;
1302
1303   add_proto qw/void vpx_highbd_sad4x4x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1304   specialize qw/vpx_highbd_sad4x4x3/;
1305
1306   # Blocks of 8
1307   add_proto qw/void vpx_highbd_sad64x64x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1308   specialize qw/vpx_highbd_sad64x64x8/;
1309
1310   add_proto qw/void vpx_highbd_sad32x32x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1311   specialize qw/vpx_highbd_sad32x32x8/;
1312
1313   add_proto qw/void vpx_highbd_sad16x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1314   specialize qw/vpx_highbd_sad16x16x8/;
1315
1316   add_proto qw/void vpx_highbd_sad16x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1317   specialize qw/vpx_highbd_sad16x8x8/;
1318
1319   add_proto qw/void vpx_highbd_sad8x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1320   specialize qw/vpx_highbd_sad8x16x8/;
1321
1322   add_proto qw/void vpx_highbd_sad8x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1323   specialize qw/vpx_highbd_sad8x8x8/;
1324
1325   add_proto qw/void vpx_highbd_sad8x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1326   specialize qw/vpx_highbd_sad8x4x8/;
1327
1328   add_proto qw/void vpx_highbd_sad4x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1329   specialize qw/vpx_highbd_sad4x8x8/;
1330
1331   add_proto qw/void vpx_highbd_sad4x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1332   specialize qw/vpx_highbd_sad4x4x8/;
1333
1334   #
1335   # Multi-block SAD, comparing a reference to N independent blocks
1336   #
1337   add_proto qw/void vpx_highbd_sad64x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1338   specialize qw/vpx_highbd_sad64x64x4d/, "$sse2_x86inc";
1339
1340   add_proto qw/void vpx_highbd_sad64x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1341   specialize qw/vpx_highbd_sad64x32x4d/, "$sse2_x86inc";
1342
1343   add_proto qw/void vpx_highbd_sad32x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1344   specialize qw/vpx_highbd_sad32x64x4d/, "$sse2_x86inc";
1345
1346   add_proto qw/void vpx_highbd_sad32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1347   specialize qw/vpx_highbd_sad32x32x4d/, "$sse2_x86inc";
1348
1349   add_proto qw/void vpx_highbd_sad32x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1350   specialize qw/vpx_highbd_sad32x16x4d/, "$sse2_x86inc";
1351
1352   add_proto qw/void vpx_highbd_sad16x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1353   specialize qw/vpx_highbd_sad16x32x4d/, "$sse2_x86inc";
1354
1355   add_proto qw/void vpx_highbd_sad16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1356   specialize qw/vpx_highbd_sad16x16x4d/, "$sse2_x86inc";
1357
1358   add_proto qw/void vpx_highbd_sad16x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1359   specialize qw/vpx_highbd_sad16x8x4d/, "$sse2_x86inc";
1360
1361   add_proto qw/void vpx_highbd_sad8x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1362   specialize qw/vpx_highbd_sad8x16x4d/, "$sse2_x86inc";
1363
1364   add_proto qw/void vpx_highbd_sad8x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1365   specialize qw/vpx_highbd_sad8x8x4d/, "$sse2_x86inc";
1366
1367   add_proto qw/void vpx_highbd_sad8x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1368   specialize qw/vpx_highbd_sad8x4x4d/, "$sse2_x86inc";
1369
1370   add_proto qw/void vpx_highbd_sad4x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1371   specialize qw/vpx_highbd_sad4x8x4d/, "$sse2_x86inc";
1372
1373   add_proto qw/void vpx_highbd_sad4x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1374   specialize qw/vpx_highbd_sad4x4x4d/, "$sse2_x86inc";
1375
1376   #
1377   # Structured Similarity (SSIM)
1378   #
1379   if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
1380     add_proto qw/void vpx_highbd_ssim_parms_8x8/, "const uint16_t *s, int sp, const uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
1381     specialize qw/vpx_highbd_ssim_parms_8x8/;
1382   }
1383 }  # CONFIG_VP9_HIGHBITDEPTH
1384 }  # CONFIG_ENCODERS
1385
1386 if (vpx_config("CONFIG_ENCODERS") eq "yes" || vpx_config("CONFIG_POSTPROC") eq "yes" || vpx_config("CONFIG_VP9_POSTPROC") eq "yes") {
1387
1388 #
1389 # Variance
1390 #
1391 add_proto qw/unsigned int vpx_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1392   specialize qw/vpx_variance64x64 sse2 avx2 neon msa/;
1393
1394 add_proto qw/unsigned int vpx_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1395   specialize qw/vpx_variance64x32 sse2 avx2 neon msa/;
1396
1397 add_proto qw/unsigned int vpx_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1398   specialize qw/vpx_variance32x64 sse2 neon msa/;
1399
1400 add_proto qw/unsigned int vpx_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1401   specialize qw/vpx_variance32x32 sse2 avx2 neon msa/;
1402
1403 add_proto qw/unsigned int vpx_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1404   specialize qw/vpx_variance32x16 sse2 avx2 msa/;
1405
1406 add_proto qw/unsigned int vpx_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1407   specialize qw/vpx_variance16x32 sse2 msa/;
1408
1409 add_proto qw/unsigned int vpx_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1410   specialize qw/vpx_variance16x16 mmx sse2 avx2 media neon msa/;
1411
1412 add_proto qw/unsigned int vpx_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1413   specialize qw/vpx_variance16x8 mmx sse2 neon msa/;
1414
1415 add_proto qw/unsigned int vpx_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1416   specialize qw/vpx_variance8x16 mmx sse2 neon msa/;
1417
1418 add_proto qw/unsigned int vpx_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1419   specialize qw/vpx_variance8x8 mmx sse2 media neon msa/;
1420
1421 add_proto qw/unsigned int vpx_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1422   specialize qw/vpx_variance8x4 sse2 msa/;
1423
1424 add_proto qw/unsigned int vpx_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1425   specialize qw/vpx_variance4x8 sse2 msa/;
1426
1427 add_proto qw/unsigned int vpx_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1428   specialize qw/vpx_variance4x4 mmx sse2 msa/;
1429
1430 #
1431 # Specialty Variance
1432 #
1433 add_proto qw/void vpx_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1434   specialize qw/vpx_get16x16var sse2 avx2 neon msa/;
1435
1436 add_proto qw/void vpx_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1437   specialize qw/vpx_get8x8var mmx sse2 neon msa/;
1438
1439 add_proto qw/unsigned int vpx_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1440   specialize qw/vpx_mse16x16 mmx sse2 avx2 media neon msa/;
1441
1442 add_proto qw/unsigned int vpx_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1443   specialize qw/vpx_mse16x8 sse2 msa/;
1444
1445 add_proto qw/unsigned int vpx_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1446   specialize qw/vpx_mse8x16 sse2 msa/;
1447
1448 add_proto qw/unsigned int vpx_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1449   specialize qw/vpx_mse8x8 sse2 msa/;
1450
1451 add_proto qw/unsigned int vpx_get_mb_ss/, "const int16_t *";
1452   specialize qw/vpx_get_mb_ss mmx sse2 msa/;
1453
1454 add_proto qw/unsigned int vpx_get4x4sse_cs/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride";
1455   specialize qw/vpx_get4x4sse_cs neon msa/;
1456
1457 add_proto qw/void vpx_comp_avg_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride";
1458
1459 #
1460 # Subpixel Variance
1461 #
1462 add_proto qw/uint32_t vpx_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1463   specialize qw/vpx_sub_pixel_variance64x64 avx2 neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
1464
1465 add_proto qw/uint32_t vpx_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1466   specialize qw/vpx_sub_pixel_variance64x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1467
1468 add_proto qw/uint32_t vpx_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1469   specialize qw/vpx_sub_pixel_variance32x64 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1470
1471 add_proto qw/uint32_t vpx_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1472   specialize qw/vpx_sub_pixel_variance32x32 avx2 neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
1473
1474 add_proto qw/uint32_t vpx_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1475   specialize qw/vpx_sub_pixel_variance32x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1476
1477 add_proto qw/uint32_t vpx_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1478   specialize qw/vpx_sub_pixel_variance16x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1479
1480 add_proto qw/uint32_t vpx_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1481   specialize qw/vpx_sub_pixel_variance16x16 mmx media neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
1482
1483 add_proto qw/uint32_t vpx_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1484   specialize qw/vpx_sub_pixel_variance16x8 mmx msa/, "$sse2_x86inc", "$ssse3_x86inc";
1485
1486 add_proto qw/uint32_t vpx_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1487   specialize qw/vpx_sub_pixel_variance8x16 mmx msa/, "$sse2_x86inc", "$ssse3_x86inc";
1488
1489 add_proto qw/uint32_t vpx_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1490   specialize qw/vpx_sub_pixel_variance8x8 mmx media neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
1491
1492 add_proto qw/uint32_t vpx_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1493   specialize qw/vpx_sub_pixel_variance8x4 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1494
1495 add_proto qw/uint32_t vpx_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1496   specialize qw/vpx_sub_pixel_variance4x8 msa/, "$sse_x86inc", "$ssse3_x86inc";
1497
1498 add_proto qw/uint32_t vpx_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1499   specialize qw/vpx_sub_pixel_variance4x4 mmx msa/, "$sse_x86inc", "$ssse3_x86inc";
1500
1501 add_proto qw/uint32_t vpx_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1502   specialize qw/vpx_sub_pixel_avg_variance64x64 avx2 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1503
1504 add_proto qw/uint32_t vpx_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1505   specialize qw/vpx_sub_pixel_avg_variance64x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1506
1507 add_proto qw/uint32_t vpx_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1508   specialize qw/vpx_sub_pixel_avg_variance32x64 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1509
1510 add_proto qw/uint32_t vpx_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1511   specialize qw/vpx_sub_pixel_avg_variance32x32 avx2 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1512
1513 add_proto qw/uint32_t vpx_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1514   specialize qw/vpx_sub_pixel_avg_variance32x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1515
1516 add_proto qw/uint32_t vpx_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1517   specialize qw/vpx_sub_pixel_avg_variance16x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1518
1519 add_proto qw/uint32_t vpx_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1520   specialize qw/vpx_sub_pixel_avg_variance16x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1521
1522 add_proto qw/uint32_t vpx_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1523   specialize qw/vpx_sub_pixel_avg_variance16x8 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1524
1525 add_proto qw/uint32_t vpx_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1526   specialize qw/vpx_sub_pixel_avg_variance8x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1527
1528 add_proto qw/uint32_t vpx_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1529   specialize qw/vpx_sub_pixel_avg_variance8x8 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1530
1531 add_proto qw/uint32_t vpx_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1532   specialize qw/vpx_sub_pixel_avg_variance8x4 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1533
1534 add_proto qw/uint32_t vpx_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1535   specialize qw/vpx_sub_pixel_avg_variance4x8 msa/, "$sse_x86inc", "$ssse3_x86inc";
1536
1537 add_proto qw/uint32_t vpx_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1538   specialize qw/vpx_sub_pixel_avg_variance4x4 msa/, "$sse_x86inc", "$ssse3_x86inc";
1539
1540 #
1541 # Specialty Subpixel
1542 #
1543 add_proto qw/uint32_t vpx_variance_halfpixvar16x16_h/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse";
1544   specialize qw/vpx_variance_halfpixvar16x16_h mmx sse2 media/;
1545
1546 add_proto qw/uint32_t vpx_variance_halfpixvar16x16_v/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse";
1547   specialize qw/vpx_variance_halfpixvar16x16_v mmx sse2 media/;
1548
1549 add_proto qw/uint32_t vpx_variance_halfpixvar16x16_hv/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse";
1550   specialize qw/vpx_variance_halfpixvar16x16_hv mmx sse2 media/;
1551
1552 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
1553   add_proto qw/unsigned int vpx_highbd_12_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1554   specialize qw/vpx_highbd_12_variance64x64 sse2/;
1555
1556   add_proto qw/unsigned int vpx_highbd_12_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1557   specialize qw/vpx_highbd_12_variance64x32 sse2/;
1558
1559   add_proto qw/unsigned int vpx_highbd_12_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1560   specialize qw/vpx_highbd_12_variance32x64 sse2/;
1561
1562   add_proto qw/unsigned int vpx_highbd_12_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1563   specialize qw/vpx_highbd_12_variance32x32 sse2/;
1564
1565   add_proto qw/unsigned int vpx_highbd_12_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1566   specialize qw/vpx_highbd_12_variance32x16 sse2/;
1567
1568   add_proto qw/unsigned int vpx_highbd_12_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1569   specialize qw/vpx_highbd_12_variance16x32 sse2/;
1570
1571   add_proto qw/unsigned int vpx_highbd_12_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1572   specialize qw/vpx_highbd_12_variance16x16 sse2/;
1573
1574   add_proto qw/unsigned int vpx_highbd_12_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1575   specialize qw/vpx_highbd_12_variance16x8 sse2/;
1576
1577   add_proto qw/unsigned int vpx_highbd_12_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1578   specialize qw/vpx_highbd_12_variance8x16 sse2/;
1579
1580   add_proto qw/unsigned int vpx_highbd_12_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1581   specialize qw/vpx_highbd_12_variance8x8 sse2/;
1582
1583   add_proto qw/unsigned int vpx_highbd_12_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1584   add_proto qw/unsigned int vpx_highbd_12_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1585   add_proto qw/unsigned int vpx_highbd_12_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1586
1587   add_proto qw/unsigned int vpx_highbd_10_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1588   specialize qw/vpx_highbd_10_variance64x64 sse2/;
1589
1590   add_proto qw/unsigned int vpx_highbd_10_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1591   specialize qw/vpx_highbd_10_variance64x32 sse2/;
1592
1593   add_proto qw/unsigned int vpx_highbd_10_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1594   specialize qw/vpx_highbd_10_variance32x64 sse2/;
1595
1596   add_proto qw/unsigned int vpx_highbd_10_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1597   specialize qw/vpx_highbd_10_variance32x32 sse2/;
1598
1599   add_proto qw/unsigned int vpx_highbd_10_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1600   specialize qw/vpx_highbd_10_variance32x16 sse2/;
1601
1602   add_proto qw/unsigned int vpx_highbd_10_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1603   specialize qw/vpx_highbd_10_variance16x32 sse2/;
1604
1605   add_proto qw/unsigned int vpx_highbd_10_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1606   specialize qw/vpx_highbd_10_variance16x16 sse2/;
1607
1608   add_proto qw/unsigned int vpx_highbd_10_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1609   specialize qw/vpx_highbd_10_variance16x8 sse2/;
1610
1611   add_proto qw/unsigned int vpx_highbd_10_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1612   specialize qw/vpx_highbd_10_variance8x16 sse2/;
1613
1614   add_proto qw/unsigned int vpx_highbd_10_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1615   specialize qw/vpx_highbd_10_variance8x8 sse2/;
1616
1617   add_proto qw/unsigned int vpx_highbd_10_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1618   add_proto qw/unsigned int vpx_highbd_10_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1619   add_proto qw/unsigned int vpx_highbd_10_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1620
1621   add_proto qw/unsigned int vpx_highbd_8_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1622   specialize qw/vpx_highbd_8_variance64x64 sse2/;
1623
1624   add_proto qw/unsigned int vpx_highbd_8_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1625   specialize qw/vpx_highbd_8_variance64x32 sse2/;
1626
1627   add_proto qw/unsigned int vpx_highbd_8_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1628   specialize qw/vpx_highbd_8_variance32x64 sse2/;
1629
1630   add_proto qw/unsigned int vpx_highbd_8_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1631   specialize qw/vpx_highbd_8_variance32x32 sse2/;
1632
1633   add_proto qw/unsigned int vpx_highbd_8_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1634   specialize qw/vpx_highbd_8_variance32x16 sse2/;
1635
1636   add_proto qw/unsigned int vpx_highbd_8_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1637   specialize qw/vpx_highbd_8_variance16x32 sse2/;
1638
1639   add_proto qw/unsigned int vpx_highbd_8_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1640   specialize qw/vpx_highbd_8_variance16x16 sse2/;
1641
1642   add_proto qw/unsigned int vpx_highbd_8_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1643   specialize qw/vpx_highbd_8_variance16x8 sse2/;
1644
1645   add_proto qw/unsigned int vpx_highbd_8_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1646   specialize qw/vpx_highbd_8_variance8x16 sse2/;
1647
1648   add_proto qw/unsigned int vpx_highbd_8_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1649   specialize qw/vpx_highbd_8_variance8x8 sse2/;
1650
1651   add_proto qw/unsigned int vpx_highbd_8_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1652   add_proto qw/unsigned int vpx_highbd_8_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1653   add_proto qw/unsigned int vpx_highbd_8_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1654
1655   add_proto qw/void vpx_highbd_8_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1656   add_proto qw/void vpx_highbd_8_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1657
1658   add_proto qw/void vpx_highbd_10_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1659   add_proto qw/void vpx_highbd_10_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1660
1661   add_proto qw/void vpx_highbd_12_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1662   add_proto qw/void vpx_highbd_12_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1663
1664   add_proto qw/unsigned int vpx_highbd_8_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1665   specialize qw/vpx_highbd_8_mse16x16 sse2/;
1666
1667   add_proto qw/unsigned int vpx_highbd_8_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1668   add_proto qw/unsigned int vpx_highbd_8_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1669   add_proto qw/unsigned int vpx_highbd_8_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1670   specialize qw/vpx_highbd_8_mse8x8 sse2/;
1671
1672   add_proto qw/unsigned int vpx_highbd_10_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1673   specialize qw/vpx_highbd_10_mse16x16 sse2/;
1674
1675   add_proto qw/unsigned int vpx_highbd_10_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1676   add_proto qw/unsigned int vpx_highbd_10_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1677   add_proto qw/unsigned int vpx_highbd_10_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1678   specialize qw/vpx_highbd_10_mse8x8 sse2/;
1679
1680   add_proto qw/unsigned int vpx_highbd_12_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1681   specialize qw/vpx_highbd_12_mse16x16 sse2/;
1682
1683   add_proto qw/unsigned int vpx_highbd_12_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1684   add_proto qw/unsigned int vpx_highbd_12_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1685   add_proto qw/unsigned int vpx_highbd_12_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1686   specialize qw/vpx_highbd_12_mse8x8 sse2/;
1687
1688   add_proto qw/void vpx_highbd_comp_avg_pred/, "uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride";
1689
1690   #
1691   # Subpixel Variance
1692   #
1693   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1694   specialize qw/vpx_highbd_12_sub_pixel_variance64x64/, "$sse2_x86inc";
1695
1696   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1697   specialize qw/vpx_highbd_12_sub_pixel_variance64x32/, "$sse2_x86inc";
1698
1699   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1700   specialize qw/vpx_highbd_12_sub_pixel_variance32x64/, "$sse2_x86inc";
1701
1702   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1703   specialize qw/vpx_highbd_12_sub_pixel_variance32x32/, "$sse2_x86inc";
1704
1705   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1706   specialize qw/vpx_highbd_12_sub_pixel_variance32x16/, "$sse2_x86inc";
1707
1708   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1709   specialize qw/vpx_highbd_12_sub_pixel_variance16x32/, "$sse2_x86inc";
1710
1711   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1712   specialize qw/vpx_highbd_12_sub_pixel_variance16x16/, "$sse2_x86inc";
1713
1714   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1715   specialize qw/vpx_highbd_12_sub_pixel_variance16x8/, "$sse2_x86inc";
1716
1717   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1718   specialize qw/vpx_highbd_12_sub_pixel_variance8x16/, "$sse2_x86inc";
1719
1720   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1721   specialize qw/vpx_highbd_12_sub_pixel_variance8x8/, "$sse2_x86inc";
1722
1723   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1724   specialize qw/vpx_highbd_12_sub_pixel_variance8x4/, "$sse2_x86inc";
1725
1726   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1727   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1728
1729   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1730   specialize qw/vpx_highbd_10_sub_pixel_variance64x64/, "$sse2_x86inc";
1731
1732   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1733   specialize qw/vpx_highbd_10_sub_pixel_variance64x32/, "$sse2_x86inc";
1734
1735   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1736   specialize qw/vpx_highbd_10_sub_pixel_variance32x64/, "$sse2_x86inc";
1737
1738   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1739   specialize qw/vpx_highbd_10_sub_pixel_variance32x32/, "$sse2_x86inc";
1740
1741   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1742   specialize qw/vpx_highbd_10_sub_pixel_variance32x16/, "$sse2_x86inc";
1743
1744   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1745   specialize qw/vpx_highbd_10_sub_pixel_variance16x32/, "$sse2_x86inc";
1746
1747   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1748   specialize qw/vpx_highbd_10_sub_pixel_variance16x16/, "$sse2_x86inc";
1749
1750   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1751   specialize qw/vpx_highbd_10_sub_pixel_variance16x8/, "$sse2_x86inc";
1752
1753   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1754   specialize qw/vpx_highbd_10_sub_pixel_variance8x16/, "$sse2_x86inc";
1755
1756   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1757   specialize qw/vpx_highbd_10_sub_pixel_variance8x8/, "$sse2_x86inc";
1758
1759   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1760   specialize qw/vpx_highbd_10_sub_pixel_variance8x4/, "$sse2_x86inc";
1761
1762   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1763   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1764
1765   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1766   specialize qw/vpx_highbd_8_sub_pixel_variance64x64/, "$sse2_x86inc";
1767
1768   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1769   specialize qw/vpx_highbd_8_sub_pixel_variance64x32/, "$sse2_x86inc";
1770
1771   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1772   specialize qw/vpx_highbd_8_sub_pixel_variance32x64/, "$sse2_x86inc";
1773
1774   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1775   specialize qw/vpx_highbd_8_sub_pixel_variance32x32/, "$sse2_x86inc";
1776
1777   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1778   specialize qw/vpx_highbd_8_sub_pixel_variance32x16/, "$sse2_x86inc";
1779
1780   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1781   specialize qw/vpx_highbd_8_sub_pixel_variance16x32/, "$sse2_x86inc";
1782
1783   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1784   specialize qw/vpx_highbd_8_sub_pixel_variance16x16/, "$sse2_x86inc";
1785
1786   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1787   specialize qw/vpx_highbd_8_sub_pixel_variance16x8/, "$sse2_x86inc";
1788
1789   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1790   specialize qw/vpx_highbd_8_sub_pixel_variance8x16/, "$sse2_x86inc";
1791
1792   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1793   specialize qw/vpx_highbd_8_sub_pixel_variance8x8/, "$sse2_x86inc";
1794
1795   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1796   specialize qw/vpx_highbd_8_sub_pixel_variance8x4/, "$sse2_x86inc";
1797
1798   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1799   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1800
1801   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1802   specialize qw/vpx_highbd_12_sub_pixel_avg_variance64x64/, "$sse2_x86inc";
1803
1804   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1805   specialize qw/vpx_highbd_12_sub_pixel_avg_variance64x32/, "$sse2_x86inc";
1806
1807   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1808   specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x64/, "$sse2_x86inc";
1809
1810   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1811   specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x32/, "$sse2_x86inc";
1812
1813   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1814   specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x16/, "$sse2_x86inc";
1815
1816   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1817   specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x32/, "$sse2_x86inc";
1818
1819   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1820   specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x16/, "$sse2_x86inc";
1821
1822   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1823   specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x8/, "$sse2_x86inc";
1824
1825   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1826   specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x16/, "$sse2_x86inc";
1827
1828   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1829   specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x8/, "$sse2_x86inc";
1830
1831   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1832   specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x4/, "$sse2_x86inc";
1833
1834   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1835   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1836
1837   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1838   specialize qw/vpx_highbd_10_sub_pixel_avg_variance64x64/, "$sse2_x86inc";
1839
1840   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1841   specialize qw/vpx_highbd_10_sub_pixel_avg_variance64x32/, "$sse2_x86inc";
1842
1843   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1844   specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x64/, "$sse2_x86inc";
1845
1846   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1847   specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x32/, "$sse2_x86inc";
1848
1849   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1850   specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x16/, "$sse2_x86inc";
1851
1852   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1853   specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x32/, "$sse2_x86inc";
1854
1855   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1856   specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x16/, "$sse2_x86inc";
1857
1858   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1859   specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x8/, "$sse2_x86inc";
1860
1861   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1862   specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x16/, "$sse2_x86inc";
1863
1864   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1865   specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x8/, "$sse2_x86inc";
1866
1867   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1868   specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x4/, "$sse2_x86inc";
1869
1870   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1871   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1872
1873   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1874   specialize qw/vpx_highbd_8_sub_pixel_avg_variance64x64/, "$sse2_x86inc";
1875
1876   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1877   specialize qw/vpx_highbd_8_sub_pixel_avg_variance64x32/, "$sse2_x86inc";
1878
1879   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1880   specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x64/, "$sse2_x86inc";
1881
1882   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1883   specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x32/, "$sse2_x86inc";
1884
1885   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1886   specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x16/, "$sse2_x86inc";
1887
1888   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1889   specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x32/, "$sse2_x86inc";
1890
1891   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1892   specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x16/, "$sse2_x86inc";
1893
1894   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1895   specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x8/, "$sse2_x86inc";
1896
1897   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1898   specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x16/, "$sse2_x86inc";
1899
1900   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1901   specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x8/, "$sse2_x86inc";
1902
1903   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1904   specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x4/, "$sse2_x86inc";
1905
1906   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1907   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1908
1909 }  # CONFIG_VP9_HIGHBITDEPTH
1910 }  # CONFIG_ENCODERS || CONFIG_POSTPROC || CONFIG_VP9_POSTPROC
1911
1912 1;