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, int count";
 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, int count";
 546 specialize qw/vpx_lpf_vertical_4 mmx neon dspr2 msa/;
 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_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
 552 specialize qw/vpx_lpf_horizontal_16 sse2 avx2 neon_asm dspr2 msa/;
 553 $vpx_lpf_horizontal_16_neon_asm=vpx_lpf_horizontal_16_neon;
 554
 555 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, int count";
 556 specialize qw/vpx_lpf_horizontal_8 sse2 neon dspr2 msa/;
 557
 558 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";
 559 specialize qw/vpx_lpf_horizontal_8_dual sse2 neon_asm dspr2 msa/;
 560 $vpx_lpf_horizontal_8_dual_neon_asm=vpx_lpf_horizontal_8_dual_neon;
 561
 562 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, int count";
 563 specialize qw/vpx_lpf_horizontal_4 mmx neon dspr2 msa/;
 564
 565 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";
 566 specialize qw/vpx_lpf_horizontal_4_dual sse2 neon dspr2 msa/;
 567
 568 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 569   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";
 570   specialize qw/vpx_highbd_lpf_vertical_16 sse2/;
 571
 572   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";
 573   specialize qw/vpx_highbd_lpf_vertical_16_dual sse2/;
 574
 575   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 count, int bd";
 576   specialize qw/vpx_highbd_lpf_vertical_8 sse2/;
 577
 578   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";
 579   specialize qw/vpx_highbd_lpf_vertical_8_dual sse2/;
 580
 581   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 count, int bd";
 582   specialize qw/vpx_highbd_lpf_vertical_4 sse2/;
 583
 584   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";
 585   specialize qw/vpx_highbd_lpf_vertical_4_dual sse2/;
 586
 587   add_proto qw/void vpx_highbd_lpf_horizontal_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
 588   specialize qw/vpx_highbd_lpf_horizontal_16 sse2/;
 589
 590   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 count, int bd";
 591   specialize qw/vpx_highbd_lpf_horizontal_8 sse2/;
 592
 593   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";
 594   specialize qw/vpx_highbd_lpf_horizontal_8_dual sse2/;
 595
 596   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 count, int bd";
 597   specialize qw/vpx_highbd_lpf_horizontal_4 sse2/;
 598
 599   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";
 600   specialize qw/vpx_highbd_lpf_horizontal_4_dual sse2/;
 601 }  # CONFIG_VP9_HIGHBITDEPTH
 602
 603 #
 604 # Encoder functions.
 605 #
 606
 607 #
 608 # Forward transform
 609 #
 610 if ((vpx_config("CONFIG_VP9_ENCODER") eq "yes") || (vpx_config("CONFIG_VP10_ENCODER") eq "yes")) {
 611 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 612   add_proto qw/void vpx_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
 613   specialize qw/vpx_fdct4x4 sse2/;
 614
 615   add_proto qw/void vpx_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
 616   specialize qw/vpx_fdct4x4_1 sse2/;
 617
 618   add_proto qw/void vpx_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
 619   specialize qw/vpx_fdct8x8 sse2/;
 620
 621   add_proto qw/void vpx_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
 622   specialize qw/vpx_fdct8x8_1 sse2/;
 623
 624   add_proto qw/void vpx_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
 625   specialize qw/vpx_fdct16x16 sse2/;
 626
 627   add_proto qw/void vpx_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
 628   specialize qw/vpx_fdct16x16_1 sse2/;
 629
 630   add_proto qw/void vpx_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
 631   specialize qw/vpx_fdct32x32 sse2/;
 632
 633   add_proto qw/void vpx_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
 634   specialize qw/vpx_fdct32x32_rd sse2/;
 635
 636   add_proto qw/void vpx_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
 637   specialize qw/vpx_fdct32x32_1 sse2/;
 638
 639   add_proto qw/void vpx_highbd_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
 640   specialize qw/vpx_highbd_fdct4x4 sse2/;
 641
 642   add_proto qw/void vpx_highbd_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
 643   specialize qw/vpx_highbd_fdct8x8 sse2/;
 644
 645   add_proto qw/void vpx_highbd_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
 646   specialize qw/vpx_highbd_fdct8x8_1/;
 647
 648   add_proto qw/void vpx_highbd_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
 649   specialize qw/vpx_highbd_fdct16x16 sse2/;
 650
 651   add_proto qw/void vpx_highbd_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
 652   specialize qw/vpx_highbd_fdct16x16_1/;
 653
 654   add_proto qw/void vpx_highbd_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
 655   specialize qw/vpx_highbd_fdct32x32 sse2/;
 656
 657   add_proto qw/void vpx_highbd_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
 658   specialize qw/vpx_highbd_fdct32x32_rd sse2/;
 659
 660   add_proto qw/void vpx_highbd_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
 661   specialize qw/vpx_highbd_fdct32x32_1/;
 662 } else {
 663   add_proto qw/void vpx_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
 664   specialize qw/vpx_fdct4x4 sse2 msa/;
 665
 666   add_proto qw/void vpx_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
 667   specialize qw/vpx_fdct4x4_1 sse2/;
 668
 669   add_proto qw/void vpx_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
 670   specialize qw/vpx_fdct8x8 sse2 neon msa/, "$ssse3_x86_64_x86inc";
 671
 672   add_proto qw/void vpx_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
 673   specialize qw/vpx_fdct8x8_1 sse2 neon msa/;
 674
 675   add_proto qw/void vpx_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
 676   specialize qw/vpx_fdct16x16 sse2 msa/;
 677
 678   add_proto qw/void vpx_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
 679   specialize qw/vpx_fdct16x16_1 sse2 msa/;
 680
 681   add_proto qw/void vpx_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
 682   specialize qw/vpx_fdct32x32 sse2 avx2 msa/;
 683
 684   add_proto qw/void vpx_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
 685   specialize qw/vpx_fdct32x32_rd sse2 avx2 msa/;
 686
 687   add_proto qw/void vpx_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
 688   specialize qw/vpx_fdct32x32_1 sse2 msa/;
 689 }  # CONFIG_VP9_HIGHBITDEPTH
 690 }  # CONFIG_VP9_ENCODER || CONFIG_VP10_ENCODER
 691
 692 #
 693 # Inverse transform
 694 if ((vpx_config("CONFIG_VP9") eq "yes") || (vpx_config("CONFIG_VP10") eq "yes")) {
 695 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 696   # Note as optimized versions of these functions are added we need to add a check to ensure
 697   # that when CONFIG_EMULATE_HARDWARE is on, it defaults to the C versions only.
 698   add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 699   specialize qw/vpx_iwht4x4_1_add/;
 700
 701   add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 702   specialize qw/vpx_iwht4x4_16_add/;
 703
 704   add_proto qw/void vpx_highbd_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 705   specialize qw/vpx_highbd_idct4x4_1_add/;
 706
 707   add_proto qw/void vpx_highbd_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 708   specialize qw/vpx_highbd_idct8x8_1_add/;
 709
 710   add_proto qw/void vpx_highbd_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 711   specialize qw/vpx_highbd_idct16x16_1_add/;
 712
 713   add_proto qw/void vpx_highbd_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 714   specialize qw/vpx_highbd_idct32x32_1024_add/;
 715
 716   add_proto qw/void vpx_highbd_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 717   specialize qw/vpx_highbd_idct32x32_34_add/;
 718
 719   add_proto qw/void vpx_highbd_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 720   specialize qw/vpx_highbd_idct32x32_1_add/;
 721
 722   add_proto qw/void vpx_highbd_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 723   specialize qw/vpx_highbd_iwht4x4_1_add/;
 724
 725   add_proto qw/void vpx_highbd_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 726   specialize qw/vpx_highbd_iwht4x4_16_add/;
 727
 728   # Force C versions if CONFIG_EMULATE_HARDWARE is 1
 729   if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
 730     add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 731     specialize qw/vpx_idct4x4_16_add/;
 732
 733     add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 734     specialize qw/vpx_idct4x4_1_add/;
 735
 736     add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 737     specialize qw/vpx_idct8x8_64_add/;
 738
 739     add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 740     specialize qw/vpx_idct8x8_12_add/;
 741
 742     add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 743     specialize qw/vpx_idct8x8_1_add/;
 744
 745     add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 746     specialize qw/vpx_idct16x16_256_add/;
 747
 748     add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 749     specialize qw/vpx_idct16x16_10_add/;
 750
 751     add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 752     specialize qw/vpx_idct16x16_1_add/;
 753
 754     add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 755     specialize qw/vpx_idct32x32_1024_add/;
 756
 757     add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 758     specialize qw/vpx_idct32x32_135_add/;
 759
 760     add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 761     specialize qw/vpx_idct32x32_34_add/;
 762
 763     add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 764     specialize qw/vpx_idct32x32_1_add/;
 765
 766     add_proto qw/void vpx_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 767     specialize qw/vpx_highbd_idct4x4_16_add/;
 768
 769     add_proto qw/void vpx_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 770     specialize qw/vpx_highbd_idct8x8_64_add/;
 771
 772     add_proto qw/void vpx_highbd_idct8x8_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 773     specialize qw/vpx_highbd_idct8x8_10_add/;
 774
 775     add_proto qw/void vpx_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 776     specialize qw/vpx_highbd_idct16x16_256_add/;
 777
 778     add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 779     specialize qw/vpx_highbd_idct16x16_10_add/;
 780   } else {
 781     add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 782     specialize qw/vpx_idct4x4_16_add sse2/;
 783
 784     add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 785     specialize qw/vpx_idct4x4_1_add sse2/;
 786
 787     add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 788     specialize qw/vpx_idct8x8_64_add sse2/;
 789
 790     add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 791     specialize qw/vpx_idct8x8_12_add sse2/;
 792
 793     add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 794     specialize qw/vpx_idct8x8_1_add sse2/;
 795
 796     add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 797     specialize qw/vpx_idct16x16_256_add sse2/;
 798
 799     add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 800     specialize qw/vpx_idct16x16_10_add sse2/;
 801
 802     add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 803     specialize qw/vpx_idct16x16_1_add sse2/;
 804
 805     add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 806     specialize qw/vpx_idct32x32_1024_add sse2/;
 807
 808     add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 809     specialize qw/vpx_idct32x32_135_add sse2/;
 810     $vpx_idct32x32_135_add_sse2=vpx_idct32x32_1024_add_sse2;
 811
 812     add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 813     specialize qw/vpx_idct32x32_34_add sse2/;
 814
 815     add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 816     specialize qw/vpx_idct32x32_1_add sse2/;
 817
 818     add_proto qw/void vpx_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 819     specialize qw/vpx_highbd_idct4x4_16_add sse2/;
 820
 821     add_proto qw/void vpx_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 822     specialize qw/vpx_highbd_idct8x8_64_add sse2/;
 823
 824     add_proto qw/void vpx_highbd_idct8x8_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 825     specialize qw/vpx_highbd_idct8x8_10_add sse2/;
 826
 827     add_proto qw/void vpx_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 828     specialize qw/vpx_highbd_idct16x16_256_add sse2/;
 829
 830     add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
 831     specialize qw/vpx_highbd_idct16x16_10_add sse2/;
 832   }  # CONFIG_EMULATE_HARDWARE
 833 } else {
 834   # Force C versions if CONFIG_EMULATE_HARDWARE is 1
 835   if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
 836     add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 837     specialize qw/vpx_idct4x4_1_add/;
 838
 839     add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 840     specialize qw/vpx_idct4x4_16_add/;
 841
 842     add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 843     specialize qw/vpx_idct8x8_1_add/;
 844
 845     add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 846     specialize qw/vpx_idct8x8_64_add/;
 847
 848     add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 849     specialize qw/vpx_idct8x8_12_add/;
 850
 851     add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 852     specialize qw/vpx_idct16x16_1_add/;
 853
 854     add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 855     specialize qw/vpx_idct16x16_256_add/;
 856
 857     add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 858     specialize qw/vpx_idct16x16_10_add/;
 859
 860     add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 861     specialize qw/vpx_idct32x32_1024_add/;
 862
 863     add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 864     specialize qw/vpx_idct32x32_135_add/;
 865
 866     add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 867     specialize qw/vpx_idct32x32_34_add/;
 868
 869     add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 870     specialize qw/vpx_idct32x32_1_add/;
 871
 872     add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 873     specialize qw/vpx_iwht4x4_1_add/;
 874
 875     add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 876     specialize qw/vpx_iwht4x4_16_add/;
 877   } else {
 878     add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 879     specialize qw/vpx_idct4x4_1_add sse2 neon dspr2 msa/;
 880
 881     add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 882     specialize qw/vpx_idct4x4_16_add sse2 neon dspr2 msa/;
 883
 884     add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 885     specialize qw/vpx_idct8x8_1_add sse2 neon dspr2 msa/;
 886
 887     add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 888     specialize qw/vpx_idct8x8_64_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
 889
 890     add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 891     specialize qw/vpx_idct8x8_12_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
 892
 893     add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 894     specialize qw/vpx_idct16x16_1_add sse2 neon dspr2 msa/;
 895
 896     add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 897     specialize qw/vpx_idct16x16_256_add sse2 neon dspr2 msa/;
 898
 899     add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 900     specialize qw/vpx_idct16x16_10_add sse2 neon dspr2 msa/;
 901
 902     add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 903     specialize qw/vpx_idct32x32_1024_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
 904
 905     add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 906     specialize qw/vpx_idct32x32_135_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
 907     # Need to add 135 eob idct32x32 implementations.
 908     $vpx_idct32x32_135_add_sse2=vpx_idct32x32_1024_add_sse2;
 909     $vpx_idct32x32_135_add_neon=vpx_idct32x32_1024_add_neon;
 910     $vpx_idct32x32_135_add_dspr2=vpx_idct32x32_1024_add_dspr2;
 911     $vpx_idct32x32_135_add_msa=vpx_idct32x32_1024_add_msa;
 912
 913     add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 914     specialize qw/vpx_idct32x32_34_add sse2 neon_asm dspr2 msa/, "$ssse3_x86_64_x86inc";
 915     # Need to add 34 eob idct32x32 neon implementation.
 916     $vpx_idct32x32_34_add_neon_asm=vpx_idct32x32_1024_add_neon;
 917
 918     add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 919     specialize qw/vpx_idct32x32_1_add sse2 neon dspr2 msa/;
 920
 921     add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 922     specialize qw/vpx_iwht4x4_1_add msa/;
 923
 924     add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
 925     specialize qw/vpx_iwht4x4_16_add msa/, "$sse2_x86inc";
 926   }  # CONFIG_EMULATE_HARDWARE
 927 }  # CONFIG_VP9_HIGHBITDEPTH
 928 }  # CONFIG_VP9 || CONFIG_VP10
 929
 930 #
 931 # Quantization
 932 #
 933 if ((vpx_config("CONFIG_VP9_ENCODER") eq "yes") || (vpx_config("CONFIG_VP10_ENCODER") eq "yes")) {
 934   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";
 935   specialize qw/vpx_quantize_b sse2/, "$ssse3_x86_64_x86inc", "$avx_x86_64_x86inc";
 936
 937   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";
 938   specialize qw/vpx_quantize_b_32x32/, "$ssse3_x86_64_x86inc", "$avx_x86_64_x86inc";
 939
 940   if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 941     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";
 942     specialize qw/vpx_highbd_quantize_b sse2/;
 943
 944     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";
 945     specialize qw/vpx_highbd_quantize_b_32x32 sse2/;
 946   }  # CONFIG_VP9_HIGHBITDEPTH
 947 }  # CONFIG_VP9_ENCODER || CONFIG_VP10_ENCODER
 948
 949 if (vpx_config("CONFIG_ENCODERS") eq "yes") {
 950 #
 951 # Block subtraction
 952 #
 953 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";
 954 specialize qw/vpx_subtract_block neon msa/, "$sse2_x86inc";
 955
 956 #
 957 # Single block SAD
 958 #
 959 add_proto qw/unsigned int vpx_sad64x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 960 specialize qw/vpx_sad64x64 avx2 neon msa/, "$sse2_x86inc";
 961
 962 add_proto qw/unsigned int vpx_sad64x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 963 specialize qw/vpx_sad64x32 avx2 msa/, "$sse2_x86inc";
 964
 965 add_proto qw/unsigned int vpx_sad32x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 966 specialize qw/vpx_sad32x64 avx2 msa/, "$sse2_x86inc";
 967
 968 add_proto qw/unsigned int vpx_sad32x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 969 specialize qw/vpx_sad32x32 avx2 neon msa/, "$sse2_x86inc";
 970
 971 add_proto qw/unsigned int vpx_sad32x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 972 specialize qw/vpx_sad32x16 avx2 msa/, "$sse2_x86inc";
 973
 974 add_proto qw/unsigned int vpx_sad16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 975 specialize qw/vpx_sad16x32 msa/, "$sse2_x86inc";
 976
 977 add_proto qw/unsigned int vpx_sad16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 978 specialize qw/vpx_sad16x16 mmx media neon msa/, "$sse2_x86inc";
 979
 980 add_proto qw/unsigned int vpx_sad16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 981 specialize qw/vpx_sad16x8 mmx neon msa/, "$sse2_x86inc";
 982
 983 add_proto qw/unsigned int vpx_sad8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 984 specialize qw/vpx_sad8x16 mmx neon msa/, "$sse2_x86inc";
 985
 986 add_proto qw/unsigned int vpx_sad8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 987 specialize qw/vpx_sad8x8 mmx neon msa/, "$sse2_x86inc";
 988
 989 add_proto qw/unsigned int vpx_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 990 specialize qw/vpx_sad8x4 msa/, "$sse2_x86inc";
 991
 992 add_proto qw/unsigned int vpx_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 993 specialize qw/vpx_sad4x8 msa/, "$sse2_x86inc";
 994
 995 add_proto qw/unsigned int vpx_sad4x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 996 specialize qw/vpx_sad4x4 mmx neon msa/, "$sse2_x86inc";
 997
 998 #
 999 # Avg
1000 #
1001 if ((vpx_config("CONFIG_VP9_ENCODER") eq "yes") || (vpx_config("CONFIG_VP10_ENCODER") eq "yes")) {
1002   add_proto qw/unsigned int vpx_avg_8x8/, "const uint8_t *, int p";
1003   specialize qw/vpx_avg_8x8 sse2 neon msa/;
1004
1005   add_proto qw/unsigned int vpx_avg_4x4/, "const uint8_t *, int p";
1006   specialize qw/vpx_avg_4x4 sse2 neon msa/;
1007
1008   add_proto qw/void vpx_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
1009   specialize qw/vpx_minmax_8x8 sse2/;
1010
1011   add_proto qw/void vpx_hadamard_8x8/, "int16_t const *src_diff, int src_stride, int16_t *coeff";
1012   specialize qw/vpx_hadamard_8x8 sse2/, "$ssse3_x86_64_x86inc";
1013
1014   add_proto qw/void vpx_hadamard_16x16/, "int16_t const *src_diff, int src_stride, int16_t *coeff";
1015   specialize qw/vpx_hadamard_16x16 sse2/;
1016
1017   add_proto qw/int vpx_satd/, "const int16_t *coeff, int length";
1018   specialize qw/vpx_satd sse2 neon/;
1019
1020   add_proto qw/void vpx_int_pro_row/, "int16_t *hbuf, uint8_t const *ref, const int ref_stride, const int height";
1021   specialize qw/vpx_int_pro_row sse2 neon/;
1022
1023   add_proto qw/int16_t vpx_int_pro_col/, "uint8_t const *ref, const int width";
1024   specialize qw/vpx_int_pro_col sse2 neon/;
1025
1026   add_proto qw/int vpx_vector_var/, "int16_t const *ref, int16_t const *src, const int bwl";
1027   specialize qw/vpx_vector_var neon sse2/;
1028 }  # CONFIG_VP9_ENCODER || CONFIG_VP10_ENCODER
1029
1030 add_proto qw/unsigned int vpx_sad64x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1031 specialize qw/vpx_sad64x64_avg avx2 msa/, "$sse2_x86inc";
1032
1033 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";
1034 specialize qw/vpx_sad64x32_avg avx2 msa/, "$sse2_x86inc";
1035
1036 add_proto qw/unsigned int vpx_sad32x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1037 specialize qw/vpx_sad32x64_avg avx2 msa/, "$sse2_x86inc";
1038
1039 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";
1040 specialize qw/vpx_sad32x32_avg avx2 msa/, "$sse2_x86inc";
1041
1042 add_proto qw/unsigned int vpx_sad32x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1043 specialize qw/vpx_sad32x16_avg avx2 msa/, "$sse2_x86inc";
1044
1045 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";
1046 specialize qw/vpx_sad16x32_avg msa/, "$sse2_x86inc";
1047
1048 add_proto qw/unsigned int vpx_sad16x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1049 specialize qw/vpx_sad16x16_avg msa/, "$sse2_x86inc";
1050
1051 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";
1052 specialize qw/vpx_sad16x8_avg msa/, "$sse2_x86inc";
1053
1054 add_proto qw/unsigned int vpx_sad8x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1055 specialize qw/vpx_sad8x16_avg msa/, "$sse2_x86inc";
1056
1057 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";
1058 specialize qw/vpx_sad8x8_avg msa/, "$sse2_x86inc";
1059
1060 add_proto qw/unsigned int vpx_sad8x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1061 specialize qw/vpx_sad8x4_avg msa/, "$sse2_x86inc";
1062
1063 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";
1064 specialize qw/vpx_sad4x8_avg msa/, "$sse2_x86inc";
1065
1066 add_proto qw/unsigned int vpx_sad4x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1067 specialize qw/vpx_sad4x4_avg msa/, "$sse2_x86inc";
1068
1069 #
1070 # Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally
1071 #
1072 # Blocks of 3
1073 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";
1074 specialize qw/vpx_sad64x64x3 msa/;
1075
1076 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";
1077 specialize qw/vpx_sad32x32x3 msa/;
1078
1079 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";
1080 specialize qw/vpx_sad16x16x3 sse3 ssse3 msa/;
1081
1082 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";
1083 specialize qw/vpx_sad16x8x3 sse3 ssse3 msa/;
1084
1085 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";
1086 specialize qw/vpx_sad8x16x3 sse3 msa/;
1087
1088 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";
1089 specialize qw/vpx_sad8x8x3 sse3 msa/;
1090
1091 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";
1092 specialize qw/vpx_sad4x4x3 sse3 msa/;
1093
1094 # Blocks of 8
1095 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";
1096 specialize qw/vpx_sad64x64x8 msa/;
1097
1098 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";
1099 specialize qw/vpx_sad32x32x8 msa/;
1100
1101 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";
1102 specialize qw/vpx_sad16x16x8 sse4_1 msa/;
1103
1104 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";
1105 specialize qw/vpx_sad16x8x8 sse4_1 msa/;
1106
1107 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";
1108 specialize qw/vpx_sad8x16x8 sse4_1 msa/;
1109
1110 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";
1111 specialize qw/vpx_sad8x8x8 sse4_1 msa/;
1112
1113 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";
1114 specialize qw/vpx_sad8x4x8 msa/;
1115
1116 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";
1117 specialize qw/vpx_sad4x8x8 msa/;
1118
1119 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";
1120 specialize qw/vpx_sad4x4x8 sse4_1 msa/;
1121
1122 #
1123 # Multi-block SAD, comparing a reference to N independent blocks
1124 #
1125 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";
1126 specialize qw/vpx_sad64x64x4d avx2 neon msa/, "$sse2_x86inc";
1127
1128 add_proto qw/void vpx_sad64x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1129 specialize qw/vpx_sad64x32x4d msa/, "$sse2_x86inc";
1130
1131 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";
1132 specialize qw/vpx_sad32x64x4d msa/, "$sse2_x86inc";
1133
1134 add_proto qw/void vpx_sad32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1135 specialize qw/vpx_sad32x32x4d avx2 neon msa/, "$sse2_x86inc";
1136
1137 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";
1138 specialize qw/vpx_sad32x16x4d msa/, "$sse2_x86inc";
1139
1140 add_proto qw/void vpx_sad16x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1141 specialize qw/vpx_sad16x32x4d msa/, "$sse2_x86inc";
1142
1143 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";
1144 specialize qw/vpx_sad16x16x4d neon msa/, "$sse2_x86inc";
1145
1146 add_proto qw/void vpx_sad16x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1147 specialize qw/vpx_sad16x8x4d msa/, "$sse2_x86inc";
1148
1149 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";
1150 specialize qw/vpx_sad8x16x4d msa/, "$sse2_x86inc";
1151
1152 add_proto qw/void vpx_sad8x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1153 specialize qw/vpx_sad8x8x4d msa/, "$sse2_x86inc";
1154
1155 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";
1156 specialize qw/vpx_sad8x4x4d msa/, "$sse2_x86inc";
1157
1158 add_proto qw/void vpx_sad4x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1159 specialize qw/vpx_sad4x8x4d msa/, "$sse2_x86inc";
1160
1161 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";
1162 specialize qw/vpx_sad4x4x4d msa/, "$sse2_x86inc";
1163
1164 #
1165 # Structured Similarity (SSIM)
1166 #
1167 if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
1168     add_proto qw/void vpx_ssim_parms_8x8/, "const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
1169     specialize qw/vpx_ssim_parms_8x8/, "$sse2_x86_64";
1170
1171     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";
1172     specialize qw/vpx_ssim_parms_16x16/, "$sse2_x86_64";
1173 }
1174
1175 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
1176   #
1177   # Block subtraction
1178   #
1179   add_proto qw/void vpx_highbd_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride, int bd";
1180   specialize qw/vpx_highbd_subtract_block/;
1181
1182   #
1183   # Single block SAD
1184   #
1185   add_proto qw/unsigned int vpx_highbd_sad64x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1186   specialize qw/vpx_highbd_sad64x64/, "$sse2_x86inc";
1187
1188   add_proto qw/unsigned int vpx_highbd_sad64x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1189   specialize qw/vpx_highbd_sad64x32/, "$sse2_x86inc";
1190
1191   add_proto qw/unsigned int vpx_highbd_sad32x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1192   specialize qw/vpx_highbd_sad32x64/, "$sse2_x86inc";
1193
1194   add_proto qw/unsigned int vpx_highbd_sad32x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1195   specialize qw/vpx_highbd_sad32x32/, "$sse2_x86inc";
1196
1197   add_proto qw/unsigned int vpx_highbd_sad32x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1198   specialize qw/vpx_highbd_sad32x16/, "$sse2_x86inc";
1199
1200   add_proto qw/unsigned int vpx_highbd_sad16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1201   specialize qw/vpx_highbd_sad16x32/, "$sse2_x86inc";
1202
1203   add_proto qw/unsigned int vpx_highbd_sad16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1204   specialize qw/vpx_highbd_sad16x16/, "$sse2_x86inc";
1205
1206   add_proto qw/unsigned int vpx_highbd_sad16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1207   specialize qw/vpx_highbd_sad16x8/, "$sse2_x86inc";
1208
1209   add_proto qw/unsigned int vpx_highbd_sad8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1210   specialize qw/vpx_highbd_sad8x16/, "$sse2_x86inc";
1211
1212   add_proto qw/unsigned int vpx_highbd_sad8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1213   specialize qw/vpx_highbd_sad8x8/, "$sse2_x86inc";
1214
1215   add_proto qw/unsigned int vpx_highbd_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1216   specialize qw/vpx_highbd_sad8x4/, "$sse2_x86inc";
1217
1218   add_proto qw/unsigned int vpx_highbd_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1219   specialize qw/vpx_highbd_sad4x8/;
1220
1221   add_proto qw/unsigned int vpx_highbd_sad4x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1222   specialize qw/vpx_highbd_sad4x4/;
1223
1224   #
1225   # Avg
1226   #
1227   add_proto qw/unsigned int vpx_highbd_avg_8x8/, "const uint8_t *, int p";
1228   specialize qw/vpx_highbd_avg_8x8/;
1229   add_proto qw/unsigned int vpx_highbd_avg_4x4/, "const uint8_t *, int p";
1230   specialize qw/vpx_highbd_avg_4x4/;
1231   add_proto qw/void vpx_highbd_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
1232   specialize qw/vpx_highbd_minmax_8x8/;
1233
1234   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";
1235   specialize qw/vpx_highbd_sad64x64_avg/, "$sse2_x86inc";
1236
1237   add_proto qw/unsigned int vpx_highbd_sad64x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1238   specialize qw/vpx_highbd_sad64x32_avg/, "$sse2_x86inc";
1239
1240   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";
1241   specialize qw/vpx_highbd_sad32x64_avg/, "$sse2_x86inc";
1242
1243   add_proto qw/unsigned int vpx_highbd_sad32x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1244   specialize qw/vpx_highbd_sad32x32_avg/, "$sse2_x86inc";
1245
1246   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";
1247   specialize qw/vpx_highbd_sad32x16_avg/, "$sse2_x86inc";
1248
1249   add_proto qw/unsigned int vpx_highbd_sad16x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1250   specialize qw/vpx_highbd_sad16x32_avg/, "$sse2_x86inc";
1251
1252   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";
1253   specialize qw/vpx_highbd_sad16x16_avg/, "$sse2_x86inc";
1254
1255   add_proto qw/unsigned int vpx_highbd_sad16x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1256   specialize qw/vpx_highbd_sad16x8_avg/, "$sse2_x86inc";
1257
1258   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";
1259   specialize qw/vpx_highbd_sad8x16_avg/, "$sse2_x86inc";
1260
1261   add_proto qw/unsigned int vpx_highbd_sad8x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1262   specialize qw/vpx_highbd_sad8x8_avg/, "$sse2_x86inc";
1263
1264   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";
1265   specialize qw/vpx_highbd_sad8x4_avg/, "$sse2_x86inc";
1266
1267   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";
1268   specialize qw/vpx_highbd_sad4x8_avg/;
1269
1270   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";
1271   specialize qw/vpx_highbd_sad4x4_avg/;
1272
1273   #
1274   # Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally
1275   #
1276   # Blocks of 3
1277   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";
1278   specialize qw/vpx_highbd_sad64x64x3/;
1279
1280   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";
1281   specialize qw/vpx_highbd_sad32x32x3/;
1282
1283   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";
1284   specialize qw/vpx_highbd_sad16x16x3/;
1285
1286   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";
1287   specialize qw/vpx_highbd_sad16x8x3/;
1288
1289   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";
1290   specialize qw/vpx_highbd_sad8x16x3/;
1291
1292   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";
1293   specialize qw/vpx_highbd_sad8x8x3/;
1294
1295   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";
1296   specialize qw/vpx_highbd_sad4x4x3/;
1297
1298   # Blocks of 8
1299   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";
1300   specialize qw/vpx_highbd_sad64x64x8/;
1301
1302   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";
1303   specialize qw/vpx_highbd_sad32x32x8/;
1304
1305   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";
1306   specialize qw/vpx_highbd_sad16x16x8/;
1307
1308   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";
1309   specialize qw/vpx_highbd_sad16x8x8/;
1310
1311   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";
1312   specialize qw/vpx_highbd_sad8x16x8/;
1313
1314   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";
1315   specialize qw/vpx_highbd_sad8x8x8/;
1316
1317   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";
1318   specialize qw/vpx_highbd_sad8x4x8/;
1319
1320   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";
1321   specialize qw/vpx_highbd_sad4x8x8/;
1322
1323   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";
1324   specialize qw/vpx_highbd_sad4x4x8/;
1325
1326   #
1327   # Multi-block SAD, comparing a reference to N independent blocks
1328   #
1329   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";
1330   specialize qw/vpx_highbd_sad64x64x4d/, "$sse2_x86inc";
1331
1332   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";
1333   specialize qw/vpx_highbd_sad64x32x4d/, "$sse2_x86inc";
1334
1335   add_proto qw/void vpx_highbd_sad32x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1336   specialize qw/vpx_highbd_sad32x64x4d/, "$sse2_x86inc";
1337
1338   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";
1339   specialize qw/vpx_highbd_sad32x32x4d/, "$sse2_x86inc";
1340
1341   add_proto qw/void vpx_highbd_sad32x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1342   specialize qw/vpx_highbd_sad32x16x4d/, "$sse2_x86inc";
1343
1344   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";
1345   specialize qw/vpx_highbd_sad16x32x4d/, "$sse2_x86inc";
1346
1347   add_proto qw/void vpx_highbd_sad16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1348   specialize qw/vpx_highbd_sad16x16x4d/, "$sse2_x86inc";
1349
1350   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";
1351   specialize qw/vpx_highbd_sad16x8x4d/, "$sse2_x86inc";
1352
1353   add_proto qw/void vpx_highbd_sad8x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1354   specialize qw/vpx_highbd_sad8x16x4d/, "$sse2_x86inc";
1355
1356   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";
1357   specialize qw/vpx_highbd_sad8x8x4d/, "$sse2_x86inc";
1358
1359   add_proto qw/void vpx_highbd_sad8x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1360   specialize qw/vpx_highbd_sad8x4x4d/, "$sse2_x86inc";
1361
1362   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";
1363   specialize qw/vpx_highbd_sad4x8x4d/, "$sse2_x86inc";
1364
1365   add_proto qw/void vpx_highbd_sad4x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1366   specialize qw/vpx_highbd_sad4x4x4d/, "$sse2_x86inc";
1367
1368   #
1369   # Structured Similarity (SSIM)
1370   #
1371   if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
1372     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";
1373     specialize qw/vpx_highbd_ssim_parms_8x8/;
1374   }
1375 }  # CONFIG_VP9_HIGHBITDEPTH
1376 }  # CONFIG_ENCODERS
1377
1378 if (vpx_config("CONFIG_ENCODERS") eq "yes" || vpx_config("CONFIG_POSTPROC") eq "yes" || vpx_config("CONFIG_VP9_POSTPROC") eq "yes") {
1379
1380 #
1381 # Variance
1382 #
1383 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";
1384   specialize qw/vpx_variance64x64 sse2 avx2 neon msa/;
1385
1386 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";
1387   specialize qw/vpx_variance64x32 sse2 avx2 neon msa/;
1388
1389 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";
1390   specialize qw/vpx_variance32x64 sse2 neon msa/;
1391
1392 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";
1393   specialize qw/vpx_variance32x32 sse2 avx2 neon msa/;
1394
1395 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";
1396   specialize qw/vpx_variance32x16 sse2 avx2 msa/;
1397
1398 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";
1399   specialize qw/vpx_variance16x32 sse2 msa/;
1400
1401 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";
1402   specialize qw/vpx_variance16x16 mmx sse2 avx2 media neon msa/;
1403
1404 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";
1405   specialize qw/vpx_variance16x8 mmx sse2 neon msa/;
1406
1407 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";
1408   specialize qw/vpx_variance8x16 mmx sse2 neon msa/;
1409
1410 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";
1411   specialize qw/vpx_variance8x8 mmx sse2 media neon msa/;
1412
1413 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";
1414   specialize qw/vpx_variance8x4 sse2 msa/;
1415
1416 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";
1417   specialize qw/vpx_variance4x8 sse2 msa/;
1418
1419 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";
1420   specialize qw/vpx_variance4x4 mmx sse2 msa/;
1421
1422 #
1423 # Specialty Variance
1424 #
1425 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";
1426   specialize qw/vpx_get16x16var sse2 avx2 neon msa/;
1427
1428 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";
1429   specialize qw/vpx_get8x8var mmx sse2 neon msa/;
1430
1431 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";
1432   specialize qw/vpx_mse16x16 mmx sse2 avx2 media neon msa/;
1433
1434 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";
1435   specialize qw/vpx_mse16x8 sse2 msa/;
1436
1437 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";
1438   specialize qw/vpx_mse8x16 sse2 msa/;
1439
1440 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";
1441   specialize qw/vpx_mse8x8 sse2 msa/;
1442
1443 add_proto qw/unsigned int vpx_get_mb_ss/, "const int16_t *";
1444   specialize qw/vpx_get_mb_ss mmx sse2 msa/;
1445
1446 add_proto qw/unsigned int vpx_get4x4sse_cs/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride";
1447   specialize qw/vpx_get4x4sse_cs neon msa/;
1448
1449 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";
1450
1451 #
1452 # Subpixel Variance
1453 #
1454 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";
1455   specialize qw/vpx_sub_pixel_variance64x64 avx2 neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
1456
1457 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";
1458   specialize qw/vpx_sub_pixel_variance64x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1459
1460 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";
1461   specialize qw/vpx_sub_pixel_variance32x64 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1462
1463 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";
1464   specialize qw/vpx_sub_pixel_variance32x32 avx2 neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
1465
1466 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";
1467   specialize qw/vpx_sub_pixel_variance32x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1468
1469 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";
1470   specialize qw/vpx_sub_pixel_variance16x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1471
1472 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";
1473   specialize qw/vpx_sub_pixel_variance16x16 mmx media neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
1474
1475 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";
1476   specialize qw/vpx_sub_pixel_variance16x8 mmx msa/, "$sse2_x86inc", "$ssse3_x86inc";
1477
1478 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";
1479   specialize qw/vpx_sub_pixel_variance8x16 mmx msa/, "$sse2_x86inc", "$ssse3_x86inc";
1480
1481 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";
1482   specialize qw/vpx_sub_pixel_variance8x8 mmx media neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
1483
1484 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";
1485   specialize qw/vpx_sub_pixel_variance8x4 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1486
1487 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";
1488   specialize qw/vpx_sub_pixel_variance4x8 msa/, "$sse_x86inc", "$ssse3_x86inc";
1489
1490 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";
1491   specialize qw/vpx_sub_pixel_variance4x4 mmx msa/, "$sse_x86inc", "$ssse3_x86inc";
1492
1493 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";
1494   specialize qw/vpx_sub_pixel_avg_variance64x64 avx2 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1495
1496 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";
1497   specialize qw/vpx_sub_pixel_avg_variance64x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1498
1499 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";
1500   specialize qw/vpx_sub_pixel_avg_variance32x64 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1501
1502 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";
1503   specialize qw/vpx_sub_pixel_avg_variance32x32 avx2 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1504
1505 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";
1506   specialize qw/vpx_sub_pixel_avg_variance32x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1507
1508 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";
1509   specialize qw/vpx_sub_pixel_avg_variance16x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1510
1511 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";
1512   specialize qw/vpx_sub_pixel_avg_variance16x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1513
1514 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";
1515   specialize qw/vpx_sub_pixel_avg_variance16x8 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1516
1517 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";
1518   specialize qw/vpx_sub_pixel_avg_variance8x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1519
1520 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";
1521   specialize qw/vpx_sub_pixel_avg_variance8x8 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1522
1523 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";
1524   specialize qw/vpx_sub_pixel_avg_variance8x4 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1525
1526 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";
1527   specialize qw/vpx_sub_pixel_avg_variance4x8 msa/, "$sse_x86inc", "$ssse3_x86inc";
1528
1529 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";
1530   specialize qw/vpx_sub_pixel_avg_variance4x4 msa/, "$sse_x86inc", "$ssse3_x86inc";
1531
1532 #
1533 # Specialty Subpixel
1534 #
1535 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";
1536   specialize qw/vpx_variance_halfpixvar16x16_h mmx sse2 media/;
1537
1538 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";
1539   specialize qw/vpx_variance_halfpixvar16x16_v mmx sse2 media/;
1540
1541 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";
1542   specialize qw/vpx_variance_halfpixvar16x16_hv mmx sse2 media/;
1543
1544 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
1545   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";
1546   specialize qw/vpx_highbd_12_variance64x64 sse2/;
1547
1548   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";
1549   specialize qw/vpx_highbd_12_variance64x32 sse2/;
1550
1551   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";
1552   specialize qw/vpx_highbd_12_variance32x64 sse2/;
1553
1554   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";
1555   specialize qw/vpx_highbd_12_variance32x32 sse2/;
1556
1557   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";
1558   specialize qw/vpx_highbd_12_variance32x16 sse2/;
1559
1560   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";
1561   specialize qw/vpx_highbd_12_variance16x32 sse2/;
1562
1563   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";
1564   specialize qw/vpx_highbd_12_variance16x16 sse2/;
1565
1566   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";
1567   specialize qw/vpx_highbd_12_variance16x8 sse2/;
1568
1569   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";
1570   specialize qw/vpx_highbd_12_variance8x16 sse2/;
1571
1572   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";
1573   specialize qw/vpx_highbd_12_variance8x8 sse2/;
1574
1575   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";
1576   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";
1577   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";
1578
1579   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";
1580   specialize qw/vpx_highbd_10_variance64x64 sse2/;
1581
1582   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";
1583   specialize qw/vpx_highbd_10_variance64x32 sse2/;
1584
1585   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";
1586   specialize qw/vpx_highbd_10_variance32x64 sse2/;
1587
1588   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";
1589   specialize qw/vpx_highbd_10_variance32x32 sse2/;
1590
1591   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";
1592   specialize qw/vpx_highbd_10_variance32x16 sse2/;
1593
1594   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";
1595   specialize qw/vpx_highbd_10_variance16x32 sse2/;
1596
1597   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";
1598   specialize qw/vpx_highbd_10_variance16x16 sse2/;
1599
1600   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";
1601   specialize qw/vpx_highbd_10_variance16x8 sse2/;
1602
1603   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";
1604   specialize qw/vpx_highbd_10_variance8x16 sse2/;
1605
1606   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";
1607   specialize qw/vpx_highbd_10_variance8x8 sse2/;
1608
1609   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";
1610   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";
1611   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";
1612
1613   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";
1614   specialize qw/vpx_highbd_8_variance64x64 sse2/;
1615
1616   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";
1617   specialize qw/vpx_highbd_8_variance64x32 sse2/;
1618
1619   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";
1620   specialize qw/vpx_highbd_8_variance32x64 sse2/;
1621
1622   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";
1623   specialize qw/vpx_highbd_8_variance32x32 sse2/;
1624
1625   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";
1626   specialize qw/vpx_highbd_8_variance32x16 sse2/;
1627
1628   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";
1629   specialize qw/vpx_highbd_8_variance16x32 sse2/;
1630
1631   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";
1632   specialize qw/vpx_highbd_8_variance16x16 sse2/;
1633
1634   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";
1635   specialize qw/vpx_highbd_8_variance16x8 sse2/;
1636
1637   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";
1638   specialize qw/vpx_highbd_8_variance8x16 sse2/;
1639
1640   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";
1641   specialize qw/vpx_highbd_8_variance8x8 sse2/;
1642
1643   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";
1644   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";
1645   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";
1646
1647   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";
1648   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";
1649
1650   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";
1651   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";
1652
1653   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";
1654   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";
1655
1656   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";
1657   specialize qw/vpx_highbd_8_mse16x16 sse2/;
1658
1659   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";
1660   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";
1661   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";
1662   specialize qw/vpx_highbd_8_mse8x8 sse2/;
1663
1664   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";
1665   specialize qw/vpx_highbd_10_mse16x16 sse2/;
1666
1667   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";
1668   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";
1669   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";
1670   specialize qw/vpx_highbd_10_mse8x8 sse2/;
1671
1672   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";
1673   specialize qw/vpx_highbd_12_mse16x16 sse2/;
1674
1675   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";
1676   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";
1677   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";
1678   specialize qw/vpx_highbd_12_mse8x8 sse2/;
1679
1680   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";
1681
1682   #
1683   # Subpixel Variance
1684   #
1685   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";
1686   specialize qw/vpx_highbd_12_sub_pixel_variance64x64/, "$sse2_x86inc";
1687
1688   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1689   specialize qw/vpx_highbd_12_sub_pixel_variance64x32/, "$sse2_x86inc";
1690
1691   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";
1692   specialize qw/vpx_highbd_12_sub_pixel_variance32x64/, "$sse2_x86inc";
1693
1694   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1695   specialize qw/vpx_highbd_12_sub_pixel_variance32x32/, "$sse2_x86inc";
1696
1697   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";
1698   specialize qw/vpx_highbd_12_sub_pixel_variance32x16/, "$sse2_x86inc";
1699
1700   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1701   specialize qw/vpx_highbd_12_sub_pixel_variance16x32/, "$sse2_x86inc";
1702
1703   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";
1704   specialize qw/vpx_highbd_12_sub_pixel_variance16x16/, "$sse2_x86inc";
1705
1706   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1707   specialize qw/vpx_highbd_12_sub_pixel_variance16x8/, "$sse2_x86inc";
1708
1709   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";
1710   specialize qw/vpx_highbd_12_sub_pixel_variance8x16/, "$sse2_x86inc";
1711
1712   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1713   specialize qw/vpx_highbd_12_sub_pixel_variance8x8/, "$sse2_x86inc";
1714
1715   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";
1716   specialize qw/vpx_highbd_12_sub_pixel_variance8x4/, "$sse2_x86inc";
1717
1718   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1719   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1720
1721   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1722   specialize qw/vpx_highbd_10_sub_pixel_variance64x64/, "$sse2_x86inc";
1723
1724   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";
1725   specialize qw/vpx_highbd_10_sub_pixel_variance64x32/, "$sse2_x86inc";
1726
1727   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1728   specialize qw/vpx_highbd_10_sub_pixel_variance32x64/, "$sse2_x86inc";
1729
1730   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";
1731   specialize qw/vpx_highbd_10_sub_pixel_variance32x32/, "$sse2_x86inc";
1732
1733   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1734   specialize qw/vpx_highbd_10_sub_pixel_variance32x16/, "$sse2_x86inc";
1735
1736   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";
1737   specialize qw/vpx_highbd_10_sub_pixel_variance16x32/, "$sse2_x86inc";
1738
1739   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1740   specialize qw/vpx_highbd_10_sub_pixel_variance16x16/, "$sse2_x86inc";
1741
1742   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";
1743   specialize qw/vpx_highbd_10_sub_pixel_variance16x8/, "$sse2_x86inc";
1744
1745   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1746   specialize qw/vpx_highbd_10_sub_pixel_variance8x16/, "$sse2_x86inc";
1747
1748   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";
1749   specialize qw/vpx_highbd_10_sub_pixel_variance8x8/, "$sse2_x86inc";
1750
1751   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1752   specialize qw/vpx_highbd_10_sub_pixel_variance8x4/, "$sse2_x86inc";
1753
1754   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";
1755   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";
1756
1757   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";
1758   specialize qw/vpx_highbd_8_sub_pixel_variance64x64/, "$sse2_x86inc";
1759
1760   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1761   specialize qw/vpx_highbd_8_sub_pixel_variance64x32/, "$sse2_x86inc";
1762
1763   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";
1764   specialize qw/vpx_highbd_8_sub_pixel_variance32x64/, "$sse2_x86inc";
1765
1766   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1767   specialize qw/vpx_highbd_8_sub_pixel_variance32x32/, "$sse2_x86inc";
1768
1769   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";
1770   specialize qw/vpx_highbd_8_sub_pixel_variance32x16/, "$sse2_x86inc";
1771
1772   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1773   specialize qw/vpx_highbd_8_sub_pixel_variance16x32/, "$sse2_x86inc";
1774
1775   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";
1776   specialize qw/vpx_highbd_8_sub_pixel_variance16x16/, "$sse2_x86inc";
1777
1778   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1779   specialize qw/vpx_highbd_8_sub_pixel_variance16x8/, "$sse2_x86inc";
1780
1781   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";
1782   specialize qw/vpx_highbd_8_sub_pixel_variance8x16/, "$sse2_x86inc";
1783
1784   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1785   specialize qw/vpx_highbd_8_sub_pixel_variance8x8/, "$sse2_x86inc";
1786
1787   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";
1788   specialize qw/vpx_highbd_8_sub_pixel_variance8x4/, "$sse2_x86inc";
1789
1790   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1791   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1792
1793   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1794   specialize qw/vpx_highbd_12_sub_pixel_avg_variance64x64/, "$sse2_x86inc";
1795
1796   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";
1797   specialize qw/vpx_highbd_12_sub_pixel_avg_variance64x32/, "$sse2_x86inc";
1798
1799   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1800   specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x64/, "$sse2_x86inc";
1801
1802   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";
1803   specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x32/, "$sse2_x86inc";
1804
1805   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1806   specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x16/, "$sse2_x86inc";
1807
1808   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";
1809   specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x32/, "$sse2_x86inc";
1810
1811   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1812   specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x16/, "$sse2_x86inc";
1813
1814   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";
1815   specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x8/, "$sse2_x86inc";
1816
1817   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1818   specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x16/, "$sse2_x86inc";
1819
1820   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";
1821   specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x8/, "$sse2_x86inc";
1822
1823   add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1824   specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x4/, "$sse2_x86inc";
1825
1826   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";
1827   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";
1828
1829   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";
1830   specialize qw/vpx_highbd_10_sub_pixel_avg_variance64x64/, "$sse2_x86inc";
1831
1832   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1833   specialize qw/vpx_highbd_10_sub_pixel_avg_variance64x32/, "$sse2_x86inc";
1834
1835   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";
1836   specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x64/, "$sse2_x86inc";
1837
1838   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1839   specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x32/, "$sse2_x86inc";
1840
1841   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";
1842   specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x16/, "$sse2_x86inc";
1843
1844   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1845   specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x32/, "$sse2_x86inc";
1846
1847   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";
1848   specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x16/, "$sse2_x86inc";
1849
1850   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1851   specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x8/, "$sse2_x86inc";
1852
1853   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";
1854   specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x16/, "$sse2_x86inc";
1855
1856   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1857   specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x8/, "$sse2_x86inc";
1858
1859   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";
1860   specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x4/, "$sse2_x86inc";
1861
1862   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1863   add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1864
1865   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1866   specialize qw/vpx_highbd_8_sub_pixel_avg_variance64x64/, "$sse2_x86inc";
1867
1868   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";
1869   specialize qw/vpx_highbd_8_sub_pixel_avg_variance64x32/, "$sse2_x86inc";
1870
1871   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1872   specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x64/, "$sse2_x86inc";
1873
1874   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";
1875   specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x32/, "$sse2_x86inc";
1876
1877   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1878   specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x16/, "$sse2_x86inc";
1879
1880   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";
1881   specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x32/, "$sse2_x86inc";
1882
1883   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1884   specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x16/, "$sse2_x86inc";
1885
1886   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";
1887   specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x8/, "$sse2_x86inc";
1888
1889   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1890   specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x16/, "$sse2_x86inc";
1891
1892   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";
1893   specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x8/, "$sse2_x86inc";
1894
1895   add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1896   specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x4/, "$sse2_x86inc";
1897
1898   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";
1899   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";
1900
1901 }  # CONFIG_VP9_HIGHBITDEPTH
1902 }  # CONFIG_ENCODERS || CONFIG_POSTPROC || CONFIG_VP9_POSTPROC
1903
1904 1;