aom_dsp_rtcd_defs.pl 111 KB
Newer Older
Johann's avatar
Johann committed
1 2 3 4 5 6 7 8 9 10
##
## Copyright (c) 2017, Alliance for Open Media. All rights reserved
##
## This source code is subject to the terms of the BSD 2 Clause License and
## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
## was not distributed with this source code in the LICENSE file, you can
## obtain it at www.aomedia.org/license/software. If the Alliance for Open
## Media Patent License 1.0 was not distributed with this source code in the
## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
##
Yaowu Xu's avatar
Yaowu Xu committed
11 12 13 14 15 16 17 18
sub aom_dsp_forward_decls() {
print <<EOF
/*
 * DSP
 */

#include "aom/aom_integer.h"
#include "aom_dsp/aom_dsp_common.h"
19
#include "av1/common/enums.h"
Cheng Chen's avatar
Cheng Chen committed
20
#include "av1/common/blockd.h"
Yaowu Xu's avatar
Yaowu Xu committed
21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53

EOF
}
forward_decls qw/aom_dsp_forward_decls/;

# optimizations which depend on multiple features
$avx2_ssse3 = '';
if ((aom_config("HAVE_AVX2") eq "yes") && (aom_config("HAVE_SSSE3") eq "yes")) {
  $avx2_ssse3 = 'avx2';
}

# functions that are 64 bit only.
$mmx_x86_64 = $sse2_x86_64 = $ssse3_x86_64 = $avx_x86_64 = $avx2_x86_64 = '';
if ($opts{arch} eq "x86_64") {
  $mmx_x86_64 = 'mmx';
  $sse2_x86_64 = 'sse2';
  $ssse3_x86_64 = 'ssse3';
  $avx_x86_64 = 'avx';
  $avx2_x86_64 = 'avx2';
}

if (aom_config("CONFIG_EXT_PARTITION") eq "yes") {
  @block_widths = (4, 8, 16, 32, 64, 128)
} else {
  @block_widths = (4, 8, 16, 32, 64)
}

@block_sizes = ();
foreach $w (@block_widths) {
  foreach $h (@block_widths) {
    push @block_sizes, [$w, $h] if ($w <= 2*$h && $h <= 2*$w) ;
  }
}
54
if (aom_config("CONFIG_EXT_PARTITION_TYPES") eq "yes") {
55 56 57 58
  push @block_sizes, [4, 16];
  push @block_sizes, [16, 4];
  push @block_sizes, [8, 32];
  push @block_sizes, [32, 8];
59 60
  push @block_sizes, [16, 64];
  push @block_sizes, [64, 16];
61 62 63 64
  if (aom_config("CONFIG_EXT_PARTITION") eq "yes") {
      push @block_sizes, [32, 128];
      push @block_sizes, [128, 32];
  }
65
}
Yaowu Xu's avatar
Yaowu Xu committed
66

67 68 69 70
@tx_dims = (2, 4, 8, 16, 32);
if (aom_config("CONFIG_TX64X64") eq "yes") {
  push @tx_dims, '64';
}
71

72 73 74 75 76
@tx_sizes = ();
foreach $w (@tx_dims) {
  push @tx_sizes, [$w, $w];
  foreach $h (@tx_dims) {
    push @tx_sizes, [$w, $h] if ($w >=4 && $h >=4 && ($w == 2*$h || $h == 2*$w));
77
    push @tx_sizes, [$w, $h] if ($w >=4 && $h >=4 && ($w == 4*$h || $h == 4*$w));
78 79 80
  }
}

81
@pred_names = qw/dc dc_top dc_left dc_128 v h d207e d63e d45e d117 d135 d153 paeth smooth smooth_v smooth_h/;
82

83 84 85
#
# Intra prediction
#
86

87 88
foreach (@tx_sizes) {
  ($w, $h) = @$_;
89 90 91
  foreach $pred_name (@pred_names) {
    add_proto "void", "aom_${pred_name}_predictor_${w}x${h}",
              "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
92 93
    add_proto "void", "aom_highbd_${pred_name}_predictor_${w}x${h}",
              "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
94 95
  }
}
96

97 98 99 100 101 102 103 104
specialize qw/aom_dc_top_predictor_4x4 msa neon sse2/;
specialize qw/aom_dc_top_predictor_4x8 sse2/;
specialize qw/aom_dc_top_predictor_8x4 sse2/;
specialize qw/aom_dc_top_predictor_8x8 neon msa sse2/;
specialize qw/aom_dc_top_predictor_8x16 sse2/;
specialize qw/aom_dc_top_predictor_16x8 sse2/;
specialize qw/aom_dc_top_predictor_16x16 neon msa sse2/;
specialize qw/aom_dc_top_predictor_16x32 sse2/;
105 106
specialize qw/aom_dc_top_predictor_32x16 sse2 avx2/;
specialize qw/aom_dc_top_predictor_32x32 msa neon sse2 avx2/;
107 108 109 110 111 112 113 114
specialize qw/aom_dc_left_predictor_4x4 msa neon sse2/;
specialize qw/aom_dc_left_predictor_4x8 sse2/;
specialize qw/aom_dc_left_predictor_8x4 sse2/;
specialize qw/aom_dc_left_predictor_8x8 neon msa sse2/;
specialize qw/aom_dc_left_predictor_8x16 sse2/;
specialize qw/aom_dc_left_predictor_16x8 sse2/;
specialize qw/aom_dc_left_predictor_16x16 neon msa sse2/;
specialize qw/aom_dc_left_predictor_16x32 sse2/;
115 116
specialize qw/aom_dc_left_predictor_32x16 sse2 avx2/;
specialize qw/aom_dc_left_predictor_32x32 msa neon sse2 avx2/;
117 118 119 120 121 122 123 124
specialize qw/aom_dc_128_predictor_4x4 msa neon sse2/;
specialize qw/aom_dc_128_predictor_4x8 sse2/;
specialize qw/aom_dc_128_predictor_8x4 sse2/;
specialize qw/aom_dc_128_predictor_8x8 neon msa sse2/;
specialize qw/aom_dc_128_predictor_8x16 sse2/;
specialize qw/aom_dc_128_predictor_16x8 sse2/;
specialize qw/aom_dc_128_predictor_16x16 neon msa sse2/;
specialize qw/aom_dc_128_predictor_16x32 sse2/;
125 126
specialize qw/aom_dc_128_predictor_32x16 sse2 avx2/;
specialize qw/aom_dc_128_predictor_32x32 msa neon sse2 avx2/;
127 128 129 130 131 132 133 134
specialize qw/aom_v_predictor_4x4 neon msa sse2/;
specialize qw/aom_v_predictor_4x8 sse2/;
specialize qw/aom_v_predictor_8x4 sse2/;
specialize qw/aom_v_predictor_8x8 neon msa sse2/;
specialize qw/aom_v_predictor_8x16 sse2/;
specialize qw/aom_v_predictor_16x8 sse2/;
specialize qw/aom_v_predictor_16x16 neon msa sse2/;
specialize qw/aom_v_predictor_16x32 sse2/;
135 136
specialize qw/aom_v_predictor_32x16 sse2 avx2/;
specialize qw/aom_v_predictor_32x32 neon msa sse2 avx2/;
137 138 139 140 141 142 143 144 145
specialize qw/aom_h_predictor_4x8 sse2/;
specialize qw/aom_h_predictor_4x4 neon dspr2 msa sse2/;
specialize qw/aom_h_predictor_8x4 sse2/;
specialize qw/aom_h_predictor_8x8 neon dspr2 msa sse2/;
specialize qw/aom_h_predictor_8x16 sse2/;
specialize qw/aom_h_predictor_16x8 sse2/;
specialize qw/aom_h_predictor_16x16 neon dspr2 msa sse2/;
specialize qw/aom_h_predictor_16x32 sse2/;
specialize qw/aom_h_predictor_32x16 sse2/;
146
specialize qw/aom_h_predictor_32x32 neon msa sse2 avx2/;
147 148 149 150 151
specialize qw/aom_paeth_predictor_4x4 ssse3/;
specialize qw/aom_paeth_predictor_4x8 ssse3/;
specialize qw/aom_paeth_predictor_8x4 ssse3/;
specialize qw/aom_paeth_predictor_8x8 ssse3/;
specialize qw/aom_paeth_predictor_8x16 ssse3/;
152 153 154 155 156
specialize qw/aom_paeth_predictor_16x8 ssse3 avx2/;
specialize qw/aom_paeth_predictor_16x16 ssse3 avx2/;
specialize qw/aom_paeth_predictor_16x32 ssse3 avx2/;
specialize qw/aom_paeth_predictor_32x16 ssse3 avx2/;
specialize qw/aom_paeth_predictor_32x32 ssse3 avx2/;
157 158 159 160 161 162 163 164 165 166 167 168 169 170 171
specialize qw/aom_paeth_predictor_16x8 ssse3/;
specialize qw/aom_paeth_predictor_16x16 ssse3/;
specialize qw/aom_paeth_predictor_16x32 ssse3/;
specialize qw/aom_paeth_predictor_32x16 ssse3/;
specialize qw/aom_paeth_predictor_32x32 ssse3/;
specialize qw/aom_smooth_predictor_4x4 ssse3/;
specialize qw/aom_smooth_predictor_4x8 ssse3/;
specialize qw/aom_smooth_predictor_8x4 ssse3/;
specialize qw/aom_smooth_predictor_8x8 ssse3/;
specialize qw/aom_smooth_predictor_8x16 ssse3/;
specialize qw/aom_smooth_predictor_16x8 ssse3/;
specialize qw/aom_smooth_predictor_16x16 ssse3/;
specialize qw/aom_smooth_predictor_16x32 ssse3/;
specialize qw/aom_smooth_predictor_32x16 ssse3/;
specialize qw/aom_smooth_predictor_32x32 ssse3/;
172

Yaowu Xu's avatar
Yaowu Xu committed
173 174 175 176 177
specialize qw/aom_d135_predictor_4x4 neon/;
specialize qw/aom_d153_predictor_4x4 ssse3/;
specialize qw/aom_d153_predictor_8x8 ssse3/;
specialize qw/aom_d153_predictor_16x16 ssse3/;
specialize qw/aom_d153_predictor_32x32 ssse3/;
178

179 180 181
specialize qw/aom_dc_predictor_4x4 dspr2 msa neon sse2/;
specialize qw/aom_dc_predictor_8x8 dspr2 neon msa sse2/;
specialize qw/aom_dc_predictor_16x16 dspr2 neon msa sse2/;
182
specialize qw/aom_dc_predictor_32x32 msa neon sse2 avx2/;
183 184 185 186 187 188 189 190 191 192


# TODO(luoyi): Need to rewrite these.
# specialize qw/aom_dc_predictor_4x8 sse2/;
# specialize qw/aom_dc_predictor_8x4 sse2/;
# specialize qw/aom_dc_predictor_8x16 sse2/;
# specialize qw/aom_dc_predictor_16x8 sse2/;
# specialize qw/aom_dc_predictor_16x32 sse2/;
# specialize qw/aom_dc_predictor_32x16 sse2 avx2/;

193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224
specialize qw/aom_d207e_predictor_4x4 sse2/;
specialize qw/aom_d207e_predictor_4x8 sse2/;
specialize qw/aom_d207e_predictor_8x4 sse2/;
specialize qw/aom_d207e_predictor_8x8 sse2/;
specialize qw/aom_d207e_predictor_8x16 sse2/;
specialize qw/aom_d207e_predictor_16x8 sse2/;
specialize qw/aom_d207e_predictor_16x16 sse2/;
specialize qw/aom_d207e_predictor_16x32 sse2/;
specialize qw/aom_d207e_predictor_32x16 sse2/;
specialize qw/aom_d207e_predictor_32x32 sse2/;

specialize qw/aom_d63e_predictor_4x4 sse2 ssse3/;
specialize qw/aom_d63e_predictor_4x8 sse2/;
specialize qw/aom_d63e_predictor_8x4 sse2/;
specialize qw/aom_d63e_predictor_8x8 sse2/;
specialize qw/aom_d63e_predictor_8x16 sse2/;
specialize qw/aom_d63e_predictor_16x8 sse2/;
specialize qw/aom_d63e_predictor_16x16 sse2/;
specialize qw/aom_d63e_predictor_16x32 sse2/;
specialize qw/aom_d63e_predictor_32x16 sse2/;
specialize qw/aom_d63e_predictor_32x32 sse2/;

specialize qw/aom_d45e_predictor_4x4 ssse3/;
specialize qw/aom_d45e_predictor_4x8 ssse3/;
specialize qw/aom_d45e_predictor_8x4 ssse3/;
specialize qw/aom_d45e_predictor_8x8 ssse3/;
specialize qw/aom_d45e_predictor_8x16 ssse3/;
specialize qw/aom_d45e_predictor_16x8 ssse3/;
specialize qw/aom_d45e_predictor_16x16 ssse3/;
specialize qw/aom_d45e_predictor_16x32 ssse3/;
specialize qw/aom_d45e_predictor_32x16 ssse3/;
specialize qw/aom_d45e_predictor_32x32 ssse3/;
Yaowu Xu's avatar
Yaowu Xu committed
225 226

  specialize qw/aom_highbd_v_predictor_4x4 sse2/;
227 228
  specialize qw/aom_highbd_v_predictor_4x8 sse2/;
  specialize qw/aom_highbd_v_predictor_8x4 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
229
  specialize qw/aom_highbd_v_predictor_8x8 sse2/;
230 231
  specialize qw/aom_highbd_v_predictor_8x16 sse2/;
  specialize qw/aom_highbd_v_predictor_16x8 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
232
  specialize qw/aom_highbd_v_predictor_16x16 sse2/;
233 234
  specialize qw/aom_highbd_v_predictor_16x32 sse2/;
  specialize qw/aom_highbd_v_predictor_32x16 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
235
  specialize qw/aom_highbd_v_predictor_32x32 sse2/;
236

237
  specialize qw/aom_highbd_dc_predictor_4x4 sse2/;
238
  specialize qw/aom_highbd_dc_predictor_8x8 sse2/;
239
  specialize qw/aom_highbd_dc_predictor_16x16 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
240
  specialize qw/aom_highbd_dc_predictor_32x32 sse2/;
241 242 243 244 245 246 247 248 249
  
  # TODO(luoyi): Need to rewrite these
  # specialize qw/aom_highbd_dc_predictor_4x8 sse2/;
  # specialize qw/aom_highbd_dc_predictor_8x4 sse2/;
  # specialize qw/aom_highbd_dc_predictor_8x16 sse2/;
  # specialize qw/aom_highbd_dc_predictor_16x8 sse2/;
  # specialize qw/aom_highbd_dc_predictor_16x32 sse2/;
  # specialize qw/aom_highbd_dc_predictor_32x16 sse2/;

250 251 252 253 254 255 256 257 258 259
  specialize qw/aom_highbd_h_predictor_4x4 sse2/;
  specialize qw/aom_highbd_h_predictor_4x8 sse2/;
  specialize qw/aom_highbd_h_predictor_8x4 sse2/;
  specialize qw/aom_highbd_h_predictor_8x8 sse2/;
  specialize qw/aom_highbd_h_predictor_8x16 sse2/;
  specialize qw/aom_highbd_h_predictor_16x8 sse2/;
  specialize qw/aom_highbd_h_predictor_16x16 sse2/;
  specialize qw/aom_highbd_h_predictor_16x32 sse2/;
  specialize qw/aom_highbd_h_predictor_32x16 sse2/;
  specialize qw/aom_highbd_h_predictor_32x32 sse2/;
260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289
  specialize qw/aom_highbd_dc_left_predictor_4x4 sse2/;
  specialize qw/aom_highbd_dc_top_predictor_4x4 sse2/;
  specialize qw/aom_highbd_dc_128_predictor_4x4 sse2/;
  specialize qw/aom_highbd_dc_left_predictor_4x8 sse2/;
  specialize qw/aom_highbd_dc_top_predictor_4x8 sse2/;
  specialize qw/aom_highbd_dc_128_predictor_4x8 sse2/;
  specialize qw/aom_highbd_dc_left_predictor_8x4 sse2/;
  specialize qw/aom_highbd_dc_top_predictor_8x4 sse2/;
  specialize qw/aom_highbd_dc_128_predictor_8x4 sse2/;
  specialize qw/aom_highbd_dc_left_predictor_8x8 sse2/;
  specialize qw/aom_highbd_dc_top_predictor_8x8 sse2/;
  specialize qw/aom_highbd_dc_128_predictor_8x8 sse2/;
  specialize qw/aom_highbd_dc_left_predictor_8x16 sse2/;
  specialize qw/aom_highbd_dc_top_predictor_8x16 sse2/;
  specialize qw/aom_highbd_dc_128_predictor_8x16 sse2/;
  specialize qw/aom_highbd_dc_left_predictor_16x8 sse2/;
  specialize qw/aom_highbd_dc_top_predictor_16x8 sse2/;
  specialize qw/aom_highbd_dc_128_predictor_16x8 sse2/;
  specialize qw/aom_highbd_dc_left_predictor_16x16 sse2/;
  specialize qw/aom_highbd_dc_top_predictor_16x16 sse2/;
  specialize qw/aom_highbd_dc_128_predictor_16x16 sse2/;
  specialize qw/aom_highbd_dc_left_predictor_16x32 sse2/;
  specialize qw/aom_highbd_dc_top_predictor_16x32 sse2/;
  specialize qw/aom_highbd_dc_128_predictor_16x32 sse2/;
  specialize qw/aom_highbd_dc_left_predictor_32x16 sse2/;
  specialize qw/aom_highbd_dc_top_predictor_32x16 sse2/;
  specialize qw/aom_highbd_dc_128_predictor_32x16 sse2/;
  specialize qw/aom_highbd_dc_left_predictor_32x32 sse2/;
  specialize qw/aom_highbd_dc_top_predictor_32x32 sse2/;
  specialize qw/aom_highbd_dc_128_predictor_32x32 sse2/;
290

291 292 293 294 295 296 297 298 299 300 301 302
  specialize qw/aom_highbd_d117_predictor_4x4 sse2/;
  specialize qw/aom_highbd_d117_predictor_8x8 ssse3/;
  specialize qw/aom_highbd_d117_predictor_16x16 ssse3/;
  specialize qw/aom_highbd_d117_predictor_32x32 ssse3/;
  specialize qw/aom_highbd_d135_predictor_4x4 sse2/;
  specialize qw/aom_highbd_d135_predictor_8x8 ssse3/;
  specialize qw/aom_highbd_d135_predictor_16x16 ssse3/;
  specialize qw/aom_highbd_d135_predictor_32x32 ssse3/;
  specialize qw/aom_highbd_d153_predictor_4x4 sse2/;
  specialize qw/aom_highbd_d153_predictor_8x8 ssse3/;
  specialize qw/aom_highbd_d153_predictor_16x16 ssse3/;
  specialize qw/aom_highbd_d153_predictor_32x32 ssse3/;
303 304 305 306 307 308 309 310 311 312 313

  specialize qw/aom_highbd_d45e_predictor_4x4 sse2/;
  specialize qw/aom_highbd_d45e_predictor_4x8 sse2/;
  specialize qw/aom_highbd_d45e_predictor_8x4 sse2/;
  specialize qw/aom_highbd_d45e_predictor_8x8 sse2/;
  specialize qw/aom_highbd_d45e_predictor_8x16 sse2/;
  specialize qw/aom_highbd_d45e_predictor_16x8 avx2/;
  specialize qw/aom_highbd_d45e_predictor_16x16 avx2/;
  specialize qw/aom_highbd_d45e_predictor_16x32 avx2/;
  specialize qw/aom_highbd_d45e_predictor_32x16 avx2/;
  specialize qw/aom_highbd_d45e_predictor_32x32 avx2/;
314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335

  specialize qw/aom_highbd_d207e_predictor_4x4 sse2/;
  specialize qw/aom_highbd_d207e_predictor_4x8 sse2/;
  specialize qw/aom_highbd_d207e_predictor_8x4 sse2/;
  specialize qw/aom_highbd_d207e_predictor_8x8 sse2/;
  specialize qw/aom_highbd_d207e_predictor_8x16 sse2/;
  specialize qw/aom_highbd_d207e_predictor_16x8 sse2/;
  specialize qw/aom_highbd_d207e_predictor_16x16 sse2/;
  specialize qw/aom_highbd_d207e_predictor_16x32 sse2/;
  specialize qw/aom_highbd_d207e_predictor_32x16 avx2/;
  specialize qw/aom_highbd_d207e_predictor_32x32 avx2/;

  specialize qw/aom_highbd_d63e_predictor_4x4 sse2/;
  specialize qw/aom_highbd_d63e_predictor_4x8 sse2/;
  specialize qw/aom_highbd_d63e_predictor_8x4 sse2/;
  specialize qw/aom_highbd_d63e_predictor_8x8 sse2/;
  specialize qw/aom_highbd_d63e_predictor_8x16 sse2/;
  specialize qw/aom_highbd_d63e_predictor_16x8 avx2/;
  specialize qw/aom_highbd_d63e_predictor_16x16 avx2/;
  specialize qw/aom_highbd_d63e_predictor_16x32 avx2/;
  specialize qw/aom_highbd_d63e_predictor_32x16 avx2/;
  specialize qw/aom_highbd_d63e_predictor_32x32 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
336 337 338 339

#
# Sub Pixel Filters
#
Fergus Simpson's avatar
Fergus Simpson committed
340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360
add_proto qw/void aom_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";
add_proto qw/void aom_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";
add_proto qw/void aom_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";
add_proto qw/void aom_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";
add_proto qw/void aom_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";
add_proto qw/void aom_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";
add_proto qw/void aom_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";
add_proto qw/void aom_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";
add_proto qw/void aom_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";
add_proto qw/void aom_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";
add_proto qw/void aom_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";
add_proto qw/void aom_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";
add_proto qw/void aom_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";
add_proto qw/void aom_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";

add_proto qw/void aom_convolve8_horiz_scale/,     "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h";
add_proto qw/void aom_convolve8_vert_scale/,      "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h";
add_proto qw/void aom_convolve8_avg_horiz_scale/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h";
add_proto qw/void aom_convolve8_avg_vert_scale/,  "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h";
add_proto qw/void aom_convolve8_scale/,           "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h";
add_proto qw/void aom_convolve8_avg_scale/,       "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h";
Yaowu Xu's avatar
Yaowu Xu committed
361 362 363 364 365 366 367 368 369 370 371

specialize qw/aom_convolve_copy       sse2      /;
specialize qw/aom_convolve_avg        sse2      /;
specialize qw/aom_convolve8           sse2 ssse3/, "$avx2_ssse3";
specialize qw/aom_convolve8_horiz     sse2 ssse3/, "$avx2_ssse3";
specialize qw/aom_convolve8_vert      sse2 ssse3/, "$avx2_ssse3";
specialize qw/aom_convolve8_avg       sse2 ssse3/;
specialize qw/aom_convolve8_avg_horiz sse2 ssse3/;
specialize qw/aom_convolve8_avg_vert  sse2 ssse3/;
specialize qw/aom_scaled_2d                ssse3/;

372 373 374 375
if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
  add_proto qw/void aom_convolve8_add_src/,       "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";
  add_proto qw/void aom_convolve8_add_src_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";
  add_proto qw/void aom_convolve8_add_src_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";
376 377 378
  add_proto qw/void aom_convolve8_add_src_hip/,       "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";
  add_proto qw/void aom_convolve8_add_src_horiz_hip/, "const uint8_t *src, ptrdiff_t src_stride, uint16_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";
  add_proto qw/void aom_convolve8_add_src_vert_hip/,  "const uint16_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";
379

380 381 382
  specialize qw/aom_convolve8_add_src ssse3/;
  specialize qw/aom_convolve8_add_src_horiz ssse3/;
  specialize qw/aom_convolve8_add_src_vert ssse3/;
383
  specialize qw/aom_convolve8_add_src_hip sse2/;
384 385
}  # CONFIG_LOOP_RESTORATION

Yaowu Xu's avatar
Yaowu Xu committed
386 387 388 389 390 391 392 393 394 395 396 397
# TODO(any): These need to be extended to up to 128x128 block sizes
if (!(aom_config("CONFIG_AV1") eq "yes" && aom_config("CONFIG_EXT_PARTITION") eq "yes")) {
  specialize qw/aom_convolve_copy       neon dspr2 msa/;
  specialize qw/aom_convolve_avg        neon dspr2 msa/;
  specialize qw/aom_convolve8           neon dspr2 msa/;
  specialize qw/aom_convolve8_horiz     neon dspr2 msa/;
  specialize qw/aom_convolve8_vert      neon dspr2 msa/;
  specialize qw/aom_convolve8_avg       neon dspr2 msa/;
  specialize qw/aom_convolve8_avg_horiz neon dspr2 msa/;
  specialize qw/aom_convolve8_avg_vert  neon dspr2 msa/;
}

398 399
add_proto qw/void aom_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";
specialize qw/aom_highbd_convolve_copy sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
400

401 402
add_proto qw/void aom_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";
specialize qw/aom_highbd_convolve_avg sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
403

404 405
add_proto qw/void aom_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";
specialize qw/aom_highbd_convolve8 avx2/, "$sse2_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
406

407 408
add_proto qw/void aom_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";
specialize qw/aom_highbd_convolve8_horiz avx2/, "$sse2_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
409

410 411
add_proto qw/void aom_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";
specialize qw/aom_highbd_convolve8_vert avx2/, "$sse2_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
412

413 414
add_proto qw/void aom_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";
specialize qw/aom_highbd_convolve8_avg avx2/, "$sse2_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
415

416 417
add_proto qw/void aom_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";
specialize qw/aom_highbd_convolve8_avg_horiz avx2/, "$sse2_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
418

419 420
add_proto qw/void aom_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";
specialize qw/aom_highbd_convolve8_avg_vert avx2/, "$sse2_x86_64";
421

422 423 424
if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
  add_proto qw/void aom_highbd_convolve8_add_src/, "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";
  add_proto qw/void aom_highbd_convolve8_add_src_hip/, "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";
425

426 427 428
  specialize qw/aom_highbd_convolve8_add_src/, "$sse2_x86_64";
  specialize qw/aom_highbd_convolve8_add_src_hip ssse3/;
}  # CONFIG_LOOP_RESTORATION
Yaowu Xu's avatar
Yaowu Xu committed
429 430 431 432 433

#
# Loopfilter
#
add_proto qw/void aom_lpf_vertical_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
434 435 436 437 438 439
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") eq "yes") {
  specialize qw/aom_lpf_vertical_16 sse2/;
} else {
  specialize qw/aom_lpf_vertical_16 sse2 neon_asm dspr2 msa/;
  $aom_lpf_vertical_16_neon_asm=aom_lpf_vertical_16_neon;
}
Yaowu Xu's avatar
Yaowu Xu committed
440 441

add_proto qw/void aom_lpf_vertical_16_dual/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
442 443 444 445
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") ne "yes") {
  specialize qw/aom_lpf_vertical_16_dual sse2 neon_asm dspr2 msa/;
  $aom_lpf_vertical_16_dual_neon_asm=aom_lpf_vertical_16_dual_neon;
}
Yaowu Xu's avatar
Yaowu Xu committed
446

447 448 449 450 451
add_proto qw/void aom_lpf_vertical_6/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") eq "yes") {
  specialize qw/aom_lpf_vertical_6 sse2/;
}

Yaowu Xu's avatar
Yaowu Xu committed
452
add_proto qw/void aom_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
453 454 455 456 457
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") eq "yes") {
  specialize qw/aom_lpf_vertical_8 sse2/;
} else {
  specialize qw/aom_lpf_vertical_8 sse2 neon dspr2 msa/;
}
Yaowu Xu's avatar
Yaowu Xu committed
458 459

add_proto qw/void aom_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";
460 461 462 463
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") ne "yes") {
  specialize qw/aom_lpf_vertical_8_dual sse2 neon_asm dspr2 msa/;
  $aom_lpf_vertical_8_dual_neon_asm=aom_lpf_vertical_8_dual_neon;
}
Yaowu Xu's avatar
Yaowu Xu committed
464 465

add_proto qw/void aom_lpf_vertical_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
466 467 468 469 470
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") eq "yes") {
  specialize qw/aom_lpf_vertical_4 sse2/;
} else {
  specialize qw/aom_lpf_vertical_4 sse2 neon dspr2 msa/;
}
Yaowu Xu's avatar
Yaowu Xu committed
471 472

add_proto qw/void aom_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";
473 474 475
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") ne "yes") {
  specialize qw/aom_lpf_vertical_4_dual sse2 neon dspr2 msa/;
}
Yaowu Xu's avatar
Yaowu Xu committed
476

James Zern's avatar
James Zern committed
477
add_proto qw/void aom_lpf_horizontal_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
478
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") eq "yes") {
James Zern's avatar
James Zern committed
479
  specialize qw/aom_lpf_horizontal_16 sse2/;
480
} else {
James Zern's avatar
James Zern committed
481 482
  specialize qw/aom_lpf_horizontal_16 sse2 avx2 neon_asm dspr2 msa/;
  $aom_lpf_horizontal_16_neon_asm=aom_lpf_horizontal_16_neon;
483
}
Yaowu Xu's avatar
Yaowu Xu committed
484

James Zern's avatar
James Zern committed
485
add_proto qw/void aom_lpf_horizontal_16_dual/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
486
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") eq "yes") {
James Zern's avatar
James Zern committed
487
  specialize qw/aom_lpf_horizontal_16_dual sse2/;
488
} else {
James Zern's avatar
James Zern committed
489 490
  specialize qw/aom_lpf_horizontal_16_dual sse2 avx2 neon_asm dspr2 msa/;
  $aom_lpf_horizontal_16_dual_neon_asm=aom_lpf_horizontal_16_dual_neon;
491
}
Yaowu Xu's avatar
Yaowu Xu committed
492

493 494 495 496 497
add_proto qw/void aom_lpf_horizontal_6/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") eq "yes") {
  specialize qw/aom_lpf_horizontal_6 sse2/;
}

Yaowu Xu's avatar
Yaowu Xu committed
498
add_proto qw/void aom_lpf_horizontal_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
499 500 501 502 503
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") eq "yes") {
  specialize qw/aom_lpf_horizontal_8 sse2/;
} else {
  specialize qw/aom_lpf_horizontal_8 sse2 neon dspr2 msa/;
}
Yaowu Xu's avatar
Yaowu Xu committed
504 505

add_proto qw/void aom_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";
506 507 508 509
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") ne "yes") {
  specialize qw/aom_lpf_horizontal_8_dual sse2 neon_asm dspr2 msa/;
  $aom_lpf_horizontal_8_dual_neon_asm=aom_lpf_horizontal_8_dual_neon;
}
Yaowu Xu's avatar
Yaowu Xu committed
510 511

add_proto qw/void aom_lpf_horizontal_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
512 513 514 515 516
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") eq "yes") {
  specialize qw/aom_lpf_horizontal_4 sse2/;
} else {
  specialize qw/aom_lpf_horizontal_4 sse2 neon dspr2 msa/;
}
Yaowu Xu's avatar
Yaowu Xu committed
517 518

add_proto qw/void aom_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";
519 520 521
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") ne "yes") {
  specialize qw/aom_lpf_horizontal_4_dual sse2 neon dspr2 msa/;
}
Yaowu Xu's avatar
Yaowu Xu committed
522 523 524 525 526

  add_proto qw/void aom_highbd_lpf_vertical_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
  specialize qw/aom_highbd_lpf_vertical_16 sse2/;

  add_proto qw/void aom_highbd_lpf_vertical_16_dual/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
Yi Luo's avatar
Yi Luo committed
527
  specialize qw/aom_highbd_lpf_vertical_16_dual sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
528 529 530 531 532

  add_proto qw/void aom_highbd_lpf_vertical_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
  specialize qw/aom_highbd_lpf_vertical_8 sse2/;

  add_proto qw/void aom_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";
Yi Luo's avatar
Yi Luo committed
533
  specialize qw/aom_highbd_lpf_vertical_8_dual sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
534 535 536 537 538

  add_proto qw/void aom_highbd_lpf_vertical_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
  specialize qw/aom_highbd_lpf_vertical_4 sse2/;

  add_proto qw/void aom_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";
Yi Luo's avatar
Yi Luo committed
539
  specialize qw/aom_highbd_lpf_vertical_4_dual sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
540

541 542
  add_proto qw/void aom_highbd_lpf_horizontal_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
  specialize qw/aom_highbd_lpf_horizontal_16 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
543

544 545
  add_proto qw/void aom_highbd_lpf_horizontal_16_dual/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
  specialize qw/aom_highbd_lpf_horizontal_16_dual sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
546 547 548 549 550

  add_proto qw/void aom_highbd_lpf_horizontal_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
  specialize qw/aom_highbd_lpf_horizontal_8 sse2/;

  add_proto qw/void aom_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";
Yi Luo's avatar
Yi Luo committed
551
  specialize qw/aom_highbd_lpf_horizontal_8_dual sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
552 553 554 555 556

  add_proto qw/void aom_highbd_lpf_horizontal_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
  specialize qw/aom_highbd_lpf_horizontal_4 sse2/;

  add_proto qw/void aom_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";
Yi Luo's avatar
Yi Luo committed
557
  specialize qw/aom_highbd_lpf_horizontal_4_dual sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
558

559 560 561 562
# Helper functions.
add_proto qw/void av1_round_shift_array/, "int32_t *arr, int size, int bit";
specialize "av1_round_shift_array", qw/sse4_1/;

Yaowu Xu's avatar
Yaowu Xu committed
563 564 565 566 567 568 569
#
# Encoder functions.
#

#
# Forward transform
#
Yushin Cho's avatar
Yushin Cho committed
570
if (aom_config("CONFIG_AV1_ENCODER") eq "yes"){
571 572 573 574 575 576 577
    add_proto qw/void aom_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
    specialize qw/aom_fdct4x4 sse2/;

    add_proto qw/void aom_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
    specialize qw/aom_fdct4x4_1 sse2/;

    add_proto qw/void aom_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
578
    specialize qw/aom_fdct8x8 sse2/, "$ssse3_x86_64";
579 580 581 582 583 584 585 586 587 588 589

    add_proto qw/void aom_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
    specialize qw/aom_fdct16x16 sse2/;

    add_proto qw/void aom_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
    specialize qw/aom_fdct32x32 sse2 avx2/;

    add_proto qw/void aom_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
    specialize qw/aom_fdct32x32_rd sse2 avx2/;

    # High bit depth
590 591
    add_proto qw/void aom_highbd_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
    specialize qw/aom_highbd_fdct4x4 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
592

593 594
    add_proto qw/void aom_highbd_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
    specialize qw/aom_highbd_fdct8x8 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
595

596 597
    add_proto qw/void aom_highbd_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
    specialize qw/aom_highbd_fdct16x16 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
598

599 600
    add_proto qw/void aom_highbd_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
    specialize qw/aom_highbd_fdct32x32 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
601

602 603
    add_proto qw/void aom_highbd_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
    specialize qw/aom_highbd_fdct32x32_rd sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
604 605 606 607 608 609 610 611 612 613 614 615 616 617 618

}  # CONFIG_AV1_ENCODER

#
# Inverse transform
if (aom_config("CONFIG_AV1") eq "yes") {
  add_proto qw/void aom_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";

  add_proto qw/void aom_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_iwht4x4_16_add sse2/;

  add_proto qw/void aom_highbd_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";

  add_proto qw/void aom_highbd_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";

James Zern's avatar
James Zern committed
619 620
  add_proto qw/void aom_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_idct4x4_16_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
621

James Zern's avatar
James Zern committed
622 623
  add_proto qw/void aom_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_idct4x4_1_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
624

James Zern's avatar
James Zern committed
625 626
  add_proto qw/void aom_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_idct8x8_64_add sse2 ssse3/;
Yaowu Xu's avatar
Yaowu Xu committed
627

James Zern's avatar
James Zern committed
628 629
  add_proto qw/void aom_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_idct8x8_12_add sse2 ssse3/;
Yaowu Xu's avatar
Yaowu Xu committed
630

James Zern's avatar
James Zern committed
631 632
  add_proto qw/void aom_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_idct8x8_1_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
633

James Zern's avatar
James Zern committed
634 635
  add_proto qw/void aom_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_idct16x16_256_add sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
636

James Zern's avatar
James Zern committed
637 638
  add_proto qw/void aom_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_idct16x16_38_add avx2/;
639

James Zern's avatar
James Zern committed
640 641
  add_proto qw/void aom_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_idct16x16_10_add sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
642

James Zern's avatar
James Zern committed
643 644
  add_proto qw/void aom_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_idct16x16_1_add sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
645

James Zern's avatar
James Zern committed
646 647
  add_proto qw/void aom_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_idct32x32_1024_add sse2 ssse3 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
648

James Zern's avatar
James Zern committed
649 650 651 652
  add_proto qw/void aom_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_idct32x32_135_add sse2 ssse3 avx2/;
  # Need to add 135 eob idct32x32 implementations.
  $aom_idct32x32_135_add_sse2=aom_idct32x32_1024_add_sse2;
Yaowu Xu's avatar
Yaowu Xu committed
653

James Zern's avatar
James Zern committed
654 655
  add_proto qw/void aom_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_idct32x32_34_add sse2 ssse3 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
656

James Zern's avatar
James Zern committed
657 658
  add_proto qw/void aom_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
  specialize qw/aom_idct32x32_1_add sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
659 660 661 662 663
}  # CONFIG_AV1

#
# Quantization
#
Thomas Davies's avatar
Thomas Davies committed
664
if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
James Zern's avatar
James Zern committed
665 666
  add_proto qw/void aom_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";
  specialize qw/aom_quantize_b sse2/, "$ssse3_x86_64", "$avx_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
667

James Zern's avatar
James Zern committed
668 669
  add_proto qw/void aom_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";
  specialize qw/aom_quantize_b_32x32/, "$ssse3_x86_64", "$avx_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
670

James Zern's avatar
James Zern committed
671
  add_proto qw/void aom_quantize_b_64x64/, "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";
Thomas Davies's avatar
Thomas Davies committed
672
}  # CONFIG_AV1_ENCODER
673

Thomas Davies's avatar
Thomas Davies committed
674 675 676
if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
  add_proto qw/void aom_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";
  specialize qw/aom_highbd_quantize_b sse2 avx2/;
677

Thomas Davies's avatar
Thomas Davies committed
678 679
  add_proto qw/void aom_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";
  specialize qw/aom_highbd_quantize_b_32x32 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
680

Thomas Davies's avatar
Thomas Davies committed
681
  add_proto qw/void aom_highbd_quantize_b_64x64/, "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";
682

Thomas Davies's avatar
Thomas Davies committed
683
}  # CONFIG_AV1_ENCODER
Yaowu Xu's avatar
Yaowu Xu committed
684 685 686 687
if (aom_config("CONFIG_AV1") eq "yes") {
  #
  # Alpha blending with mask
  #
688
  add_proto qw/void aom_blend_a64_d32_mask/, "int32_t *dst, uint32_t dst_stride, const int32_t *src0, uint32_t src0_stride, const int32_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int h, int w, int suby, int subx";
Yaowu Xu's avatar
Yaowu Xu committed
689 690 691 692 693 694 695 696 697 698 699 700 701
  add_proto qw/void aom_blend_a64_mask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int h, int w, int suby, int subx";
  add_proto qw/void aom_blend_a64_hmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w";
  add_proto qw/void aom_blend_a64_vmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w";
  specialize "aom_blend_a64_mask", qw/sse4_1/;
  specialize "aom_blend_a64_hmask", qw/sse4_1/;
  specialize "aom_blend_a64_vmask", qw/sse4_1/;

    add_proto qw/void aom_highbd_blend_a64_mask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int h, int w, int suby, int subx, int bd";
    add_proto qw/void aom_highbd_blend_a64_hmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd";
    add_proto qw/void aom_highbd_blend_a64_vmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd";
    specialize "aom_highbd_blend_a64_mask", qw/sse4_1/;
    specialize "aom_highbd_blend_a64_hmask", qw/sse4_1/;
    specialize "aom_highbd_blend_a64_vmask", qw/sse4_1/;
702

Yaowu Xu's avatar
Yaowu Xu committed
703 704
}  # CONFIG_AV1

705
if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
706
  #
James Zern's avatar
James Zern committed
707
  # Block subtraction
Yaowu Xu's avatar
Yaowu Xu committed
708
  #
James Zern's avatar
James Zern committed
709 710 711 712 713 714 715 716 717 718 719 720
  add_proto qw/void aom_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";
  specialize qw/aom_subtract_block neon msa sse2/;

  if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
    #
    # Sum of Squares
    #
    add_proto qw/uint64_t aom_sum_squares_2d_i16/, "const int16_t *src, int stride, int width, int height";
    specialize qw/aom_sum_squares_2d_i16 sse2/;

    add_proto qw/uint64_t aom_sum_squares_i16/, "const int16_t *src, uint32_t N";
    specialize qw/aom_sum_squares_i16 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
721 722
  }

James Zern's avatar
James Zern committed
723

Yaowu Xu's avatar
Yaowu Xu committed
724
  #
James Zern's avatar
James Zern committed
725
  # Avg
Yaowu Xu's avatar
Yaowu Xu committed
726
  #
James Zern's avatar
James Zern committed
727 728 729 730 731 732 733
  if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
    #
    # Avg
    #
    specialize qw/aom_avg_8x8 sse2 neon msa/;
      add_proto qw/void aom_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";
      specialize qw/aom_highbd_subtract_block sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
734

James Zern's avatar
James Zern committed
735 736 737 738 739 740
    #
    # Minmax
    #
    add_proto qw/void aom_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
    specialize qw/aom_minmax_8x8 sse2 neon/;
      add_proto qw/void aom_highbd_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
Yaowu Xu's avatar
Yaowu Xu committed
741

James Zern's avatar
James Zern committed
742 743
    add_proto qw/void aom_hadamard_8x8/, "const int16_t *src_diff, int src_stride, int16_t *coeff";
    specialize qw/aom_hadamard_8x8 sse2 neon/, "$ssse3_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
744

James Zern's avatar
James Zern committed
745 746
    add_proto qw/void aom_hadamard_16x16/, "const int16_t *src_diff, int src_stride, int16_t *coeff";
    specialize qw/aom_hadamard_16x16 sse2 neon/;
Yaowu Xu's avatar
Yaowu Xu committed
747

James Zern's avatar
James Zern committed
748 749
    add_proto qw/int aom_satd/, "const int16_t *coeff, int length";
    specialize qw/aom_satd sse2 neon/;
Yaowu Xu's avatar
Yaowu Xu committed
750

James Zern's avatar
James Zern committed
751 752
    add_proto qw/void aom_int_pro_row/, "int16_t *hbuf, const uint8_t *ref, int ref_stride, int height";
    specialize qw/aom_int_pro_row sse2 neon/;
Yaowu Xu's avatar
Yaowu Xu committed
753

James Zern's avatar
James Zern committed
754 755
    add_proto qw/int16_t aom_int_pro_col/, "const uint8_t *ref, int width";
    specialize qw/aom_int_pro_col sse2 neon/;
Yaowu Xu's avatar
Yaowu Xu committed
756

James Zern's avatar
James Zern committed
757 758 759
    add_proto qw/int aom_vector_var/, "const int16_t *ref, const int16_t *src, int bwl";
    specialize qw/aom_vector_var neon sse2/;
  }  # CONFIG_AV1_ENCODER
Yaowu Xu's avatar
Yaowu Xu committed
760

James Zern's avatar
James Zern committed
761 762 763
  #
  # Single block SAD / Single block Avg SAD
  #
Yaowu Xu's avatar
Yaowu Xu committed
764 765
  foreach (@block_sizes) {
    ($w, $h) = @$_;
James Zern's avatar
James Zern committed
766 767
    add_proto qw/unsigned int/, "aom_sad${w}x${h}", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
    add_proto qw/unsigned int/, "aom_sad${w}x${h}_avg", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
Cheng Chen's avatar
Cheng Chen committed
768 769 770
    if (aom_config("CONFIG_JNT_COMP") eq "yes") {
      add_proto qw/unsigned int/, "aom_jnt_sad${w}x${h}_avg", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const JNT_COMP_PARAMS *jcp_param";
    }
Yaowu Xu's avatar
Yaowu Xu committed
771 772
  }

James Zern's avatar
James Zern committed
773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805
  specialize qw/aom_sad128x128    avx2          sse2/;
  specialize qw/aom_sad128x64     avx2          sse2/;
  specialize qw/aom_sad64x128     avx2          sse2/;
  specialize qw/aom_sad64x64      avx2 neon msa sse2/;
  specialize qw/aom_sad64x32      avx2      msa sse2/;
  specialize qw/aom_sad32x64      avx2      msa sse2/;
  specialize qw/aom_sad32x32      avx2 neon msa sse2/;
  specialize qw/aom_sad32x16      avx2      msa sse2/;
  specialize qw/aom_sad16x32                msa sse2/;
  specialize qw/aom_sad16x16           neon msa sse2/;
  specialize qw/aom_sad16x8            neon msa sse2/;
  specialize qw/aom_sad8x16            neon msa sse2/;
  specialize qw/aom_sad8x8             neon msa sse2/;
  specialize qw/aom_sad8x4                  msa sse2/;
  specialize qw/aom_sad4x8                  msa sse2/;
  specialize qw/aom_sad4x4             neon msa sse2/;

  specialize qw/aom_sad128x128_avg avx2     sse2/;
  specialize qw/aom_sad128x64_avg  avx2     sse2/;
  specialize qw/aom_sad64x128_avg  avx2     sse2/;
  specialize qw/aom_sad64x64_avg   avx2 msa sse2/;
  specialize qw/aom_sad64x32_avg   avx2 msa sse2/;
  specialize qw/aom_sad32x64_avg   avx2 msa sse2/;
  specialize qw/aom_sad32x32_avg   avx2 msa sse2/;
  specialize qw/aom_sad32x16_avg   avx2 msa sse2/;
  specialize qw/aom_sad16x32_avg        msa sse2/;
  specialize qw/aom_sad16x16_avg        msa sse2/;
  specialize qw/aom_sad16x8_avg         msa sse2/;
  specialize qw/aom_sad8x16_avg         msa sse2/;
  specialize qw/aom_sad8x8_avg          msa sse2/;
  specialize qw/aom_sad8x4_avg          msa sse2/;
  specialize qw/aom_sad4x8_avg          msa sse2/;
  specialize qw/aom_sad4x4_avg          msa sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
806

807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822
  if (aom_config("CONFIG_EXT_PARTITION_TYPES") eq "yes") {
    specialize qw/aom_sad4x16      sse2/;
    specialize qw/aom_sad16x4      sse2/;
    specialize qw/aom_sad8x32      sse2/;
    specialize qw/aom_sad32x8      sse2/;
    specialize qw/aom_sad16x64     sse2/;
    specialize qw/aom_sad64x16     sse2/;

    specialize qw/aom_sad4x16_avg  sse2/;
    specialize qw/aom_sad16x4_avg  sse2/;
    specialize qw/aom_sad8x32_avg  sse2/;
    specialize qw/aom_sad32x8_avg  sse2/;
    specialize qw/aom_sad16x64_avg sse2/;
    specialize qw/aom_sad64x16_avg sse2/;
  }

823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866
  if (aom_config("CONFIG_JNT_COMP") eq "yes") {
    specialize qw/aom_jnt_sad128x128_avg ssse3/;
    specialize qw/aom_jnt_sad128x64_avg  ssse3/;
    specialize qw/aom_jnt_sad64x128_avg  ssse3/;
    specialize qw/aom_jnt_sad64x64_avg   ssse3/;
    specialize qw/aom_jnt_sad64x32_avg   ssse3/;
    specialize qw/aom_jnt_sad32x64_avg   ssse3/;
    specialize qw/aom_jnt_sad32x32_avg   ssse3/;
    specialize qw/aom_jnt_sad32x16_avg   ssse3/;
    specialize qw/aom_jnt_sad16x32_avg   ssse3/;
    specialize qw/aom_jnt_sad16x16_avg   ssse3/;
    specialize qw/aom_jnt_sad16x8_avg    ssse3/;
    specialize qw/aom_jnt_sad8x16_avg    ssse3/;
    specialize qw/aom_jnt_sad8x8_avg     ssse3/;
    specialize qw/aom_jnt_sad8x4_avg     ssse3/;
    specialize qw/aom_jnt_sad4x8_avg     ssse3/;
    specialize qw/aom_jnt_sad4x4_avg     ssse3/;

    if (aom_config("CONFIG_EXT_PARTITION_TYPES") eq "yes") {
      specialize qw/aom_jnt_sad4x16_avg     ssse3/;
      specialize qw/aom_jnt_sad16x4_avg     ssse3/;
      specialize qw/aom_jnt_sad8x32_avg     ssse3/;
      specialize qw/aom_jnt_sad32x8_avg     ssse3/;
      specialize qw/aom_jnt_sad16x64_avg     ssse3/;
      specialize qw/aom_jnt_sad64x16_avg     ssse3/;
      specialize qw/aom_jnt_sad32x128_avg     ssse3/;
      specialize qw/aom_jnt_sad128x32_avg     ssse3/;
    }

    add_proto qw/unsigned int/, "aom_sad4xh", "const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, int width, int height";
    add_proto qw/unsigned int/, "aom_sad8xh", "const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, int width, int height";
    add_proto qw/unsigned int/, "aom_sad16xh", "const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, int width, int height";
    add_proto qw/unsigned int/, "aom_sad32xh", "const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, int width, int height";
    add_proto qw/unsigned int/, "aom_sad64xh", "const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, int width, int height";
    add_proto qw/unsigned int/, "aom_sad128xh", "const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, int width, int height";

    specialize qw/aom_sad4xh   sse2/;
    specialize qw/aom_sad8xh   sse2/;
    specialize qw/aom_sad16xh  sse2/;
    specialize qw/aom_sad32xh  sse2/;
    specialize qw/aom_sad64xh  sse2/;
    specialize qw/aom_sad128xh sse2/;
  }

867

Yaowu Xu's avatar
Yaowu Xu committed
868 869
    foreach (@block_sizes) {
      ($w, $h) = @$_;
James Zern's avatar
James Zern committed
870 871 872 873 874 875
      add_proto qw/unsigned int/, "aom_highbd_sad${w}x${h}", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
      add_proto qw/unsigned int/, "aom_highbd_sad${w}x${h}_avg", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
      if ($w != 128 && $h != 128 && $w != 4) {
        specialize "aom_highbd_sad${w}x${h}", qw/sse2/;
        specialize "aom_highbd_sad${w}x${h}_avg", qw/sse2/;
      }
876 877 878
      if (aom_config("CONFIG_JNT_COMP") eq "yes") {
        add_proto qw/unsigned int/, "aom_highbd_jnt_sad${w}x${h}_avg", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const JNT_COMP_PARAMS* jcp_param";
      }
Yaowu Xu's avatar
Yaowu Xu committed
879
    }
James Zern's avatar
James Zern committed
880 881 882
    specialize qw/aom_highbd_sad128x128 avx2/;
    specialize qw/aom_highbd_sad128x64  avx2/;
    specialize qw/aom_highbd_sad64x128  avx2/;
883 884 885 886 887 888 889 890 891
    specialize qw/aom_highbd_sad64x64   avx2 sse2/;
    specialize qw/aom_highbd_sad64x32   avx2 sse2/;
    specialize qw/aom_highbd_sad32x64   avx2 sse2/;
    specialize qw/aom_highbd_sad32x32   avx2 sse2/;
    specialize qw/aom_highbd_sad32x16   avx2 sse2/;
    specialize qw/aom_highbd_sad16x32   avx2 sse2/;
    specialize qw/aom_highbd_sad16x16   avx2 sse2/;
    specialize qw/aom_highbd_sad16x8    avx2 sse2/;
    specialize qw/aom_highbd_sad8x4     sse2/;
James Zern's avatar
James Zern committed
892 893 894 895

    specialize qw/aom_highbd_sad128x128_avg avx2/;
    specialize qw/aom_highbd_sad128x64_avg  avx2/;
    specialize qw/aom_highbd_sad64x128_avg  avx2/;
896 897 898 899 900 901 902 903 904
    specialize qw/aom_highbd_sad64x64_avg   avx2 sse2/;
    specialize qw/aom_highbd_sad64x32_avg   avx2 sse2/;
    specialize qw/aom_highbd_sad32x64_avg   avx2 sse2/;
    specialize qw/aom_highbd_sad32x32_avg   avx2 sse2/;
    specialize qw/aom_highbd_sad32x16_avg   avx2 sse2/;
    specialize qw/aom_highbd_sad16x32_avg   avx2 sse2/;
    specialize qw/aom_highbd_sad16x16_avg   avx2 sse2/;
    specialize qw/aom_highbd_sad16x8_avg    avx2 sse2/;
    specialize qw/aom_highbd_sad8x4_avg     sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
905

906 907 908 909 910 911 912 913 914 915 916 917 918
    if (aom_config("CONFIG_EXT_PARTITION_TYPES") eq "yes") {
      specialize qw/aom_highbd_sad16x4       sse2/;
      specialize qw/aom_highbd_sad8x32       sse2/;
      specialize qw/aom_highbd_sad32x8       sse2/;
      specialize qw/aom_highbd_sad16x64      sse2/;
      specialize qw/aom_highbd_sad64x16      sse2/;

      specialize qw/aom_highbd_sad16x4_avg   sse2/;
      specialize qw/aom_highbd_sad8x32_avg   sse2/;
      specialize qw/aom_highbd_sad32x8_avg   sse2/;
      specialize qw/aom_highbd_sad16x64_avg  sse2/;
      specialize qw/aom_highbd_sad64x16_avg  sse2/;
    }
James Zern's avatar
James Zern committed
919 920 921
  #
  # Masked SAD
  #
922 923 924 925 926 927
  foreach (@block_sizes) {
    ($w, $h) = @$_;
    add_proto qw/unsigned int/, "aom_masked_sad${w}x${h}", "const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask";
    specialize "aom_masked_sad${w}x${h}", qw/ssse3/;
  }

928

James Zern's avatar
James Zern committed
929 930
    foreach (@block_sizes) {
      ($w, $h) = @$_;
931 932
      add_proto qw/unsigned int/, "aom_highbd_masked_sad${w}x${h}", "const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask";
      specialize "aom_highbd_masked_sad${w}x${h}", qw/ssse3/;
James Zern's avatar
James Zern committed
933
    }
934

Yaowu Xu's avatar
Yaowu Xu committed
935

James Zern's avatar
James Zern committed
936 937 938
  #
  # OBMC SAD
  #
939 940 941 942 943 944 945 946
  foreach (@block_sizes) {
    ($w, $h) = @$_;
    add_proto qw/unsigned int/, "aom_obmc_sad${w}x${h}", "const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask";
    if (! (($w == 128 && $h == 32) || ($w == 32 && $h == 128))) {
       specialize "aom_obmc_sad${w}x${h}", qw/sse4_1/;
    }
  }

947

Yaowu Xu's avatar
Yaowu Xu committed
948 949
    foreach (@block_sizes) {
      ($w, $h) = @$_;
950
      add_proto qw/unsigned int/, "aom_highbd_obmc_sad${w}x${h}", "const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask";
951
      if (! (($w == 128 && $h == 32) || ($w == 32 && $h == 128))) {
952
        specialize "aom_highbd_obmc_sad${w}x${h}", qw/sse4_1/;
James Zern's avatar
James Zern committed
953
      }
Yaowu Xu's avatar
Yaowu Xu committed
954
    }
955

Yaowu Xu's avatar
Yaowu Xu committed
956

James Zern's avatar
James Zern committed
957 958 959 960 961 962 963 964 965 966 967 968
  #
  # Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally
  #
  # Blocks of 3
  foreach $s (@block_widths) {
    add_proto qw/void/, "aom_sad${s}x${s}x3", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
  }
  specialize qw/aom_sad64x64x3            msa/;
  specialize qw/aom_sad32x32x3            msa/;
  specialize qw/aom_sad16x16x3 sse3 ssse3 msa/;
  specialize qw/aom_sad8x8x3   sse3       msa/;
  specialize qw/aom_sad4x4x3   sse3       msa/;
Yaowu Xu's avatar
Yaowu Xu committed
969

James Zern's avatar
James Zern committed
970 971 972 973 974 975
  add_proto qw/void/, "aom_sad16x8x3", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
  specialize qw/aom_sad16x8x3 sse3 ssse3 msa/;
  add_proto qw/void/, "aom_sad8x16x3", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
  specialize qw/aom_sad8x16x3 sse3 msa/;

  # Blocks of 8
Yaowu Xu's avatar
Yaowu Xu committed
976
  foreach $s (@block_widths) {
James Zern's avatar
James Zern committed
977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993
    add_proto qw/void/, "aom_sad${s}x${s}x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
  }
  specialize qw/aom_sad64x64x8        msa/;
  specialize qw/aom_sad32x32x8        msa/;
  specialize qw/aom_sad16x16x8 sse4_1 msa/;
  specialize qw/aom_sad8x8x8   sse4_1 msa/;
  specialize qw/aom_sad4x4x8   sse4_1 msa/;

  add_proto qw/void/, "aom_sad16x8x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
  specialize qw/aom_sad16x8x8 sse4_1 msa/;
  add_proto qw/void/, "aom_sad8x16x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
  specialize qw/aom_sad8x16x8 sse4_1 msa/;
  add_proto qw/void/, "aom_sad8x4x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
  specialize qw/aom_sad8x4x8 msa/;
  add_proto qw/void/, "aom_sad4x8x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
  specialize qw/aom_sad4x8x8 msa/;

994 995

  foreach $s (@block_widths) {
Yaowu Xu's avatar
Yaowu Xu committed
996
    # Blocks of 3
997
    add_proto qw/void/, "aom_highbd_sad${s}x${s}x3", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
Yaowu Xu's avatar
Yaowu Xu committed
998
    # Blocks of 8
999
    add_proto qw/void/, "aom_highbd_sad${s}x${s}x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
Yaowu Xu's avatar
Yaowu Xu committed
1000
  }
1001 1002 1003 1004 1005 1006 1007 1008 1009
  # Blocks of 3
  add_proto qw/void/, "aom_highbd_sad16x8x3", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
  add_proto qw/void/, "aom_highbd_sad8x16x3", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
  # Blocks of 8
  add_proto qw/void/, "aom_highbd_sad16x8x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
  add_proto qw/void/, "aom_highbd_sad8x16x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
  add_proto qw/void/, "aom_highbd_sad8x4x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
  add_proto qw/void/, "aom_highbd_sad4x8x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";

Yaowu Xu's avatar
Yaowu Xu committed
1010 1011 1012 1013 1014 1015

  #
  # Multi-block SAD, comparing a reference to N independent blocks
  #
  foreach (@block_sizes) {
    ($w, $h) = @$_;
James Zern's avatar
James Zern committed
1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035
    add_proto qw/void/, "aom_sad${w}x${h}x4d", "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
  }

  specialize qw/aom_sad128x128x4d avx2          sse2/;
  specialize qw/aom_sad128x64x4d  avx2          sse2/;
  specialize qw/aom_sad64x128x4d  avx2          sse2/;
  specialize qw/aom_sad64x64x4d   avx2 neon msa sse2/;
  specialize qw/aom_sad64x32x4d   avx2      msa sse2/;
  specialize qw/aom_sad32x64x4d   avx2      msa sse2/;
  specialize qw/aom_sad32x32x4d   avx2 neon msa sse2/;
  specialize qw/aom_sad32x16x4d             msa sse2/;
  specialize qw/aom_sad16x32x4d             msa sse2/;
  specialize qw/aom_sad16x16x4d        neon msa sse2/;
  specialize qw/aom_sad16x8x4d              msa sse2/;
  specialize qw/aom_sad8x16x4d              msa sse2/;
  specialize qw/aom_sad8x8x4d               msa sse2/;
  specialize qw/aom_sad8x4x4d               msa sse2/;
  specialize qw/aom_sad4x8x4d               msa sse2/;
  specialize qw/aom_sad4x4x4d               msa sse2/;

1036 1037 1038 1039 1040 1041 1042 1043 1044
  if (aom_config("CONFIG_EXT_PARTITION_TYPES") eq "yes") {
    specialize qw/aom_sad4x16x4d  sse2/;
    specialize qw/aom_sad16x4x4d  sse2/;
    specialize qw/aom_sad8x32x4d  sse2/;
    specialize qw/aom_sad32x8x4d  sse2/;
    specialize qw/aom_sad16x64x4d sse2/;
    specialize qw/aom_sad64x16x4d sse2/;
  }

James Zern's avatar
James Zern committed
1045 1046 1047 1048 1049 1050 1051 1052 1053
    #
    # Multi-block SAD, comparing a reference to N independent blocks
    #
    foreach (@block_sizes) {
      ($w, $h) = @$_;
      add_proto qw/void/, "aom_highbd_sad${w}x${h}x4d", "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
      if ($w != 128 && $h != 128) {
        specialize "aom_highbd_sad${w}x${h}x4d", qw/sse2/;
      }
Yaowu Xu's avatar
Yaowu Xu committed
1054
    }
James Zern's avatar
James Zern committed
1055 1056 1057
    specialize qw/aom_highbd_sad128x128x4d avx2/;
    specialize qw/aom_highbd_sad128x64x4d  avx2/;
    specialize qw/aom_highbd_sad64x128x4d  avx2/;
1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079
    specialize qw/aom_highbd_sad64x64x4d   sse2 avx2/;
    specialize qw/aom_highbd_sad64x32x4d   sse2 avx2/;
    specialize qw/aom_highbd_sad32x64x4d   sse2 avx2/;
    specialize qw/aom_highbd_sad32x32x4d   sse2 avx2/;
    specialize qw/aom_highbd_sad32x16x4d   sse2 avx2/;
    specialize qw/aom_highbd_sad16x32x4d   sse2 avx2/;
    specialize qw/aom_highbd_sad16x16x4d   sse2 avx2/;
    specialize qw/aom_highbd_sad16x8x4d    sse2 avx2/;
    specialize qw/aom_highbd_sad8x16x4d    sse2/;
    specialize qw/aom_highbd_sad8x8x4d     sse2/;
    specialize qw/aom_highbd_sad8x4x4d     sse2/;
    specialize qw/aom_highbd_sad4x8x4d     sse2/;
    specialize qw/aom_highbd_sad4x4x4d     sse2/;

    if (aom_config("CONFIG_EXT_PARTITION_TYPES") eq "yes") {
      specialize qw/aom_highbd_sad4x16x4d  sse2/;
      specialize qw/aom_highbd_sad16x4x4d  sse2/;
      specialize qw/aom_highbd_sad8x32x4d  sse2/;
      specialize qw/aom_highbd_sad32x8x4d  sse2/;
      specialize qw/aom_highbd_sad16x64x4d sse2/;
      specialize qw/aom_highbd_sad64x16x4d sse2/;
    }
1080

Yaowu Xu's avatar
Yaowu Xu committed
1081

James Zern's avatar
James Zern committed
1082 1083 1084 1085 1086 1087
  #
  # Structured Similarity (SSIM)
  #
  if (aom_config("CONFIG_INTERNAL_STATS") eq "yes") {
    add_proto qw/void aom_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";
    specialize qw/aom_ssim_parms_8x8/, "$sse2_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
1088

James Zern's avatar
James Zern committed
1089 1090
    add_proto qw/void aom_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";
    specialize qw/aom_ssim_parms_16x16/, "$sse2_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
1091

1092 1093
    add_proto qw/void aom_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";

Yaowu Xu's avatar
Yaowu Xu committed
1094
  }
1095
}  # CONFIG_AV1_ENCODER
Yaowu Xu's avatar
Yaowu Xu committed
1096

1097
if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
1098

James Zern's avatar
James Zern committed
1099 1100 1101 1102
  #
  # Specialty Variance
  #
  add_proto qw/void aom_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
Yaowu Xu's avatar
Yaowu Xu committed
1103

James Zern's avatar
James Zern committed
1104
  add_proto qw/void aom_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
Yaowu Xu's avatar
Yaowu Xu committed
1105

James Zern's avatar
James Zern committed
1106 1107
  specialize qw/aom_get16x16var sse2 avx2 neon msa/;
  specialize qw/aom_get8x8var   sse2      neon msa/;
Yaowu Xu's avatar
Yaowu Xu committed
1108 1109