1 ## 2 ## Copyright (c) 2017 The WebM project authors. All Rights Reserved. 3 ## 4 ## Use of this source code is governed by a BSD-style license 5 ## that can be found in the LICENSE file in the root of the source 6 ## tree. An additional intellectual property rights grant can be found 7 ## in the file PATENTS. All contributing project authors may 8 ## be found in the AUTHORS file in the root of the source tree. 9 ## 10 11 sub vpx_dsp_forward_decls() { 12 print <<EOF 13 /* 14 * DSP 15 */ 16 17 #include "vpx/vpx_integer.h" 18 #include "vpx_dsp/vpx_dsp_common.h" 19 #include "vpx_dsp/vpx_filter.h" 20 21 EOF 22 } 23 forward_decls qw/vpx_dsp_forward_decls/; 24 25 # functions that are 64 bit only. 26 $mmx_x86_64 = $sse2_x86_64 = $ssse3_x86_64 = $avx_x86_64 = $avx2_x86_64 = ''; 27 if ($opts{arch} eq "x86_64") { 28 $mmx_x86_64 = 'mmx'; 29 $sse2_x86_64 = 'sse2'; 30 $ssse3_x86_64 = 'ssse3'; 31 $avx_x86_64 = 'avx'; 32 $avx2_x86_64 = 'avx2'; 33 $avx512_x86_64 = 'avx512'; 34 } 35 36 # 37 # Intra prediction 38 # 39 40 add_proto qw/void vpx_d207_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 41 specialize qw/vpx_d207_predictor_4x4 sse2/; 42 43 add_proto qw/void vpx_d45_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 44 specialize qw/vpx_d45_predictor_4x4 neon sse2/; 45 46 add_proto qw/void vpx_d45e_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 47 48 add_proto qw/void vpx_d63_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 49 specialize qw/vpx_d63_predictor_4x4 ssse3/; 50 51 add_proto qw/void vpx_d63e_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 52 53 add_proto qw/void vpx_h_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 54 specialize qw/vpx_h_predictor_4x4 neon dspr2 msa sse2 vsx/; 55 56 add_proto qw/void vpx_he_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 57 58 add_proto qw/void vpx_d117_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 59 60 add_proto qw/void vpx_d135_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 61 specialize qw/vpx_d135_predictor_4x4 neon/; 62 63 add_proto qw/void vpx_d153_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 64 specialize qw/vpx_d153_predictor_4x4 ssse3/; 65 66 add_proto qw/void vpx_v_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 67 specialize qw/vpx_v_predictor_4x4 neon msa sse2/; 68 69 add_proto qw/void vpx_ve_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 70 71 add_proto qw/void vpx_tm_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 72 specialize qw/vpx_tm_predictor_4x4 neon dspr2 msa sse2 vsx/; 73 74 add_proto qw/void vpx_dc_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 75 specialize qw/vpx_dc_predictor_4x4 dspr2 msa neon sse2/; 76 77 add_proto qw/void vpx_dc_top_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 78 specialize qw/vpx_dc_top_predictor_4x4 msa neon sse2/; 79 80 add_proto qw/void vpx_dc_left_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 81 specialize qw/vpx_dc_left_predictor_4x4 msa neon sse2/; 82 83 add_proto qw/void vpx_dc_128_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 84 specialize qw/vpx_dc_128_predictor_4x4 msa neon sse2/; 85 86 add_proto qw/void vpx_d207_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 87 specialize qw/vpx_d207_predictor_8x8 ssse3/; 88 89 add_proto qw/void vpx_d45_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 90 specialize qw/vpx_d45_predictor_8x8 neon sse2 vsx/; 91 92 add_proto qw/void vpx_d63_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 93 specialize qw/vpx_d63_predictor_8x8 ssse3 vsx/; 94 95 add_proto qw/void vpx_h_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 96 specialize qw/vpx_h_predictor_8x8 neon dspr2 msa sse2 vsx/; 97 98 add_proto qw/void vpx_d117_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 99 100 add_proto qw/void vpx_d135_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 101 specialize qw/vpx_d135_predictor_8x8 neon/; 102 103 add_proto qw/void vpx_d153_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 104 specialize qw/vpx_d153_predictor_8x8 ssse3/; 105 106 add_proto qw/void vpx_v_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 107 specialize qw/vpx_v_predictor_8x8 neon msa sse2/; 108 109 add_proto qw/void vpx_tm_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 110 specialize qw/vpx_tm_predictor_8x8 neon dspr2 msa sse2 vsx/; 111 112 add_proto qw/void vpx_dc_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 113 specialize qw/vpx_dc_predictor_8x8 dspr2 neon msa sse2 vsx/; 114 115 add_proto qw/void vpx_dc_top_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 116 specialize qw/vpx_dc_top_predictor_8x8 neon msa sse2/; 117 118 add_proto qw/void vpx_dc_left_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 119 specialize qw/vpx_dc_left_predictor_8x8 neon msa sse2/; 120 121 add_proto qw/void vpx_dc_128_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 122 specialize qw/vpx_dc_128_predictor_8x8 neon msa sse2/; 123 124 add_proto qw/void vpx_d207_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 125 specialize qw/vpx_d207_predictor_16x16 ssse3/; 126 127 add_proto qw/void vpx_d45_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 128 specialize qw/vpx_d45_predictor_16x16 neon ssse3 vsx/; 129 130 add_proto qw/void vpx_d63_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 131 specialize qw/vpx_d63_predictor_16x16 ssse3 vsx/; 132 133 add_proto qw/void vpx_h_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 134 specialize qw/vpx_h_predictor_16x16 neon dspr2 msa sse2 vsx/; 135 136 add_proto qw/void vpx_d117_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 137 138 add_proto qw/void vpx_d135_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 139 specialize qw/vpx_d135_predictor_16x16 neon/; 140 141 add_proto qw/void vpx_d153_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 142 specialize qw/vpx_d153_predictor_16x16 ssse3/; 143 144 add_proto qw/void vpx_v_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 145 specialize qw/vpx_v_predictor_16x16 neon msa sse2 vsx/; 146 147 add_proto qw/void vpx_tm_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 148 specialize qw/vpx_tm_predictor_16x16 neon msa sse2 vsx/; 149 150 add_proto qw/void vpx_dc_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 151 specialize qw/vpx_dc_predictor_16x16 dspr2 neon msa sse2 vsx/; 152 153 add_proto qw/void vpx_dc_top_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 154 specialize qw/vpx_dc_top_predictor_16x16 neon msa sse2 vsx/; 155 156 add_proto qw/void vpx_dc_left_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 157 specialize qw/vpx_dc_left_predictor_16x16 neon msa sse2 vsx/; 158 159 add_proto qw/void vpx_dc_128_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 160 specialize qw/vpx_dc_128_predictor_16x16 neon msa sse2 vsx/; 161 162 add_proto qw/void vpx_d207_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 163 specialize qw/vpx_d207_predictor_32x32 ssse3/; 164 165 add_proto qw/void vpx_d45_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 166 specialize qw/vpx_d45_predictor_32x32 neon ssse3 vsx/; 167 168 add_proto qw/void vpx_d63_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 169 specialize qw/vpx_d63_predictor_32x32 ssse3 vsx/; 170 171 add_proto qw/void vpx_h_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 172 specialize qw/vpx_h_predictor_32x32 neon msa sse2 vsx/; 173 174 add_proto qw/void vpx_d117_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 175 176 add_proto qw/void vpx_d135_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 177 specialize qw/vpx_d135_predictor_32x32 neon/; 178 179 add_proto qw/void vpx_d153_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 180 specialize qw/vpx_d153_predictor_32x32 ssse3/; 181 182 add_proto qw/void vpx_v_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 183 specialize qw/vpx_v_predictor_32x32 neon msa sse2 vsx/; 184 185 add_proto qw/void vpx_tm_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 186 specialize qw/vpx_tm_predictor_32x32 neon msa sse2 vsx/; 187 188 add_proto qw/void vpx_dc_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 189 specialize qw/vpx_dc_predictor_32x32 msa neon sse2 vsx/; 190 191 add_proto qw/void vpx_dc_top_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 192 specialize qw/vpx_dc_top_predictor_32x32 msa neon sse2 vsx/; 193 194 add_proto qw/void vpx_dc_left_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 195 specialize qw/vpx_dc_left_predictor_32x32 msa neon sse2 vsx/; 196 197 add_proto qw/void vpx_dc_128_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; 198 specialize qw/vpx_dc_128_predictor_32x32 msa neon sse2 vsx/; 199 200 # High bitdepth functions 201 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") { 202 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"; 203 specialize qw/vpx_highbd_d207_predictor_4x4 sse2/; 204 205 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"; 206 specialize qw/vpx_highbd_d45_predictor_4x4 neon ssse3/; 207 208 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"; 209 specialize qw/vpx_highbd_d63_predictor_4x4 sse2/; 210 211 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"; 212 specialize qw/vpx_highbd_h_predictor_4x4 neon sse2/; 213 214 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"; 215 specialize qw/vpx_highbd_d117_predictor_4x4 sse2/; 216 217 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"; 218 specialize qw/vpx_highbd_d135_predictor_4x4 neon sse2/; 219 220 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"; 221 specialize qw/vpx_highbd_d153_predictor_4x4 sse2/; 222 223 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"; 224 specialize qw/vpx_highbd_v_predictor_4x4 neon sse2/; 225 226 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"; 227 specialize qw/vpx_highbd_tm_predictor_4x4 neon sse2/; 228 229 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"; 230 specialize qw/vpx_highbd_dc_predictor_4x4 neon sse2/; 231 232 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"; 233 specialize qw/vpx_highbd_dc_top_predictor_4x4 neon sse2/; 234 235 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"; 236 specialize qw/vpx_highbd_dc_left_predictor_4x4 neon sse2/; 237 238 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"; 239 specialize qw/vpx_highbd_dc_128_predictor_4x4 neon sse2/; 240 241 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"; 242 specialize qw/vpx_highbd_d207_predictor_8x8 ssse3/; 243 244 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"; 245 specialize qw/vpx_highbd_d45_predictor_8x8 neon ssse3/; 246 247 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"; 248 specialize qw/vpx_highbd_d63_predictor_8x8 ssse3/; 249 250 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"; 251 specialize qw/vpx_highbd_h_predictor_8x8 neon sse2/; 252 253 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"; 254 specialize qw/vpx_highbd_d117_predictor_8x8 ssse3/; 255 256 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"; 257 specialize qw/vpx_highbd_d135_predictor_8x8 neon ssse3/; 258 259 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"; 260 specialize qw/vpx_highbd_d153_predictor_8x8 ssse3/; 261 262 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"; 263 specialize qw/vpx_highbd_v_predictor_8x8 neon sse2/; 264 265 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"; 266 specialize qw/vpx_highbd_tm_predictor_8x8 neon sse2/; 267 268 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"; 269 specialize qw/vpx_highbd_dc_predictor_8x8 neon sse2/; 270 271 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"; 272 specialize qw/vpx_highbd_dc_top_predictor_8x8 neon sse2/; 273 274 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"; 275 specialize qw/vpx_highbd_dc_left_predictor_8x8 neon sse2/; 276 277 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"; 278 specialize qw/vpx_highbd_dc_128_predictor_8x8 neon sse2/; 279 280 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"; 281 specialize qw/vpx_highbd_d207_predictor_16x16 ssse3/; 282 283 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"; 284 specialize qw/vpx_highbd_d45_predictor_16x16 neon ssse3/; 285 286 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"; 287 specialize qw/vpx_highbd_d63_predictor_16x16 ssse3/; 288 289 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"; 290 specialize qw/vpx_highbd_h_predictor_16x16 neon sse2/; 291 292 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"; 293 specialize qw/vpx_highbd_d117_predictor_16x16 ssse3/; 294 295 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"; 296 specialize qw/vpx_highbd_d135_predictor_16x16 neon ssse3/; 297 298 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"; 299 specialize qw/vpx_highbd_d153_predictor_16x16 ssse3/; 300 301 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"; 302 specialize qw/vpx_highbd_v_predictor_16x16 neon sse2/; 303 304 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"; 305 specialize qw/vpx_highbd_tm_predictor_16x16 neon sse2/; 306 307 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"; 308 specialize qw/vpx_highbd_dc_predictor_16x16 neon sse2/; 309 310 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"; 311 specialize qw/vpx_highbd_dc_top_predictor_16x16 neon sse2/; 312 313 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"; 314 specialize qw/vpx_highbd_dc_left_predictor_16x16 neon sse2/; 315 316 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"; 317 specialize qw/vpx_highbd_dc_128_predictor_16x16 neon sse2/; 318 319 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"; 320 specialize qw/vpx_highbd_d207_predictor_32x32 ssse3/; 321 322 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"; 323 specialize qw/vpx_highbd_d45_predictor_32x32 neon ssse3/; 324 325 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"; 326 specialize qw/vpx_highbd_d63_predictor_32x32 ssse3/; 327 328 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"; 329 specialize qw/vpx_highbd_h_predictor_32x32 neon sse2/; 330 331 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"; 332 specialize qw/vpx_highbd_d117_predictor_32x32 ssse3/; 333 334 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"; 335 specialize qw/vpx_highbd_d135_predictor_32x32 neon ssse3/; 336 337 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"; 338 specialize qw/vpx_highbd_d153_predictor_32x32 ssse3/; 339 340 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"; 341 specialize qw/vpx_highbd_v_predictor_32x32 neon sse2/; 342 343 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"; 344 specialize qw/vpx_highbd_tm_predictor_32x32 neon sse2/; 345 346 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"; 347 specialize qw/vpx_highbd_dc_predictor_32x32 neon sse2/; 348 349 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"; 350 specialize qw/vpx_highbd_dc_top_predictor_32x32 neon sse2/; 351 352 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"; 353 specialize qw/vpx_highbd_dc_left_predictor_32x32 neon sse2/; 354 355 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"; 356 specialize qw/vpx_highbd_dc_128_predictor_32x32 neon sse2/; 357 } # CONFIG_VP9_HIGHBITDEPTH 358 359 # 360 # Sub Pixel Filters 361 # 362 add_proto qw/void vpx_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 363 specialize qw/vpx_convolve_copy neon dspr2 msa sse2 vsx/; 364 365 add_proto qw/void vpx_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 366 specialize qw/vpx_convolve_avg neon dspr2 msa sse2 vsx/; 367 368 add_proto qw/void vpx_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 369 specialize qw/vpx_convolve8 sse2 ssse3 avx2 neon dspr2 msa vsx/; 370 371 add_proto qw/void vpx_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 372 specialize qw/vpx_convolve8_horiz sse2 ssse3 avx2 neon dspr2 msa vsx/; 373 374 add_proto qw/void vpx_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 375 specialize qw/vpx_convolve8_vert sse2 ssse3 avx2 neon dspr2 msa vsx/; 376 377 add_proto qw/void vpx_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 378 specialize qw/vpx_convolve8_avg sse2 ssse3 avx2 neon dspr2 msa vsx/; 379 380 add_proto qw/void vpx_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 381 specialize qw/vpx_convolve8_avg_horiz sse2 ssse3 avx2 neon dspr2 msa vsx/; 382 383 add_proto qw/void vpx_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 384 specialize qw/vpx_convolve8_avg_vert sse2 ssse3 avx2 neon dspr2 msa vsx/; 385 386 add_proto qw/void vpx_scaled_2d/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 387 specialize qw/vpx_scaled_2d ssse3 neon msa/; 388 389 add_proto qw/void vpx_scaled_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 390 391 add_proto qw/void vpx_scaled_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 392 393 add_proto qw/void vpx_scaled_avg_2d/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 394 395 add_proto qw/void vpx_scaled_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 396 397 add_proto qw/void vpx_scaled_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; 398 399 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") { 400 # 401 # Sub Pixel Filters 402 # 403 add_proto qw/void vpx_highbd_convolve_copy/, "const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h, int bps"; 404 specialize qw/vpx_highbd_convolve_copy sse2 avx2 neon/; 405 406 add_proto qw/void vpx_highbd_convolve_avg/, "const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h, int bps"; 407 specialize qw/vpx_highbd_convolve_avg sse2 avx2 neon/; 408 409 add_proto qw/void vpx_highbd_convolve8/, "const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h, int bps"; 410 specialize qw/vpx_highbd_convolve8 avx2 neon/, "$sse2_x86_64"; 411 412 add_proto qw/void vpx_highbd_convolve8_horiz/, "const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h, int bps"; 413 specialize qw/vpx_highbd_convolve8_horiz avx2 neon/, "$sse2_x86_64"; 414 415 add_proto qw/void vpx_highbd_convolve8_vert/, "const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h, int bps"; 416 specialize qw/vpx_highbd_convolve8_vert avx2 neon/, "$sse2_x86_64"; 417 418 add_proto qw/void vpx_highbd_convolve8_avg/, "const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h, int bps"; 419 specialize qw/vpx_highbd_convolve8_avg avx2 neon/, "$sse2_x86_64"; 420 421 add_proto qw/void vpx_highbd_convolve8_avg_horiz/, "const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h, int bps"; 422 specialize qw/vpx_highbd_convolve8_avg_horiz avx2 neon/, "$sse2_x86_64"; 423 424 add_proto qw/void vpx_highbd_convolve8_avg_vert/, "const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h, int bps"; 425 specialize qw/vpx_highbd_convolve8_avg_vert avx2 neon/, "$sse2_x86_64"; 426 } # CONFIG_VP9_HIGHBITDEPTH 427 428 # 429 # Loopfilter 430 # 431 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"; 432 specialize qw/vpx_lpf_vertical_16 sse2 neon dspr2 msa/; 433 434 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"; 435 specialize qw/vpx_lpf_vertical_16_dual sse2 neon dspr2 msa/; 436 437 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"; 438 specialize qw/vpx_lpf_vertical_8 sse2 neon dspr2 msa/; 439 440 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"; 441 specialize qw/vpx_lpf_vertical_8_dual sse2 neon dspr2 msa/; 442 443 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"; 444 specialize qw/vpx_lpf_vertical_4 sse2 neon dspr2 msa/; 445 446 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"; 447 specialize qw/vpx_lpf_vertical_4_dual sse2 neon dspr2 msa/; 448 449 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"; 450 specialize qw/vpx_lpf_horizontal_16 sse2 avx2 neon dspr2 msa/; 451 452 add_proto qw/void vpx_lpf_horizontal_16_dual/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh"; 453 specialize qw/vpx_lpf_horizontal_16_dual sse2 avx2 neon dspr2 msa/; 454 455 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"; 456 specialize qw/vpx_lpf_horizontal_8 sse2 neon dspr2 msa/; 457 458 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"; 459 specialize qw/vpx_lpf_horizontal_8_dual sse2 neon dspr2 msa/; 460 461 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"; 462 specialize qw/vpx_lpf_horizontal_4 sse2 neon dspr2 msa/; 463 464 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"; 465 specialize qw/vpx_lpf_horizontal_4_dual sse2 neon dspr2 msa/; 466 467 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") { 468 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"; 469 specialize qw/vpx_highbd_lpf_vertical_16 sse2 neon/; 470 471 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"; 472 specialize qw/vpx_highbd_lpf_vertical_16_dual sse2 neon/; 473 474 add_proto qw/void vpx_highbd_lpf_vertical_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd"; 475 specialize qw/vpx_highbd_lpf_vertical_8 sse2 neon/; 476 477 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"; 478 specialize qw/vpx_highbd_lpf_vertical_8_dual sse2 neon/; 479 480 add_proto qw/void vpx_highbd_lpf_vertical_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd"; 481 specialize qw/vpx_highbd_lpf_vertical_4 sse2 neon/; 482 483 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"; 484 specialize qw/vpx_highbd_lpf_vertical_4_dual sse2 neon/; 485 486 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 bd"; 487 specialize qw/vpx_highbd_lpf_horizontal_16 sse2 neon/; 488 489 add_proto qw/void vpx_highbd_lpf_horizontal_16_dual/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd"; 490 specialize qw/vpx_highbd_lpf_horizontal_16_dual sse2 neon/; 491 492 add_proto qw/void vpx_highbd_lpf_horizontal_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd"; 493 specialize qw/vpx_highbd_lpf_horizontal_8 sse2 neon/; 494 495 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"; 496 specialize qw/vpx_highbd_lpf_horizontal_8_dual sse2 neon/; 497 498 add_proto qw/void vpx_highbd_lpf_horizontal_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd"; 499 specialize qw/vpx_highbd_lpf_horizontal_4 sse2 neon/; 500 501 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"; 502 specialize qw/vpx_highbd_lpf_horizontal_4_dual sse2 neon/; 503 } # CONFIG_VP9_HIGHBITDEPTH 504 505 # 506 # Encoder functions. 507 # 508 509 # 510 # Forward transform 511 # 512 if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") { 513 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") { 514 add_proto qw/void vpx_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride"; 515 specialize qw/vpx_fdct4x4 neon sse2/; 516 517 add_proto qw/void vpx_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride"; 518 specialize qw/vpx_fdct4x4_1 sse2 neon/; 519 520 add_proto qw/void vpx_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride"; 521 specialize qw/vpx_fdct8x8 neon sse2/; 522 523 add_proto qw/void vpx_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride"; 524 specialize qw/vpx_fdct8x8_1 neon sse2 msa/; 525 526 add_proto qw/void vpx_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride"; 527 specialize qw/vpx_fdct16x16 neon sse2/; 528 529 add_proto qw/void vpx_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride"; 530 specialize qw/vpx_fdct16x16_1 sse2 neon/; 531 532 add_proto qw/void vpx_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride"; 533 specialize qw/vpx_fdct32x32 neon sse2/; 534 535 add_proto qw/void vpx_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride"; 536 specialize qw/vpx_fdct32x32_rd neon sse2/; 537 538 add_proto qw/void vpx_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride"; 539 specialize qw/vpx_fdct32x32_1 sse2 neon/; 540 541 add_proto qw/void vpx_highbd_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride"; 542 specialize qw/vpx_highbd_fdct4x4 sse2/; 543 544 add_proto qw/void vpx_highbd_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride"; 545 specialize qw/vpx_highbd_fdct8x8 sse2/; 546 547 add_proto qw/void vpx_highbd_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride"; 548 specialize qw/vpx_highbd_fdct8x8_1 neon/; 549 $vpx_highbd_fdct8x8_1_neon=vpx_fdct8x8_1_neon; 550 551 add_proto qw/void vpx_highbd_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride"; 552 specialize qw/vpx_highbd_fdct16x16 sse2/; 553 554 add_proto qw/void vpx_highbd_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride"; 555 556 add_proto qw/void vpx_highbd_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride"; 557 specialize qw/vpx_highbd_fdct32x32 sse2/; 558 559 add_proto qw/void vpx_highbd_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride"; 560 specialize qw/vpx_highbd_fdct32x32_rd sse2/; 561 562 add_proto qw/void vpx_highbd_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride"; 563 } else { 564 add_proto qw/void vpx_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride"; 565 specialize qw/vpx_fdct4x4 neon sse2 msa/; 566 567 add_proto qw/void vpx_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride"; 568 specialize qw/vpx_fdct4x4_1 sse2 neon/; 569 570 add_proto qw/void vpx_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride"; 571 specialize qw/vpx_fdct8x8 sse2 neon msa/, "$ssse3_x86_64"; 572 573 add_proto qw/void vpx_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride"; 574 specialize qw/vpx_fdct8x8_1 sse2 neon msa/; 575 576 add_proto qw/void vpx_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride"; 577 specialize qw/vpx_fdct16x16 neon sse2 msa/; 578 579 add_proto qw/void vpx_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride"; 580 specialize qw/vpx_fdct16x16_1 sse2 neon msa/; 581 582 add_proto qw/void vpx_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride"; 583 specialize qw/vpx_fdct32x32 neon sse2 avx2 msa/; 584 585 add_proto qw/void vpx_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride"; 586 specialize qw/vpx_fdct32x32_rd sse2 avx2 neon msa/; 587 588 add_proto qw/void vpx_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride"; 589 specialize qw/vpx_fdct32x32_1 sse2 neon msa/; 590 } # CONFIG_VP9_HIGHBITDEPTH 591 } # CONFIG_VP9_ENCODER 592 593 # 594 # Inverse transform 595 if (vpx_config("CONFIG_VP9") eq "yes") { 596 597 add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 598 add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 599 add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 600 add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 601 add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 602 add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 603 add_proto qw/void vpx_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 604 add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 605 add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 606 add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 607 add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 608 add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 609 add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 610 add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 611 add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride"; 612 613 if (vpx_config("CONFIG_EMULATE_HARDWARE") ne "yes") { 614 # Note that there are more specializations appended when 615 # CONFIG_VP9_HIGHBITDEPTH is off. 616 specialize qw/vpx_idct4x4_16_add neon sse2 vsx/; 617 specialize qw/vpx_idct4x4_1_add neon sse2/; 618 specialize qw/vpx_idct8x8_64_add neon sse2 vsx/; 619 specialize qw/vpx_idct8x8_12_add neon sse2 ssse3/; 620 specialize qw/vpx_idct8x8_1_add neon sse2/; 621 specialize qw/vpx_idct16x16_256_add neon sse2 vsx/; 622 specialize qw/vpx_idct16x16_38_add neon sse2/; 623 specialize qw/vpx_idct16x16_10_add neon sse2/; 624 specialize qw/vpx_idct16x16_1_add neon sse2/; 625 specialize qw/vpx_idct32x32_1024_add neon sse2 vsx/; 626 specialize qw/vpx_idct32x32_135_add neon sse2 ssse3/; 627 specialize qw/vpx_idct32x32_34_add neon sse2 ssse3/; 628 specialize qw/vpx_idct32x32_1_add neon sse2/; 629 630 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") ne "yes") { 631 # Note that these specializations are appended to the above ones. 632 specialize qw/vpx_idct4x4_16_add dspr2 msa/; 633 specialize qw/vpx_idct4x4_1_add dspr2 msa/; 634 specialize qw/vpx_idct8x8_64_add dspr2 msa/; 635 specialize qw/vpx_idct8x8_12_add dspr2 msa/; 636 specialize qw/vpx_idct8x8_1_add dspr2 msa/; 637 specialize qw/vpx_idct16x16_256_add dspr2 msa/; 638 specialize qw/vpx_idct16x16_38_add dspr2 msa/; 639 $vpx_idct16x16_38_add_dspr2=vpx_idct16x16_256_add_dspr2; 640 $vpx_idct16x16_38_add_msa=vpx_idct16x16_256_add_msa; 641 specialize qw/vpx_idct16x16_10_add dspr2 msa/; 642 specialize qw/vpx_idct16x16_1_add dspr2 msa/; 643 specialize qw/vpx_idct32x32_1024_add dspr2 msa/; 644 specialize qw/vpx_idct32x32_135_add dspr2 msa/; 645 $vpx_idct32x32_135_add_dspr2=vpx_idct32x32_1024_add_dspr2; 646 $vpx_idct32x32_135_add_msa=vpx_idct32x32_1024_add_msa; 647 specialize qw/vpx_idct32x32_34_add dspr2 msa/; 648 specialize qw/vpx_idct32x32_1_add dspr2 msa/; 649 specialize qw/vpx_iwht4x4_16_add msa sse2/; 650 specialize qw/vpx_iwht4x4_1_add msa/; 651 } # !CONFIG_VP9_HIGHBITDEPTH 652 } # !CONFIG_EMULATE_HARDWARE 653 654 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") { 655 # Note as optimized versions of these functions are added we need to add a check to ensure 656 # that when CONFIG_EMULATE_HARDWARE is on, it defaults to the C versions only. 657 specialize qw/vpx_iwht4x4_16_add sse2/; 658 659 add_proto qw/void vpx_highbd_idct4x4_16_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 660 add_proto qw/void vpx_highbd_idct4x4_1_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 661 specialize qw/vpx_highbd_idct4x4_1_add neon sse2/; 662 663 add_proto qw/void vpx_highbd_idct8x8_64_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 664 add_proto qw/void vpx_highbd_idct8x8_12_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 665 add_proto qw/void vpx_highbd_idct8x8_1_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 666 specialize qw/vpx_highbd_idct8x8_1_add neon sse2/; 667 668 add_proto qw/void vpx_highbd_idct16x16_256_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 669 add_proto qw/void vpx_highbd_idct16x16_38_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 670 add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 671 add_proto qw/void vpx_highbd_idct16x16_1_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 672 specialize qw/vpx_highbd_idct16x16_1_add neon sse2/; 673 674 add_proto qw/void vpx_highbd_idct32x32_1024_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 675 add_proto qw/void vpx_highbd_idct32x32_135_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 676 add_proto qw/void vpx_highbd_idct32x32_34_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 677 add_proto qw/void vpx_highbd_idct32x32_1_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 678 specialize qw/vpx_highbd_idct32x32_1_add neon sse2/; 679 680 add_proto qw/void vpx_highbd_iwht4x4_16_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 681 add_proto qw/void vpx_highbd_iwht4x4_1_add/, "const tran_low_t *input, uint16_t *dest, int stride, int bd"; 682 683 if (vpx_config("CONFIG_EMULATE_HARDWARE") ne "yes") { 684 specialize qw/vpx_highbd_idct4x4_16_add neon sse2 sse4_1/; 685 specialize qw/vpx_highbd_idct8x8_64_add neon sse2 sse4_1/; 686 specialize qw/vpx_highbd_idct8x8_12_add neon sse2 sse4_1/; 687 specialize qw/vpx_highbd_idct16x16_256_add neon sse2 sse4_1/; 688 specialize qw/vpx_highbd_idct16x16_38_add neon sse2 sse4_1/; 689 specialize qw/vpx_highbd_idct16x16_10_add neon sse2 sse4_1/; 690 specialize qw/vpx_highbd_idct32x32_1024_add neon sse2 sse4_1/; 691 specialize qw/vpx_highbd_idct32x32_135_add neon sse2 sse4_1/; 692 specialize qw/vpx_highbd_idct32x32_34_add neon sse2 sse4_1/; 693 } # !CONFIG_EMULATE_HARDWARE 694 } # CONFIG_VP9_HIGHBITDEPTH 695 } # CONFIG_VP9 696 697 # 698 # Quantization 699 # 700 if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") { 701 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"; 702 specialize qw/vpx_quantize_b neon sse2 ssse3 avx/; 703 704 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"; 705 specialize qw/vpx_quantize_b_32x32 neon ssse3 avx/; 706 707 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") { 708 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"; 709 specialize qw/vpx_highbd_quantize_b sse2/; 710 711 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"; 712 specialize qw/vpx_highbd_quantize_b_32x32 sse2/; 713 } # CONFIG_VP9_HIGHBITDEPTH 714 } # CONFIG_VP9_ENCODER 715 716 if (vpx_config("CONFIG_ENCODERS") eq "yes") { 717 # 718 # Block subtraction 719 # 720 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"; 721 specialize qw/vpx_subtract_block neon msa mmi sse2/; 722 723 # 724 # Single block SAD 725 # 726 add_proto qw/unsigned int vpx_sad64x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 727 specialize qw/vpx_sad64x64 neon avx2 msa sse2 vsx mmi/; 728 729 add_proto qw/unsigned int vpx_sad64x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 730 specialize qw/vpx_sad64x32 neon avx2 msa sse2 vsx mmi/; 731 732 add_proto qw/unsigned int vpx_sad32x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 733 specialize qw/vpx_sad32x64 neon avx2 msa sse2 vsx mmi/; 734 735 add_proto qw/unsigned int vpx_sad32x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 736 specialize qw/vpx_sad32x32 neon avx2 msa sse2 vsx mmi/; 737 738 add_proto qw/unsigned int vpx_sad32x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 739 specialize qw/vpx_sad32x16 neon avx2 msa sse2 vsx mmi/; 740 741 add_proto qw/unsigned int vpx_sad16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 742 specialize qw/vpx_sad16x32 neon msa sse2 vsx mmi/; 743 744 add_proto qw/unsigned int vpx_sad16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 745 specialize qw/vpx_sad16x16 neon msa sse2 vsx mmi/; 746 747 add_proto qw/unsigned int vpx_sad16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 748 specialize qw/vpx_sad16x8 neon msa sse2 vsx mmi/; 749 750 add_proto qw/unsigned int vpx_sad8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 751 specialize qw/vpx_sad8x16 neon msa sse2 mmi/; 752 753 add_proto qw/unsigned int vpx_sad8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 754 specialize qw/vpx_sad8x8 neon msa sse2 mmi/; 755 756 add_proto qw/unsigned int vpx_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 757 specialize qw/vpx_sad8x4 neon msa sse2 mmi/; 758 759 add_proto qw/unsigned int vpx_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 760 specialize qw/vpx_sad4x8 neon msa sse2 mmi/; 761 762 add_proto qw/unsigned int vpx_sad4x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 763 specialize qw/vpx_sad4x4 neon msa sse2 mmi/; 764 765 # 766 # Avg 767 # 768 if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") { 769 add_proto qw/unsigned int vpx_avg_8x8/, "const uint8_t *, int p"; 770 specialize qw/vpx_avg_8x8 sse2 neon msa/; 771 772 add_proto qw/unsigned int vpx_avg_4x4/, "const uint8_t *, int p"; 773 specialize qw/vpx_avg_4x4 sse2 neon msa/; 774 775 add_proto qw/void vpx_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max"; 776 specialize qw/vpx_minmax_8x8 sse2 neon msa/; 777 778 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") { 779 add_proto qw/void vpx_hadamard_8x8/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff"; 780 specialize qw/vpx_hadamard_8x8 sse2 neon vsx/, "$ssse3_x86_64"; 781 782 add_proto qw/void vpx_hadamard_16x16/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff"; 783 specialize qw/vpx_hadamard_16x16 avx2 sse2 neon vsx/; 784 785 add_proto qw/int vpx_satd/, "const tran_low_t *coeff, int length"; 786 specialize qw/vpx_satd avx2 sse2 neon/; 787 } else { 788 add_proto qw/void vpx_hadamard_8x8/, "const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff"; 789 specialize qw/vpx_hadamard_8x8 sse2 neon msa vsx/, "$ssse3_x86_64"; 790 791 add_proto qw/void vpx_hadamard_16x16/, "const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff"; 792 specialize qw/vpx_hadamard_16x16 avx2 sse2 neon msa vsx/; 793 794 add_proto qw/int vpx_satd/, "const int16_t *coeff, int length"; 795 specialize qw/vpx_satd avx2 sse2 neon msa/; 796 } 797 798 add_proto qw/void vpx_int_pro_row/, "int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height"; 799 specialize qw/vpx_int_pro_row sse2 neon msa/; 800 801 add_proto qw/int16_t vpx_int_pro_col/, "const uint8_t *ref, const int width"; 802 specialize qw/vpx_int_pro_col sse2 neon msa/; 803 804 add_proto qw/int vpx_vector_var/, "const int16_t *ref, const int16_t *src, const int bwl"; 805 specialize qw/vpx_vector_var neon sse2 msa/; 806 } # CONFIG_VP9_ENCODER 807 808 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"; 809 specialize qw/vpx_sad64x64_avg neon avx2 msa sse2 vsx mmi/; 810 811 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"; 812 specialize qw/vpx_sad64x32_avg neon avx2 msa sse2 vsx mmi/; 813 814 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"; 815 specialize qw/vpx_sad32x64_avg neon avx2 msa sse2 vsx mmi/; 816 817 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"; 818 specialize qw/vpx_sad32x32_avg neon avx2 msa sse2 vsx mmi/; 819 820 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"; 821 specialize qw/vpx_sad32x16_avg neon avx2 msa sse2 vsx mmi/; 822 823 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"; 824 specialize qw/vpx_sad16x32_avg neon msa sse2 vsx mmi/; 825 826 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"; 827 specialize qw/vpx_sad16x16_avg neon msa sse2 vsx mmi/; 828 829 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"; 830 specialize qw/vpx_sad16x8_avg neon msa sse2 vsx mmi/; 831 832 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"; 833 specialize qw/vpx_sad8x16_avg neon msa sse2 mmi/; 834 835 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"; 836 specialize qw/vpx_sad8x8_avg neon msa sse2 mmi/; 837 838 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"; 839 specialize qw/vpx_sad8x4_avg neon msa sse2 mmi/; 840 841 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"; 842 specialize qw/vpx_sad4x8_avg neon msa sse2 mmi/; 843 844 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"; 845 specialize qw/vpx_sad4x4_avg neon msa sse2 mmi/; 846 847 # 848 # Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally 849 # 850 # Blocks of 3 851 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"; 852 specialize qw/vpx_sad16x16x3 sse3 ssse3 msa mmi/; 853 854 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"; 855 specialize qw/vpx_sad16x8x3 sse3 ssse3 msa mmi/; 856 857 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"; 858 specialize qw/vpx_sad8x16x3 sse3 msa mmi/; 859 860 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"; 861 specialize qw/vpx_sad8x8x3 sse3 msa mmi/; 862 863 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"; 864 specialize qw/vpx_sad4x4x3 sse3 msa mmi/; 865 866 # Blocks of 8 867 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"; 868 specialize qw/vpx_sad16x16x8 sse4_1 msa mmi/; 869 870 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"; 871 specialize qw/vpx_sad16x8x8 sse4_1 msa mmi/; 872 873 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"; 874 specialize qw/vpx_sad8x16x8 sse4_1 msa mmi/; 875 876 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"; 877 specialize qw/vpx_sad8x8x8 sse4_1 msa mmi/; 878 879 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"; 880 specialize qw/vpx_sad4x4x8 sse4_1 msa mmi/; 881 882 # 883 # Multi-block SAD, comparing a reference to N independent blocks 884 # 885 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"; 886 specialize qw/vpx_sad64x64x4d avx512 avx2 neon msa sse2 vsx mmi/; 887 888 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"; 889 specialize qw/vpx_sad64x32x4d neon msa sse2 vsx mmi/; 890 891 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"; 892 specialize qw/vpx_sad32x64x4d neon msa sse2 vsx mmi/; 893 894 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"; 895 specialize qw/vpx_sad32x32x4d avx2 neon msa sse2 vsx mmi/; 896 897 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"; 898 specialize qw/vpx_sad32x16x4d neon msa sse2 vsx mmi/; 899 900 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"; 901 specialize qw/vpx_sad16x32x4d neon msa sse2 vsx mmi/; 902 903 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"; 904 specialize qw/vpx_sad16x16x4d neon msa sse2 vsx mmi/; 905 906 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"; 907 specialize qw/vpx_sad16x8x4d neon msa sse2 vsx mmi/; 908 909 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"; 910 specialize qw/vpx_sad8x16x4d neon msa sse2 mmi/; 911 912 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"; 913 specialize qw/vpx_sad8x8x4d neon msa sse2 mmi/; 914 915 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"; 916 specialize qw/vpx_sad8x4x4d neon msa sse2 mmi/; 917 918 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"; 919 specialize qw/vpx_sad4x8x4d neon msa sse2 mmi/; 920 921 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"; 922 specialize qw/vpx_sad4x4x4d neon msa sse2 mmi/; 923 924 add_proto qw/uint64_t vpx_sum_squares_2d_i16/, "const int16_t *src, int stride, int size"; 925 specialize qw/vpx_sum_squares_2d_i16 sse2 msa/; 926 927 # 928 # Structured Similarity (SSIM) 929 # 930 if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") { 931 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"; 932 specialize qw/vpx_ssim_parms_8x8/, "$sse2_x86_64"; 933 934 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"; 935 specialize qw/vpx_ssim_parms_16x16/, "$sse2_x86_64"; 936 } 937 938 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") { 939 # 940 # Block subtraction 941 # 942 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"; 943 944 # 945 # Single block SAD 946 # 947 add_proto qw/unsigned int vpx_highbd_sad64x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 948 specialize qw/vpx_highbd_sad64x64 sse2/; 949 950 add_proto qw/unsigned int vpx_highbd_sad64x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 951 specialize qw/vpx_highbd_sad64x32 sse2/; 952 953 add_proto qw/unsigned int vpx_highbd_sad32x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 954 specialize qw/vpx_highbd_sad32x64 sse2/; 955 956 add_proto qw/unsigned int vpx_highbd_sad32x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 957 specialize qw/vpx_highbd_sad32x32 sse2/; 958 959 add_proto qw/unsigned int vpx_highbd_sad32x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 960 specialize qw/vpx_highbd_sad32x16 sse2/; 961 962 add_proto qw/unsigned int vpx_highbd_sad16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 963 specialize qw/vpx_highbd_sad16x32 sse2/; 964 965 add_proto qw/unsigned int vpx_highbd_sad16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 966 specialize qw/vpx_highbd_sad16x16 sse2/; 967 968 add_proto qw/unsigned int vpx_highbd_sad16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 969 specialize qw/vpx_highbd_sad16x8 sse2/; 970 971 add_proto qw/unsigned int vpx_highbd_sad8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 972 specialize qw/vpx_highbd_sad8x16 sse2/; 973 974 add_proto qw/unsigned int vpx_highbd_sad8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 975 specialize qw/vpx_highbd_sad8x8 sse2/; 976 977 add_proto qw/unsigned int vpx_highbd_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 978 specialize qw/vpx_highbd_sad8x4 sse2/; 979 980 add_proto qw/unsigned int vpx_highbd_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 981 982 add_proto qw/unsigned int vpx_highbd_sad4x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; 983 984 # 985 # Avg 986 # 987 add_proto qw/unsigned int vpx_highbd_avg_8x8/, "const uint8_t *, int p"; 988 add_proto qw/unsigned int vpx_highbd_avg_4x4/, "const uint8_t *, int p"; 989 add_proto qw/void vpx_highbd_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max"; 990 991 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"; 992 specialize qw/vpx_highbd_sad64x64_avg sse2/; 993 994 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"; 995 specialize qw/vpx_highbd_sad64x32_avg sse2/; 996 997 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"; 998 specialize qw/vpx_highbd_sad32x64_avg sse2/; 999 1000 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"; 1001 specialize qw/vpx_highbd_sad32x32_avg sse2/; 1002 1003 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"; 1004 specialize qw/vpx_highbd_sad32x16_avg sse2/; 1005 1006 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"; 1007 specialize qw/vpx_highbd_sad16x32_avg sse2/; 1008 1009 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"; 1010 specialize qw/vpx_highbd_sad16x16_avg sse2/; 1011 1012 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"; 1013 specialize qw/vpx_highbd_sad16x8_avg sse2/; 1014 1015 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"; 1016 specialize qw/vpx_highbd_sad8x16_avg sse2/; 1017 1018 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"; 1019 specialize qw/vpx_highbd_sad8x8_avg sse2/; 1020 1021 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"; 1022 specialize qw/vpx_highbd_sad8x4_avg sse2/; 1023 1024 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"; 1025 1026 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"; 1027 1028 # 1029 # Multi-block SAD, comparing a reference to N independent blocks 1030 # 1031 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"; 1032 specialize qw/vpx_highbd_sad64x64x4d sse2/; 1033 1034 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"; 1035 specialize qw/vpx_highbd_sad64x32x4d sse2/; 1036 1037 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"; 1038 specialize qw/vpx_highbd_sad32x64x4d sse2/; 1039 1040 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"; 1041 specialize qw/vpx_highbd_sad32x32x4d sse2/; 1042 1043 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"; 1044 specialize qw/vpx_highbd_sad32x16x4d sse2/; 1045 1046 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"; 1047 specialize qw/vpx_highbd_sad16x32x4d sse2/; 1048 1049 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"; 1050 specialize qw/vpx_highbd_sad16x16x4d sse2/; 1051 1052 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"; 1053 specialize qw/vpx_highbd_sad16x8x4d sse2/; 1054 1055 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"; 1056 specialize qw/vpx_highbd_sad8x16x4d sse2/; 1057 1058 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"; 1059 specialize qw/vpx_highbd_sad8x8x4d sse2/; 1060 1061 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"; 1062 specialize qw/vpx_highbd_sad8x4x4d sse2/; 1063 1064 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"; 1065 specialize qw/vpx_highbd_sad4x8x4d sse2/; 1066 1067 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"; 1068 specialize qw/vpx_highbd_sad4x4x4d sse2/; 1069 1070 # 1071 # Structured Similarity (SSIM) 1072 # 1073 if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") { 1074 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"; 1075 } 1076 } # CONFIG_VP9_HIGHBITDEPTH 1077 } # CONFIG_ENCODERS 1078 1079 if (vpx_config("CONFIG_ENCODERS") eq "yes" || vpx_config("CONFIG_POSTPROC") eq "yes" || vpx_config("CONFIG_VP9_POSTPROC") eq "yes") { 1080 1081 # 1082 # Variance 1083 # 1084 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"; 1085 specialize qw/vpx_variance64x64 sse2 avx2 neon msa mmi/; 1086 1087 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"; 1088 specialize qw/vpx_variance64x32 sse2 avx2 neon msa mmi/; 1089 1090 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"; 1091 specialize qw/vpx_variance32x64 sse2 neon msa mmi/; 1092 1093 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"; 1094 specialize qw/vpx_variance32x32 sse2 avx2 neon msa mmi/; 1095 1096 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"; 1097 specialize qw/vpx_variance32x16 sse2 avx2 neon msa mmi/; 1098 1099 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"; 1100 specialize qw/vpx_variance16x32 sse2 neon msa mmi/; 1101 1102 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"; 1103 specialize qw/vpx_variance16x16 sse2 avx2 neon msa mmi/; 1104 1105 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"; 1106 specialize qw/vpx_variance16x8 sse2 neon msa mmi/; 1107 1108 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"; 1109 specialize qw/vpx_variance8x16 sse2 neon msa mmi/; 1110 1111 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"; 1112 specialize qw/vpx_variance8x8 sse2 neon msa mmi/; 1113 1114 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"; 1115 specialize qw/vpx_variance8x4 sse2 neon msa mmi/; 1116 1117 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"; 1118 specialize qw/vpx_variance4x8 sse2 neon msa mmi/; 1119 1120 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"; 1121 specialize qw/vpx_variance4x4 sse2 neon msa mmi/; 1122 1123 # 1124 # Specialty Variance 1125 # 1126 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"; 1127 specialize qw/vpx_get16x16var sse2 avx2 neon msa/; 1128 1129 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"; 1130 specialize qw/vpx_get8x8var sse2 neon msa/; 1131 1132 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"; 1133 specialize qw/vpx_mse16x16 sse2 avx2 neon msa mmi/; 1134 1135 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"; 1136 specialize qw/vpx_mse16x8 sse2 msa mmi/; 1137 1138 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"; 1139 specialize qw/vpx_mse8x16 sse2 msa mmi/; 1140 1141 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"; 1142 specialize qw/vpx_mse8x8 sse2 msa mmi/; 1143 1144 add_proto qw/unsigned int vpx_get_mb_ss/, "const int16_t *"; 1145 specialize qw/vpx_get_mb_ss sse2 msa vsx/; 1146 1147 add_proto qw/unsigned int vpx_get4x4sse_cs/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride"; 1148 specialize qw/vpx_get4x4sse_cs neon msa vsx/; 1149 1150 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"; 1151 specialize qw/vpx_comp_avg_pred neon sse2 vsx/; 1152 1153 # 1154 # Subpixel Variance 1155 # 1156 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"; 1157 specialize qw/vpx_sub_pixel_variance64x64 avx2 neon msa mmi sse2 ssse3/; 1158 1159 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"; 1160 specialize qw/vpx_sub_pixel_variance64x32 neon msa mmi sse2 ssse3/; 1161 1162 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"; 1163 specialize qw/vpx_sub_pixel_variance32x64 neon msa mmi sse2 ssse3/; 1164 1165 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"; 1166 specialize qw/vpx_sub_pixel_variance32x32 avx2 neon msa mmi sse2 ssse3/; 1167 1168 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"; 1169 specialize qw/vpx_sub_pixel_variance32x16 neon msa mmi sse2 ssse3/; 1170 1171 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"; 1172 specialize qw/vpx_sub_pixel_variance16x32 neon msa mmi sse2 ssse3/; 1173 1174 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"; 1175 specialize qw/vpx_sub_pixel_variance16x16 neon msa mmi sse2 ssse3/; 1176 1177 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"; 1178 specialize qw/vpx_sub_pixel_variance16x8 neon msa mmi sse2 ssse3/; 1179 1180 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"; 1181 specialize qw/vpx_sub_pixel_variance8x16 neon msa mmi sse2 ssse3/; 1182 1183 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"; 1184 specialize qw/vpx_sub_pixel_variance8x8 neon msa mmi sse2 ssse3/; 1185 1186 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"; 1187 specialize qw/vpx_sub_pixel_variance8x4 neon msa mmi sse2 ssse3/; 1188 1189 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"; 1190 specialize qw/vpx_sub_pixel_variance4x8 neon msa mmi sse2 ssse3/; 1191 1192 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"; 1193 specialize qw/vpx_sub_pixel_variance4x4 neon msa mmi sse2 ssse3/; 1194 1195 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"; 1196 specialize qw/vpx_sub_pixel_avg_variance64x64 neon avx2 msa mmi sse2 ssse3/; 1197 1198 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"; 1199 specialize qw/vpx_sub_pixel_avg_variance64x32 neon msa mmi sse2 ssse3/; 1200 1201 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"; 1202 specialize qw/vpx_sub_pixel_avg_variance32x64 neon msa mmi sse2 ssse3/; 1203 1204 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"; 1205 specialize qw/vpx_sub_pixel_avg_variance32x32 neon avx2 msa mmi sse2 ssse3/; 1206 1207 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"; 1208 specialize qw/vpx_sub_pixel_avg_variance32x16 neon msa mmi sse2 ssse3/; 1209 1210 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"; 1211 specialize qw/vpx_sub_pixel_avg_variance16x32 neon msa mmi sse2 ssse3/; 1212 1213 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"; 1214 specialize qw/vpx_sub_pixel_avg_variance16x16 neon msa mmi sse2 ssse3/; 1215 1216 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"; 1217 specialize qw/vpx_sub_pixel_avg_variance16x8 neon msa mmi sse2 ssse3/; 1218 1219 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"; 1220 specialize qw/vpx_sub_pixel_avg_variance8x16 neon msa mmi sse2 ssse3/; 1221 1222 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"; 1223 specialize qw/vpx_sub_pixel_avg_variance8x8 neon msa mmi sse2 ssse3/; 1224 1225 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"; 1226 specialize qw/vpx_sub_pixel_avg_variance8x4 neon msa mmi sse2 ssse3/; 1227 1228 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"; 1229 specialize qw/vpx_sub_pixel_avg_variance4x8 neon msa mmi sse2 ssse3/; 1230 1231 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"; 1232 specialize qw/vpx_sub_pixel_avg_variance4x4 neon msa mmi sse2 ssse3/; 1233 1234 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") { 1235 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"; 1236 specialize qw/vpx_highbd_12_variance64x64 sse2/; 1237 1238 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"; 1239 specialize qw/vpx_highbd_12_variance64x32 sse2/; 1240 1241 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"; 1242 specialize qw/vpx_highbd_12_variance32x64 sse2/; 1243 1244 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"; 1245 specialize qw/vpx_highbd_12_variance32x32 sse2/; 1246 1247 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"; 1248 specialize qw/vpx_highbd_12_variance32x16 sse2/; 1249 1250 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"; 1251 specialize qw/vpx_highbd_12_variance16x32 sse2/; 1252 1253 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"; 1254 specialize qw/vpx_highbd_12_variance16x16 sse2/; 1255 1256 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"; 1257 specialize qw/vpx_highbd_12_variance16x8 sse2/; 1258 1259 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"; 1260 specialize qw/vpx_highbd_12_variance8x16 sse2/; 1261 1262 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"; 1263 specialize qw/vpx_highbd_12_variance8x8 sse2/; 1264 1265 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"; 1266 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"; 1267 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"; 1268 1269 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"; 1270 specialize qw/vpx_highbd_10_variance64x64 sse2/; 1271 1272 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"; 1273 specialize qw/vpx_highbd_10_variance64x32 sse2/; 1274 1275 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"; 1276 specialize qw/vpx_highbd_10_variance32x64 sse2/; 1277 1278 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"; 1279 specialize qw/vpx_highbd_10_variance32x32 sse2/; 1280 1281 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"; 1282 specialize qw/vpx_highbd_10_variance32x16 sse2/; 1283 1284 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"; 1285 specialize qw/vpx_highbd_10_variance16x32 sse2/; 1286 1287 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"; 1288 specialize qw/vpx_highbd_10_variance16x16 sse2/; 1289 1290 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"; 1291 specialize qw/vpx_highbd_10_variance16x8 sse2/; 1292 1293 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"; 1294 specialize qw/vpx_highbd_10_variance8x16 sse2/; 1295 1296 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"; 1297 specialize qw/vpx_highbd_10_variance8x8 sse2/; 1298 1299 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"; 1300 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"; 1301 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"; 1302 1303 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"; 1304 specialize qw/vpx_highbd_8_variance64x64 sse2/; 1305 1306 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"; 1307 specialize qw/vpx_highbd_8_variance64x32 sse2/; 1308 1309 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"; 1310 specialize qw/vpx_highbd_8_variance32x64 sse2/; 1311 1312 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"; 1313 specialize qw/vpx_highbd_8_variance32x32 sse2/; 1314 1315 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"; 1316 specialize qw/vpx_highbd_8_variance32x16 sse2/; 1317 1318 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"; 1319 specialize qw/vpx_highbd_8_variance16x32 sse2/; 1320 1321 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"; 1322 specialize qw/vpx_highbd_8_variance16x16 sse2/; 1323 1324 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"; 1325 specialize qw/vpx_highbd_8_variance16x8 sse2/; 1326 1327 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"; 1328 specialize qw/vpx_highbd_8_variance8x16 sse2/; 1329 1330 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"; 1331 specialize qw/vpx_highbd_8_variance8x8 sse2/; 1332 1333 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"; 1334 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"; 1335 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"; 1336 1337 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"; 1338 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"; 1339 1340 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"; 1341 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"; 1342 1343 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"; 1344 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"; 1345 1346 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"; 1347 specialize qw/vpx_highbd_8_mse16x16 sse2/; 1348 1349 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"; 1350 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"; 1351 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"; 1352 specialize qw/vpx_highbd_8_mse8x8 sse2/; 1353 1354 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"; 1355 specialize qw/vpx_highbd_10_mse16x16 sse2/; 1356 1357 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"; 1358 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"; 1359 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"; 1360 specialize qw/vpx_highbd_10_mse8x8 sse2/; 1361 1362 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"; 1363 specialize qw/vpx_highbd_12_mse16x16 sse2/; 1364 1365 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"; 1366 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"; 1367 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"; 1368 specialize qw/vpx_highbd_12_mse8x8 sse2/; 1369 1370 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"; 1371 1372 # 1373 # Subpixel Variance 1374 # 1375 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"; 1376 specialize qw/vpx_highbd_12_sub_pixel_variance64x64 sse2/; 1377 1378 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"; 1379 specialize qw/vpx_highbd_12_sub_pixel_variance64x32 sse2/; 1380 1381 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"; 1382 specialize qw/vpx_highbd_12_sub_pixel_variance32x64 sse2/; 1383 1384 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"; 1385 specialize qw/vpx_highbd_12_sub_pixel_variance32x32 sse2/; 1386 1387 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"; 1388 specialize qw/vpx_highbd_12_sub_pixel_variance32x16 sse2/; 1389 1390 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"; 1391 specialize qw/vpx_highbd_12_sub_pixel_variance16x32 sse2/; 1392 1393 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"; 1394 specialize qw/vpx_highbd_12_sub_pixel_variance16x16 sse2/; 1395 1396 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"; 1397 specialize qw/vpx_highbd_12_sub_pixel_variance16x8 sse2/; 1398 1399 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"; 1400 specialize qw/vpx_highbd_12_sub_pixel_variance8x16 sse2/; 1401 1402 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"; 1403 specialize qw/vpx_highbd_12_sub_pixel_variance8x8 sse2/; 1404 1405 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"; 1406 specialize qw/vpx_highbd_12_sub_pixel_variance8x4 sse2/; 1407 1408 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"; 1409 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"; 1410 1411 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"; 1412 specialize qw/vpx_highbd_10_sub_pixel_variance64x64 sse2/; 1413 1414 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"; 1415 specialize qw/vpx_highbd_10_sub_pixel_variance64x32 sse2/; 1416 1417 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"; 1418 specialize qw/vpx_highbd_10_sub_pixel_variance32x64 sse2/; 1419 1420 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"; 1421 specialize qw/vpx_highbd_10_sub_pixel_variance32x32 sse2/; 1422 1423 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"; 1424 specialize qw/vpx_highbd_10_sub_pixel_variance32x16 sse2/; 1425 1426 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"; 1427 specialize qw/vpx_highbd_10_sub_pixel_variance16x32 sse2/; 1428 1429 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"; 1430 specialize qw/vpx_highbd_10_sub_pixel_variance16x16 sse2/; 1431 1432 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"; 1433 specialize qw/vpx_highbd_10_sub_pixel_variance16x8 sse2/; 1434 1435 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"; 1436 specialize qw/vpx_highbd_10_sub_pixel_variance8x16 sse2/; 1437 1438 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"; 1439 specialize qw/vpx_highbd_10_sub_pixel_variance8x8 sse2/; 1440 1441 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"; 1442 specialize qw/vpx_highbd_10_sub_pixel_variance8x4 sse2/; 1443 1444 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"; 1445 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"; 1446 1447 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"; 1448 specialize qw/vpx_highbd_8_sub_pixel_variance64x64 sse2/; 1449 1450 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"; 1451 specialize qw/vpx_highbd_8_sub_pixel_variance64x32 sse2/; 1452 1453 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"; 1454 specialize qw/vpx_highbd_8_sub_pixel_variance32x64 sse2/; 1455 1456 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"; 1457 specialize qw/vpx_highbd_8_sub_pixel_variance32x32 sse2/; 1458 1459 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"; 1460 specialize qw/vpx_highbd_8_sub_pixel_variance32x16 sse2/; 1461 1462 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"; 1463 specialize qw/vpx_highbd_8_sub_pixel_variance16x32 sse2/; 1464 1465 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"; 1466 specialize qw/vpx_highbd_8_sub_pixel_variance16x16 sse2/; 1467 1468 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"; 1469 specialize qw/vpx_highbd_8_sub_pixel_variance16x8 sse2/; 1470 1471 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"; 1472 specialize qw/vpx_highbd_8_sub_pixel_variance8x16 sse2/; 1473 1474 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"; 1475 specialize qw/vpx_highbd_8_sub_pixel_variance8x8 sse2/; 1476 1477 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"; 1478 specialize qw/vpx_highbd_8_sub_pixel_variance8x4 sse2/; 1479 1480 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"; 1481 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"; 1482 1483 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"; 1484 specialize qw/vpx_highbd_12_sub_pixel_avg_variance64x64 sse2/; 1485 1486 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"; 1487 specialize qw/vpx_highbd_12_sub_pixel_avg_variance64x32 sse2/; 1488 1489 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"; 1490 specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x64 sse2/; 1491 1492 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"; 1493 specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x32 sse2/; 1494 1495 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"; 1496 specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x16 sse2/; 1497 1498 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"; 1499 specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x32 sse2/; 1500 1501 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"; 1502 specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x16 sse2/; 1503 1504 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"; 1505 specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x8 sse2/; 1506 1507 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"; 1508 specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x16 sse2/; 1509 1510 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"; 1511 specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x8 sse2/; 1512 1513 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"; 1514 specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x4 sse2/; 1515 1516 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"; 1517 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"; 1518 1519 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"; 1520 specialize qw/vpx_highbd_10_sub_pixel_avg_variance64x64 sse2/; 1521 1522 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"; 1523 specialize qw/vpx_highbd_10_sub_pixel_avg_variance64x32 sse2/; 1524 1525 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"; 1526 specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x64 sse2/; 1527 1528 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"; 1529 specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x32 sse2/; 1530 1531 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"; 1532 specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x16 sse2/; 1533 1534 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"; 1535 specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x32 sse2/; 1536 1537 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"; 1538 specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x16 sse2/; 1539 1540 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"; 1541 specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x8 sse2/; 1542 1543 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"; 1544 specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x16 sse2/; 1545 1546 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"; 1547 specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x8 sse2/; 1548 1549 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"; 1550 specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x4 sse2/; 1551 1552 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"; 1553 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"; 1554 1555 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"; 1556 specialize qw/vpx_highbd_8_sub_pixel_avg_variance64x64 sse2/; 1557 1558 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"; 1559 specialize qw/vpx_highbd_8_sub_pixel_avg_variance64x32 sse2/; 1560 1561 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"; 1562 specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x64 sse2/; 1563 1564 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"; 1565 specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x32 sse2/; 1566 1567 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"; 1568 specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x16 sse2/; 1569 1570 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"; 1571 specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x32 sse2/; 1572 1573 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"; 1574 specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x16 sse2/; 1575 1576 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"; 1577 specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x8 sse2/; 1578 1579 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"; 1580 specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x16 sse2/; 1581 1582 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"; 1583 specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x8 sse2/; 1584 1585 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"; 1586 specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x4 sse2/; 1587 1588 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"; 1589 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"; 1590 1591 } # CONFIG_VP9_HIGHBITDEPTH 1592 1593 # 1594 # Post Processing 1595 # 1596 if (vpx_config("CONFIG_POSTPROC") eq "yes" || vpx_config("CONFIG_VP9_POSTPROC") eq "yes") { 1597 add_proto qw/void vpx_plane_add_noise/, "uint8_t *start, const int8_t *noise, int blackclamp, int whiteclamp, int width, int height, int pitch"; 1598 specialize qw/vpx_plane_add_noise sse2 msa/; 1599 1600 add_proto qw/void vpx_mbpost_proc_down/, "unsigned char *dst, int pitch, int rows, int cols,int flimit"; 1601 specialize qw/vpx_mbpost_proc_down sse2 neon msa/; 1602 1603 add_proto qw/void vpx_mbpost_proc_across_ip/, "unsigned char *dst, int pitch, int rows, int cols,int flimit"; 1604 specialize qw/vpx_mbpost_proc_across_ip sse2 neon msa/; 1605 1606 add_proto qw/void vpx_post_proc_down_and_across_mb_row/, "unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size"; 1607 specialize qw/vpx_post_proc_down_and_across_mb_row sse2 neon msa/; 1608 1609 } 1610 1611 } # CONFIG_ENCODERS || CONFIG_POSTPROC || CONFIG_VP9_POSTPROC 1612 1613 1; 1614