Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 1 | /* |
Krishna Rapaka | 7319db5 | 2021-09-28 20:35:29 -0700 | [diff] [blame] | 2 | * Copyright (c) 2021, Alliance for Open Media. All rights reserved |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 3 | * |
Vibhoothi | 41c6dd7 | 2021-10-12 18:48:26 +0000 | [diff] [blame] | 4 | * This source code is subject to the terms of the BSD 3-Clause Clear License |
| 5 | * and the Alliance for Open Media Patent License 1.0. If the BSD 3-Clause Clear |
| 6 | * License was not distributed with this source code in the LICENSE file, you |
| 7 | * can obtain it at aomedia.org/license/software-license/bsd-3-c-c/. If the |
| 8 | * Alliance for Open Media Patent License 1.0 was not distributed with this |
| 9 | * source code in the PATENTS file, you can obtain it at |
| 10 | * aomedia.org/license/patent-license/. |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 11 | */ |
| 12 | |
| 13 | #include "aom_dsp/mips/common_dspr2.h" |
| 14 | |
| 15 | #if HAVE_DSPR2 |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 16 | void aom_h_predictor_16x16_dspr2(uint8_t *dst, ptrdiff_t stride, |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 17 | const uint8_t *above, const uint8_t *left) { |
| 18 | int32_t tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8; |
| 19 | int32_t tmp9, tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16; |
| 20 | |
Tom Finegan | 3b41136 | 2017-09-13 12:12:56 -0700 | [diff] [blame] | 21 | (void)above; |
| 22 | |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 23 | __asm__ __volatile__( |
| 24 | "lb %[tmp1], (%[left]) \n\t" |
| 25 | "lb %[tmp2], 1(%[left]) \n\t" |
| 26 | "lb %[tmp3], 2(%[left]) \n\t" |
| 27 | "lb %[tmp4], 3(%[left]) \n\t" |
| 28 | "lb %[tmp5], 4(%[left]) \n\t" |
| 29 | "lb %[tmp6], 5(%[left]) \n\t" |
| 30 | "lb %[tmp7], 6(%[left]) \n\t" |
| 31 | "lb %[tmp8], 7(%[left]) \n\t" |
| 32 | "lb %[tmp9], 8(%[left]) \n\t" |
| 33 | "lb %[tmp10], 9(%[left]) \n\t" |
| 34 | "lb %[tmp11], 10(%[left]) \n\t" |
| 35 | "lb %[tmp12], 11(%[left]) \n\t" |
| 36 | "lb %[tmp13], 12(%[left]) \n\t" |
| 37 | "lb %[tmp14], 13(%[left]) \n\t" |
| 38 | "lb %[tmp15], 14(%[left]) \n\t" |
| 39 | "lb %[tmp16], 15(%[left]) \n\t" |
| 40 | |
| 41 | "replv.qb %[tmp1], %[tmp1] \n\t" |
| 42 | "replv.qb %[tmp2], %[tmp2] \n\t" |
| 43 | "replv.qb %[tmp3], %[tmp3] \n\t" |
| 44 | "replv.qb %[tmp4], %[tmp4] \n\t" |
| 45 | "replv.qb %[tmp5], %[tmp5] \n\t" |
| 46 | "replv.qb %[tmp6], %[tmp6] \n\t" |
| 47 | "replv.qb %[tmp7], %[tmp7] \n\t" |
| 48 | "replv.qb %[tmp8], %[tmp8] \n\t" |
| 49 | "replv.qb %[tmp9], %[tmp9] \n\t" |
| 50 | "replv.qb %[tmp10], %[tmp10] \n\t" |
| 51 | "replv.qb %[tmp11], %[tmp11] \n\t" |
| 52 | "replv.qb %[tmp12], %[tmp12] \n\t" |
| 53 | "replv.qb %[tmp13], %[tmp13] \n\t" |
| 54 | "replv.qb %[tmp14], %[tmp14] \n\t" |
| 55 | "replv.qb %[tmp15], %[tmp15] \n\t" |
| 56 | "replv.qb %[tmp16], %[tmp16] \n\t" |
| 57 | |
| 58 | "sw %[tmp1], (%[dst]) \n\t" |
| 59 | "sw %[tmp1], 4(%[dst]) \n\t" |
| 60 | "sw %[tmp1], 8(%[dst]) \n\t" |
| 61 | "sw %[tmp1], 12(%[dst]) \n\t" |
| 62 | |
| 63 | "add %[dst], %[dst], %[stride] \n\t" |
| 64 | "sw %[tmp2], (%[dst]) \n\t" |
| 65 | "sw %[tmp2], 4(%[dst]) \n\t" |
| 66 | "sw %[tmp2], 8(%[dst]) \n\t" |
| 67 | "sw %[tmp2], 12(%[dst]) \n\t" |
| 68 | |
| 69 | "add %[dst], %[dst], %[stride] \n\t" |
| 70 | "sw %[tmp3], (%[dst]) \n\t" |
| 71 | "sw %[tmp3], 4(%[dst]) \n\t" |
| 72 | "sw %[tmp3], 8(%[dst]) \n\t" |
| 73 | "sw %[tmp3], 12(%[dst]) \n\t" |
| 74 | |
| 75 | "add %[dst], %[dst], %[stride] \n\t" |
| 76 | "sw %[tmp4], (%[dst]) \n\t" |
| 77 | "sw %[tmp4], 4(%[dst]) \n\t" |
| 78 | "sw %[tmp4], 8(%[dst]) \n\t" |
| 79 | "sw %[tmp4], 12(%[dst]) \n\t" |
| 80 | |
| 81 | "add %[dst], %[dst], %[stride] \n\t" |
| 82 | "sw %[tmp5], (%[dst]) \n\t" |
| 83 | "sw %[tmp5], 4(%[dst]) \n\t" |
| 84 | "sw %[tmp5], 8(%[dst]) \n\t" |
| 85 | "sw %[tmp5], 12(%[dst]) \n\t" |
| 86 | |
| 87 | "add %[dst], %[dst], %[stride] \n\t" |
| 88 | "sw %[tmp6], (%[dst]) \n\t" |
| 89 | "sw %[tmp6], 4(%[dst]) \n\t" |
| 90 | "sw %[tmp6], 8(%[dst]) \n\t" |
| 91 | "sw %[tmp6], 12(%[dst]) \n\t" |
| 92 | |
| 93 | "add %[dst], %[dst], %[stride] \n\t" |
| 94 | "sw %[tmp7], (%[dst]) \n\t" |
| 95 | "sw %[tmp7], 4(%[dst]) \n\t" |
| 96 | "sw %[tmp7], 8(%[dst]) \n\t" |
| 97 | "sw %[tmp7], 12(%[dst]) \n\t" |
| 98 | |
| 99 | "add %[dst], %[dst], %[stride] \n\t" |
| 100 | "sw %[tmp8], (%[dst]) \n\t" |
| 101 | "sw %[tmp8], 4(%[dst]) \n\t" |
| 102 | "sw %[tmp8], 8(%[dst]) \n\t" |
| 103 | "sw %[tmp8], 12(%[dst]) \n\t" |
| 104 | |
| 105 | "add %[dst], %[dst], %[stride] \n\t" |
| 106 | "sw %[tmp9], (%[dst]) \n\t" |
| 107 | "sw %[tmp9], 4(%[dst]) \n\t" |
| 108 | "sw %[tmp9], 8(%[dst]) \n\t" |
| 109 | "sw %[tmp9], 12(%[dst]) \n\t" |
| 110 | |
| 111 | "add %[dst], %[dst], %[stride] \n\t" |
| 112 | "sw %[tmp10], (%[dst]) \n\t" |
| 113 | "sw %[tmp10], 4(%[dst]) \n\t" |
| 114 | "sw %[tmp10], 8(%[dst]) \n\t" |
| 115 | "sw %[tmp10], 12(%[dst]) \n\t" |
| 116 | |
| 117 | "add %[dst], %[dst], %[stride] \n\t" |
| 118 | "sw %[tmp11], (%[dst]) \n\t" |
| 119 | "sw %[tmp11], 4(%[dst]) \n\t" |
| 120 | "sw %[tmp11], 8(%[dst]) \n\t" |
| 121 | "sw %[tmp11], 12(%[dst]) \n\t" |
| 122 | |
| 123 | "add %[dst], %[dst], %[stride] \n\t" |
| 124 | "sw %[tmp12], (%[dst]) \n\t" |
| 125 | "sw %[tmp12], 4(%[dst]) \n\t" |
| 126 | "sw %[tmp12], 8(%[dst]) \n\t" |
| 127 | "sw %[tmp12], 12(%[dst]) \n\t" |
| 128 | |
| 129 | "add %[dst], %[dst], %[stride] \n\t" |
| 130 | "sw %[tmp13], (%[dst]) \n\t" |
| 131 | "sw %[tmp13], 4(%[dst]) \n\t" |
| 132 | "sw %[tmp13], 8(%[dst]) \n\t" |
| 133 | "sw %[tmp13], 12(%[dst]) \n\t" |
| 134 | |
| 135 | "add %[dst], %[dst], %[stride] \n\t" |
| 136 | "sw %[tmp14], (%[dst]) \n\t" |
| 137 | "sw %[tmp14], 4(%[dst]) \n\t" |
| 138 | "sw %[tmp14], 8(%[dst]) \n\t" |
| 139 | "sw %[tmp14], 12(%[dst]) \n\t" |
| 140 | |
| 141 | "add %[dst], %[dst], %[stride] \n\t" |
| 142 | "sw %[tmp15], (%[dst]) \n\t" |
| 143 | "sw %[tmp15], 4(%[dst]) \n\t" |
| 144 | "sw %[tmp15], 8(%[dst]) \n\t" |
| 145 | "sw %[tmp15], 12(%[dst]) \n\t" |
| 146 | |
| 147 | "add %[dst], %[dst], %[stride] \n\t" |
| 148 | "sw %[tmp16], (%[dst]) \n\t" |
| 149 | "sw %[tmp16], 4(%[dst]) \n\t" |
| 150 | "sw %[tmp16], 8(%[dst]) \n\t" |
| 151 | "sw %[tmp16], 12(%[dst]) \n\t" |
| 152 | |
| 153 | : [tmp1] "=&r"(tmp1), [tmp2] "=&r"(tmp2), [tmp3] "=&r"(tmp3), |
| 154 | [tmp4] "=&r"(tmp4), [tmp5] "=&r"(tmp5), [tmp7] "=&r"(tmp7), |
| 155 | [tmp6] "=&r"(tmp6), [tmp8] "=&r"(tmp8), [tmp9] "=&r"(tmp9), |
| 156 | [tmp10] "=&r"(tmp10), [tmp11] "=&r"(tmp11), [tmp12] "=&r"(tmp12), |
| 157 | [tmp13] "=&r"(tmp13), [tmp14] "=&r"(tmp14), [tmp15] "=&r"(tmp15), |
| 158 | [tmp16] "=&r"(tmp16) |
| 159 | : [left] "r"(left), [dst] "r"(dst), [stride] "r"(stride)); |
| 160 | } |
| 161 | |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 162 | void aom_dc_predictor_16x16_dspr2(uint8_t *dst, ptrdiff_t stride, |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 163 | const uint8_t *above, const uint8_t *left) { |
| 164 | int32_t expected_dc; |
| 165 | int32_t average; |
| 166 | int32_t tmp, above1, above_l1, above_r1, left1, left_r1, left_l1; |
| 167 | int32_t above2, left2; |
| 168 | |
| 169 | __asm__ __volatile__( |
| 170 | "lw %[above1], (%[above]) \n\t" |
| 171 | "lw %[above2], 4(%[above]) \n\t" |
| 172 | "lw %[left1], (%[left]) \n\t" |
| 173 | "lw %[left2], 4(%[left]) \n\t" |
| 174 | |
| 175 | "preceu.ph.qbl %[above_l1], %[above1] \n\t" |
| 176 | "preceu.ph.qbr %[above_r1], %[above1] \n\t" |
| 177 | "preceu.ph.qbl %[left_l1], %[left1] \n\t" |
| 178 | "preceu.ph.qbr %[left_r1], %[left1] \n\t" |
| 179 | |
| 180 | "addu.ph %[average], %[above_r1], %[above_l1] \n\t" |
| 181 | "addu.ph %[average], %[average], %[left_l1] \n\t" |
| 182 | "addu.ph %[average], %[average], %[left_r1] \n\t" |
| 183 | |
| 184 | "preceu.ph.qbl %[above_l1], %[above2] \n\t" |
| 185 | "preceu.ph.qbr %[above_r1], %[above2] \n\t" |
| 186 | "preceu.ph.qbl %[left_l1], %[left2] \n\t" |
| 187 | "preceu.ph.qbr %[left_r1], %[left2] \n\t" |
| 188 | |
| 189 | "addu.ph %[average], %[average], %[above_l1] \n\t" |
| 190 | "addu.ph %[average], %[average], %[above_r1] \n\t" |
| 191 | "addu.ph %[average], %[average], %[left_l1] \n\t" |
| 192 | "addu.ph %[average], %[average], %[left_r1] \n\t" |
| 193 | |
| 194 | "lw %[above1], 8(%[above]) \n\t" |
| 195 | "lw %[above2], 12(%[above]) \n\t" |
| 196 | "lw %[left1], 8(%[left]) \n\t" |
| 197 | "lw %[left2], 12(%[left]) \n\t" |
| 198 | |
| 199 | "preceu.ph.qbl %[above_l1], %[above1] \n\t" |
| 200 | "preceu.ph.qbr %[above_r1], %[above1] \n\t" |
| 201 | "preceu.ph.qbl %[left_l1], %[left1] \n\t" |
| 202 | "preceu.ph.qbr %[left_r1], %[left1] \n\t" |
| 203 | |
| 204 | "addu.ph %[average], %[average], %[above_l1] \n\t" |
| 205 | "addu.ph %[average], %[average], %[above_r1] \n\t" |
| 206 | "addu.ph %[average], %[average], %[left_l1] \n\t" |
| 207 | "addu.ph %[average], %[average], %[left_r1] \n\t" |
| 208 | |
| 209 | "preceu.ph.qbl %[above_l1], %[above2] \n\t" |
| 210 | "preceu.ph.qbr %[above_r1], %[above2] \n\t" |
| 211 | "preceu.ph.qbl %[left_l1], %[left2] \n\t" |
| 212 | "preceu.ph.qbr %[left_r1], %[left2] \n\t" |
| 213 | |
| 214 | "addu.ph %[average], %[average], %[above_l1] \n\t" |
| 215 | "addu.ph %[average], %[average], %[above_r1] \n\t" |
| 216 | "addu.ph %[average], %[average], %[left_l1] \n\t" |
| 217 | "addu.ph %[average], %[average], %[left_r1] \n\t" |
| 218 | |
| 219 | "addiu %[average], %[average], 16 \n\t" |
| 220 | "srl %[tmp], %[average], 16 \n\t" |
| 221 | "addu.ph %[average], %[tmp], %[average] \n\t" |
| 222 | "srl %[expected_dc], %[average], 5 \n\t" |
| 223 | "replv.qb %[expected_dc], %[expected_dc] \n\t" |
| 224 | |
| 225 | "sw %[expected_dc], (%[dst]) \n\t" |
| 226 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 227 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 228 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 229 | |
| 230 | "add %[dst], %[dst], %[stride] \n\t" |
| 231 | "sw %[expected_dc], (%[dst]) \n\t" |
| 232 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 233 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 234 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 235 | |
| 236 | "add %[dst], %[dst], %[stride] \n\t" |
| 237 | "sw %[expected_dc], (%[dst]) \n\t" |
| 238 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 239 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 240 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 241 | |
| 242 | "add %[dst], %[dst], %[stride] \n\t" |
| 243 | "sw %[expected_dc], (%[dst]) \n\t" |
| 244 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 245 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 246 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 247 | |
| 248 | "add %[dst], %[dst], %[stride] \n\t" |
| 249 | "sw %[expected_dc], (%[dst]) \n\t" |
| 250 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 251 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 252 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 253 | |
| 254 | "add %[dst], %[dst], %[stride] \n\t" |
| 255 | "sw %[expected_dc], (%[dst]) \n\t" |
| 256 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 257 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 258 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 259 | |
| 260 | "add %[dst], %[dst], %[stride] \n\t" |
| 261 | "sw %[expected_dc], (%[dst]) \n\t" |
| 262 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 263 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 264 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 265 | |
| 266 | "add %[dst], %[dst], %[stride] \n\t" |
| 267 | "sw %[expected_dc], (%[dst]) \n\t" |
| 268 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 269 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 270 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 271 | |
| 272 | "add %[dst], %[dst], %[stride] \n\t" |
| 273 | "sw %[expected_dc], (%[dst]) \n\t" |
| 274 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 275 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 276 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 277 | |
| 278 | "add %[dst], %[dst], %[stride] \n\t" |
| 279 | "sw %[expected_dc], (%[dst]) \n\t" |
| 280 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 281 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 282 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 283 | |
| 284 | "add %[dst], %[dst], %[stride] \n\t" |
| 285 | "sw %[expected_dc], (%[dst]) \n\t" |
| 286 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 287 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 288 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 289 | |
| 290 | "add %[dst], %[dst], %[stride] \n\t" |
| 291 | "sw %[expected_dc], (%[dst]) \n\t" |
| 292 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 293 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 294 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 295 | |
| 296 | "add %[dst], %[dst], %[stride] \n\t" |
| 297 | "sw %[expected_dc], (%[dst]) \n\t" |
| 298 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 299 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 300 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 301 | |
| 302 | "add %[dst], %[dst], %[stride] \n\t" |
| 303 | "sw %[expected_dc], (%[dst]) \n\t" |
| 304 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 305 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 306 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 307 | |
| 308 | "add %[dst], %[dst], %[stride] \n\t" |
| 309 | "sw %[expected_dc], (%[dst]) \n\t" |
| 310 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 311 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 312 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 313 | |
| 314 | "add %[dst], %[dst], %[stride] \n\t" |
| 315 | "sw %[expected_dc], (%[dst]) \n\t" |
| 316 | "sw %[expected_dc], 4(%[dst]) \n\t" |
| 317 | "sw %[expected_dc], 8(%[dst]) \n\t" |
| 318 | "sw %[expected_dc], 12(%[dst]) \n\t" |
| 319 | |
| 320 | : [left1] "=&r"(left1), [above1] "=&r"(above1), [left_l1] "=&r"(left_l1), |
| 321 | [above_l1] "=&r"(above_l1), [left_r1] "=&r"(left_r1), |
| 322 | [above_r1] "=&r"(above_r1), [above2] "=&r"(above2), |
| 323 | [left2] "=&r"(left2), [average] "=&r"(average), [tmp] "=&r"(tmp), |
| 324 | [expected_dc] "=&r"(expected_dc) |
| 325 | : [above] "r"(above), [left] "r"(left), [dst] "r"(dst), |
| 326 | [stride] "r"(stride)); |
| 327 | } |
| 328 | #endif // #if HAVE_DSPR2 |