1 /* 2 * Copyright (c) 2016 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 #include <arm_neon.h> 12 13 #include "./vpx_config.h" 14 #include "./vpx_dsp_rtcd.h" 15 #include "vpx/vpx_integer.h" 16 17 //------------------------------------------------------------------------------ 18 // DC 4x4 19 20 static INLINE uint16x4_t dc_sum_4(const uint16_t *ref) { 21 const uint16x4_t ref_u16 = vld1_u16(ref); 22 const uint16x4_t p0 = vpadd_u16(ref_u16, ref_u16); 23 return vpadd_u16(p0, p0); 24 } 25 26 static INLINE void dc_store_4x4(uint16_t *dst, ptrdiff_t stride, 27 const uint16x4_t dc) { 28 const uint16x4_t dc_dup = vdup_lane_u16(dc, 0); 29 int i; 30 for (i = 0; i < 4; ++i, dst += stride) { 31 vst1_u16(dst, dc_dup); 32 } 33 } 34 35 void vpx_highbd_dc_predictor_4x4_neon(uint16_t *dst, ptrdiff_t stride, 36 const uint16_t *above, 37 const uint16_t *left, int bd) { 38 const uint16x4_t a = vld1_u16(above); 39 const uint16x4_t l = vld1_u16(left); 40 uint16x4_t sum; 41 uint16x4_t dc; 42 (void)bd; 43 sum = vadd_u16(a, l); 44 sum = vpadd_u16(sum, sum); 45 sum = vpadd_u16(sum, sum); 46 dc = vrshr_n_u16(sum, 3); 47 dc_store_4x4(dst, stride, dc); 48 } 49 50 void vpx_highbd_dc_left_predictor_4x4_neon(uint16_t *dst, ptrdiff_t stride, 51 const uint16_t *above, 52 const uint16_t *left, int bd) { 53 const uint16x4_t sum = dc_sum_4(left); 54 const uint16x4_t dc = vrshr_n_u16(sum, 2); 55 (void)above; 56 (void)bd; 57 dc_store_4x4(dst, stride, dc); 58 } 59 60 void vpx_highbd_dc_top_predictor_4x4_neon(uint16_t *dst, ptrdiff_t stride, 61 const uint16_t *above, 62 const uint16_t *left, int bd) { 63 const uint16x4_t sum = dc_sum_4(above); 64 const uint16x4_t dc = vrshr_n_u16(sum, 2); 65 (void)left; 66 (void)bd; 67 dc_store_4x4(dst, stride, dc); 68 } 69 70 void vpx_highbd_dc_128_predictor_4x4_neon(uint16_t *dst, ptrdiff_t stride, 71 const uint16_t *above, 72 const uint16_t *left, int bd) { 73 const uint16x4_t dc = vdup_n_u16(1 << (bd - 1)); 74 (void)above; 75 (void)left; 76 dc_store_4x4(dst, stride, dc); 77 } 78 79 //------------------------------------------------------------------------------ 80 // DC 8x8 81 82 static INLINE uint16x4_t dc_sum_8(const uint16_t *ref) { 83 const uint16x8_t ref_u16 = vld1q_u16(ref); 84 uint16x4_t sum = vadd_u16(vget_low_u16(ref_u16), vget_high_u16(ref_u16)); 85 sum = vpadd_u16(sum, sum); 86 return vpadd_u16(sum, sum); 87 } 88 89 static INLINE void dc_store_8x8(uint16_t *dst, ptrdiff_t stride, 90 const uint16x4_t dc) { 91 const uint16x8_t dc_dup = vdupq_lane_u16(dc, 0); 92 int i; 93 for (i = 0; i < 8; ++i, dst += stride) { 94 vst1q_u16(dst, dc_dup); 95 } 96 } 97 98 void vpx_highbd_dc_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, 99 const uint16_t *above, 100 const uint16_t *left, int bd) { 101 const uint16x8_t above_u16 = vld1q_u16(above); 102 const uint16x8_t left_u16 = vld1q_u16(left); 103 const uint16x8_t p0 = vaddq_u16(above_u16, left_u16); 104 uint16x4_t sum = vadd_u16(vget_low_u16(p0), vget_high_u16(p0)); 105 uint16x4_t dc; 106 (void)bd; 107 sum = vpadd_u16(sum, sum); 108 sum = vpadd_u16(sum, sum); 109 dc = vrshr_n_u16(sum, 4); 110 dc_store_8x8(dst, stride, dc); 111 } 112 113 void vpx_highbd_dc_left_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, 114 const uint16_t *above, 115 const uint16_t *left, int bd) { 116 const uint16x4_t sum = dc_sum_8(left); 117 const uint16x4_t dc = vrshr_n_u16(sum, 3); 118 (void)above; 119 (void)bd; 120 dc_store_8x8(dst, stride, dc); 121 } 122 123 void vpx_highbd_dc_top_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, 124 const uint16_t *above, 125 const uint16_t *left, int bd) { 126 const uint16x4_t sum = dc_sum_8(above); 127 const uint16x4_t dc = vrshr_n_u16(sum, 3); 128 (void)left; 129 (void)bd; 130 dc_store_8x8(dst, stride, dc); 131 } 132 133 void vpx_highbd_dc_128_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, 134 const uint16_t *above, 135 const uint16_t *left, int bd) { 136 const uint16x4_t dc = vdup_n_u16(1 << (bd - 1)); 137 (void)above; 138 (void)left; 139 dc_store_8x8(dst, stride, dc); 140 } 141 142 //------------------------------------------------------------------------------ 143 // DC 16x16 144 145 static INLINE uint16x4_t dc_sum_16(const uint16_t *ref) { 146 const uint16x8x2_t ref_u16 = vld2q_u16(ref); 147 const uint16x8_t p0 = vaddq_u16(ref_u16.val[0], ref_u16.val[1]); 148 uint16x4_t sum = vadd_u16(vget_low_u16(p0), vget_high_u16(p0)); 149 sum = vpadd_u16(sum, sum); 150 return vpadd_u16(sum, sum); 151 } 152 153 static INLINE void dc_store_16x16(uint16_t *dst, ptrdiff_t stride, 154 const uint16x4_t dc) { 155 uint16x8x2_t dc_dup; 156 int i; 157 dc_dup.val[0] = dc_dup.val[1] = vdupq_lane_u16(dc, 0); 158 for (i = 0; i < 16; ++i, dst += stride) { 159 vst2q_u16(dst, dc_dup); 160 } 161 } 162 163 void vpx_highbd_dc_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, 164 const uint16_t *above, 165 const uint16_t *left, int bd) { 166 const uint16x8x2_t a = vld2q_u16(above); 167 const uint16x8x2_t l = vld2q_u16(left); 168 const uint16x8_t pa = vaddq_u16(a.val[0], a.val[1]); 169 const uint16x8_t pl = vaddq_u16(l.val[0], l.val[1]); 170 const uint16x8_t pal0 = vaddq_u16(pa, pl); 171 uint16x4_t pal1 = vadd_u16(vget_low_u16(pal0), vget_high_u16(pal0)); 172 uint32x2_t sum; 173 uint16x4_t dc; 174 (void)bd; 175 pal1 = vpadd_u16(pal1, pal1); 176 sum = vpaddl_u16(pal1); 177 dc = vreinterpret_u16_u32(vrshr_n_u32(sum, 5)); 178 dc_store_16x16(dst, stride, dc); 179 } 180 181 void vpx_highbd_dc_left_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, 182 const uint16_t *above, 183 const uint16_t *left, int bd) { 184 const uint16x4_t sum = dc_sum_16(left); 185 const uint16x4_t dc = vrshr_n_u16(sum, 4); 186 (void)above; 187 (void)bd; 188 dc_store_16x16(dst, stride, dc); 189 } 190 191 void vpx_highbd_dc_top_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, 192 const uint16_t *above, 193 const uint16_t *left, int bd) { 194 const uint16x4_t sum = dc_sum_16(above); 195 const uint16x4_t dc = vrshr_n_u16(sum, 4); 196 (void)left; 197 (void)bd; 198 dc_store_16x16(dst, stride, dc); 199 } 200 201 void vpx_highbd_dc_128_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, 202 const uint16_t *above, 203 const uint16_t *left, int bd) { 204 const uint16x4_t dc = vdup_n_u16(1 << (bd - 1)); 205 (void)above; 206 (void)left; 207 dc_store_16x16(dst, stride, dc); 208 } 209 210 //------------------------------------------------------------------------------ 211 // DC 32x32 212 213 static INLINE uint32x2_t dc_sum_32(const uint16_t *ref) { 214 const uint16x8x4_t r = vld4q_u16(ref); 215 const uint16x8_t p0 = vaddq_u16(r.val[0], r.val[1]); 216 const uint16x8_t p1 = vaddq_u16(r.val[2], r.val[3]); 217 const uint16x8_t p2 = vaddq_u16(p0, p1); 218 uint16x4_t sum = vadd_u16(vget_low_u16(p2), vget_high_u16(p2)); 219 sum = vpadd_u16(sum, sum); 220 return vpaddl_u16(sum); 221 } 222 223 static INLINE void dc_store_32x32(uint16_t *dst, ptrdiff_t stride, 224 const uint16x4_t dc) { 225 uint16x8x2_t dc_dup; 226 int i; 227 dc_dup.val[0] = dc_dup.val[1] = vdupq_lane_u16(dc, 0); 228 229 for (i = 0; i < 32; ++i) { 230 vst2q_u16(dst, dc_dup); 231 dst += 16; 232 vst2q_u16(dst, dc_dup); 233 dst += stride - 16; 234 } 235 } 236 237 void vpx_highbd_dc_predictor_32x32_neon(uint16_t *dst, ptrdiff_t stride, 238 const uint16_t *above, 239 const uint16_t *left, int bd) { 240 const uint16x8x4_t a = vld4q_u16(above); 241 const uint16x8x4_t l = vld4q_u16(left); 242 const uint16x8_t pa0 = vaddq_u16(a.val[0], a.val[1]); 243 const uint16x8_t pa1 = vaddq_u16(a.val[2], a.val[3]); 244 const uint16x8_t pl0 = vaddq_u16(l.val[0], l.val[1]); 245 const uint16x8_t pl1 = vaddq_u16(l.val[2], l.val[3]); 246 const uint16x8_t pa = vaddq_u16(pa0, pa1); 247 const uint16x8_t pl = vaddq_u16(pl0, pl1); 248 const uint16x8_t pal0 = vaddq_u16(pa, pl); 249 const uint16x4_t pal1 = vadd_u16(vget_low_u16(pal0), vget_high_u16(pal0)); 250 uint32x2_t sum = vpaddl_u16(pal1); 251 uint16x4_t dc; 252 (void)bd; 253 sum = vpadd_u32(sum, sum); 254 dc = vreinterpret_u16_u32(vrshr_n_u32(sum, 6)); 255 dc_store_32x32(dst, stride, dc); 256 } 257 258 void vpx_highbd_dc_left_predictor_32x32_neon(uint16_t *dst, ptrdiff_t stride, 259 const uint16_t *above, 260 const uint16_t *left, int bd) { 261 const uint32x2_t sum = dc_sum_32(left); 262 const uint16x4_t dc = vreinterpret_u16_u32(vrshr_n_u32(sum, 5)); 263 (void)above; 264 (void)bd; 265 dc_store_32x32(dst, stride, dc); 266 } 267 268 void vpx_highbd_dc_top_predictor_32x32_neon(uint16_t *dst, ptrdiff_t stride, 269 const uint16_t *above, 270 const uint16_t *left, int bd) { 271 const uint32x2_t sum = dc_sum_32(above); 272 const uint16x4_t dc = vreinterpret_u16_u32(vrshr_n_u32(sum, 5)); 273 (void)left; 274 (void)bd; 275 dc_store_32x32(dst, stride, dc); 276 } 277 278 void vpx_highbd_dc_128_predictor_32x32_neon(uint16_t *dst, ptrdiff_t stride, 279 const uint16_t *above, 280 const uint16_t *left, int bd) { 281 const uint16x4_t dc = vdup_n_u16(1 << (bd - 1)); 282 (void)above; 283 (void)left; 284 dc_store_32x32(dst, stride, dc); 285 } 286 287 // ----------------------------------------------------------------------------- 288 289 void vpx_highbd_d45_predictor_4x4_neon(uint16_t *dst, ptrdiff_t stride, 290 const uint16_t *above, 291 const uint16_t *left, int bd) { 292 const uint16x8_t ABCDEFGH = vld1q_u16(above); 293 const uint16x8_t BCDEFGH0 = vld1q_u16(above + 1); 294 const uint16x8_t CDEFGH00 = vld1q_u16(above + 2); 295 const uint16x8_t avg1 = vhaddq_u16(ABCDEFGH, CDEFGH00); 296 const uint16x8_t avg2 = vrhaddq_u16(avg1, BCDEFGH0); 297 const uint16x4_t avg2_low = vget_low_u16(avg2); 298 const uint16x4_t avg2_high = vget_high_u16(avg2); 299 const uint16x4_t r1 = vext_u16(avg2_low, avg2_high, 1); 300 const uint16x4_t r2 = vext_u16(avg2_low, avg2_high, 2); 301 const uint16x4_t r3 = vext_u16(avg2_low, avg2_high, 3); 302 (void)left; 303 (void)bd; 304 vst1_u16(dst, avg2_low); 305 dst += stride; 306 vst1_u16(dst, r1); 307 dst += stride; 308 vst1_u16(dst, r2); 309 dst += stride; 310 vst1_u16(dst, r3); 311 vst1q_lane_u16(dst + 3, ABCDEFGH, 7); 312 } 313 314 static INLINE void d45_store_8(uint16_t **dst, const ptrdiff_t stride, 315 const uint16x8_t above_right, uint16x8_t *row) { 316 *row = vextq_u16(*row, above_right, 1); 317 vst1q_u16(*dst, *row); 318 *dst += stride; 319 } 320 321 void vpx_highbd_d45_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, 322 const uint16_t *above, 323 const uint16_t *left, int bd) { 324 const uint16x8_t A0 = vld1q_u16(above); 325 const uint16x8_t above_right = vdupq_lane_u16(vget_high_u16(A0), 3); 326 const uint16x8_t A1 = vld1q_u16(above + 1); 327 const uint16x8_t A2 = vld1q_u16(above + 2); 328 const uint16x8_t avg1 = vhaddq_u16(A0, A2); 329 uint16x8_t row = vrhaddq_u16(avg1, A1); 330 (void)left; 331 (void)bd; 332 333 vst1q_u16(dst, row); 334 dst += stride; 335 d45_store_8(&dst, stride, above_right, &row); 336 d45_store_8(&dst, stride, above_right, &row); 337 d45_store_8(&dst, stride, above_right, &row); 338 d45_store_8(&dst, stride, above_right, &row); 339 d45_store_8(&dst, stride, above_right, &row); 340 d45_store_8(&dst, stride, above_right, &row); 341 vst1q_u16(dst, above_right); 342 } 343 344 static INLINE void d45_store_16(uint16_t **dst, const ptrdiff_t stride, 345 const uint16x8_t above_right, uint16x8_t *row_0, 346 uint16x8_t *row_1) { 347 *row_0 = vextq_u16(*row_0, *row_1, 1); 348 *row_1 = vextq_u16(*row_1, above_right, 1); 349 vst1q_u16(*dst, *row_0); 350 *dst += 8; 351 vst1q_u16(*dst, *row_1); 352 *dst += stride - 8; 353 } 354 355 void vpx_highbd_d45_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, 356 const uint16_t *above, 357 const uint16_t *left, int bd) { 358 const uint16x8_t A0_0 = vld1q_u16(above); 359 const uint16x8_t A0_1 = vld1q_u16(above + 8); 360 const uint16x8_t above_right = vdupq_lane_u16(vget_high_u16(A0_1), 3); 361 const uint16x8_t A1_0 = vld1q_u16(above + 1); 362 const uint16x8_t A1_1 = vld1q_u16(above + 9); 363 const uint16x8_t A2_0 = vld1q_u16(above + 2); 364 const uint16x8_t A2_1 = vld1q_u16(above + 10); 365 const uint16x8_t avg_0 = vhaddq_u16(A0_0, A2_0); 366 const uint16x8_t avg_1 = vhaddq_u16(A0_1, A2_1); 367 uint16x8_t row_0 = vrhaddq_u16(avg_0, A1_0); 368 uint16x8_t row_1 = vrhaddq_u16(avg_1, A1_1); 369 (void)left; 370 (void)bd; 371 372 vst1q_u16(dst, row_0); 373 vst1q_u16(dst + 8, row_1); 374 dst += stride; 375 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 376 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 377 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 378 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 379 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 380 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 381 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 382 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 383 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 384 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 385 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 386 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 387 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 388 d45_store_16(&dst, stride, above_right, &row_0, &row_1); 389 vst1q_u16(dst, above_right); 390 vst1q_u16(dst + 8, above_right); 391 } 392 393 void vpx_highbd_d45_predictor_32x32_neon(uint16_t *dst, ptrdiff_t stride, 394 const uint16_t *above, 395 const uint16_t *left, int bd) { 396 const uint16x8_t A0_0 = vld1q_u16(above); 397 const uint16x8_t A0_1 = vld1q_u16(above + 8); 398 const uint16x8_t A0_2 = vld1q_u16(above + 16); 399 const uint16x8_t A0_3 = vld1q_u16(above + 24); 400 const uint16x8_t above_right = vdupq_lane_u16(vget_high_u16(A0_3), 3); 401 const uint16x8_t A1_0 = vld1q_u16(above + 1); 402 const uint16x8_t A1_1 = vld1q_u16(above + 9); 403 const uint16x8_t A1_2 = vld1q_u16(above + 17); 404 const uint16x8_t A1_3 = vld1q_u16(above + 25); 405 const uint16x8_t A2_0 = vld1q_u16(above + 2); 406 const uint16x8_t A2_1 = vld1q_u16(above + 10); 407 const uint16x8_t A2_2 = vld1q_u16(above + 18); 408 const uint16x8_t A2_3 = vld1q_u16(above + 26); 409 const uint16x8_t avg_0 = vhaddq_u16(A0_0, A2_0); 410 const uint16x8_t avg_1 = vhaddq_u16(A0_1, A2_1); 411 const uint16x8_t avg_2 = vhaddq_u16(A0_2, A2_2); 412 const uint16x8_t avg_3 = vhaddq_u16(A0_3, A2_3); 413 uint16x8_t row_0 = vrhaddq_u16(avg_0, A1_0); 414 uint16x8_t row_1 = vrhaddq_u16(avg_1, A1_1); 415 uint16x8_t row_2 = vrhaddq_u16(avg_2, A1_2); 416 uint16x8_t row_3 = vrhaddq_u16(avg_3, A1_3); 417 int i; 418 (void)left; 419 (void)bd; 420 421 vst1q_u16(dst, row_0); 422 dst += 8; 423 vst1q_u16(dst, row_1); 424 dst += 8; 425 vst1q_u16(dst, row_2); 426 dst += 8; 427 vst1q_u16(dst, row_3); 428 dst += stride - 24; 429 430 for (i = 0; i < 30; ++i) { 431 row_0 = vextq_u16(row_0, row_1, 1); 432 row_1 = vextq_u16(row_1, row_2, 1); 433 row_2 = vextq_u16(row_2, row_3, 1); 434 row_3 = vextq_u16(row_3, above_right, 1); 435 vst1q_u16(dst, row_0); 436 dst += 8; 437 vst1q_u16(dst, row_1); 438 dst += 8; 439 vst1q_u16(dst, row_2); 440 dst += 8; 441 vst1q_u16(dst, row_3); 442 dst += stride - 24; 443 } 444 445 vst1q_u16(dst, above_right); 446 dst += 8; 447 vst1q_u16(dst, above_right); 448 dst += 8; 449 vst1q_u16(dst, above_right); 450 dst += 8; 451 vst1q_u16(dst, above_right); 452 } 453 454 // ----------------------------------------------------------------------------- 455 456 void vpx_highbd_d135_predictor_4x4_neon(uint16_t *dst, ptrdiff_t stride, 457 const uint16_t *above, 458 const uint16_t *left, int bd) { 459 const uint16x8_t XA0123___ = vld1q_u16(above - 1); 460 const uint16x4_t L0123 = vld1_u16(left); 461 const uint16x4_t L3210 = vrev64_u16(L0123); 462 const uint16x8_t L____3210 = vcombine_u16(L0123, L3210); 463 const uint16x8_t L3210XA012 = vcombine_u16(L3210, vget_low_u16(XA0123___)); 464 const uint16x8_t L210XA0123 = vextq_u16(L____3210, XA0123___, 5); 465 const uint16x8_t L10XA0123_ = vextq_u16(L____3210, XA0123___, 6); 466 const uint16x8_t avg1 = vhaddq_u16(L3210XA012, L10XA0123_); 467 const uint16x8_t avg2 = vrhaddq_u16(avg1, L210XA0123); 468 const uint16x4_t row_0 = vget_low_u16(avg2); 469 const uint16x4_t row_1 = vget_high_u16(avg2); 470 const uint16x4_t r0 = vext_u16(row_0, row_1, 3); 471 const uint16x4_t r1 = vext_u16(row_0, row_1, 2); 472 const uint16x4_t r2 = vext_u16(row_0, row_1, 1); 473 (void)bd; 474 vst1_u16(dst, r0); 475 dst += stride; 476 vst1_u16(dst, r1); 477 dst += stride; 478 vst1_u16(dst, r2); 479 dst += stride; 480 vst1_u16(dst, row_0); 481 } 482 483 void vpx_highbd_d135_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, 484 const uint16_t *above, 485 const uint16_t *left, int bd) { 486 const uint16x8_t XA0123456 = vld1q_u16(above - 1); 487 const uint16x8_t A01234567 = vld1q_u16(above); 488 const uint16x8_t A1234567_ = vld1q_u16(above + 1); 489 const uint16x8_t L01234567 = vld1q_u16(left); 490 const uint16x4_t L3210 = vrev64_u16(vget_low_u16(L01234567)); 491 const uint16x4_t L7654 = vrev64_u16(vget_high_u16(L01234567)); 492 const uint16x8_t L76543210 = vcombine_u16(L7654, L3210); 493 const uint16x8_t L6543210X = vextq_u16(L76543210, XA0123456, 1); 494 const uint16x8_t L543210XA0 = vextq_u16(L76543210, XA0123456, 2); 495 const uint16x8_t avg_0 = vhaddq_u16(L76543210, L543210XA0); 496 const uint16x8_t avg_1 = vhaddq_u16(XA0123456, A1234567_); 497 const uint16x8_t row_0 = vrhaddq_u16(avg_0, L6543210X); 498 const uint16x8_t row_1 = vrhaddq_u16(avg_1, A01234567); 499 const uint16x8_t r0 = vextq_u16(row_0, row_1, 7); 500 const uint16x8_t r1 = vextq_u16(row_0, row_1, 6); 501 const uint16x8_t r2 = vextq_u16(row_0, row_1, 5); 502 const uint16x8_t r3 = vextq_u16(row_0, row_1, 4); 503 const uint16x8_t r4 = vextq_u16(row_0, row_1, 3); 504 const uint16x8_t r5 = vextq_u16(row_0, row_1, 2); 505 const uint16x8_t r6 = vextq_u16(row_0, row_1, 1); 506 (void)bd; 507 vst1q_u16(dst, r0); 508 dst += stride; 509 vst1q_u16(dst, r1); 510 dst += stride; 511 vst1q_u16(dst, r2); 512 dst += stride; 513 vst1q_u16(dst, r3); 514 dst += stride; 515 vst1q_u16(dst, r4); 516 dst += stride; 517 vst1q_u16(dst, r5); 518 dst += stride; 519 vst1q_u16(dst, r6); 520 dst += stride; 521 vst1q_u16(dst, row_0); 522 } 523 524 static INLINE void d135_store_16(uint16_t **dst, const ptrdiff_t stride, 525 const uint16x8_t row_0, 526 const uint16x8_t row_1) { 527 vst1q_u16(*dst, row_0); 528 *dst += 8; 529 vst1q_u16(*dst, row_1); 530 *dst += stride - 8; 531 } 532 533 void vpx_highbd_d135_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, 534 const uint16_t *above, 535 const uint16_t *left, int bd) { 536 const uint16x8_t L01234567 = vld1q_u16(left); 537 const uint16x8_t L89abcdef = vld1q_u16(left + 8); 538 const uint16x4_t L3210 = vrev64_u16(vget_low_u16(L01234567)); 539 const uint16x4_t L7654 = vrev64_u16(vget_high_u16(L01234567)); 540 const uint16x4_t Lba98 = vrev64_u16(vget_low_u16(L89abcdef)); 541 const uint16x4_t Lfedc = vrev64_u16(vget_high_u16(L89abcdef)); 542 const uint16x8_t L76543210 = vcombine_u16(L7654, L3210); 543 const uint16x8_t Lfedcba98 = vcombine_u16(Lfedc, Lba98); 544 const uint16x8_t Ledcba987 = vextq_u16(Lfedcba98, L76543210, 1); 545 const uint16x8_t Ldcba9876 = vextq_u16(Lfedcba98, L76543210, 2); 546 const uint16x8_t avg_0 = vhaddq_u16(Lfedcba98, Ldcba9876); 547 const uint16x8_t row_0 = vrhaddq_u16(avg_0, Ledcba987); 548 549 const uint16x8_t XA0123456 = vld1q_u16(above - 1); 550 const uint16x8_t L6543210X = vextq_u16(L76543210, XA0123456, 1); 551 const uint16x8_t L543210XA0 = vextq_u16(L76543210, XA0123456, 2); 552 const uint16x8_t avg_1 = vhaddq_u16(L76543210, L543210XA0); 553 const uint16x8_t row_1 = vrhaddq_u16(avg_1, L6543210X); 554 555 const uint16x8_t A01234567 = vld1q_u16(above); 556 const uint16x8_t A12345678 = vld1q_u16(above + 1); 557 const uint16x8_t avg_2 = vhaddq_u16(XA0123456, A12345678); 558 const uint16x8_t row_2 = vrhaddq_u16(avg_2, A01234567); 559 560 const uint16x8_t A789abcde = vld1q_u16(above + 7); 561 const uint16x8_t A89abcdef = vld1q_u16(above + 8); 562 const uint16x8_t A9abcdef_ = vld1q_u16(above + 9); 563 const uint16x8_t avg_3 = vhaddq_u16(A789abcde, A9abcdef_); 564 const uint16x8_t row_3 = vrhaddq_u16(avg_3, A89abcdef); 565 566 const uint16x8_t r0_0 = vextq_u16(row_1, row_2, 7); 567 const uint16x8_t r0_1 = vextq_u16(row_2, row_3, 7); 568 const uint16x8_t r1_0 = vextq_u16(row_1, row_2, 6); 569 const uint16x8_t r1_1 = vextq_u16(row_2, row_3, 6); 570 const uint16x8_t r2_0 = vextq_u16(row_1, row_2, 5); 571 const uint16x8_t r2_1 = vextq_u16(row_2, row_3, 5); 572 const uint16x8_t r3_0 = vextq_u16(row_1, row_2, 4); 573 const uint16x8_t r3_1 = vextq_u16(row_2, row_3, 4); 574 const uint16x8_t r4_0 = vextq_u16(row_1, row_2, 3); 575 const uint16x8_t r4_1 = vextq_u16(row_2, row_3, 3); 576 const uint16x8_t r5_0 = vextq_u16(row_1, row_2, 2); 577 const uint16x8_t r5_1 = vextq_u16(row_2, row_3, 2); 578 const uint16x8_t r6_0 = vextq_u16(row_1, row_2, 1); 579 const uint16x8_t r6_1 = vextq_u16(row_2, row_3, 1); 580 const uint16x8_t r8_0 = vextq_u16(row_0, row_1, 7); 581 const uint16x8_t r9_0 = vextq_u16(row_0, row_1, 6); 582 const uint16x8_t ra_0 = vextq_u16(row_0, row_1, 5); 583 const uint16x8_t rb_0 = vextq_u16(row_0, row_1, 4); 584 const uint16x8_t rc_0 = vextq_u16(row_0, row_1, 3); 585 const uint16x8_t rd_0 = vextq_u16(row_0, row_1, 2); 586 const uint16x8_t re_0 = vextq_u16(row_0, row_1, 1); 587 (void)bd; 588 589 d135_store_16(&dst, stride, r0_0, r0_1); 590 d135_store_16(&dst, stride, r1_0, r1_1); 591 d135_store_16(&dst, stride, r2_0, r2_1); 592 d135_store_16(&dst, stride, r3_0, r3_1); 593 d135_store_16(&dst, stride, r4_0, r4_1); 594 d135_store_16(&dst, stride, r5_0, r5_1); 595 d135_store_16(&dst, stride, r6_0, r6_1); 596 d135_store_16(&dst, stride, row_1, row_2); 597 d135_store_16(&dst, stride, r8_0, r0_0); 598 d135_store_16(&dst, stride, r9_0, r1_0); 599 d135_store_16(&dst, stride, ra_0, r2_0); 600 d135_store_16(&dst, stride, rb_0, r3_0); 601 d135_store_16(&dst, stride, rc_0, r4_0); 602 d135_store_16(&dst, stride, rd_0, r5_0); 603 d135_store_16(&dst, stride, re_0, r6_0); 604 vst1q_u16(dst, row_0); 605 dst += 8; 606 vst1q_u16(dst, row_1); 607 } 608 609 void vpx_highbd_d135_predictor_32x32_neon(uint16_t *dst, ptrdiff_t stride, 610 const uint16_t *above, 611 const uint16_t *left, int bd) { 612 const uint16x8_t LL01234567 = vld1q_u16(left + 16); 613 const uint16x8_t LL89abcdef = vld1q_u16(left + 24); 614 const uint16x4_t LL3210 = vrev64_u16(vget_low_u16(LL01234567)); 615 const uint16x4_t LL7654 = vrev64_u16(vget_high_u16(LL01234567)); 616 const uint16x4_t LLba98 = vrev64_u16(vget_low_u16(LL89abcdef)); 617 const uint16x4_t LLfedc = vrev64_u16(vget_high_u16(LL89abcdef)); 618 const uint16x8_t LL76543210 = vcombine_u16(LL7654, LL3210); 619 const uint16x8_t LLfedcba98 = vcombine_u16(LLfedc, LLba98); 620 const uint16x8_t LLedcba987 = vextq_u16(LLfedcba98, LL76543210, 1); 621 const uint16x8_t LLdcba9876 = vextq_u16(LLfedcba98, LL76543210, 2); 622 const uint16x8_t avg_0 = vhaddq_u16(LLfedcba98, LLdcba9876); 623 uint16x8_t row_0 = vrhaddq_u16(avg_0, LLedcba987); 624 625 const uint16x8_t LU01234567 = vld1q_u16(left); 626 const uint16x8_t LU89abcdef = vld1q_u16(left + 8); 627 const uint16x4_t LU3210 = vrev64_u16(vget_low_u16(LU01234567)); 628 const uint16x4_t LU7654 = vrev64_u16(vget_high_u16(LU01234567)); 629 const uint16x4_t LUba98 = vrev64_u16(vget_low_u16(LU89abcdef)); 630 const uint16x4_t LUfedc = vrev64_u16(vget_high_u16(LU89abcdef)); 631 const uint16x8_t LU76543210 = vcombine_u16(LU7654, LU3210); 632 const uint16x8_t LUfedcba98 = vcombine_u16(LUfedc, LUba98); 633 const uint16x8_t LL6543210Uf = vextq_u16(LL76543210, LUfedcba98, 1); 634 const uint16x8_t LL543210Ufe = vextq_u16(LL76543210, LUfedcba98, 2); 635 const uint16x8_t avg_1 = vhaddq_u16(LL76543210, LL543210Ufe); 636 uint16x8_t row_1 = vrhaddq_u16(avg_1, LL6543210Uf); 637 638 const uint16x8_t LUedcba987 = vextq_u16(LUfedcba98, LU76543210, 1); 639 const uint16x8_t LUdcba9876 = vextq_u16(LUfedcba98, LU76543210, 2); 640 const uint16x8_t avg_2 = vhaddq_u16(LUfedcba98, LUdcba9876); 641 uint16x8_t row_2 = vrhaddq_u16(avg_2, LUedcba987); 642 643 const uint16x8_t XAL0123456 = vld1q_u16(above - 1); 644 const uint16x8_t LU6543210X = vextq_u16(LU76543210, XAL0123456, 1); 645 const uint16x8_t LU543210XA0 = vextq_u16(LU76543210, XAL0123456, 2); 646 const uint16x8_t avg_3 = vhaddq_u16(LU76543210, LU543210XA0); 647 uint16x8_t row_3 = vrhaddq_u16(avg_3, LU6543210X); 648 649 const uint16x8_t AL01234567 = vld1q_u16(above); 650 const uint16x8_t AL12345678 = vld1q_u16(above + 1); 651 const uint16x8_t avg_4 = vhaddq_u16(XAL0123456, AL12345678); 652 uint16x8_t row_4 = vrhaddq_u16(avg_4, AL01234567); 653 654 const uint16x8_t AL789abcde = vld1q_u16(above + 7); 655 const uint16x8_t AL89abcdef = vld1q_u16(above + 8); 656 const uint16x8_t AL9abcdefg = vld1q_u16(above + 9); 657 const uint16x8_t avg_5 = vhaddq_u16(AL789abcde, AL9abcdefg); 658 uint16x8_t row_5 = vrhaddq_u16(avg_5, AL89abcdef); 659 660 const uint16x8_t ALfR0123456 = vld1q_u16(above + 15); 661 const uint16x8_t AR01234567 = vld1q_u16(above + 16); 662 const uint16x8_t AR12345678 = vld1q_u16(above + 17); 663 const uint16x8_t avg_6 = vhaddq_u16(ALfR0123456, AR12345678); 664 uint16x8_t row_6 = vrhaddq_u16(avg_6, AR01234567); 665 666 const uint16x8_t AR789abcde = vld1q_u16(above + 23); 667 const uint16x8_t AR89abcdef = vld1q_u16(above + 24); 668 const uint16x8_t AR9abcdef_ = vld1q_u16(above + 25); 669 const uint16x8_t avg_7 = vhaddq_u16(AR789abcde, AR9abcdef_); 670 uint16x8_t row_7 = vrhaddq_u16(avg_7, AR89abcdef); 671 int i, j; 672 (void)bd; 673 674 dst += 31 * stride; 675 for (i = 0; i < 4; ++i) { 676 for (j = 0; j < 8; ++j) { 677 vst1q_u16(dst, row_0); 678 dst += 8; 679 vst1q_u16(dst, row_1); 680 dst += 8; 681 vst1q_u16(dst, row_2); 682 dst += 8; 683 vst1q_u16(dst, row_3); 684 dst -= stride + 24; 685 row_0 = vextq_u16(row_0, row_1, 1); 686 row_1 = vextq_u16(row_1, row_2, 1); 687 row_2 = vextq_u16(row_2, row_3, 1); 688 row_3 = vextq_u16(row_3, row_4, 1); 689 row_4 = vextq_u16(row_4, row_4, 1); 690 } 691 row_4 = row_5; 692 row_5 = row_6; 693 row_6 = row_7; 694 } 695 } 696 697 //------------------------------------------------------------------------------ 698 699 void vpx_highbd_v_predictor_4x4_neon(uint16_t *dst, ptrdiff_t stride, 700 const uint16_t *above, 701 const uint16_t *left, int bd) { 702 const uint16x4_t row = vld1_u16(above); 703 int i; 704 (void)left; 705 (void)bd; 706 707 for (i = 0; i < 4; i++, dst += stride) { 708 vst1_u16(dst, row); 709 } 710 } 711 712 void vpx_highbd_v_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, 713 const uint16_t *above, 714 const uint16_t *left, int bd) { 715 const uint16x8_t row = vld1q_u16(above); 716 int i; 717 (void)left; 718 (void)bd; 719 720 for (i = 0; i < 8; i++, dst += stride) { 721 vst1q_u16(dst, row); 722 } 723 } 724 725 void vpx_highbd_v_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, 726 const uint16_t *above, 727 const uint16_t *left, int bd) { 728 const uint16x8x2_t row = vld2q_u16(above); 729 int i; 730 (void)left; 731 (void)bd; 732 733 for (i = 0; i < 16; i++, dst += stride) { 734 vst2q_u16(dst, row); 735 } 736 } 737 738 void vpx_highbd_v_predictor_32x32_neon(uint16_t *dst, ptrdiff_t stride, 739 const uint16_t *above, 740 const uint16_t *left, int bd) { 741 const uint16x8x2_t row0 = vld2q_u16(above); 742 const uint16x8x2_t row1 = vld2q_u16(above + 16); 743 int i; 744 (void)left; 745 (void)bd; 746 747 for (i = 0; i < 32; i++) { 748 vst2q_u16(dst, row0); 749 dst += 16; 750 vst2q_u16(dst, row1); 751 dst += stride - 16; 752 } 753 } 754 755 // ----------------------------------------------------------------------------- 756 757 void vpx_highbd_h_predictor_4x4_neon(uint16_t *dst, ptrdiff_t stride, 758 const uint16_t *above, 759 const uint16_t *left, int bd) { 760 const uint16x4_t left_u16 = vld1_u16(left); 761 uint16x4_t row; 762 (void)above; 763 (void)bd; 764 765 row = vdup_lane_u16(left_u16, 0); 766 vst1_u16(dst, row); 767 dst += stride; 768 row = vdup_lane_u16(left_u16, 1); 769 vst1_u16(dst, row); 770 dst += stride; 771 row = vdup_lane_u16(left_u16, 2); 772 vst1_u16(dst, row); 773 dst += stride; 774 row = vdup_lane_u16(left_u16, 3); 775 vst1_u16(dst, row); 776 } 777 778 void vpx_highbd_h_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, 779 const uint16_t *above, 780 const uint16_t *left, int bd) { 781 const uint16x8_t left_u16 = vld1q_u16(left); 782 const uint16x4_t left_low = vget_low_u16(left_u16); 783 const uint16x4_t left_high = vget_high_u16(left_u16); 784 uint16x8_t row; 785 (void)above; 786 (void)bd; 787 788 row = vdupq_lane_u16(left_low, 0); 789 vst1q_u16(dst, row); 790 dst += stride; 791 row = vdupq_lane_u16(left_low, 1); 792 vst1q_u16(dst, row); 793 dst += stride; 794 row = vdupq_lane_u16(left_low, 2); 795 vst1q_u16(dst, row); 796 dst += stride; 797 row = vdupq_lane_u16(left_low, 3); 798 vst1q_u16(dst, row); 799 dst += stride; 800 row = vdupq_lane_u16(left_high, 0); 801 vst1q_u16(dst, row); 802 dst += stride; 803 row = vdupq_lane_u16(left_high, 1); 804 vst1q_u16(dst, row); 805 dst += stride; 806 row = vdupq_lane_u16(left_high, 2); 807 vst1q_u16(dst, row); 808 dst += stride; 809 row = vdupq_lane_u16(left_high, 3); 810 vst1q_u16(dst, row); 811 } 812 813 static INLINE void h_store_16(uint16_t **dst, const ptrdiff_t stride, 814 const uint16x8_t row) { 815 // Note: vst1q is faster than vst2q 816 vst1q_u16(*dst, row); 817 *dst += 8; 818 vst1q_u16(*dst, row); 819 *dst += stride - 8; 820 } 821 822 void vpx_highbd_h_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, 823 const uint16_t *above, 824 const uint16_t *left, int bd) { 825 int i; 826 (void)above; 827 (void)bd; 828 829 for (i = 0; i < 2; i++, left += 8) { 830 const uint16x8_t left_u16q = vld1q_u16(left); 831 const uint16x4_t left_low = vget_low_u16(left_u16q); 832 const uint16x4_t left_high = vget_high_u16(left_u16q); 833 uint16x8_t row; 834 835 row = vdupq_lane_u16(left_low, 0); 836 h_store_16(&dst, stride, row); 837 row = vdupq_lane_u16(left_low, 1); 838 h_store_16(&dst, stride, row); 839 row = vdupq_lane_u16(left_low, 2); 840 h_store_16(&dst, stride, row); 841 row = vdupq_lane_u16(left_low, 3); 842 h_store_16(&dst, stride, row); 843 row = vdupq_lane_u16(left_high, 0); 844 h_store_16(&dst, stride, row); 845 row = vdupq_lane_u16(left_high, 1); 846 h_store_16(&dst, stride, row); 847 row = vdupq_lane_u16(left_high, 2); 848 h_store_16(&dst, stride, row); 849 row = vdupq_lane_u16(left_high, 3); 850 h_store_16(&dst, stride, row); 851 } 852 } 853 854 static INLINE void h_store_32(uint16_t **dst, const ptrdiff_t stride, 855 const uint16x8_t row) { 856 // Note: vst1q is faster than vst2q 857 vst1q_u16(*dst, row); 858 *dst += 8; 859 vst1q_u16(*dst, row); 860 *dst += 8; 861 vst1q_u16(*dst, row); 862 *dst += 8; 863 vst1q_u16(*dst, row); 864 *dst += stride - 24; 865 } 866 867 void vpx_highbd_h_predictor_32x32_neon(uint16_t *dst, ptrdiff_t stride, 868 const uint16_t *above, 869 const uint16_t *left, int bd) { 870 int i; 871 (void)above; 872 (void)bd; 873 874 for (i = 0; i < 4; i++, left += 8) { 875 const uint16x8_t left_u16q = vld1q_u16(left); 876 const uint16x4_t left_low = vget_low_u16(left_u16q); 877 const uint16x4_t left_high = vget_high_u16(left_u16q); 878 uint16x8_t row; 879 880 row = vdupq_lane_u16(left_low, 0); 881 h_store_32(&dst, stride, row); 882 row = vdupq_lane_u16(left_low, 1); 883 h_store_32(&dst, stride, row); 884 row = vdupq_lane_u16(left_low, 2); 885 h_store_32(&dst, stride, row); 886 row = vdupq_lane_u16(left_low, 3); 887 h_store_32(&dst, stride, row); 888 row = vdupq_lane_u16(left_high, 0); 889 h_store_32(&dst, stride, row); 890 row = vdupq_lane_u16(left_high, 1); 891 h_store_32(&dst, stride, row); 892 row = vdupq_lane_u16(left_high, 2); 893 h_store_32(&dst, stride, row); 894 row = vdupq_lane_u16(left_high, 3); 895 h_store_32(&dst, stride, row); 896 } 897 } 898 899 // ----------------------------------------------------------------------------- 900 901 void vpx_highbd_tm_predictor_4x4_neon(uint16_t *dst, ptrdiff_t stride, 902 const uint16_t *above, 903 const uint16_t *left, int bd) { 904 const int16x8_t max = vmovq_n_s16((1 << bd) - 1); 905 const int16x8_t top_left = vld1q_dup_s16((const int16_t *)(above - 1)); 906 const int16x4_t above_s16d = vld1_s16((const int16_t *)above); 907 const int16x8_t above_s16 = vcombine_s16(above_s16d, above_s16d); 908 const int16x4_t left_s16 = vld1_s16((const int16_t *)left); 909 const int16x8_t sub = vsubq_s16(above_s16, top_left); 910 int16x8_t sum; 911 uint16x8_t row; 912 913 sum = vcombine_s16(vdup_lane_s16(left_s16, 0), vdup_lane_s16(left_s16, 1)); 914 sum = vaddq_s16(sum, sub); 915 sum = vminq_s16(sum, max); 916 row = vqshluq_n_s16(sum, 0); 917 vst1_u16(dst, vget_low_u16(row)); 918 dst += stride; 919 vst1_u16(dst, vget_high_u16(row)); 920 dst += stride; 921 922 sum = vcombine_s16(vdup_lane_s16(left_s16, 2), vdup_lane_s16(left_s16, 3)); 923 sum = vaddq_s16(sum, sub); 924 sum = vminq_s16(sum, max); 925 row = vqshluq_n_s16(sum, 0); 926 vst1_u16(dst, vget_low_u16(row)); 927 dst += stride; 928 vst1_u16(dst, vget_high_u16(row)); 929 } 930 931 static INLINE void tm_8_kernel(uint16_t **dst, const ptrdiff_t stride, 932 const int16x8_t left_dup, const int16x8_t sub, 933 const int16x8_t max) { 934 uint16x8_t row; 935 int16x8_t sum = vaddq_s16(left_dup, sub); 936 sum = vminq_s16(sum, max); 937 row = vqshluq_n_s16(sum, 0); 938 vst1q_u16(*dst, row); 939 *dst += stride; 940 } 941 942 void vpx_highbd_tm_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, 943 const uint16_t *above, 944 const uint16_t *left, int bd) { 945 const int16x8_t max = vmovq_n_s16((1 << bd) - 1); 946 const int16x8_t top_left = vld1q_dup_s16((const int16_t *)(above - 1)); 947 const int16x8_t above_s16 = vld1q_s16((const int16_t *)above); 948 const int16x8_t left_s16 = vld1q_s16((const int16_t *)left); 949 const int16x8_t sub = vsubq_s16(above_s16, top_left); 950 int16x4_t left_s16d; 951 int16x8_t left_dup; 952 int i; 953 954 left_s16d = vget_low_s16(left_s16); 955 956 for (i = 0; i < 2; i++, left_s16d = vget_high_s16(left_s16)) { 957 left_dup = vdupq_lane_s16(left_s16d, 0); 958 tm_8_kernel(&dst, stride, left_dup, sub, max); 959 960 left_dup = vdupq_lane_s16(left_s16d, 1); 961 tm_8_kernel(&dst, stride, left_dup, sub, max); 962 963 left_dup = vdupq_lane_s16(left_s16d, 2); 964 tm_8_kernel(&dst, stride, left_dup, sub, max); 965 966 left_dup = vdupq_lane_s16(left_s16d, 3); 967 tm_8_kernel(&dst, stride, left_dup, sub, max); 968 } 969 } 970 971 static INLINE void tm_16_kernel(uint16_t **dst, const ptrdiff_t stride, 972 const int16x8_t left_dup, const int16x8_t sub0, 973 const int16x8_t sub1, const int16x8_t max) { 974 uint16x8_t row0, row1; 975 int16x8_t sum0 = vaddq_s16(left_dup, sub0); 976 int16x8_t sum1 = vaddq_s16(left_dup, sub1); 977 sum0 = vminq_s16(sum0, max); 978 sum1 = vminq_s16(sum1, max); 979 row0 = vqshluq_n_s16(sum0, 0); 980 row1 = vqshluq_n_s16(sum1, 0); 981 vst1q_u16(*dst, row0); 982 *dst += 8; 983 vst1q_u16(*dst, row1); 984 *dst += stride - 8; 985 } 986 987 void vpx_highbd_tm_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, 988 const uint16_t *above, 989 const uint16_t *left, int bd) { 990 const int16x8_t max = vmovq_n_s16((1 << bd) - 1); 991 const int16x8_t top_left = vld1q_dup_s16((const int16_t *)(above - 1)); 992 const int16x8_t above0 = vld1q_s16((const int16_t *)above); 993 const int16x8_t above1 = vld1q_s16((const int16_t *)(above + 8)); 994 const int16x8_t sub0 = vsubq_s16(above0, top_left); 995 const int16x8_t sub1 = vsubq_s16(above1, top_left); 996 int16x8_t left_dup; 997 int i, j; 998 999 for (j = 0; j < 2; j++, left += 8) { 1000 const int16x8_t left_s16q = vld1q_s16((const int16_t *)left); 1001 int16x4_t left_s16d = vget_low_s16(left_s16q); 1002 for (i = 0; i < 2; i++, left_s16d = vget_high_s16(left_s16q)) { 1003 left_dup = vdupq_lane_s16(left_s16d, 0); 1004 tm_16_kernel(&dst, stride, left_dup, sub0, sub1, max); 1005 1006 left_dup = vdupq_lane_s16(left_s16d, 1); 1007 tm_16_kernel(&dst, stride, left_dup, sub0, sub1, max); 1008 1009 left_dup = vdupq_lane_s16(left_s16d, 2); 1010 tm_16_kernel(&dst, stride, left_dup, sub0, sub1, max); 1011 1012 left_dup = vdupq_lane_s16(left_s16d, 3); 1013 tm_16_kernel(&dst, stride, left_dup, sub0, sub1, max); 1014 } 1015 } 1016 } 1017 1018 static INLINE void tm_32_kernel(uint16_t **dst, const ptrdiff_t stride, 1019 const int16x8_t left_dup, const int16x8_t sub0, 1020 const int16x8_t sub1, const int16x8_t sub2, 1021 const int16x8_t sub3, const int16x8_t max) { 1022 uint16x8_t row0, row1, row2, row3; 1023 int16x8_t sum0 = vaddq_s16(left_dup, sub0); 1024 int16x8_t sum1 = vaddq_s16(left_dup, sub1); 1025 int16x8_t sum2 = vaddq_s16(left_dup, sub2); 1026 int16x8_t sum3 = vaddq_s16(left_dup, sub3); 1027 sum0 = vminq_s16(sum0, max); 1028 sum1 = vminq_s16(sum1, max); 1029 sum2 = vminq_s16(sum2, max); 1030 sum3 = vminq_s16(sum3, max); 1031 row0 = vqshluq_n_s16(sum0, 0); 1032 row1 = vqshluq_n_s16(sum1, 0); 1033 row2 = vqshluq_n_s16(sum2, 0); 1034 row3 = vqshluq_n_s16(sum3, 0); 1035 vst1q_u16(*dst, row0); 1036 *dst += 8; 1037 vst1q_u16(*dst, row1); 1038 *dst += 8; 1039 vst1q_u16(*dst, row2); 1040 *dst += 8; 1041 vst1q_u16(*dst, row3); 1042 *dst += stride - 24; 1043 } 1044 1045 void vpx_highbd_tm_predictor_32x32_neon(uint16_t *dst, ptrdiff_t stride, 1046 const uint16_t *above, 1047 const uint16_t *left, int bd) { 1048 const int16x8_t max = vmovq_n_s16((1 << bd) - 1); 1049 const int16x8_t top_left = vld1q_dup_s16((const int16_t *)(above - 1)); 1050 const int16x8_t above0 = vld1q_s16((const int16_t *)above); 1051 const int16x8_t above1 = vld1q_s16((const int16_t *)(above + 8)); 1052 const int16x8_t above2 = vld1q_s16((const int16_t *)(above + 16)); 1053 const int16x8_t above3 = vld1q_s16((const int16_t *)(above + 24)); 1054 const int16x8_t sub0 = vsubq_s16(above0, top_left); 1055 const int16x8_t sub1 = vsubq_s16(above1, top_left); 1056 const int16x8_t sub2 = vsubq_s16(above2, top_left); 1057 const int16x8_t sub3 = vsubq_s16(above3, top_left); 1058 int16x8_t left_dup; 1059 int i, j; 1060 1061 for (i = 0; i < 4; i++, left += 8) { 1062 const int16x8_t left_s16q = vld1q_s16((const int16_t *)left); 1063 int16x4_t left_s16d = vget_low_s16(left_s16q); 1064 for (j = 0; j < 2; j++, left_s16d = vget_high_s16(left_s16q)) { 1065 left_dup = vdupq_lane_s16(left_s16d, 0); 1066 tm_32_kernel(&dst, stride, left_dup, sub0, sub1, sub2, sub3, max); 1067 1068 left_dup = vdupq_lane_s16(left_s16d, 1); 1069 tm_32_kernel(&dst, stride, left_dup, sub0, sub1, sub2, sub3, max); 1070 1071 left_dup = vdupq_lane_s16(left_s16d, 2); 1072 tm_32_kernel(&dst, stride, left_dup, sub0, sub1, sub2, sub3, max); 1073 1074 left_dup = vdupq_lane_s16(left_s16d, 3); 1075 tm_32_kernel(&dst, stride, left_dup, sub0, sub1, sub2, sub3, max); 1076 } 1077 } 1078 } 1079
