highbd_convolve8_sve.c (19839B)
1 /* 2 * Copyright (c) 2024, Alliance for Open Media. All rights reserved. 3 * 4 * This source code is subject to the terms of the BSD 2 Clause License and 5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License 6 * was not distributed with this source code in the LICENSE file, you can 7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open 8 * Media Patent License 1.0 was not distributed with this source code in the 9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent. 10 */ 11 12 #include <arm_neon.h> 13 #include <assert.h> 14 #include <stdint.h> 15 16 #include "config/aom_config.h" 17 #include "config/aom_dsp_rtcd.h" 18 19 #include "aom_dsp/arm/aom_neon_sve_bridge.h" 20 #include "aom_dsp/arm/aom_filter.h" 21 #include "aom_dsp/arm/highbd_convolve8_neon.h" 22 #include "aom_dsp/arm/mem_neon.h" 23 #include "aom_dsp/arm/transpose_neon.h" 24 25 static inline uint16x4_t highbd_convolve8_4_h(int16x8_t s[4], int16x8_t filter, 26 uint16x4_t max) { 27 int64x2_t sum[4]; 28 29 sum[0] = aom_sdotq_s16(vdupq_n_s64(0), s[0], filter); 30 sum[1] = aom_sdotq_s16(vdupq_n_s64(0), s[1], filter); 31 sum[2] = aom_sdotq_s16(vdupq_n_s64(0), s[2], filter); 32 sum[3] = aom_sdotq_s16(vdupq_n_s64(0), s[3], filter); 33 34 int64x2_t sum01 = vpaddq_s64(sum[0], sum[1]); 35 int64x2_t sum23 = vpaddq_s64(sum[2], sum[3]); 36 37 int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); 38 39 uint16x4_t res = vqrshrun_n_s32(sum0123, FILTER_BITS); 40 return vmin_u16(res, max); 41 } 42 43 static inline uint16x8_t highbd_convolve8_8_h(int16x8_t s[8], int16x8_t filter, 44 uint16x8_t max) { 45 int64x2_t sum[8]; 46 47 sum[0] = aom_sdotq_s16(vdupq_n_s64(0), s[0], filter); 48 sum[1] = aom_sdotq_s16(vdupq_n_s64(0), s[1], filter); 49 sum[2] = aom_sdotq_s16(vdupq_n_s64(0), s[2], filter); 50 sum[3] = aom_sdotq_s16(vdupq_n_s64(0), s[3], filter); 51 sum[4] = aom_sdotq_s16(vdupq_n_s64(0), s[4], filter); 52 sum[5] = aom_sdotq_s16(vdupq_n_s64(0), s[5], filter); 53 sum[6] = aom_sdotq_s16(vdupq_n_s64(0), s[6], filter); 54 sum[7] = aom_sdotq_s16(vdupq_n_s64(0), s[7], filter); 55 56 int64x2_t sum01 = vpaddq_s64(sum[0], sum[1]); 57 int64x2_t sum23 = vpaddq_s64(sum[2], sum[3]); 58 int64x2_t sum45 = vpaddq_s64(sum[4], sum[5]); 59 int64x2_t sum67 = vpaddq_s64(sum[6], sum[7]); 60 61 int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); 62 int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum45), vmovn_s64(sum67)); 63 64 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(sum0123, FILTER_BITS), 65 vqrshrun_n_s32(sum4567, FILTER_BITS)); 66 return vminq_u16(res, max); 67 } 68 69 static inline void highbd_convolve8_horiz_8tap_sve( 70 const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, 71 ptrdiff_t dst_stride, const int16_t *filter_x, int width, int height, 72 int bd) { 73 const int16x8_t filter = vld1q_s16(filter_x); 74 75 if (width == 4) { 76 const uint16x4_t max = vdup_n_u16((1 << bd) - 1); 77 const int16_t *s = (const int16_t *)src; 78 uint16_t *d = dst; 79 80 do { 81 int16x8_t s0[4], s1[4], s2[4], s3[4]; 82 load_s16_8x4(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3]); 83 load_s16_8x4(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3]); 84 load_s16_8x4(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3]); 85 load_s16_8x4(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3]); 86 87 uint16x4_t d0 = highbd_convolve8_4_h(s0, filter, max); 88 uint16x4_t d1 = highbd_convolve8_4_h(s1, filter, max); 89 uint16x4_t d2 = highbd_convolve8_4_h(s2, filter, max); 90 uint16x4_t d3 = highbd_convolve8_4_h(s3, filter, max); 91 92 store_u16_4x4(d, dst_stride, d0, d1, d2, d3); 93 94 s += 4 * src_stride; 95 d += 4 * dst_stride; 96 height -= 4; 97 } while (height > 0); 98 } else { 99 do { 100 const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); 101 const int16_t *s = (const int16_t *)src; 102 uint16_t *d = dst; 103 int w = width; 104 105 do { 106 int16x8_t s0[8], s1[8], s2[8], s3[8]; 107 load_s16_8x8(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3], 108 &s0[4], &s0[5], &s0[6], &s0[7]); 109 load_s16_8x8(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3], 110 &s1[4], &s1[5], &s1[6], &s1[7]); 111 load_s16_8x8(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3], 112 &s2[4], &s2[5], &s2[6], &s2[7]); 113 load_s16_8x8(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3], 114 &s3[4], &s3[5], &s3[6], &s3[7]); 115 116 uint16x8_t d0 = highbd_convolve8_8_h(s0, filter, max); 117 uint16x8_t d1 = highbd_convolve8_8_h(s1, filter, max); 118 uint16x8_t d2 = highbd_convolve8_8_h(s2, filter, max); 119 uint16x8_t d3 = highbd_convolve8_8_h(s3, filter, max); 120 121 store_u16_8x4(d, dst_stride, d0, d1, d2, d3); 122 123 s += 8; 124 d += 8; 125 w -= 8; 126 } while (w != 0); 127 src += 4 * src_stride; 128 dst += 4 * dst_stride; 129 height -= 4; 130 } while (height > 0); 131 } 132 } 133 134 // clang-format off 135 DECLARE_ALIGNED(16, static const uint16_t, kDotProdTbl[16]) = { 136 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6, 137 }; 138 139 DECLARE_ALIGNED(16, static const uint16_t, kDeinterleaveTbl[8]) = { 140 0, 2, 4, 6, 1, 3, 5, 7, 141 }; 142 // clang-format on 143 144 static inline uint16x4_t highbd_convolve4_4_h(int16x8_t s, int16x8_t filter, 145 uint16x8x2_t permute_tbl, 146 uint16x4_t max) { 147 int16x8_t permuted_samples0 = aom_tbl_s16(s, permute_tbl.val[0]); 148 int16x8_t permuted_samples1 = aom_tbl_s16(s, permute_tbl.val[1]); 149 150 int64x2_t sum0 = 151 aom_svdot_lane_s16(vdupq_n_s64(0), permuted_samples0, filter, 0); 152 int64x2_t sum1 = 153 aom_svdot_lane_s16(vdupq_n_s64(0), permuted_samples1, filter, 0); 154 155 int32x4_t res_s32 = vcombine_s32(vmovn_s64(sum0), vmovn_s64(sum1)); 156 uint16x4_t res = vqrshrun_n_s32(res_s32, FILTER_BITS); 157 158 return vmin_u16(res, max); 159 } 160 161 static inline uint16x8_t highbd_convolve4_8_h(int16x8_t s[4], int16x8_t filter, 162 uint16x8_t idx, uint16x8_t max) { 163 int64x2_t sum04 = aom_svdot_lane_s16(vdupq_n_s64(0), s[0], filter, 0); 164 int64x2_t sum15 = aom_svdot_lane_s16(vdupq_n_s64(0), s[1], filter, 0); 165 int64x2_t sum26 = aom_svdot_lane_s16(vdupq_n_s64(0), s[2], filter, 0); 166 int64x2_t sum37 = aom_svdot_lane_s16(vdupq_n_s64(0), s[3], filter, 0); 167 168 int32x4_t res0 = vcombine_s32(vmovn_s64(sum04), vmovn_s64(sum15)); 169 int32x4_t res1 = vcombine_s32(vmovn_s64(sum26), vmovn_s64(sum37)); 170 171 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(res0, FILTER_BITS), 172 vqrshrun_n_s32(res1, FILTER_BITS)); 173 174 res = aom_tbl_u16(res, idx); 175 176 return vminq_u16(res, max); 177 } 178 179 static inline void highbd_convolve8_horiz_4tap_sve( 180 const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, 181 ptrdiff_t dst_stride, const int16_t *filter_x, int width, int height, 182 int bd) { 183 const int16x8_t filter = vcombine_s16(vld1_s16(filter_x + 2), vdup_n_s16(0)); 184 185 if (width == 4) { 186 const uint16x4_t max = vdup_n_u16((1 << bd) - 1); 187 uint16x8x2_t permute_tbl = vld1q_u16_x2(kDotProdTbl); 188 189 const int16_t *s = (const int16_t *)src; 190 uint16_t *d = dst; 191 192 do { 193 int16x8_t s0, s1, s2, s3; 194 load_s16_8x4(s, src_stride, &s0, &s1, &s2, &s3); 195 196 uint16x4_t d0 = highbd_convolve4_4_h(s0, filter, permute_tbl, max); 197 uint16x4_t d1 = highbd_convolve4_4_h(s1, filter, permute_tbl, max); 198 uint16x4_t d2 = highbd_convolve4_4_h(s2, filter, permute_tbl, max); 199 uint16x4_t d3 = highbd_convolve4_4_h(s3, filter, permute_tbl, max); 200 201 store_u16_4x4(d, dst_stride, d0, d1, d2, d3); 202 203 s += 4 * src_stride; 204 d += 4 * dst_stride; 205 height -= 4; 206 } while (height > 0); 207 } else { 208 const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); 209 uint16x8_t idx = vld1q_u16(kDeinterleaveTbl); 210 211 do { 212 const int16_t *s = (const int16_t *)src; 213 uint16_t *d = dst; 214 int w = width; 215 216 do { 217 int16x8_t s0[4], s1[4], s2[4], s3[4]; 218 load_s16_8x4(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3]); 219 load_s16_8x4(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3]); 220 load_s16_8x4(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3]); 221 load_s16_8x4(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3]); 222 223 uint16x8_t d0 = highbd_convolve4_8_h(s0, filter, idx, max); 224 uint16x8_t d1 = highbd_convolve4_8_h(s1, filter, idx, max); 225 uint16x8_t d2 = highbd_convolve4_8_h(s2, filter, idx, max); 226 uint16x8_t d3 = highbd_convolve4_8_h(s3, filter, idx, max); 227 228 store_u16_8x4(d, dst_stride, d0, d1, d2, d3); 229 230 s += 8; 231 d += 8; 232 w -= 8; 233 } while (w != 0); 234 src += 4 * src_stride; 235 dst += 4 * dst_stride; 236 height -= 4; 237 } while (height > 0); 238 } 239 } 240 241 void aom_highbd_convolve8_horiz_sve(const uint8_t *src8, ptrdiff_t src_stride, 242 uint8_t *dst8, ptrdiff_t dst_stride, 243 const int16_t *filter_x, int x_step_q4, 244 const int16_t *filter_y, int y_step_q4, 245 int width, int height, int bd) { 246 assert(x_step_q4 == 16); 247 assert(width >= 4 && height >= 4); 248 (void)filter_y; 249 (void)x_step_q4; 250 (void)y_step_q4; 251 252 const uint16_t *src = CONVERT_TO_SHORTPTR(src8); 253 uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); 254 255 src -= SUBPEL_TAPS / 2 - 1; 256 257 const int filter_taps = get_filter_taps_convolve8(filter_x); 258 259 if (filter_taps == 2) { 260 highbd_convolve8_horiz_2tap_neon(src + 3, src_stride, dst, dst_stride, 261 filter_x, width, height, bd); 262 } else if (filter_taps == 4) { 263 highbd_convolve8_horiz_4tap_sve(src + 2, src_stride, dst, dst_stride, 264 filter_x, width, height, bd); 265 } else { 266 highbd_convolve8_horiz_8tap_sve(src, src_stride, dst, dst_stride, filter_x, 267 width, height, bd); 268 } 269 } 270 271 DECLARE_ALIGNED(16, static const uint8_t, kDotProdMergeBlockTbl[48]) = { 272 // Shift left and insert new last column in transposed 4x4 block. 273 2, 3, 4, 5, 6, 7, 16, 17, 10, 11, 12, 13, 14, 15, 24, 25, 274 // Shift left and insert two new columns in transposed 4x4 block. 275 4, 5, 6, 7, 16, 17, 18, 19, 12, 13, 14, 15, 24, 25, 26, 27, 276 // Shift left and insert three new columns in transposed 4x4 block. 277 6, 7, 16, 17, 18, 19, 20, 21, 14, 15, 24, 25, 26, 27, 28, 29 278 }; 279 280 static inline void aom_tbl2x4_s16(int16x8_t t0[4], int16x8_t t1[4], 281 uint8x16_t tbl, int16x8_t res[4]) { 282 int8x16x2_t samples0 = { vreinterpretq_s8_s16(t0[0]), 283 vreinterpretq_s8_s16(t1[0]) }; 284 int8x16x2_t samples1 = { vreinterpretq_s8_s16(t0[1]), 285 vreinterpretq_s8_s16(t1[1]) }; 286 int8x16x2_t samples2 = { vreinterpretq_s8_s16(t0[2]), 287 vreinterpretq_s8_s16(t1[2]) }; 288 int8x16x2_t samples3 = { vreinterpretq_s8_s16(t0[3]), 289 vreinterpretq_s8_s16(t1[3]) }; 290 291 res[0] = vreinterpretq_s16_s8(vqtbl2q_s8(samples0, tbl)); 292 res[1] = vreinterpretq_s16_s8(vqtbl2q_s8(samples1, tbl)); 293 res[2] = vreinterpretq_s16_s8(vqtbl2q_s8(samples2, tbl)); 294 res[3] = vreinterpretq_s16_s8(vqtbl2q_s8(samples3, tbl)); 295 } 296 297 static inline void aom_tbl2x2_s16(int16x8_t t0[2], int16x8_t t1[2], 298 uint8x16_t tbl, int16x8_t res[2]) { 299 int8x16x2_t samples0 = { vreinterpretq_s8_s16(t0[0]), 300 vreinterpretq_s8_s16(t1[0]) }; 301 int8x16x2_t samples1 = { vreinterpretq_s8_s16(t0[1]), 302 vreinterpretq_s8_s16(t1[1]) }; 303 304 res[0] = vreinterpretq_s16_s8(vqtbl2q_s8(samples0, tbl)); 305 res[1] = vreinterpretq_s16_s8(vqtbl2q_s8(samples1, tbl)); 306 } 307 308 static inline uint16x4_t highbd_convolve8_4_v(int16x8_t samples_lo[2], 309 int16x8_t samples_hi[2], 310 int16x8_t filter, 311 uint16x4_t max) { 312 int64x2_t sum[2]; 313 314 sum[0] = aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[0], filter, 0); 315 sum[0] = aom_svdot_lane_s16(sum[0], samples_hi[0], filter, 1); 316 317 sum[1] = aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[1], filter, 0); 318 sum[1] = aom_svdot_lane_s16(sum[1], samples_hi[1], filter, 1); 319 320 int32x4_t res_s32 = vcombine_s32(vmovn_s64(sum[0]), vmovn_s64(sum[1])); 321 322 uint16x4_t res = vqrshrun_n_s32(res_s32, FILTER_BITS); 323 324 return vmin_u16(res, max); 325 } 326 327 static inline uint16x8_t highbd_convolve8_8_v(int16x8_t samples_lo[4], 328 int16x8_t samples_hi[4], 329 int16x8_t filter, 330 uint16x8_t max) { 331 int64x2_t sum[4]; 332 333 sum[0] = aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[0], filter, 0); 334 sum[0] = aom_svdot_lane_s16(sum[0], samples_hi[0], filter, 1); 335 336 sum[1] = aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[1], filter, 0); 337 sum[1] = aom_svdot_lane_s16(sum[1], samples_hi[1], filter, 1); 338 339 sum[2] = aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[2], filter, 0); 340 sum[2] = aom_svdot_lane_s16(sum[2], samples_hi[2], filter, 1); 341 342 sum[3] = aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[3], filter, 0); 343 sum[3] = aom_svdot_lane_s16(sum[3], samples_hi[3], filter, 1); 344 345 int32x4_t res0 = vcombine_s32(vmovn_s64(sum[0]), vmovn_s64(sum[1])); 346 int32x4_t res1 = vcombine_s32(vmovn_s64(sum[2]), vmovn_s64(sum[3])); 347 348 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(res0, FILTER_BITS), 349 vqrshrun_n_s32(res1, FILTER_BITS)); 350 351 return vminq_u16(res, max); 352 } 353 354 static inline void highbd_convolve8_vert_8tap_sve( 355 const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, 356 ptrdiff_t dst_stride, const int16_t *filter_y, int width, int height, 357 int bd) { 358 const int16x8_t y_filter = vld1q_s16(filter_y); 359 360 uint8x16_t merge_block_tbl[3]; 361 merge_block_tbl[0] = vld1q_u8(kDotProdMergeBlockTbl); 362 merge_block_tbl[1] = vld1q_u8(kDotProdMergeBlockTbl + 16); 363 merge_block_tbl[2] = vld1q_u8(kDotProdMergeBlockTbl + 32); 364 365 if (width == 4) { 366 const uint16x4_t max = vdup_n_u16((1 << bd) - 1); 367 int16_t *s = (int16_t *)src; 368 369 int16x4_t s0, s1, s2, s3, s4, s5, s6; 370 load_s16_4x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); 371 s += 7 * src_stride; 372 373 // This operation combines a conventional transpose and the sample permute 374 // required before computing the dot product. 375 int16x8_t s0123[2], s1234[2], s2345[2], s3456[2]; 376 transpose_concat_elems_s16_4x4(s0, s1, s2, s3, s0123); 377 transpose_concat_elems_s16_4x4(s1, s2, s3, s4, s1234); 378 transpose_concat_elems_s16_4x4(s2, s3, s4, s5, s2345); 379 transpose_concat_elems_s16_4x4(s3, s4, s5, s6, s3456); 380 381 do { 382 int16x4_t s7, s8, s9, s10; 383 load_s16_4x4(s, src_stride, &s7, &s8, &s9, &s10); 384 385 int16x8_t s4567[2], s5678[2], s6789[2], s78910[2]; 386 387 // Transpose and shuffle the 4 lines that were loaded. 388 transpose_concat_elems_s16_4x4(s7, s8, s9, s10, s78910); 389 390 // Merge new data into block from previous iteration. 391 aom_tbl2x2_s16(s3456, s78910, merge_block_tbl[0], s4567); 392 aom_tbl2x2_s16(s3456, s78910, merge_block_tbl[1], s5678); 393 aom_tbl2x2_s16(s3456, s78910, merge_block_tbl[2], s6789); 394 395 uint16x4_t d0 = highbd_convolve8_4_v(s0123, s4567, y_filter, max); 396 uint16x4_t d1 = highbd_convolve8_4_v(s1234, s5678, y_filter, max); 397 uint16x4_t d2 = highbd_convolve8_4_v(s2345, s6789, y_filter, max); 398 uint16x4_t d3 = highbd_convolve8_4_v(s3456, s78910, y_filter, max); 399 400 store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); 401 402 // Prepare block for next iteration - re-using as much as possible. 403 // Shuffle everything up four rows. 404 s0123[0] = s4567[0]; 405 s0123[1] = s4567[1]; 406 s1234[0] = s5678[0]; 407 s1234[1] = s5678[1]; 408 s2345[0] = s6789[0]; 409 s2345[1] = s6789[1]; 410 s3456[0] = s78910[0]; 411 s3456[1] = s78910[1]; 412 413 s += 4 * src_stride; 414 dst += 4 * dst_stride; 415 height -= 4; 416 } while (height != 0); 417 } else { 418 const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); 419 do { 420 int h = height; 421 int16_t *s = (int16_t *)src; 422 uint16_t *d = dst; 423 424 int16x8_t s0, s1, s2, s3, s4, s5, s6; 425 load_s16_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); 426 s += 7 * src_stride; 427 428 // This operation combines a conventional transpose and the sample permute 429 // required before computing the dot product. 430 int16x8_t s0123[4], s1234[4], s2345[4], s3456[4]; 431 transpose_concat_elems_s16_8x4(s0, s1, s2, s3, s0123); 432 transpose_concat_elems_s16_8x4(s1, s2, s3, s4, s1234); 433 transpose_concat_elems_s16_8x4(s2, s3, s4, s5, s2345); 434 transpose_concat_elems_s16_8x4(s3, s4, s5, s6, s3456); 435 436 do { 437 int16x8_t s7, s8, s9, s10; 438 load_s16_8x4(s, src_stride, &s7, &s8, &s9, &s10); 439 440 int16x8_t s4567[4], s5678[4], s6789[4], s78910[4]; 441 442 // Transpose and shuffle the 4 lines that were loaded. 443 transpose_concat_elems_s16_8x4(s7, s8, s9, s10, s78910); 444 445 // Merge new data into block from previous iteration. 446 aom_tbl2x4_s16(s3456, s78910, merge_block_tbl[0], s4567); 447 aom_tbl2x4_s16(s3456, s78910, merge_block_tbl[1], s5678); 448 aom_tbl2x4_s16(s3456, s78910, merge_block_tbl[2], s6789); 449 450 uint16x8_t d0 = highbd_convolve8_8_v(s0123, s4567, y_filter, max); 451 uint16x8_t d1 = highbd_convolve8_8_v(s1234, s5678, y_filter, max); 452 uint16x8_t d2 = highbd_convolve8_8_v(s2345, s6789, y_filter, max); 453 uint16x8_t d3 = highbd_convolve8_8_v(s3456, s78910, y_filter, max); 454 455 store_u16_8x4(d, dst_stride, d0, d1, d2, d3); 456 457 // Prepare block for next iteration - re-using as much as possible. 458 // Shuffle everything up four rows. 459 s0123[0] = s4567[0]; 460 s0123[1] = s4567[1]; 461 s0123[2] = s4567[2]; 462 s0123[3] = s4567[3]; 463 464 s1234[0] = s5678[0]; 465 s1234[1] = s5678[1]; 466 s1234[2] = s5678[2]; 467 s1234[3] = s5678[3]; 468 469 s2345[0] = s6789[0]; 470 s2345[1] = s6789[1]; 471 s2345[2] = s6789[2]; 472 s2345[3] = s6789[3]; 473 474 s3456[0] = s78910[0]; 475 s3456[1] = s78910[1]; 476 s3456[2] = s78910[2]; 477 s3456[3] = s78910[3]; 478 479 s += 4 * src_stride; 480 d += 4 * dst_stride; 481 h -= 4; 482 } while (h != 0); 483 src += 8; 484 dst += 8; 485 width -= 8; 486 } while (width != 0); 487 } 488 } 489 490 void aom_highbd_convolve8_vert_sve(const uint8_t *src8, ptrdiff_t src_stride, 491 uint8_t *dst8, ptrdiff_t dst_stride, 492 const int16_t *filter_x, int x_step_q4, 493 const int16_t *filter_y, int y_step_q4, 494 int width, int height, int bd) { 495 assert(y_step_q4 == 16); 496 assert(width >= 4 && height >= 4); 497 (void)filter_x; 498 (void)y_step_q4; 499 (void)x_step_q4; 500 501 const uint16_t *src = CONVERT_TO_SHORTPTR(src8); 502 uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); 503 504 src -= (SUBPEL_TAPS / 2 - 1) * src_stride; 505 506 const int filter_taps = get_filter_taps_convolve8(filter_y); 507 508 if (filter_taps == 2) { 509 highbd_convolve8_vert_2tap_neon(src + 3 * src_stride, src_stride, dst, 510 dst_stride, filter_y, width, height, bd); 511 } else if (filter_taps == 4) { 512 highbd_convolve8_vert_4tap_neon(src + 2 * src_stride, src_stride, dst, 513 dst_stride, filter_y, width, height, bd); 514 } else { 515 highbd_convolve8_vert_8tap_sve(src, src_stride, dst, dst_stride, filter_y, 516 width, height, bd); 517 } 518 }