av1_convolve_scale_neon_i8mm.c (16078B)
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 <assert.h> 13 #include <arm_neon.h> 14 #include <stddef.h> 15 #include <stdint.h> 16 17 #include "config/aom_config.h" 18 #include "config/av1_rtcd.h" 19 20 #include "aom_dsp/aom_dsp_common.h" 21 #include "aom_dsp/aom_filter.h" 22 #include "aom_dsp/arm/mem_neon.h" 23 #include "aom_dsp/arm/transpose_neon.h" 24 #include "aom_ports/mem.h" 25 #include "av1/common/arm/convolve_scale_neon.h" 26 #include "av1/common/convolve.h" 27 #include "av1/common/enums.h" 28 #include "av1/common/filter.h" 29 30 // clang-format off 31 DECLARE_ALIGNED(16, static const uint8_t, kScale2DotProdPermuteTbl[32]) = { 32 0, 1, 2, 3, 2, 3, 4, 5, 4, 5, 6, 7, 6, 7, 8, 9, 33 4, 5, 6, 7, 6, 7, 8, 9, 8, 9, 10, 11, 10, 11, 12, 13 34 }; 35 // clang-format on 36 37 static inline int16x4_t convolve8_4_h(const uint8x8_t s0, const uint8x8_t s1, 38 const uint8x8_t s2, const uint8x8_t s3, 39 const int8x8_t filter, 40 const int32x4_t horiz_const) { 41 const int8x16_t filters = vcombine_s8(filter, filter); 42 43 uint8x16_t s01 = vcombine_u8(s0, s1); 44 uint8x16_t s23 = vcombine_u8(s2, s3); 45 46 int32x4_t sum01 = vusdotq_s32(horiz_const, s01, filters); 47 int32x4_t sum23 = vusdotq_s32(horiz_const, s23, filters); 48 49 int32x4_t sum = vpaddq_s32(sum01, sum23); 50 51 // We halved the filter values so -1 from right shift. 52 return vshrn_n_s32(sum, ROUND0_BITS - 1); 53 } 54 55 static inline int16x8_t convolve8_8_h(const uint8x8_t s0, const uint8x8_t s1, 56 const uint8x8_t s2, const uint8x8_t s3, 57 const uint8x8_t s4, const uint8x8_t s5, 58 const uint8x8_t s6, const uint8x8_t s7, 59 const int8x8_t filter, 60 const int32x4_t horiz_const) { 61 const int8x16_t filters = vcombine_s8(filter, filter); 62 63 uint8x16_t s01 = vcombine_u8(s0, s1); 64 uint8x16_t s23 = vcombine_u8(s2, s3); 65 uint8x16_t s45 = vcombine_u8(s4, s5); 66 uint8x16_t s67 = vcombine_u8(s6, s7); 67 68 int32x4_t sum01 = vusdotq_s32(horiz_const, s01, filters); 69 int32x4_t sum23 = vusdotq_s32(horiz_const, s23, filters); 70 int32x4_t sum45 = vusdotq_s32(horiz_const, s45, filters); 71 int32x4_t sum67 = vusdotq_s32(horiz_const, s67, filters); 72 73 int32x4_t sum0123 = vpaddq_s32(sum01, sum23); 74 int32x4_t sum4567 = vpaddq_s32(sum45, sum67); 75 76 // We halved the filter values so -1 from right shift. 77 return vcombine_s16(vshrn_n_s32(sum0123, ROUND0_BITS - 1), 78 vshrn_n_s32(sum4567, ROUND0_BITS - 1)); 79 } 80 81 static inline void convolve_horiz_scale_neon_i8mm(const uint8_t *src, 82 int src_stride, int16_t *dst, 83 int dst_stride, int w, int h, 84 const int16_t *x_filter, 85 const int subpel_x_qn, 86 const int x_step_qn) { 87 DECLARE_ALIGNED(16, int16_t, temp[8 * 8]); 88 const int bd = 8; 89 // A shim of 1 << (ROUND0_BITS - 1) enables us to use non-rounding 90 // shifts - which are generally faster than rounding shifts on modern CPUs. 91 // Divide the total by 4: we halved the filter values and will use a pairwise 92 // add in the convolution kernel. 93 const int32x4_t horiz_offset = vdupq_n_s32( 94 ((1 << (bd + FILTER_BITS - 1)) + (1 << (ROUND0_BITS - 1))) >> 2); 95 96 if (w == 4) { 97 do { 98 int x_qn = subpel_x_qn; 99 100 // Process a 4x4 tile. 101 for (int r = 0; r < 4; r++) { 102 const uint8_t *const s = &src[x_qn >> SCALE_SUBPEL_BITS]; 103 104 const ptrdiff_t filter_offset = 105 SUBPEL_TAPS * ((x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS); 106 // Filter values are all even so halve them to fit in int8_t. 107 const int8x8_t filter = 108 vshrn_n_s16(vld1q_s16(x_filter + filter_offset), 1); 109 110 uint8x8_t t0, t1, t2, t3; 111 load_u8_8x4(s, src_stride, &t0, &t1, &t2, &t3); 112 113 int16x4_t d0 = convolve8_4_h(t0, t1, t2, t3, filter, horiz_offset); 114 115 vst1_s16(&temp[r * 4], d0); 116 x_qn += x_step_qn; 117 } 118 119 // Transpose the 4x4 result tile and store. 120 int16x4_t d0, d1, d2, d3; 121 load_s16_4x4(temp, 4, &d0, &d1, &d2, &d3); 122 123 transpose_elems_inplace_s16_4x4(&d0, &d1, &d2, &d3); 124 125 store_s16_4x4(dst, dst_stride, d0, d1, d2, d3); 126 127 dst += 4 * dst_stride; 128 src += 4 * src_stride; 129 h -= 4; 130 } while (h > 0); 131 } else { 132 do { 133 int x_qn = subpel_x_qn; 134 int16_t *d = dst; 135 int width = w; 136 137 do { 138 // Process an 8x8 tile. 139 for (int r = 0; r < 8; r++) { 140 const uint8_t *const s = &src[(x_qn >> SCALE_SUBPEL_BITS)]; 141 142 const ptrdiff_t filter_offset = 143 SUBPEL_TAPS * ((x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS); 144 // Filter values are all even so halve them to fit in int8_t. 145 const int8x8_t filter = 146 vshrn_n_s16(vld1q_s16(x_filter + filter_offset), 1); 147 148 uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7; 149 load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7); 150 151 int16x8_t d0 = convolve8_8_h(t0, t1, t2, t3, t4, t5, t6, t7, filter, 152 horiz_offset); 153 154 vst1q_s16(&temp[r * 8], d0); 155 156 x_qn += x_step_qn; 157 } 158 159 // Transpose the 8x8 result tile and store. 160 int16x8_t d0, d1, d2, d3, d4, d5, d6, d7; 161 load_s16_8x8(temp, 8, &d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7); 162 163 transpose_elems_inplace_s16_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7); 164 165 store_s16_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7); 166 167 d += 8; 168 width -= 8; 169 } while (width != 0); 170 171 dst += 8 * dst_stride; 172 src += 8 * src_stride; 173 h -= 8; 174 } while (h > 0); 175 } 176 } 177 178 static inline int16x4_t convolve8_4_h_scale_2(uint8x16_t samples, 179 const int8x8_t filters, 180 const int32x4_t horiz_const, 181 const uint8x16x2_t permute_tbl) { 182 // Permute samples ready for dot product. 183 // { 0, 1, 2, 3, 2, 3, 4, 5, 4, 5, 6, 7, 6, 7, 8, 9 } 184 // { 4, 5, 6, 7, 6, 7, 8, 9, 8, 9, 10, 11, 10, 11, 12, 13 } 185 uint8x16_t perm_samples[2] = { vqtbl1q_u8(samples, permute_tbl.val[0]), 186 vqtbl1q_u8(samples, permute_tbl.val[1]) }; 187 188 int32x4_t sum = vusdotq_lane_s32(horiz_const, perm_samples[0], filters, 0); 189 sum = vusdotq_lane_s32(sum, perm_samples[1], filters, 1); 190 191 // We halved the filter values so -1 from right shift. 192 return vshrn_n_s32(sum, ROUND0_BITS - 1); 193 } 194 195 static inline int16x8_t convolve8_8_h_scale_2(uint8x16_t samples[2], 196 const int8x8_t filters, 197 const int32x4_t horiz_const, 198 const uint8x16x2_t permute_tbl) { 199 // Permute samples ready for dot product. 200 // { 0, 1, 2, 3, 2, 3, 4, 5, 4, 5, 6, 7, 6, 7, 8, 9 } 201 // { 4, 5, 6, 7, 6, 7, 8, 9, 8, 9, 10, 11, 10, 11, 12, 13 } 202 uint8x16_t perm_samples[4] = { vqtbl1q_u8(samples[0], permute_tbl.val[0]), 203 vqtbl1q_u8(samples[0], permute_tbl.val[1]), 204 vqtbl1q_u8(samples[1], permute_tbl.val[0]), 205 vqtbl1q_u8(samples[1], permute_tbl.val[1]) }; 206 207 // First 4 output values. 208 int32x4_t sum0123 = 209 vusdotq_lane_s32(horiz_const, perm_samples[0], filters, 0); 210 sum0123 = vusdotq_lane_s32(sum0123, perm_samples[1], filters, 1); 211 212 // Second 4 output values. 213 int32x4_t sum4567 = 214 vusdotq_lane_s32(horiz_const, perm_samples[2], filters, 0); 215 sum4567 = vusdotq_lane_s32(sum4567, perm_samples[3], filters, 1); 216 217 // We halved the filter values so -1 from right shift. 218 return vcombine_s16(vshrn_n_s32(sum0123, ROUND0_BITS - 1), 219 vshrn_n_s32(sum4567, ROUND0_BITS - 1)); 220 } 221 222 static inline void convolve_horiz_scale_2_neon_i8mm( 223 const uint8_t *src, int src_stride, int16_t *dst, int dst_stride, int w, 224 int h, const int16_t *x_filter) { 225 const int bd = 8; 226 // A shim of 1 << (ROUND0_BITS - 1) enables us to use non-rounding 227 // shifts - which are generally faster than rounding shifts on modern CPUs. 228 // The additional -1 is needed because we are halving the filter values. 229 const int32x4_t horiz_offset = 230 vdupq_n_s32((1 << (bd + FILTER_BITS - 2)) + (1 << (ROUND0_BITS - 2))); 231 232 const uint8x16x2_t permute_tbl = vld1q_u8_x2(kScale2DotProdPermuteTbl); 233 // Filter values are all even so halve them to fit in int8_t. 234 const int8x8_t filter = vshrn_n_s16(vld1q_s16(x_filter), 1); 235 236 if (w == 4) { 237 do { 238 const uint8_t *s = src; 239 int16_t *d = dst; 240 int width = w; 241 242 do { 243 uint8x16_t s0, s1, s2, s3; 244 load_u8_16x4(s, src_stride, &s0, &s1, &s2, &s3); 245 246 int16x4_t d0 = 247 convolve8_4_h_scale_2(s0, filter, horiz_offset, permute_tbl); 248 int16x4_t d1 = 249 convolve8_4_h_scale_2(s1, filter, horiz_offset, permute_tbl); 250 int16x4_t d2 = 251 convolve8_4_h_scale_2(s2, filter, horiz_offset, permute_tbl); 252 int16x4_t d3 = 253 convolve8_4_h_scale_2(s3, filter, horiz_offset, permute_tbl); 254 255 store_s16_4x4(d, dst_stride, d0, d1, d2, d3); 256 257 s += 8; 258 d += 4; 259 width -= 4; 260 } while (width != 0); 261 262 dst += 4 * dst_stride; 263 src += 4 * src_stride; 264 h -= 4; 265 } while (h > 0); 266 } else { 267 do { 268 const uint8_t *s = src; 269 int16_t *d = dst; 270 int width = w; 271 272 do { 273 uint8x16_t s0[2], s1[2], s2[2], s3[2]; 274 load_u8_16x4(s, src_stride, &s0[0], &s1[0], &s2[0], &s3[0]); 275 load_u8_16x4(s + 8, src_stride, &s0[1], &s1[1], &s2[1], &s3[1]); 276 277 int16x8_t d0 = 278 convolve8_8_h_scale_2(s0, filter, horiz_offset, permute_tbl); 279 int16x8_t d1 = 280 convolve8_8_h_scale_2(s1, filter, horiz_offset, permute_tbl); 281 int16x8_t d2 = 282 convolve8_8_h_scale_2(s2, filter, horiz_offset, permute_tbl); 283 int16x8_t d3 = 284 convolve8_8_h_scale_2(s3, filter, horiz_offset, permute_tbl); 285 286 store_s16_8x4(d, dst_stride, d0, d1, d2, d3); 287 288 s += 16; 289 d += 8; 290 width -= 8; 291 } while (width != 0); 292 293 dst += 4 * dst_stride; 294 src += 4 * src_stride; 295 h -= 4; 296 } while (h > 0); 297 } 298 } 299 300 void av1_convolve_2d_scale_neon_i8mm(const uint8_t *src, int src_stride, 301 uint8_t *dst, int dst_stride, int w, int h, 302 const InterpFilterParams *filter_params_x, 303 const InterpFilterParams *filter_params_y, 304 const int subpel_x_qn, const int x_step_qn, 305 const int subpel_y_qn, const int y_step_qn, 306 ConvolveParams *conv_params) { 307 if (w < 4 || h < 4) { 308 av1_convolve_2d_scale_c(src, src_stride, dst, dst_stride, w, h, 309 filter_params_x, filter_params_y, subpel_x_qn, 310 x_step_qn, subpel_y_qn, y_step_qn, conv_params); 311 return; 312 } 313 314 // For the interpolation 8-tap filters are used. 315 assert(filter_params_y->taps <= 8 && filter_params_x->taps <= 8); 316 317 DECLARE_ALIGNED(32, int16_t, 318 im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]); 319 int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) + 320 filter_params_y->taps; 321 int im_stride = MAX_SB_SIZE; 322 CONV_BUF_TYPE *dst16 = conv_params->dst; 323 const int dst16_stride = conv_params->dst_stride; 324 325 // Account for needing filter_taps / 2 - 1 lines prior and filter_taps / 2 326 // lines post both horizontally and vertically. 327 const ptrdiff_t horiz_offset = filter_params_x->taps / 2 - 1; 328 const ptrdiff_t vert_offset = (filter_params_y->taps / 2 - 1) * src_stride; 329 330 // Horizontal filter 331 if (x_step_qn != 2 * (1 << SCALE_SUBPEL_BITS)) { 332 convolve_horiz_scale_neon_i8mm( 333 src - horiz_offset - vert_offset, src_stride, im_block, im_stride, w, 334 im_h, filter_params_x->filter_ptr, subpel_x_qn, x_step_qn); 335 } else { 336 assert(subpel_x_qn < (1 << SCALE_SUBPEL_BITS)); 337 // The filter index is calculated using the 338 // ((subpel_x_qn + x * x_step_qn) & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS 339 // equation, where the values of x are from 0 to w. If x_step_qn is a 340 // multiple of SCALE_SUBPEL_MASK we can leave it out of the equation. 341 const ptrdiff_t filter_offset = 342 SUBPEL_TAPS * ((subpel_x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS); 343 const int16_t *x_filter = filter_params_x->filter_ptr + filter_offset; 344 345 // The source index is calculated using the (subpel_x_qn + x * x_step_qn) >> 346 // SCALE_SUBPEL_BITS, where the values of x are from 0 to w. If subpel_x_qn 347 // < (1 << SCALE_SUBPEL_BITS) and x_step_qn % (1 << SCALE_SUBPEL_BITS) == 0, 348 // the source index can be determined using the value x * (x_step_qn / 349 // (1 << SCALE_SUBPEL_BITS)). 350 convolve_horiz_scale_2_neon_i8mm(src - horiz_offset - vert_offset, 351 src_stride, im_block, im_stride, w, im_h, 352 x_filter); 353 } 354 355 // Vertical filter 356 if (filter_params_y->interp_filter == MULTITAP_SHARP) { 357 if (UNLIKELY(conv_params->is_compound)) { 358 if (conv_params->do_average) { 359 if (conv_params->use_dist_wtd_comp_avg) { 360 compound_dist_wtd_convolve_vert_scale_8tap_neon( 361 im_block, im_stride, dst, dst_stride, dst16, dst16_stride, w, h, 362 filter_params_y->filter_ptr, conv_params, subpel_y_qn, y_step_qn); 363 } else { 364 compound_avg_convolve_vert_scale_8tap_neon( 365 im_block, im_stride, dst, dst_stride, dst16, dst16_stride, w, h, 366 filter_params_y->filter_ptr, subpel_y_qn, y_step_qn); 367 } 368 } else { 369 compound_convolve_vert_scale_8tap_neon( 370 im_block, im_stride, dst16, dst16_stride, w, h, 371 filter_params_y->filter_ptr, subpel_y_qn, y_step_qn); 372 } 373 } else { 374 convolve_vert_scale_8tap_neon(im_block, im_stride, dst, dst_stride, w, h, 375 filter_params_y->filter_ptr, subpel_y_qn, 376 y_step_qn); 377 } 378 } else { 379 if (UNLIKELY(conv_params->is_compound)) { 380 if (conv_params->do_average) { 381 if (conv_params->use_dist_wtd_comp_avg) { 382 compound_dist_wtd_convolve_vert_scale_6tap_neon( 383 im_block + im_stride, im_stride, dst, dst_stride, dst16, 384 dst16_stride, w, h, filter_params_y->filter_ptr, conv_params, 385 subpel_y_qn, y_step_qn); 386 } else { 387 compound_avg_convolve_vert_scale_6tap_neon( 388 im_block + im_stride, im_stride, dst, dst_stride, dst16, 389 dst16_stride, w, h, filter_params_y->filter_ptr, subpel_y_qn, 390 y_step_qn); 391 } 392 } else { 393 compound_convolve_vert_scale_6tap_neon( 394 im_block + im_stride, im_stride, dst16, dst16_stride, w, h, 395 filter_params_y->filter_ptr, subpel_y_qn, y_step_qn); 396 } 397 } else { 398 convolve_vert_scale_6tap_neon( 399 im_block + im_stride, im_stride, dst, dst_stride, w, h, 400 filter_params_y->filter_ptr, subpel_y_qn, y_step_qn); 401 } 402 } 403 }