1 /* 2 * Copyright (c) 2016, 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 "av1/common/cfl.h" 13 #include "av1/common/common_data.h" 14 #include "av1/common/onyxc_int.h" 15 16 #include "config/av1_rtcd.h" 17 18 void cfl_init(CFL_CTX *cfl, const SequenceHeader *seq_params) { 19 assert(block_size_wide[CFL_MAX_BLOCK_SIZE] == CFL_BUF_LINE); 20 assert(block_size_high[CFL_MAX_BLOCK_SIZE] == CFL_BUF_LINE); 21 22 memset(&cfl->recon_buf_q3, 0, sizeof(cfl->recon_buf_q3)); 23 memset(&cfl->ac_buf_q3, 0, sizeof(cfl->ac_buf_q3)); 24 cfl->subsampling_x = seq_params->subsampling_x; 25 cfl->subsampling_y = seq_params->subsampling_y; 26 cfl->are_parameters_computed = 0; 27 cfl->store_y = 0; 28 // The DC_PRED cache is disabled by default and is only enabled in 29 // cfl_rd_pick_alpha 30 cfl->use_dc_pred_cache = 0; 31 cfl->dc_pred_is_cached[CFL_PRED_U] = 0; 32 cfl->dc_pred_is_cached[CFL_PRED_V] = 0; 33 } 34 35 void cfl_store_dc_pred(MACROBLOCKD *const xd, const uint8_t *input, 36 CFL_PRED_TYPE pred_plane, int width) { 37 assert(pred_plane < CFL_PRED_PLANES); 38 assert(width <= CFL_BUF_LINE); 39 40 if (is_cur_buf_hbd(xd)) { 41 uint16_t *const input_16 = CONVERT_TO_SHORTPTR(input); 42 memcpy(xd->cfl.dc_pred_cache[pred_plane], input_16, width << 1); 43 return; 44 } 45 46 memcpy(xd->cfl.dc_pred_cache[pred_plane], input, width); 47 } 48 49 static void cfl_load_dc_pred_lbd(const int16_t *dc_pred_cache, uint8_t *dst, 50 int dst_stride, int width, int height) { 51 for (int j = 0; j < height; j++) { 52 memcpy(dst, dc_pred_cache, width); 53 dst += dst_stride; 54 } 55 } 56 57 static void cfl_load_dc_pred_hbd(const int16_t *dc_pred_cache, uint16_t *dst, 58 int dst_stride, int width, int height) { 59 const size_t num_bytes = width << 1; 60 for (int j = 0; j < height; j++) { 61 memcpy(dst, dc_pred_cache, num_bytes); 62 dst += dst_stride; 63 } 64 } 65 void cfl_load_dc_pred(MACROBLOCKD *const xd, uint8_t *dst, int dst_stride, 66 TX_SIZE tx_size, CFL_PRED_TYPE pred_plane) { 67 const int width = tx_size_wide[tx_size]; 68 const int height = tx_size_high[tx_size]; 69 assert(pred_plane < CFL_PRED_PLANES); 70 assert(width <= CFL_BUF_LINE); 71 assert(height <= CFL_BUF_LINE); 72 if (is_cur_buf_hbd(xd)) { 73 uint16_t *dst_16 = CONVERT_TO_SHORTPTR(dst); 74 cfl_load_dc_pred_hbd(xd->cfl.dc_pred_cache[pred_plane], dst_16, dst_stride, 75 width, height); 76 return; 77 } 78 cfl_load_dc_pred_lbd(xd->cfl.dc_pred_cache[pred_plane], dst, dst_stride, 79 width, height); 80 } 81 82 // Due to frame boundary issues, it is possible that the total area covered by 83 // chroma exceeds that of luma. When this happens, we fill the missing pixels by 84 // repeating the last columns and/or rows. 85 static INLINE void cfl_pad(CFL_CTX *cfl, int width, int height) { 86 const int diff_width = width - cfl->buf_width; 87 const int diff_height = height - cfl->buf_height; 88 89 if (diff_width > 0) { 90 const int min_height = height - diff_height; 91 uint16_t *recon_buf_q3 = cfl->recon_buf_q3 + (width - diff_width); 92 for (int j = 0; j < min_height; j++) { 93 const uint16_t last_pixel = recon_buf_q3[-1]; 94 assert(recon_buf_q3 + diff_width <= cfl->recon_buf_q3 + CFL_BUF_SQUARE); 95 for (int i = 0; i < diff_width; i++) { 96 recon_buf_q3[i] = last_pixel; 97 } 98 recon_buf_q3 += CFL_BUF_LINE; 99 } 100 cfl->buf_width = width; 101 } 102 if (diff_height > 0) { 103 uint16_t *recon_buf_q3 = 104 cfl->recon_buf_q3 + ((height - diff_height) * CFL_BUF_LINE); 105 for (int j = 0; j < diff_height; j++) { 106 const uint16_t *last_row_q3 = recon_buf_q3 - CFL_BUF_LINE; 107 assert(recon_buf_q3 + width <= cfl->recon_buf_q3 + CFL_BUF_SQUARE); 108 for (int i = 0; i < width; i++) { 109 recon_buf_q3[i] = last_row_q3[i]; 110 } 111 recon_buf_q3 += CFL_BUF_LINE; 112 } 113 cfl->buf_height = height; 114 } 115 } 116 117 static void subtract_average_c(const uint16_t *src, int16_t *dst, int width, 118 int height, int round_offset, int num_pel_log2) { 119 int sum = round_offset; 120 const uint16_t *recon = src; 121 for (int j = 0; j < height; j++) { 122 for (int i = 0; i < width; i++) { 123 sum += recon[i]; 124 } 125 recon += CFL_BUF_LINE; 126 } 127 const int avg = sum >> num_pel_log2; 128 for (int j = 0; j < height; j++) { 129 for (int i = 0; i < width; i++) { 130 dst[i] = src[i] - avg; 131 } 132 src += CFL_BUF_LINE; 133 dst += CFL_BUF_LINE; 134 } 135 } 136 137 CFL_SUB_AVG_FN(c) 138 139 static INLINE int cfl_idx_to_alpha(int alpha_idx, int joint_sign, 140 CFL_PRED_TYPE pred_type) { 141 const int alpha_sign = (pred_type == CFL_PRED_U) ? CFL_SIGN_U(joint_sign) 142 : CFL_SIGN_V(joint_sign); 143 if (alpha_sign == CFL_SIGN_ZERO) return 0; 144 const int abs_alpha_q3 = 145 (pred_type == CFL_PRED_U) ? CFL_IDX_U(alpha_idx) : CFL_IDX_V(alpha_idx); 146 return (alpha_sign == CFL_SIGN_POS) ? abs_alpha_q3 + 1 : -abs_alpha_q3 - 1; 147 } 148 149 static INLINE void cfl_predict_lbd_c(const int16_t *ac_buf_q3, uint8_t *dst, 150 int dst_stride, int alpha_q3, int width, 151 int height) { 152 for (int j = 0; j < height; j++) { 153 for (int i = 0; i < width; i++) { 154 dst[i] = clip_pixel(get_scaled_luma_q0(alpha_q3, ac_buf_q3[i]) + dst[i]); 155 } 156 dst += dst_stride; 157 ac_buf_q3 += CFL_BUF_LINE; 158 } 159 } 160 161 CFL_PREDICT_FN(c, lbd) 162 163 void cfl_predict_hbd_c(const int16_t *ac_buf_q3, uint16_t *dst, int dst_stride, 164 int alpha_q3, int bit_depth, int width, int height) { 165 for (int j = 0; j < height; j++) { 166 for (int i = 0; i < width; i++) { 167 dst[i] = clip_pixel_highbd( 168 get_scaled_luma_q0(alpha_q3, ac_buf_q3[i]) + dst[i], bit_depth); 169 } 170 dst += dst_stride; 171 ac_buf_q3 += CFL_BUF_LINE; 172 } 173 } 174 175 CFL_PREDICT_FN(c, hbd) 176 177 static void cfl_compute_parameters(MACROBLOCKD *const xd, TX_SIZE tx_size) { 178 CFL_CTX *const cfl = &xd->cfl; 179 // Do not call cfl_compute_parameters multiple time on the same values. 180 assert(cfl->are_parameters_computed == 0); 181 182 cfl_pad(cfl, tx_size_wide[tx_size], tx_size_high[tx_size]); 183 get_subtract_average_fn(tx_size)(cfl->recon_buf_q3, cfl->ac_buf_q3); 184 cfl->are_parameters_computed = 1; 185 } 186 187 void cfl_predict_block(MACROBLOCKD *const xd, uint8_t *dst, int dst_stride, 188 TX_SIZE tx_size, int plane) { 189 CFL_CTX *const cfl = &xd->cfl; 190 MB_MODE_INFO *mbmi = xd->mi[0]; 191 assert(is_cfl_allowed(xd)); 192 193 if (!cfl->are_parameters_computed) cfl_compute_parameters(xd, tx_size); 194 195 const int alpha_q3 = 196 cfl_idx_to_alpha(mbmi->cfl_alpha_idx, mbmi->cfl_alpha_signs, plane - 1); 197 assert((tx_size_high[tx_size] - 1) * CFL_BUF_LINE + tx_size_wide[tx_size] <= 198 CFL_BUF_SQUARE); 199 if (is_cur_buf_hbd(xd)) { 200 uint16_t *dst_16 = CONVERT_TO_SHORTPTR(dst); 201 get_predict_hbd_fn(tx_size)(cfl->ac_buf_q3, dst_16, dst_stride, alpha_q3, 202 xd->bd); 203 return; 204 } 205 get_predict_lbd_fn(tx_size)(cfl->ac_buf_q3, dst, dst_stride, alpha_q3); 206 } 207 208 static void cfl_luma_subsampling_420_lbd_c(const uint8_t *input, 209 int input_stride, 210 uint16_t *output_q3, int width, 211 int height) { 212 for (int j = 0; j < height; j += 2) { 213 for (int i = 0; i < width; i += 2) { 214 const int bot = i + input_stride; 215 output_q3[i >> 1] = 216 (input[i] + input[i + 1] + input[bot] + input[bot + 1]) << 1; 217 } 218 input += input_stride << 1; 219 output_q3 += CFL_BUF_LINE; 220 } 221 } 222 223 static void cfl_luma_subsampling_422_lbd_c(const uint8_t *input, 224 int input_stride, 225 uint16_t *output_q3, int width, 226 int height) { 227 assert((height - 1) * CFL_BUF_LINE + width <= CFL_BUF_SQUARE); 228 for (int j = 0; j < height; j++) { 229 for (int i = 0; i < width; i += 2) { 230 output_q3[i >> 1] = (input[i] + input[i + 1]) << 2; 231 } 232 input += input_stride; 233 output_q3 += CFL_BUF_LINE; 234 } 235 } 236 237 static void cfl_luma_subsampling_444_lbd_c(const uint8_t *input, 238 int input_stride, 239 uint16_t *output_q3, int width, 240 int height) { 241 assert((height - 1) * CFL_BUF_LINE + width <= CFL_BUF_SQUARE); 242 for (int j = 0; j < height; j++) { 243 for (int i = 0; i < width; i++) { 244 output_q3[i] = input[i] << 3; 245 } 246 input += input_stride; 247 output_q3 += CFL_BUF_LINE; 248 } 249 } 250 251 static void cfl_luma_subsampling_420_hbd_c(const uint16_t *input, 252 int input_stride, 253 uint16_t *output_q3, int width, 254 int height) { 255 for (int j = 0; j < height; j += 2) { 256 for (int i = 0; i < width; i += 2) { 257 const int bot = i + input_stride; 258 output_q3[i >> 1] = 259 (input[i] + input[i + 1] + input[bot] + input[bot + 1]) << 1; 260 } 261 input += input_stride << 1; 262 output_q3 += CFL_BUF_LINE; 263 } 264 } 265 266 static void cfl_luma_subsampling_422_hbd_c(const uint16_t *input, 267 int input_stride, 268 uint16_t *output_q3, int width, 269 int height) { 270 assert((height - 1) * CFL_BUF_LINE + width <= CFL_BUF_SQUARE); 271 for (int j = 0; j < height; j++) { 272 for (int i = 0; i < width; i += 2) { 273 output_q3[i >> 1] = (input[i] + input[i + 1]) << 2; 274 } 275 input += input_stride; 276 output_q3 += CFL_BUF_LINE; 277 } 278 } 279 280 static void cfl_luma_subsampling_444_hbd_c(const uint16_t *input, 281 int input_stride, 282 uint16_t *output_q3, int width, 283 int height) { 284 assert((height - 1) * CFL_BUF_LINE + width <= CFL_BUF_SQUARE); 285 for (int j = 0; j < height; j++) { 286 for (int i = 0; i < width; i++) { 287 output_q3[i] = input[i] << 3; 288 } 289 input += input_stride; 290 output_q3 += CFL_BUF_LINE; 291 } 292 } 293 294 CFL_GET_SUBSAMPLE_FUNCTION(c) 295 296 static INLINE cfl_subsample_hbd_fn cfl_subsampling_hbd(TX_SIZE tx_size, 297 int sub_x, int sub_y) { 298 if (sub_x == 1) { 299 if (sub_y == 1) { 300 return cfl_get_luma_subsampling_420_hbd(tx_size); 301 } 302 return cfl_get_luma_subsampling_422_hbd(tx_size); 303 } 304 return cfl_get_luma_subsampling_444_hbd(tx_size); 305 } 306 307 static INLINE cfl_subsample_lbd_fn cfl_subsampling_lbd(TX_SIZE tx_size, 308 int sub_x, int sub_y) { 309 if (sub_x == 1) { 310 if (sub_y == 1) { 311 return cfl_get_luma_subsampling_420_lbd(tx_size); 312 } 313 return cfl_get_luma_subsampling_422_lbd(tx_size); 314 } 315 return cfl_get_luma_subsampling_444_lbd(tx_size); 316 } 317 318 static void cfl_store(CFL_CTX *cfl, const uint8_t *input, int input_stride, 319 int row, int col, TX_SIZE tx_size, int use_hbd) { 320 const int width = tx_size_wide[tx_size]; 321 const int height = tx_size_high[tx_size]; 322 const int tx_off_log2 = tx_size_wide_log2[0]; 323 const int sub_x = cfl->subsampling_x; 324 const int sub_y = cfl->subsampling_y; 325 const int store_row = row << (tx_off_log2 - sub_y); 326 const int store_col = col << (tx_off_log2 - sub_x); 327 const int store_height = height >> sub_y; 328 const int store_width = width >> sub_x; 329 330 // Invalidate current parameters 331 cfl->are_parameters_computed = 0; 332 333 // Store the surface of the pixel buffer that was written to, this way we 334 // can manage chroma overrun (e.g. when the chroma surfaces goes beyond the 335 // frame boundary) 336 if (col == 0 && row == 0) { 337 cfl->buf_width = store_width; 338 cfl->buf_height = store_height; 339 } else { 340 cfl->buf_width = OD_MAXI(store_col + store_width, cfl->buf_width); 341 cfl->buf_height = OD_MAXI(store_row + store_height, cfl->buf_height); 342 } 343 344 // Check that we will remain inside the pixel buffer. 345 assert(store_row + store_height <= CFL_BUF_LINE); 346 assert(store_col + store_width <= CFL_BUF_LINE); 347 348 // Store the input into the CfL pixel buffer 349 uint16_t *recon_buf_q3 = 350 cfl->recon_buf_q3 + (store_row * CFL_BUF_LINE + store_col); 351 352 if (use_hbd) { 353 cfl_subsampling_hbd(tx_size, sub_x, sub_y)(CONVERT_TO_SHORTPTR(input), 354 input_stride, recon_buf_q3); 355 } else { 356 cfl_subsampling_lbd(tx_size, sub_x, sub_y)(input, input_stride, 357 recon_buf_q3); 358 } 359 } 360 361 // Adjust the row and column of blocks smaller than 8X8, as chroma-referenced 362 // and non-chroma-referenced blocks are stored together in the CfL buffer. 363 static INLINE void sub8x8_adjust_offset(const CFL_CTX *cfl, int *row_out, 364 int *col_out) { 365 // Increment row index for bottom: 8x4, 16x4 or both bottom 4x4s. 366 if ((cfl->mi_row & 0x01) && cfl->subsampling_y) { 367 assert(*row_out == 0); 368 (*row_out)++; 369 } 370 371 // Increment col index for right: 4x8, 4x16 or both right 4x4s. 372 if ((cfl->mi_col & 0x01) && cfl->subsampling_x) { 373 assert(*col_out == 0); 374 (*col_out)++; 375 } 376 } 377 378 void cfl_store_tx(MACROBLOCKD *const xd, int row, int col, TX_SIZE tx_size, 379 BLOCK_SIZE bsize) { 380 CFL_CTX *const cfl = &xd->cfl; 381 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y]; 382 uint8_t *dst = 383 &pd->dst.buf[(row * pd->dst.stride + col) << tx_size_wide_log2[0]]; 384 385 if (block_size_high[bsize] == 4 || block_size_wide[bsize] == 4) { 386 // Only dimensions of size 4 can have an odd offset. 387 assert(!((col & 1) && tx_size_wide[tx_size] != 4)); 388 assert(!((row & 1) && tx_size_high[tx_size] != 4)); 389 sub8x8_adjust_offset(cfl, &row, &col); 390 } 391 cfl_store(cfl, dst, pd->dst.stride, row, col, tx_size, is_cur_buf_hbd(xd)); 392 } 393 394 void cfl_store_block(MACROBLOCKD *const xd, BLOCK_SIZE bsize, TX_SIZE tx_size) { 395 CFL_CTX *const cfl = &xd->cfl; 396 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y]; 397 int row = 0; 398 int col = 0; 399 400 if (block_size_high[bsize] == 4 || block_size_wide[bsize] == 4) { 401 sub8x8_adjust_offset(cfl, &row, &col); 402 } 403 const int width = max_intra_block_width(xd, bsize, AOM_PLANE_Y, tx_size); 404 const int height = max_intra_block_height(xd, bsize, AOM_PLANE_Y, tx_size); 405 tx_size = get_tx_size(width, height); 406 cfl_store(cfl, pd->dst.buf, pd->dst.stride, row, col, tx_size, 407 is_cur_buf_hbd(xd)); 408 } 409