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 "config/aom_config.h" 13 #include "config/av1_rtcd.h" 14 #include "config/aom_dsp_rtcd.h" 15 16 #include "aom_dsp/bitwriter.h" 17 #include "aom_dsp/quantize.h" 18 #include "aom_mem/aom_mem.h" 19 #include "aom_ports/mem.h" 20 21 #if CONFIG_BITSTREAM_DEBUG || CONFIG_MISMATCH_DEBUG 22 #include "aom_util/debug_util.h" 23 #endif // CONFIG_BITSTREAM_DEBUG || CONFIG_MISMATCH_DEBUG 24 25 #include "av1/common/cfl.h" 26 #include "av1/common/idct.h" 27 #include "av1/common/reconinter.h" 28 #include "av1/common/reconintra.h" 29 #include "av1/common/scan.h" 30 31 #include "av1/encoder/av1_quantize.h" 32 #include "av1/encoder/encodemb.h" 33 #include "av1/encoder/encodetxb.h" 34 #include "av1/encoder/hybrid_fwd_txfm.h" 35 #include "av1/encoder/rd.h" 36 #include "av1/encoder/rdopt.h" 37 38 // Check if one needs to use c version subtraction. 39 static int check_subtract_block_size(int w, int h) { return w < 4 || h < 4; } 40 41 static void subtract_block(const MACROBLOCKD *xd, int rows, int cols, 42 int16_t *diff, ptrdiff_t diff_stride, 43 const uint8_t *src8, ptrdiff_t src_stride, 44 const uint8_t *pred8, ptrdiff_t pred_stride) { 45 if (check_subtract_block_size(rows, cols)) { 46 if (is_cur_buf_hbd(xd)) { 47 aom_highbd_subtract_block_c(rows, cols, diff, diff_stride, src8, 48 src_stride, pred8, pred_stride, xd->bd); 49 return; 50 } 51 aom_subtract_block_c(rows, cols, diff, diff_stride, src8, src_stride, pred8, 52 pred_stride); 53 54 return; 55 } 56 57 if (is_cur_buf_hbd(xd)) { 58 aom_highbd_subtract_block(rows, cols, diff, diff_stride, src8, src_stride, 59 pred8, pred_stride, xd->bd); 60 return; 61 } 62 aom_subtract_block(rows, cols, diff, diff_stride, src8, src_stride, pred8, 63 pred_stride); 64 } 65 66 void av1_subtract_txb(MACROBLOCK *x, int plane, BLOCK_SIZE plane_bsize, 67 int blk_col, int blk_row, TX_SIZE tx_size) { 68 MACROBLOCKD *const xd = &x->e_mbd; 69 struct macroblock_plane *const p = &x->plane[plane]; 70 const struct macroblockd_plane *const pd = &x->e_mbd.plane[plane]; 71 const int diff_stride = block_size_wide[plane_bsize]; 72 const int src_stride = p->src.stride; 73 const int dst_stride = pd->dst.stride; 74 const int tx1d_width = tx_size_wide[tx_size]; 75 const int tx1d_height = tx_size_high[tx_size]; 76 uint8_t *dst = 77 &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]]; 78 uint8_t *src = 79 &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]]; 80 int16_t *src_diff = 81 &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]]; 82 subtract_block(xd, tx1d_height, tx1d_width, src_diff, diff_stride, src, 83 src_stride, dst, dst_stride); 84 } 85 86 void av1_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) { 87 struct macroblock_plane *const p = &x->plane[plane]; 88 const struct macroblockd_plane *const pd = &x->e_mbd.plane[plane]; 89 const BLOCK_SIZE plane_bsize = 90 get_plane_block_size(bsize, pd->subsampling_x, pd->subsampling_y); 91 const int bw = block_size_wide[plane_bsize]; 92 const int bh = block_size_high[plane_bsize]; 93 const MACROBLOCKD *xd = &x->e_mbd; 94 95 subtract_block(xd, bh, bw, p->src_diff, bw, p->src.buf, p->src.stride, 96 pd->dst.buf, pd->dst.stride); 97 } 98 99 int av1_optimize_b(const struct AV1_COMP *cpi, MACROBLOCK *mb, int plane, 100 int block, TX_SIZE tx_size, TX_TYPE tx_type, 101 const TXB_CTX *const txb_ctx, int fast_mode, 102 int *rate_cost) { 103 MACROBLOCKD *const xd = &mb->e_mbd; 104 struct macroblock_plane *const p = &mb->plane[plane]; 105 const int eob = p->eobs[block]; 106 const int segment_id = xd->mi[0]->segment_id; 107 108 if (eob == 0 || !cpi->optimize_seg_arr[segment_id] || 109 xd->lossless[segment_id]) { 110 *rate_cost = av1_cost_skip_txb(mb, txb_ctx, plane, tx_size); 111 return eob; 112 } 113 114 return av1_optimize_txb_new(cpi, mb, plane, block, tx_size, tx_type, txb_ctx, 115 rate_cost, cpi->oxcf.sharpness, fast_mode); 116 } 117 118 enum { 119 QUANT_FUNC_LOWBD = 0, 120 QUANT_FUNC_HIGHBD = 1, 121 QUANT_FUNC_TYPES = 2 122 } UENUM1BYTE(QUANT_FUNC); 123 124 static AV1_QUANT_FACADE 125 quant_func_list[AV1_XFORM_QUANT_TYPES][QUANT_FUNC_TYPES] = { 126 { av1_quantize_fp_facade, av1_highbd_quantize_fp_facade }, 127 { av1_quantize_b_facade, av1_highbd_quantize_b_facade }, 128 { av1_quantize_dc_facade, av1_highbd_quantize_dc_facade }, 129 { NULL, NULL } 130 }; 131 132 void av1_xform_quant(const AV1_COMMON *cm, MACROBLOCK *x, int plane, int block, 133 int blk_row, int blk_col, BLOCK_SIZE plane_bsize, 134 TX_SIZE tx_size, TX_TYPE tx_type, 135 AV1_XFORM_QUANT xform_quant_idx) { 136 MACROBLOCKD *const xd = &x->e_mbd; 137 MB_MODE_INFO *const mbmi = xd->mi[0]; 138 const struct macroblock_plane *const p = &x->plane[plane]; 139 const struct macroblockd_plane *const pd = &xd->plane[plane]; 140 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type); 141 142 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block); 143 tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block); 144 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block); 145 uint16_t *const eob = &p->eobs[block]; 146 const int diff_stride = block_size_wide[plane_bsize]; 147 int seg_id = mbmi->segment_id; 148 const TX_SIZE qm_tx_size = av1_get_adjusted_tx_size(tx_size); 149 // Use a flat matrix (i.e. no weighting) for 1D and Identity transforms 150 const qm_val_t *qmatrix = 151 IS_2D_TRANSFORM(tx_type) ? pd->seg_qmatrix[seg_id][qm_tx_size] 152 : cm->gqmatrix[NUM_QM_LEVELS - 1][0][qm_tx_size]; 153 const qm_val_t *iqmatrix = 154 IS_2D_TRANSFORM(tx_type) 155 ? pd->seg_iqmatrix[seg_id][qm_tx_size] 156 : cm->giqmatrix[NUM_QM_LEVELS - 1][0][qm_tx_size]; 157 158 const int src_offset = (blk_row * diff_stride + blk_col); 159 const int16_t *src_diff = &p->src_diff[src_offset << tx_size_wide_log2[0]]; 160 QUANT_PARAM qparam; 161 qparam.log_scale = av1_get_tx_scale(tx_size); 162 qparam.tx_size = tx_size; 163 qparam.qmatrix = qmatrix; 164 qparam.iqmatrix = iqmatrix; 165 qparam.use_quant_b_adapt = cm->use_quant_b_adapt; 166 TxfmParam txfm_param; 167 txfm_param.tx_type = tx_type; 168 txfm_param.tx_size = tx_size; 169 txfm_param.lossless = xd->lossless[mbmi->segment_id]; 170 txfm_param.tx_set_type = av1_get_ext_tx_set_type( 171 txfm_param.tx_size, is_inter_block(mbmi), cm->reduced_tx_set_used); 172 173 txfm_param.bd = xd->bd; 174 txfm_param.is_hbd = is_cur_buf_hbd(xd); 175 176 av1_fwd_txfm(src_diff, coeff, diff_stride, &txfm_param); 177 178 if (xform_quant_idx != AV1_XFORM_QUANT_SKIP_QUANT) { 179 const int n_coeffs = av1_get_max_eob(tx_size); 180 if (LIKELY(!x->skip_block)) { 181 quant_func_list[xform_quant_idx][txfm_param.is_hbd]( 182 coeff, n_coeffs, p, qcoeff, dqcoeff, eob, scan_order, &qparam); 183 } else { 184 av1_quantize_skip(n_coeffs, qcoeff, dqcoeff, eob); 185 } 186 } 187 // NOTE: optimize_b_following is true means av1_optimze_b will be called 188 // When the condition of doing optimize_b is changed, 189 // this flag need update simultaneously 190 const int optimize_b_following = 191 (xform_quant_idx != AV1_XFORM_QUANT_FP) || (txfm_param.lossless); 192 if (optimize_b_following) { 193 p->txb_entropy_ctx[block] = 194 (uint8_t)av1_get_txb_entropy_context(qcoeff, scan_order, *eob); 195 } else { 196 p->txb_entropy_ctx[block] = 0; 197 } 198 return; 199 } 200 201 static void encode_block(int plane, int block, int blk_row, int blk_col, 202 BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg, 203 int mi_row, int mi_col, RUN_TYPE dry_run) { 204 (void)mi_row; 205 (void)mi_col; 206 (void)dry_run; 207 struct encode_b_args *const args = arg; 208 const AV1_COMMON *const cm = &args->cpi->common; 209 MACROBLOCK *const x = args->x; 210 MACROBLOCKD *const xd = &x->e_mbd; 211 MB_MODE_INFO *mbmi = xd->mi[0]; 212 struct macroblock_plane *const p = &x->plane[plane]; 213 struct macroblockd_plane *const pd = &xd->plane[plane]; 214 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block); 215 uint8_t *dst; 216 ENTROPY_CONTEXT *a, *l; 217 int dummy_rate_cost = 0; 218 219 const int bw = block_size_wide[plane_bsize] >> tx_size_wide_log2[0]; 220 dst = &pd->dst 221 .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]]; 222 223 a = &args->ta[blk_col]; 224 l = &args->tl[blk_row]; 225 226 if (!is_blk_skip(x, plane, blk_row * bw + blk_col) && !mbmi->skip_mode) { 227 TX_TYPE tx_type = av1_get_tx_type(pd->plane_type, xd, blk_row, blk_col, 228 tx_size, cm->reduced_tx_set_used); 229 if (args->enable_optimize_b != NO_TRELLIS_OPT) { 230 av1_xform_quant( 231 cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size, tx_type, 232 USE_B_QUANT_NO_TRELLIS && 233 (args->enable_optimize_b == FINAL_PASS_TRELLIS_OPT) 234 ? AV1_XFORM_QUANT_B 235 : AV1_XFORM_QUANT_FP); 236 TXB_CTX txb_ctx; 237 get_txb_ctx(plane_bsize, tx_size, plane, a, l, &txb_ctx); 238 av1_optimize_b(args->cpi, x, plane, block, tx_size, tx_type, &txb_ctx, 239 args->cpi->sf.trellis_eob_fast, &dummy_rate_cost); 240 } else { 241 av1_xform_quant( 242 cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size, tx_type, 243 USE_B_QUANT_NO_TRELLIS ? AV1_XFORM_QUANT_B : AV1_XFORM_QUANT_FP); 244 } 245 } else { 246 p->eobs[block] = 0; 247 p->txb_entropy_ctx[block] = 0; 248 } 249 250 av1_set_txb_context(x, plane, block, tx_size, a, l); 251 252 if (p->eobs[block]) { 253 *(args->skip) = 0; 254 255 TX_TYPE tx_type = av1_get_tx_type(pd->plane_type, xd, blk_row, blk_col, 256 tx_size, cm->reduced_tx_set_used); 257 av1_inverse_transform_block(xd, dqcoeff, plane, tx_type, tx_size, dst, 258 pd->dst.stride, p->eobs[block], 259 cm->reduced_tx_set_used); 260 } 261 262 // TODO(debargha, jingning): Temporarily disable txk_type check for eob=0 263 // case. It is possible that certain collision in hash index would cause 264 // the assertion failure. To further optimize the rate-distortion 265 // performance, we need to re-visit this part and enable this assert 266 // again. 267 if (p->eobs[block] == 0 && plane == 0) { 268 #if 0 269 if (args->cpi->oxcf.aq_mode == NO_AQ && 270 args->cpi->oxcf.deltaq_mode == NO_DELTA_Q) { 271 // TODO(jingning,angiebird,huisu (at) google.com): enable txk_check when 272 // enable_optimize_b is true to detect potential RD bug. 273 const uint8_t disable_txk_check = args->enable_optimize_b; 274 if (!disable_txk_check) { 275 assert(mbmi->txk_type[av1_get_txk_type_index(plane_bsize, blk_row, 276 blk_col)] == DCT_DCT); 277 } 278 } 279 #endif 280 update_txk_array(mbmi->txk_type, plane_bsize, blk_row, blk_col, tx_size, 281 DCT_DCT); 282 } 283 284 #if CONFIG_MISMATCH_DEBUG 285 if (dry_run == OUTPUT_ENABLED) { 286 int pixel_c, pixel_r; 287 BLOCK_SIZE bsize = txsize_to_bsize[tx_size]; 288 int blk_w = block_size_wide[bsize]; 289 int blk_h = block_size_high[bsize]; 290 mi_to_pixel_loc(&pixel_c, &pixel_r, mi_col, mi_row, blk_col, blk_row, 291 pd->subsampling_x, pd->subsampling_y); 292 mismatch_record_block_tx(dst, pd->dst.stride, cm->current_frame.order_hint, 293 plane, pixel_c, pixel_r, blk_w, blk_h, 294 xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH); 295 } 296 #endif 297 } 298 299 static void encode_block_inter(int plane, int block, int blk_row, int blk_col, 300 BLOCK_SIZE plane_bsize, TX_SIZE tx_size, 301 void *arg, int mi_row, int mi_col, 302 RUN_TYPE dry_run) { 303 (void)mi_row; 304 (void)mi_col; 305 struct encode_b_args *const args = arg; 306 MACROBLOCK *const x = args->x; 307 MACROBLOCKD *const xd = &x->e_mbd; 308 MB_MODE_INFO *const mbmi = xd->mi[0]; 309 const struct macroblockd_plane *const pd = &xd->plane[plane]; 310 const int max_blocks_high = max_block_high(xd, plane_bsize, plane); 311 const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane); 312 313 if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return; 314 315 const TX_SIZE plane_tx_size = 316 plane ? av1_get_max_uv_txsize(mbmi->sb_type, pd->subsampling_x, 317 pd->subsampling_y) 318 : mbmi->inter_tx_size[av1_get_txb_size_index(plane_bsize, blk_row, 319 blk_col)]; 320 if (!plane) { 321 assert(tx_size_wide[tx_size] >= tx_size_wide[plane_tx_size] && 322 tx_size_high[tx_size] >= tx_size_high[plane_tx_size]); 323 } 324 325 if (tx_size == plane_tx_size || plane) { 326 encode_block(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg, 327 mi_row, mi_col, dry_run); 328 } else { 329 assert(tx_size < TX_SIZES_ALL); 330 const TX_SIZE sub_txs = sub_tx_size_map[tx_size]; 331 assert(IMPLIES(tx_size <= TX_4X4, sub_txs == tx_size)); 332 assert(IMPLIES(tx_size > TX_4X4, sub_txs < tx_size)); 333 // This is the square transform block partition entry point. 334 const int bsw = tx_size_wide_unit[sub_txs]; 335 const int bsh = tx_size_high_unit[sub_txs]; 336 const int step = bsh * bsw; 337 assert(bsw > 0 && bsh > 0); 338 339 for (int row = 0; row < tx_size_high_unit[tx_size]; row += bsh) { 340 for (int col = 0; col < tx_size_wide_unit[tx_size]; col += bsw) { 341 const int offsetr = blk_row + row; 342 const int offsetc = blk_col + col; 343 344 if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue; 345 346 encode_block_inter(plane, block, offsetr, offsetc, plane_bsize, sub_txs, 347 arg, mi_row, mi_col, dry_run); 348 block += step; 349 } 350 } 351 } 352 } 353 354 void av1_foreach_transformed_block_in_plane( 355 const MACROBLOCKD *const xd, BLOCK_SIZE bsize, int plane, 356 foreach_transformed_block_visitor visit, void *arg) { 357 const struct macroblockd_plane *const pd = &xd->plane[plane]; 358 // block and transform sizes, in number of 4x4 blocks log 2 ("*_b") 359 // 4x4=0, 8x8=2, 16x16=4, 32x32=6, 64x64=8 360 // transform size varies per plane, look it up in a common way. 361 const TX_SIZE tx_size = av1_get_tx_size(plane, xd); 362 const BLOCK_SIZE plane_bsize = 363 get_plane_block_size(bsize, pd->subsampling_x, pd->subsampling_y); 364 const uint8_t txw_unit = tx_size_wide_unit[tx_size]; 365 const uint8_t txh_unit = tx_size_high_unit[tx_size]; 366 const int step = txw_unit * txh_unit; 367 int i = 0, r, c; 368 369 // If mb_to_right_edge is < 0 we are in a situation in which 370 // the current block size extends into the UMV and we won't 371 // visit the sub blocks that are wholly within the UMV. 372 const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane); 373 const int max_blocks_high = max_block_high(xd, plane_bsize, plane); 374 375 int blk_row, blk_col; 376 377 const BLOCK_SIZE max_unit_bsize = 378 get_plane_block_size(BLOCK_64X64, pd->subsampling_x, pd->subsampling_y); 379 int mu_blocks_wide = block_size_wide[max_unit_bsize] >> tx_size_wide_log2[0]; 380 int mu_blocks_high = block_size_high[max_unit_bsize] >> tx_size_high_log2[0]; 381 mu_blocks_wide = AOMMIN(max_blocks_wide, mu_blocks_wide); 382 mu_blocks_high = AOMMIN(max_blocks_high, mu_blocks_high); 383 384 // Keep track of the row and column of the blocks we use so that we know 385 // if we are in the unrestricted motion border. 386 for (r = 0; r < max_blocks_high; r += mu_blocks_high) { 387 const int unit_height = AOMMIN(mu_blocks_high + r, max_blocks_high); 388 // Skip visiting the sub blocks that are wholly within the UMV. 389 for (c = 0; c < max_blocks_wide; c += mu_blocks_wide) { 390 const int unit_width = AOMMIN(mu_blocks_wide + c, max_blocks_wide); 391 for (blk_row = r; blk_row < unit_height; blk_row += txh_unit) { 392 for (blk_col = c; blk_col < unit_width; blk_col += txw_unit) { 393 visit(plane, i, blk_row, blk_col, plane_bsize, tx_size, arg); 394 i += step; 395 } 396 } 397 } 398 } 399 } 400 401 void av1_foreach_transformed_block(const MACROBLOCKD *const xd, 402 BLOCK_SIZE bsize, int mi_row, int mi_col, 403 foreach_transformed_block_visitor visit, 404 void *arg, const int num_planes) { 405 for (int plane = 0; plane < num_planes; ++plane) { 406 if (!is_chroma_reference(mi_row, mi_col, bsize, 407 xd->plane[plane].subsampling_x, 408 xd->plane[plane].subsampling_y)) 409 continue; 410 av1_foreach_transformed_block_in_plane(xd, bsize, plane, visit, arg); 411 } 412 } 413 414 typedef struct encode_block_pass1_args { 415 AV1_COMMON *cm; 416 MACROBLOCK *x; 417 } encode_block_pass1_args; 418 419 static void encode_block_pass1(int plane, int block, int blk_row, int blk_col, 420 BLOCK_SIZE plane_bsize, TX_SIZE tx_size, 421 void *arg) { 422 encode_block_pass1_args *args = (encode_block_pass1_args *)arg; 423 AV1_COMMON *cm = args->cm; 424 MACROBLOCK *const x = args->x; 425 MACROBLOCKD *const xd = &x->e_mbd; 426 struct macroblock_plane *const p = &x->plane[plane]; 427 struct macroblockd_plane *const pd = &xd->plane[plane]; 428 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block); 429 TxfmParam txfm_param; 430 uint8_t *dst; 431 dst = &pd->dst 432 .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]]; 433 av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size, 434 DCT_DCT, AV1_XFORM_QUANT_B); 435 436 if (p->eobs[block] > 0) { 437 txfm_param.bd = xd->bd; 438 txfm_param.is_hbd = is_cur_buf_hbd(xd); 439 txfm_param.tx_type = DCT_DCT; 440 txfm_param.tx_size = tx_size; 441 txfm_param.eob = p->eobs[block]; 442 txfm_param.lossless = xd->lossless[xd->mi[0]->segment_id]; 443 txfm_param.tx_set_type = av1_get_ext_tx_set_type( 444 txfm_param.tx_size, is_inter_block(xd->mi[0]), cm->reduced_tx_set_used); 445 if (txfm_param.is_hbd) { 446 av1_highbd_inv_txfm_add(dqcoeff, dst, pd->dst.stride, &txfm_param); 447 return; 448 } 449 av1_inv_txfm_add(dqcoeff, dst, pd->dst.stride, &txfm_param); 450 } 451 } 452 453 void av1_encode_sby_pass1(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize) { 454 encode_block_pass1_args args = { cm, x }; 455 av1_subtract_plane(x, bsize, 0); 456 av1_foreach_transformed_block_in_plane(&x->e_mbd, bsize, 0, 457 encode_block_pass1, &args); 458 } 459 460 void av1_encode_sb(const struct AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, 461 int mi_row, int mi_col, RUN_TYPE dry_run) { 462 (void)dry_run; 463 const AV1_COMMON *const cm = &cpi->common; 464 const int num_planes = av1_num_planes(cm); 465 MACROBLOCKD *const xd = &x->e_mbd; 466 struct optimize_ctx ctx; 467 MB_MODE_INFO *mbmi = xd->mi[0]; 468 struct encode_b_args arg = { cpi, 469 x, 470 &ctx, 471 &mbmi->skip, 472 NULL, 473 NULL, 474 cpi->optimize_seg_arr[mbmi->segment_id] }; 475 int plane; 476 477 mbmi->skip = 1; 478 479 if (x->skip) return; 480 481 for (plane = 0; plane < num_planes; ++plane) { 482 const int subsampling_x = xd->plane[plane].subsampling_x; 483 const int subsampling_y = xd->plane[plane].subsampling_y; 484 485 if (!is_chroma_reference(mi_row, mi_col, bsize, subsampling_x, 486 subsampling_y)) 487 continue; 488 489 const BLOCK_SIZE bsizec = 490 scale_chroma_bsize(bsize, subsampling_x, subsampling_y); 491 492 // TODO(jingning): Clean this up. 493 const struct macroblockd_plane *const pd = &xd->plane[plane]; 494 const BLOCK_SIZE plane_bsize = 495 get_plane_block_size(bsizec, pd->subsampling_x, pd->subsampling_y); 496 const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0]; 497 const int mi_height = block_size_high[plane_bsize] >> tx_size_high_log2[0]; 498 const TX_SIZE max_tx_size = get_vartx_max_txsize(xd, plane_bsize, plane); 499 500 const BLOCK_SIZE txb_size = txsize_to_bsize[max_tx_size]; 501 const int bw = block_size_wide[txb_size] >> tx_size_wide_log2[0]; 502 const int bh = block_size_high[txb_size] >> tx_size_high_log2[0]; 503 int idx, idy; 504 int block = 0; 505 int step = tx_size_wide_unit[max_tx_size] * tx_size_high_unit[max_tx_size]; 506 av1_get_entropy_contexts(bsizec, pd, ctx.ta[plane], ctx.tl[plane]); 507 508 av1_subtract_plane(x, bsizec, plane); 509 510 arg.ta = ctx.ta[plane]; 511 arg.tl = ctx.tl[plane]; 512 513 const BLOCK_SIZE max_unit_bsize = 514 get_plane_block_size(BLOCK_64X64, pd->subsampling_x, pd->subsampling_y); 515 int mu_blocks_wide = 516 block_size_wide[max_unit_bsize] >> tx_size_wide_log2[0]; 517 int mu_blocks_high = 518 block_size_high[max_unit_bsize] >> tx_size_high_log2[0]; 519 520 mu_blocks_wide = AOMMIN(mi_width, mu_blocks_wide); 521 mu_blocks_high = AOMMIN(mi_height, mu_blocks_high); 522 523 for (idy = 0; idy < mi_height; idy += mu_blocks_high) { 524 for (idx = 0; idx < mi_width; idx += mu_blocks_wide) { 525 int blk_row, blk_col; 526 const int unit_height = AOMMIN(mu_blocks_high + idy, mi_height); 527 const int unit_width = AOMMIN(mu_blocks_wide + idx, mi_width); 528 for (blk_row = idy; blk_row < unit_height; blk_row += bh) { 529 for (blk_col = idx; blk_col < unit_width; blk_col += bw) { 530 encode_block_inter(plane, block, blk_row, blk_col, plane_bsize, 531 max_tx_size, &arg, mi_row, mi_col, dry_run); 532 block += step; 533 } 534 } 535 } 536 } 537 } 538 } 539 540 static void encode_block_intra_and_set_context(int plane, int block, 541 int blk_row, int blk_col, 542 BLOCK_SIZE plane_bsize, 543 TX_SIZE tx_size, void *arg) { 544 av1_encode_block_intra(plane, block, blk_row, blk_col, plane_bsize, tx_size, 545 arg); 546 547 struct encode_b_args *const args = arg; 548 MACROBLOCK *x = args->x; 549 ENTROPY_CONTEXT *a = &args->ta[blk_col]; 550 ENTROPY_CONTEXT *l = &args->tl[blk_row]; 551 av1_set_txb_context(x, plane, block, tx_size, a, l); 552 } 553 554 void av1_encode_block_intra(int plane, int block, int blk_row, int blk_col, 555 BLOCK_SIZE plane_bsize, TX_SIZE tx_size, 556 void *arg) { 557 struct encode_b_args *const args = arg; 558 const AV1_COMMON *const cm = &args->cpi->common; 559 MACROBLOCK *const x = args->x; 560 MACROBLOCKD *const xd = &x->e_mbd; 561 MB_MODE_INFO *mbmi = xd->mi[0]; 562 struct macroblock_plane *const p = &x->plane[plane]; 563 struct macroblockd_plane *const pd = &xd->plane[plane]; 564 tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block); 565 PLANE_TYPE plane_type = get_plane_type(plane); 566 const TX_TYPE tx_type = av1_get_tx_type(plane_type, xd, blk_row, blk_col, 567 tx_size, cm->reduced_tx_set_used); 568 uint16_t *eob = &p->eobs[block]; 569 const int dst_stride = pd->dst.stride; 570 uint8_t *dst = 571 &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]]; 572 int dummy_rate_cost = 0; 573 574 av1_predict_intra_block_facade(cm, xd, plane, blk_col, blk_row, tx_size); 575 576 const int bw = block_size_wide[plane_bsize] >> tx_size_wide_log2[0]; 577 if (plane == 0 && is_blk_skip(x, plane, blk_row * bw + blk_col)) { 578 *eob = 0; 579 p->txb_entropy_ctx[block] = 0; 580 } else { 581 av1_subtract_txb(x, plane, plane_bsize, blk_col, blk_row, tx_size); 582 583 const ENTROPY_CONTEXT *a = &args->ta[blk_col]; 584 const ENTROPY_CONTEXT *l = &args->tl[blk_row]; 585 if (args->enable_optimize_b != NO_TRELLIS_OPT) { 586 av1_xform_quant( 587 cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size, tx_type, 588 USE_B_QUANT_NO_TRELLIS && 589 (args->enable_optimize_b == FINAL_PASS_TRELLIS_OPT) 590 ? AV1_XFORM_QUANT_B 591 : AV1_XFORM_QUANT_FP); 592 TXB_CTX txb_ctx; 593 get_txb_ctx(plane_bsize, tx_size, plane, a, l, &txb_ctx); 594 av1_optimize_b(args->cpi, x, plane, block, tx_size, tx_type, &txb_ctx, 595 args->cpi->sf.trellis_eob_fast, &dummy_rate_cost); 596 } else { 597 av1_xform_quant( 598 cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size, tx_type, 599 USE_B_QUANT_NO_TRELLIS ? AV1_XFORM_QUANT_B : AV1_XFORM_QUANT_FP); 600 } 601 } 602 603 if (*eob) { 604 av1_inverse_transform_block(xd, dqcoeff, plane, tx_type, tx_size, dst, 605 dst_stride, *eob, cm->reduced_tx_set_used); 606 } 607 608 // TODO(jingning): Temporarily disable txk_type check for eob=0 case. 609 // It is possible that certain collision in hash index would cause 610 // the assertion failure. To further optimize the rate-distortion 611 // performance, we need to re-visit this part and enable this assert 612 // again. 613 if (*eob == 0 && plane == 0) { 614 #if 0 615 if (args->cpi->oxcf.aq_mode == NO_AQ 616 && args->cpi->oxcf.deltaq_mode == NO_DELTA_Q) { 617 assert(mbmi->txk_type[av1_get_txk_type_index(plane_bsize, blk_row, 618 blk_col)] == DCT_DCT); 619 } 620 #endif 621 update_txk_array(mbmi->txk_type, plane_bsize, blk_row, blk_col, tx_size, 622 DCT_DCT); 623 } 624 625 // For intra mode, skipped blocks are so rare that transmitting skip=1 is 626 // very expensive. 627 *(args->skip) = 0; 628 629 if (plane == AOM_PLANE_Y && xd->cfl.store_y) { 630 cfl_store_tx(xd, blk_row, blk_col, tx_size, plane_bsize); 631 } 632 } 633 634 void av1_encode_intra_block_plane(const struct AV1_COMP *cpi, MACROBLOCK *x, 635 BLOCK_SIZE bsize, int plane, 636 int enable_optimize_b, int mi_row, 637 int mi_col) { 638 const MACROBLOCKD *const xd = &x->e_mbd; 639 ENTROPY_CONTEXT ta[MAX_MIB_SIZE] = { 0 }; 640 ENTROPY_CONTEXT tl[MAX_MIB_SIZE] = { 0 }; 641 642 struct encode_b_args arg = { 643 cpi, x, NULL, &(xd->mi[0]->skip), ta, tl, enable_optimize_b 644 }; 645 646 if (!is_chroma_reference(mi_row, mi_col, bsize, 647 xd->plane[plane].subsampling_x, 648 xd->plane[plane].subsampling_y)) 649 return; 650 651 if (enable_optimize_b) { 652 const struct macroblockd_plane *const pd = &xd->plane[plane]; 653 av1_get_entropy_contexts(bsize, pd, ta, tl); 654 } 655 av1_foreach_transformed_block_in_plane( 656 xd, bsize, plane, encode_block_intra_and_set_context, &arg); 657 } 658