encodetxb.c 67.3 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12
/*
 * Copyright (c) 2017, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */

#include "av1/common/scan.h"
Angie Chiang's avatar
Angie Chiang committed
13
#include "av1/common/blockd.h"
14
#include "av1/common/idct.h"
Angie Chiang's avatar
Angie Chiang committed
15
#include "av1/common/pred_common.h"
16 17
#include "av1/encoder/bitstream.h"
#include "av1/encoder/encodeframe.h"
18
#include "av1/encoder/cost.h"
19
#include "av1/encoder/encodetxb.h"
Angie Chiang's avatar
Angie Chiang committed
20
#include "av1/encoder/rdopt.h"
21
#include "av1/encoder/subexp.h"
Angie Chiang's avatar
Angie Chiang committed
22
#include "av1/encoder/tokenize.h"
23

Angie Chiang's avatar
Angie Chiang committed
24 25
#define TEST_OPTIMIZE_TXB 0

26
void av1_alloc_txb_buf(AV1_COMP *cpi) {
27
#if 0
28 29 30 31 32 33 34 35 36 37 38 39
  AV1_COMMON *cm = &cpi->common;
  int mi_block_size = 1 << MI_SIZE_LOG2;
  // TODO(angiebird): Make sure cm->subsampling_x/y is set correctly, and then
  // use precise buffer size according to cm->subsampling_x/y
  int pixel_stride = mi_block_size * cm->mi_cols;
  int pixel_height = mi_block_size * cm->mi_rows;
  int i;
  for (i = 0; i < MAX_MB_PLANE; ++i) {
    CHECK_MEM_ERROR(
        cm, cpi->tcoeff_buf[i],
        aom_malloc(sizeof(*cpi->tcoeff_buf[i]) * pixel_stride * pixel_height));
  }
40 41 42
#else
  (void)cpi;
#endif
43 44 45
}

void av1_free_txb_buf(AV1_COMP *cpi) {
46
#if 0
47 48 49 50
  int i;
  for (i = 0; i < MAX_MB_PLANE; ++i) {
    aom_free(cpi->tcoeff_buf[i]);
  }
51 52 53
#else
  (void)cpi;
#endif
54 55
}

56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72
static void write_golomb(aom_writer *w, int level) {
  int x = level + 1;
  int i = x;
  int length = 0;

  while (i) {
    i >>= 1;
    ++length;
  }
  assert(length > 0);

  for (i = 0; i < length - 1; ++i) aom_write_bit(w, 0);

  for (i = length - 1; i >= 0; --i) aom_write_bit(w, (x >> i) & 0x01);
}

void av1_write_coeffs_txb(const AV1_COMMON *const cm, MACROBLOCKD *xd,
Jingning Han's avatar
Jingning Han committed
73
                          aom_writer *w, int block, int plane, TX_SIZE tx_size,
74 75 76 77 78
                          const tran_low_t *tcoeff, uint16_t eob,
                          TXB_CTX *txb_ctx) {
  aom_prob *nz_map;
  aom_prob *eob_flag;
  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
79
  const PLANE_TYPE plane_type = get_plane_type(plane);
80
  const TX_SIZE txs_ctx = get_txsize_context(tx_size);
81 82 83 84 85 86 87
  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
  const SCAN_ORDER *const scan_order =
      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
  const int16_t *scan = scan_order->scan;
  int c;
  int is_nz;
  const int bwl = b_width_log2_lookup[txsize_to_bsize[tx_size]] + 2;
88
  const int height = tx_size_high[tx_size];
89
  const int seg_eob = tx_size_2d[tx_size];
90 91 92
  uint8_t txb_mask[32 * 32] = { 0 };
  uint16_t update_eob = 0;

93
  aom_write(w, eob == 0, cm->fc->txb_skip[txs_ctx][txb_ctx->txb_skip_ctx]);
94 95

  if (eob == 0) return;
Angie Chiang's avatar
Angie Chiang committed
96
#if CONFIG_TXK_SEL
97
  av1_write_tx_type(cm, xd, block, plane, w);
Angie Chiang's avatar
Angie Chiang committed
98
#endif
99

100 101
  nz_map = cm->fc->nz_map[txs_ctx][plane_type];
  eob_flag = cm->fc->eob_flag[txs_ctx][plane_type];
102 103

  for (c = 0; c < eob; ++c) {
104
    int coeff_ctx = get_nz_map_ctx(tcoeff, txb_mask, scan[c], bwl, height);
105
    int eob_ctx = get_eob_ctx(tcoeff, scan[c], txs_ctx);
106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121

    tran_low_t v = tcoeff[scan[c]];
    is_nz = (v != 0);

    if (c == seg_eob - 1) break;

    aom_write(w, is_nz, nz_map[coeff_ctx]);

    if (is_nz) {
      aom_write(w, c == (eob - 1), eob_flag[eob_ctx]);
    }
    txb_mask[scan[c]] = 1;
  }

  int i;
  for (i = 0; i < NUM_BASE_LEVELS; ++i) {
122
    aom_prob *coeff_base = cm->fc->coeff_base[txs_ctx][plane_type][i];
123 124 125 126 127 128 129 130 131 132

    update_eob = 0;
    for (c = eob - 1; c >= 0; --c) {
      tran_low_t v = tcoeff[scan[c]];
      tran_low_t level = abs(v);
      int sign = (v < 0) ? 1 : 0;
      int ctx;

      if (level <= i) continue;

133
      ctx = get_base_ctx(tcoeff, scan[c], bwl, height, i + 1);
134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164

      if (level == i + 1) {
        aom_write(w, 1, coeff_base[ctx]);
        if (c == 0) {
          aom_write(w, sign, cm->fc->dc_sign[plane_type][txb_ctx->dc_sign_ctx]);
        } else {
          aom_write_bit(w, sign);
        }
        continue;
      }
      aom_write(w, 0, coeff_base[ctx]);
      update_eob = AOMMAX(update_eob, c);
    }
  }

  for (c = update_eob; c >= 0; --c) {
    tran_low_t v = tcoeff[scan[c]];
    tran_low_t level = abs(v);
    int sign = (v < 0) ? 1 : 0;
    int idx;
    int ctx;

    if (level <= NUM_BASE_LEVELS) continue;

    if (c == 0) {
      aom_write(w, sign, cm->fc->dc_sign[plane_type][txb_ctx->dc_sign_ctx]);
    } else {
      aom_write_bit(w, sign);
    }

    // level is above 1.
165
    ctx = get_br_ctx(tcoeff, scan[c], bwl, height);
166 167
    for (idx = 0; idx < COEFF_BASE_RANGE; ++idx) {
      if (level == (idx + 1 + NUM_BASE_LEVELS)) {
168
        aom_write(w, 1, cm->fc->coeff_lps[txs_ctx][plane_type][ctx]);
169 170
        break;
      }
171
      aom_write(w, 0, cm->fc->coeff_lps[txs_ctx][plane_type][ctx]);
172 173 174 175 176 177 178
    }
    if (idx < COEFF_BASE_RANGE) continue;

    // use 0-th order Golomb code to handle the residual level.
    write_golomb(w, level - COEFF_BASE_RANGE - 1 - NUM_BASE_LEVELS);
  }
}
179

Angie Chiang's avatar
Angie Chiang committed
180 181 182 183 184 185 186 187
void av1_write_coeffs_mb(const AV1_COMMON *const cm, MACROBLOCK *x,
                         aom_writer *w, int plane) {
  MACROBLOCKD *xd = &x->e_mbd;
  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
  BLOCK_SIZE bsize = mbmi->sb_type;
  struct macroblockd_plane *pd = &xd->plane[plane];

#if CONFIG_CB4X4
188
#if CONFIG_CHROMA_2X2
Angie Chiang's avatar
Angie Chiang committed
189
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
190 191 192 193
#else
  const BLOCK_SIZE plane_bsize =
      AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
#endif  // CONFIG_CHROMA_2X2
Angie Chiang's avatar
Angie Chiang committed
194 195 196 197
#else
  const BLOCK_SIZE plane_bsize =
      get_plane_block_size(AOMMAX(bsize, BLOCK_8X8), pd);
#endif
198 199
  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
Angie Chiang's avatar
Angie Chiang committed
200 201 202 203 204 205
  TX_SIZE tx_size = get_tx_size(plane, xd);
  const int bkw = tx_size_wide_unit[tx_size];
  const int bkh = tx_size_high_unit[tx_size];
  const int step = tx_size_wide_unit[tx_size] * tx_size_high_unit[tx_size];
  int row, col;
  int block = 0;
206 207
  for (row = 0; row < max_blocks_high; row += bkh) {
    for (col = 0; col < max_blocks_wide; col += bkw) {
Angie Chiang's avatar
Angie Chiang committed
208 209 210 211
      tran_low_t *tcoeff = BLOCK_OFFSET(x->mbmi_ext->tcoeff[plane], block);
      uint16_t eob = x->mbmi_ext->eobs[plane][block];
      TXB_CTX txb_ctx = { x->mbmi_ext->txb_skip_ctx[plane][block],
                          x->mbmi_ext->dc_sign_ctx[plane][block] };
Jingning Han's avatar
Jingning Han committed
212 213
      av1_write_coeffs_txb(cm, xd, w, block, plane, tx_size, tcoeff, eob,
                           &txb_ctx);
Angie Chiang's avatar
Angie Chiang committed
214 215 216 217 218
      block += step;
    }
  }
}

219 220
static INLINE void get_base_ctx_set(const tran_low_t *tcoeffs,
                                    int c,  // raster order
221
                                    const int bwl, const int height,
222 223 224 225 226 227 228 229 230 231 232 233 234 235
                                    int ctx_set[NUM_BASE_LEVELS]) {
  const int row = c >> bwl;
  const int col = c - (row << bwl);
  const int stride = 1 << bwl;
  int mag[NUM_BASE_LEVELS] = { 0 };
  int idx;
  tran_low_t abs_coeff;
  int i;

  for (idx = 0; idx < BASE_CONTEXT_POSITION_NUM; ++idx) {
    int ref_row = row + base_ref_offset[idx][0];
    int ref_col = col + base_ref_offset[idx][1];
    int pos = (ref_row << bwl) + ref_col;

236
    if (ref_row < 0 || ref_col < 0 || ref_row >= height || ref_col >= stride)
237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262
      continue;

    abs_coeff = abs(tcoeffs[pos]);

    for (i = 0; i < NUM_BASE_LEVELS; ++i) {
      ctx_set[i] += abs_coeff > i;
      if (base_ref_offset[idx][0] >= 0 && base_ref_offset[idx][1] >= 0)
        mag[i] |= abs_coeff > (i + 1);
    }
  }

  for (i = 0; i < NUM_BASE_LEVELS; ++i) {
    ctx_set[i] = (ctx_set[i] + 1) >> 1;

    if (row == 0 && col == 0)
      ctx_set[i] = (ctx_set[i] << 1) + mag[i];
    else if (row == 0)
      ctx_set[i] = 8 + (ctx_set[i] << 1) + mag[i];
    else if (col == 0)
      ctx_set[i] = 18 + (ctx_set[i] << 1) + mag[i];
    else
      ctx_set[i] = 28 + (ctx_set[i] << 1) + mag[i];
  }
  return;
}

263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288
static INLINE int get_br_cost(tran_low_t abs_qc, int ctx,
                              const aom_prob *coeff_lps) {
  const tran_low_t min_level = 1 + NUM_BASE_LEVELS;
  const tran_low_t max_level = 1 + NUM_BASE_LEVELS + COEFF_BASE_RANGE;
  if (abs_qc >= min_level) {
    const int cost0 = av1_cost_bit(coeff_lps[ctx], 0);
    const int cost1 = av1_cost_bit(coeff_lps[ctx], 1);
    if (abs_qc >= max_level)
      return COEFF_BASE_RANGE * cost0;
    else
      return (abs_qc - min_level) * cost0 + cost1;
  } else {
    return 0;
  }
}

static INLINE int get_base_cost(tran_low_t abs_qc, int ctx,
                                aom_prob (*coeff_base)[COEFF_BASE_CONTEXTS],
                                int base_idx) {
  const int level = base_idx + 1;
  if (abs_qc < level)
    return 0;
  else
    return av1_cost_bit(coeff_base[base_idx][ctx], abs_qc == level);
}

289
int av1_cost_coeffs_txb(const AV1_COMP *const cpi, MACROBLOCK *x, int plane,
Jingning Han's avatar
Jingning Han committed
290
                        int block, TX_SIZE tx_size, TXB_CTX *txb_ctx) {
291
  const AV1_COMMON *const cm = &cpi->common;
292
  MACROBLOCKD *const xd = &x->e_mbd;
293
  TX_SIZE txs_ctx = get_txsize_context(tx_size);
294 295 296 297 298 299 300
  const PLANE_TYPE plane_type = get_plane_type(plane);
  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
  const struct macroblock_plane *p = &x->plane[plane];
  const int eob = p->eobs[block];
  const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
  int c, cost;
301
  const int seg_eob = AOMMIN(eob, tx_size_2d[tx_size] - 1);
302
  int txb_skip_ctx = txb_ctx->txb_skip_ctx;
303
  aom_prob *nz_map = xd->fc->nz_map[txs_ctx][plane_type];
304 305

  const int bwl = b_width_log2_lookup[txsize_to_bsize[tx_size]] + 2;
306 307
  const int height = tx_size_high[tx_size];

308 309 310 311
  // txb_mask is only initialized for once here. After that, it will be set when
  // coding zero map and then reset when coding level 1 info.
  uint8_t txb_mask[32 * 32] = { 0 };
  aom_prob(*coeff_base)[COEFF_BASE_CONTEXTS] =
312
      xd->fc->coeff_base[txs_ctx][plane_type];
313 314 315 316 317 318 319 320

  const SCAN_ORDER *const scan_order =
      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
  const int16_t *scan = scan_order->scan;

  cost = 0;

  if (eob == 0) {
321
    cost = av1_cost_bit(xd->fc->txb_skip[txs_ctx][txb_skip_ctx], 1);
322 323 324
    return cost;
  }

325
  cost = av1_cost_bit(xd->fc->txb_skip[txs_ctx][txb_skip_ctx], 0);
326

Angie Chiang's avatar
Angie Chiang committed
327
#if CONFIG_TXK_SEL
328
  cost += av1_tx_type_cost(cpi, xd, mbmi->sb_type, plane, tx_size, tx_type);
Angie Chiang's avatar
Angie Chiang committed
329
#endif
330

331 332 333 334 335 336
  for (c = 0; c < eob; ++c) {
    tran_low_t v = qcoeff[scan[c]];
    int is_nz = (v != 0);
    int level = abs(v);

    if (c < seg_eob) {
337
      int coeff_ctx = get_nz_map_ctx(qcoeff, txb_mask, scan[c], bwl, height);
338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353
      cost += av1_cost_bit(nz_map[coeff_ctx], is_nz);
    }

    if (is_nz) {
      int ctx_ls[NUM_BASE_LEVELS] = { 0 };
      int sign = (v < 0) ? 1 : 0;

      // sign bit cost
      if (c == 0) {
        int dc_sign_ctx = txb_ctx->dc_sign_ctx;

        cost += av1_cost_bit(xd->fc->dc_sign[plane_type][dc_sign_ctx], sign);
      } else {
        cost += av1_cost_bit(128, sign);
      }

354
      get_base_ctx_set(qcoeff, scan[c], bwl, height, ctx_ls);
355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370

      int i;
      for (i = 0; i < NUM_BASE_LEVELS; ++i) {
        if (level <= i) continue;

        if (level == i + 1) {
          cost += av1_cost_bit(coeff_base[i][ctx_ls[i]], 1);
          continue;
        }
        cost += av1_cost_bit(coeff_base[i][ctx_ls[i]], 0);
      }

      if (level > NUM_BASE_LEVELS) {
        int idx;
        int ctx;

371
        ctx = get_br_ctx(qcoeff, scan[c], bwl, height);
372 373 374 375

        for (idx = 0; idx < COEFF_BASE_RANGE; ++idx) {
          if (level == (idx + 1 + NUM_BASE_LEVELS)) {
            cost +=
376
                av1_cost_bit(xd->fc->coeff_lps[txs_ctx][plane_type][ctx], 1);
377 378
            break;
          }
379
          cost += av1_cost_bit(xd->fc->coeff_lps[txs_ctx][plane_type][ctx], 0);
380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400
        }

        if (idx >= COEFF_BASE_RANGE) {
          // residual cost
          int r = level - COEFF_BASE_RANGE - NUM_BASE_LEVELS;
          int ri = r;
          int length = 0;

          while (ri) {
            ri >>= 1;
            ++length;
          }

          for (ri = 0; ri < length - 1; ++ri) cost += av1_cost_bit(128, 0);

          for (ri = length - 1; ri >= 0; --ri)
            cost += av1_cost_bit(128, (r >> ri) & 0x01);
        }
      }

      if (c < seg_eob) {
401
        int eob_ctx = get_eob_ctx(qcoeff, scan[c], txs_ctx);
402
        cost += av1_cost_bit(xd->fc->eob_flag[txs_ctx][plane_type][eob_ctx],
403 404 405 406 407 408 409 410 411
                             c == (eob - 1));
      }
    }

    txb_mask[scan[c]] = 1;
  }

  return cost;
}
Angie Chiang's avatar
Angie Chiang committed
412

413 414 415 416 417 418 419 420
static INLINE int has_base(tran_low_t qc, int base_idx) {
  const int level = base_idx + 1;
  return abs(qc) >= level;
}

static void gen_base_count_mag_arr(int (*base_count_arr)[MAX_TX_SQUARE],
                                   int (*base_mag_arr)[2],
                                   const tran_low_t *qcoeff, int stride,
421
                                   int height, int eob, const int16_t *scan) {
422 423 424 425 426 427
  for (int c = 0; c < eob; ++c) {
    const int coeff_idx = scan[c];  // raster order
    if (!has_base(qcoeff[coeff_idx], 0)) continue;
    const int row = coeff_idx / stride;
    const int col = coeff_idx % stride;
    int *mag = base_mag_arr[coeff_idx];
428
    get_mag(mag, qcoeff, stride, height, row, col, base_ref_offset,
429 430 431 432
            BASE_CONTEXT_POSITION_NUM);
    for (int i = 0; i < NUM_BASE_LEVELS; ++i) {
      if (!has_base(qcoeff[coeff_idx], i)) continue;
      int *count = base_count_arr[i] + coeff_idx;
433 434
      *count = get_level_count(qcoeff, stride, height, row, col, i,
                               base_ref_offset, BASE_CONTEXT_POSITION_NUM);
435 436 437 438 439
    }
  }
}

static void gen_nz_count_arr(int(*nz_count_arr), const tran_low_t *qcoeff,
440
                             int stride, int height, int eob,
441 442 443 444 445 446 447
                             const SCAN_ORDER *scan_order) {
  const int16_t *scan = scan_order->scan;
  const int16_t *iscan = scan_order->iscan;
  for (int c = 0; c < eob; ++c) {
    const int coeff_idx = scan[c];  // raster order
    const int row = coeff_idx / stride;
    const int col = coeff_idx % stride;
448 449
    nz_count_arr[coeff_idx] =
        get_nz_count(qcoeff, stride, height, row, col, iscan);
450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490
  }
}

static void gen_nz_ctx_arr(int (*nz_ctx_arr)[2], int(*nz_count_arr),
                           const tran_low_t *qcoeff, int bwl, int eob,
                           const SCAN_ORDER *scan_order) {
  const int16_t *scan = scan_order->scan;
  const int16_t *iscan = scan_order->iscan;
  for (int c = 0; c < eob; ++c) {
    const int coeff_idx = scan[c];  // raster order
    const int count = nz_count_arr[coeff_idx];
    nz_ctx_arr[coeff_idx][0] =
        get_nz_map_ctx_from_count(count, qcoeff, coeff_idx, bwl, iscan);
  }
}

static void gen_base_ctx_arr(int (*base_ctx_arr)[MAX_TX_SQUARE][2],
                             int (*base_count_arr)[MAX_TX_SQUARE],
                             int (*base_mag_arr)[2], const tran_low_t *qcoeff,
                             int stride, int eob, const int16_t *scan) {
  (void)qcoeff;
  for (int i = 0; i < NUM_BASE_LEVELS; ++i) {
    for (int c = 0; c < eob; ++c) {
      const int coeff_idx = scan[c];  // raster order
      if (!has_base(qcoeff[coeff_idx], i)) continue;
      const int row = coeff_idx / stride;
      const int col = coeff_idx % stride;
      const int count = base_count_arr[i][coeff_idx];
      const int *mag = base_mag_arr[coeff_idx];
      const int level = i + 1;
      base_ctx_arr[i][coeff_idx][0] =
          get_base_ctx_from_count_mag(row, col, count, mag[0], level);
    }
  }
}

static INLINE int has_br(tran_low_t qc) {
  return abs(qc) >= 1 + NUM_BASE_LEVELS;
}

static void gen_br_count_mag_arr(int *br_count_arr, int (*br_mag_arr)[2],
491 492
                                 const tran_low_t *qcoeff, int stride,
                                 int height, int eob, const int16_t *scan) {
493 494 495 496 497 498 499
  for (int c = 0; c < eob; ++c) {
    const int coeff_idx = scan[c];  // raster order
    if (!has_br(qcoeff[coeff_idx])) continue;
    const int row = coeff_idx / stride;
    const int col = coeff_idx % stride;
    int *count = br_count_arr + coeff_idx;
    int *mag = br_mag_arr[coeff_idx];
500
    *count = get_level_count(qcoeff, stride, height, row, col, NUM_BASE_LEVELS,
501
                             br_ref_offset, BR_CONTEXT_POSITION_NUM);
502
    get_mag(mag, qcoeff, stride, height, row, col, br_ref_offset,
503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522
            BR_CONTEXT_POSITION_NUM);
  }
}

static void gen_br_ctx_arr(int (*br_ctx_arr)[2], const int *br_count_arr,
                           int (*br_mag_arr)[2], const tran_low_t *qcoeff,
                           int stride, int eob, const int16_t *scan) {
  (void)qcoeff;
  for (int c = 0; c < eob; ++c) {
    const int coeff_idx = scan[c];  // raster order
    if (!has_br(qcoeff[coeff_idx])) continue;
    const int row = coeff_idx / stride;
    const int col = coeff_idx % stride;
    const int count = br_count_arr[coeff_idx];
    const int *mag = br_mag_arr[coeff_idx];
    br_ctx_arr[coeff_idx][0] =
        get_br_ctx_from_count_mag(row, col, count, mag[0]);
  }
}

523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551
static INLINE int get_sign_bit_cost(tran_low_t qc, int coeff_idx,
                                    const aom_prob *dc_sign_prob,
                                    int dc_sign_ctx) {
  const int sign = (qc < 0) ? 1 : 0;
  // sign bit cost
  if (coeff_idx == 0) {
    return av1_cost_bit(dc_sign_prob[dc_sign_ctx], sign);
  } else {
    return av1_cost_bit(128, sign);
  }
}
static INLINE int get_golomb_cost(int abs_qc) {
  if (abs_qc >= 1 + NUM_BASE_LEVELS + COEFF_BASE_RANGE) {
    // residual cost
    int r = abs_qc - COEFF_BASE_RANGE - NUM_BASE_LEVELS;
    int ri = r;
    int length = 0;

    while (ri) {
      ri >>= 1;
      ++length;
    }

    return av1_cost_literal(2 * length - 1);
  } else {
    return 0;
  }
}

552 553 554 555
// TODO(angiebird): add static once this function is called
void gen_txb_cache(TxbCache *txb_cache, TxbInfo *txb_info) {
  const int16_t *scan = txb_info->scan_order->scan;
  gen_nz_count_arr(txb_cache->nz_count_arr, txb_info->qcoeff, txb_info->stride,
556
                   txb_info->height, txb_info->eob, txb_info->scan_order);
557 558 559 560
  gen_nz_ctx_arr(txb_cache->nz_ctx_arr, txb_cache->nz_count_arr,
                 txb_info->qcoeff, txb_info->bwl, txb_info->eob,
                 txb_info->scan_order);
  gen_base_count_mag_arr(txb_cache->base_count_arr, txb_cache->base_mag_arr,
561 562
                         txb_info->qcoeff, txb_info->stride, txb_info->height,
                         txb_info->eob, scan);
563 564 565 566
  gen_base_ctx_arr(txb_cache->base_ctx_arr, txb_cache->base_count_arr,
                   txb_cache->base_mag_arr, txb_info->qcoeff, txb_info->stride,
                   txb_info->eob, scan);
  gen_br_count_mag_arr(txb_cache->br_count_arr, txb_cache->br_mag_arr,
567 568
                       txb_info->qcoeff, txb_info->stride, txb_info->height,
                       txb_info->eob, scan);
569 570 571 572 573
  gen_br_ctx_arr(txb_cache->br_ctx_arr, txb_cache->br_count_arr,
                 txb_cache->br_mag_arr, txb_info->qcoeff, txb_info->stride,
                 txb_info->eob, scan);
}

574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604
static INLINE aom_prob get_level_prob(int level, int coeff_idx,
                                      const TxbCache *txb_cache,
                                      const TxbProbs *txb_probs) {
  if (level == 0) {
    const int ctx = txb_cache->nz_ctx_arr[coeff_idx][0];
    return txb_probs->nz_map[ctx];
  } else if (level >= 1 && level < 1 + NUM_BASE_LEVELS) {
    const int idx = level - 1;
    const int ctx = txb_cache->base_ctx_arr[idx][coeff_idx][0];
    return txb_probs->coeff_base[idx][ctx];
  } else if (level >= 1 + NUM_BASE_LEVELS &&
             level < 1 + NUM_BASE_LEVELS + COEFF_BASE_RANGE) {
    const int ctx = txb_cache->br_ctx_arr[coeff_idx][0];
    return txb_probs->coeff_lps[ctx];
  } else if (level >= 1 + NUM_BASE_LEVELS + COEFF_BASE_RANGE) {
    printf("get_level_prob does not support golomb\n");
    assert(0);
    return 0;
  } else {
    assert(0);
    return 0;
  }
}

static INLINE tran_low_t get_lower_coeff(tran_low_t qc) {
  if (qc == 0) {
    return 0;
  }
  return qc > 0 ? qc - 1 : qc + 1;
}

Angie Chiang's avatar
Angie Chiang committed
605 606 607 608 609 610 611
static INLINE void update_mag_arr(int *mag_arr, int abs_qc) {
  if (mag_arr[0] == abs_qc) {
    mag_arr[1] -= 1;
    assert(mag_arr[1] >= 0);
  }
}

Angie Chiang's avatar
Angie Chiang committed
612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766
static INLINE int get_mag_from_mag_arr(const int *mag_arr) {
  int mag;
  if (mag_arr[1] > 0) {
    mag = mag_arr[0];
  } else if (mag_arr[0] > 0) {
    mag = mag_arr[0] - 1;
  } else {
    // no neighbor
    assert(mag_arr[0] == 0 && mag_arr[1] == 0);
    mag = 0;
  }
  return mag;
}

static int neighbor_level_down_update(int *new_count, int *new_mag, int count,
                                      const int *mag, int coeff_idx,
                                      tran_low_t abs_nb_coeff, int nb_coeff_idx,
                                      int level, const TxbInfo *txb_info) {
  *new_count = count;
  *new_mag = get_mag_from_mag_arr(mag);

  int update = 0;
  // check if br_count changes
  if (abs_nb_coeff == level) {
    update = 1;
    *new_count -= 1;
    assert(*new_count >= 0);
  }
  const int row = coeff_idx >> txb_info->bwl;
  const int col = coeff_idx - (row << txb_info->bwl);
  const int nb_row = nb_coeff_idx >> txb_info->bwl;
  const int nb_col = nb_coeff_idx - (nb_row << txb_info->bwl);

  // check if mag changes
  if (nb_row >= row && nb_col >= col) {
    if (abs_nb_coeff == mag[0]) {
      assert(mag[1] > 0);
      if (mag[1] == 1) {
        // the nb is the only qc with max mag
        *new_mag -= 1;
        assert(*new_mag >= 0);
        update = 1;
      }
    }
  }
  return update;
}

static int try_neighbor_level_down_br(int coeff_idx, int nb_coeff_idx,
                                      const TxbCache *txb_cache,
                                      const TxbProbs *txb_probs,
                                      const TxbInfo *txb_info) {
  const tran_low_t qc = txb_info->qcoeff[coeff_idx];
  const tran_low_t abs_qc = abs(qc);
  const int level = NUM_BASE_LEVELS + 1;
  if (abs_qc < level) return 0;

  const tran_low_t nb_coeff = txb_info->qcoeff[nb_coeff_idx];
  const tran_low_t abs_nb_coeff = abs(nb_coeff);
  const int count = txb_cache->br_count_arr[coeff_idx];
  const int *mag = txb_cache->br_mag_arr[coeff_idx];
  int new_count;
  int new_mag;
  const int update =
      neighbor_level_down_update(&new_count, &new_mag, count, mag, coeff_idx,
                                 abs_nb_coeff, nb_coeff_idx, level, txb_info);
  if (update) {
    const int row = coeff_idx >> txb_info->bwl;
    const int col = coeff_idx - (row << txb_info->bwl);
    const int ctx = txb_cache->br_ctx_arr[coeff_idx][0];
    const int org_cost = get_br_cost(abs_qc, ctx, txb_probs->coeff_lps);

    const int new_ctx = get_br_ctx_from_count_mag(row, col, new_count, new_mag);
    const int new_cost = get_br_cost(abs_qc, new_ctx, txb_probs->coeff_lps);
    const int cost_diff = -org_cost + new_cost;
    return cost_diff;
  } else {
    return 0;
  }
}

static int try_neighbor_level_down_base(int coeff_idx, int nb_coeff_idx,
                                        const TxbCache *txb_cache,
                                        const TxbProbs *txb_probs,
                                        const TxbInfo *txb_info) {
  const tran_low_t qc = txb_info->qcoeff[coeff_idx];
  const tran_low_t abs_qc = abs(qc);

  int cost_diff = 0;
  for (int base_idx = 0; base_idx < NUM_BASE_LEVELS; ++base_idx) {
    const int level = base_idx + 1;
    if (abs_qc < level) continue;

    const tran_low_t nb_coeff = txb_info->qcoeff[nb_coeff_idx];
    const tran_low_t abs_nb_coeff = abs(nb_coeff);

    const int count = txb_cache->base_count_arr[base_idx][coeff_idx];
    const int *mag = txb_cache->base_mag_arr[coeff_idx];
    int new_count;
    int new_mag;
    const int update =
        neighbor_level_down_update(&new_count, &new_mag, count, mag, coeff_idx,
                                   abs_nb_coeff, nb_coeff_idx, level, txb_info);
    if (update) {
      const int row = coeff_idx >> txb_info->bwl;
      const int col = coeff_idx - (row << txb_info->bwl);
      const int ctx = txb_cache->base_ctx_arr[base_idx][coeff_idx][0];
      const int org_cost =
          get_base_cost(abs_qc, ctx, txb_probs->coeff_base, base_idx);

      const int new_ctx =
          get_base_ctx_from_count_mag(row, col, new_count, new_mag, level);
      const int new_cost =
          get_base_cost(abs_qc, new_ctx, txb_probs->coeff_base, base_idx);
      cost_diff += -org_cost + new_cost;
    }
  }
  return cost_diff;
}

static int try_neighbor_level_down_nz(int coeff_idx, int nb_coeff_idx,
                                      const TxbCache *txb_cache,
                                      const TxbProbs *txb_probs,
                                      TxbInfo *txb_info) {
  // assume eob doesn't change
  const tran_low_t qc = txb_info->qcoeff[coeff_idx];
  const tran_low_t abs_qc = abs(qc);
  const tran_low_t nb_coeff = txb_info->qcoeff[nb_coeff_idx];
  const tran_low_t abs_nb_coeff = abs(nb_coeff);
  if (abs_nb_coeff != 1) return 0;
  const int16_t *iscan = txb_info->scan_order->iscan;
  const int scan_idx = iscan[coeff_idx];
  if (scan_idx == txb_info->seg_eob) return 0;
  const int nb_scan_idx = iscan[nb_coeff_idx];
  if (nb_scan_idx < scan_idx) {
    const int count = txb_cache->nz_count_arr[coeff_idx];
    assert(count > 0);
    txb_info->qcoeff[nb_coeff_idx] = get_lower_coeff(nb_coeff);
    const int new_ctx = get_nz_map_ctx_from_count(
        count - 1, txb_info->qcoeff, coeff_idx, txb_info->bwl, iscan);
    txb_info->qcoeff[nb_coeff_idx] = nb_coeff;
    const int ctx = txb_cache->nz_ctx_arr[coeff_idx][0];
    const int is_nz = abs_qc > 0;
    const int org_cost = av1_cost_bit(txb_probs->nz_map[ctx], is_nz);
    const int new_cost = av1_cost_bit(txb_probs->nz_map[new_ctx], is_nz);
    const int cost_diff = new_cost - org_cost;
    return cost_diff;
  } else {
    return 0;
  }
}

static int try_self_level_down(tran_low_t *low_coeff, int coeff_idx,
                               const TxbCache *txb_cache,
                               const TxbProbs *txb_probs, TxbInfo *txb_info) {
767 768 769 770 771 772 773 774 775 776
  const tran_low_t qc = txb_info->qcoeff[coeff_idx];
  if (qc == 0) {
    *low_coeff = 0;
    return 0;
  }
  const tran_low_t abs_qc = abs(qc);
  *low_coeff = get_lower_coeff(qc);
  int cost_diff;
  if (*low_coeff == 0) {
    const int scan_idx = txb_info->scan_order->iscan[coeff_idx];
Angie Chiang's avatar
Angie Chiang committed
777
    const aom_prob level_prob =
778
        get_level_prob(abs_qc, coeff_idx, txb_cache, txb_probs);
Angie Chiang's avatar
Angie Chiang committed
779
    const aom_prob low_level_prob =
780 781 782 783 784 785 786 787 788 789 790 791 792 793 794
        get_level_prob(abs(*low_coeff), coeff_idx, txb_cache, txb_probs);
    if (scan_idx < txb_info->seg_eob) {
      // When level-0, we code the binary of abs_qc > level
      // but when level-k k > 0 we code the binary of abs_qc == level
      // That's why wee need this special treatment for level-0 map
      // TODO(angiebird): make leve-0 consistent to other levels
      cost_diff = -av1_cost_bit(level_prob, 1) +
                  av1_cost_bit(low_level_prob, 0) -
                  av1_cost_bit(low_level_prob, 1);
    } else {
      cost_diff = -av1_cost_bit(level_prob, 1);
    }

    if (scan_idx < txb_info->seg_eob) {
      const int eob_ctx =
795
          get_eob_ctx(txb_info->qcoeff, coeff_idx, txb_info->txs_ctx);
796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822
      cost_diff -= av1_cost_bit(txb_probs->eob_flag[eob_ctx],
                                scan_idx == (txb_info->eob - 1));
    }

    const int sign_cost = get_sign_bit_cost(
        qc, coeff_idx, txb_probs->dc_sign_prob, txb_info->txb_ctx->dc_sign_ctx);
    cost_diff -= sign_cost;
  } else if (abs_qc < 1 + NUM_BASE_LEVELS + COEFF_BASE_RANGE) {
    const aom_prob level_prob =
        get_level_prob(abs_qc, coeff_idx, txb_cache, txb_probs);
    const aom_prob low_level_prob =
        get_level_prob(abs(*low_coeff), coeff_idx, txb_cache, txb_probs);
    cost_diff = -av1_cost_bit(level_prob, 1) + av1_cost_bit(low_level_prob, 1) -
                av1_cost_bit(low_level_prob, 0);
  } else if (abs_qc == 1 + NUM_BASE_LEVELS + COEFF_BASE_RANGE) {
    const aom_prob low_level_prob =
        get_level_prob(abs(*low_coeff), coeff_idx, txb_cache, txb_probs);
    cost_diff = -get_golomb_cost(abs_qc) + av1_cost_bit(low_level_prob, 1) -
                av1_cost_bit(low_level_prob, 0);
  } else {
    assert(abs_qc > 1 + NUM_BASE_LEVELS + COEFF_BASE_RANGE);
    const tran_low_t abs_low_coeff = abs(*low_coeff);
    cost_diff = -get_golomb_cost(abs_qc) + get_golomb_cost(abs_low_coeff);
  }
  return cost_diff;
}

Angie Chiang's avatar
Angie Chiang committed
823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866
#define COST_MAP_SIZE 5
#define COST_MAP_OFFSET 2

static INLINE int check_nz_neighbor(tran_low_t qc) { return abs(qc) == 1; }

static INLINE int check_base_neighbor(tran_low_t qc) {
  return abs(qc) <= 1 + NUM_BASE_LEVELS;
}

static INLINE int check_br_neighbor(tran_low_t qc) {
  return abs(qc) > BR_MAG_OFFSET;
}

// TODO(angiebird): add static to this function once it's called
int try_level_down(int coeff_idx, const TxbCache *txb_cache,
                   const TxbProbs *txb_probs, TxbInfo *txb_info,
                   int (*cost_map)[COST_MAP_SIZE]) {
  if (cost_map) {
    for (int i = 0; i < COST_MAP_SIZE; ++i) av1_zero(cost_map[i]);
  }

  tran_low_t qc = txb_info->qcoeff[coeff_idx];
  tran_low_t low_coeff;
  if (qc == 0) return 0;
  int accu_cost_diff = 0;

  const int16_t *iscan = txb_info->scan_order->iscan;
  const int eob = txb_info->eob;
  const int scan_idx = iscan[coeff_idx];
  if (scan_idx < eob) {
    const int cost_diff = try_self_level_down(&low_coeff, coeff_idx, txb_cache,
                                              txb_probs, txb_info);
    if (cost_map)
      cost_map[0 + COST_MAP_OFFSET][0 + COST_MAP_OFFSET] = cost_diff;
    accu_cost_diff += cost_diff;
  }

  const int row = coeff_idx >> txb_info->bwl;
  const int col = coeff_idx - (row << txb_info->bwl);
  if (check_nz_neighbor(qc)) {
    for (int i = 0; i < SIG_REF_OFFSET_NUM; ++i) {
      const int nb_row = row - sig_ref_offset[i][0];
      const int nb_col = col - sig_ref_offset[i][1];
      const int nb_coeff_idx = nb_row * txb_info->stride + nb_col;
867

868 869 870 871
      if (!(nb_row >= 0 && nb_col >= 0 && nb_row < txb_info->height &&
            nb_col < txb_info->stride))
        continue;

Angie Chiang's avatar
Angie Chiang committed
872
      const int nb_scan_idx = iscan[nb_coeff_idx];
873
      if (nb_scan_idx < eob) {
Angie Chiang's avatar
Angie Chiang committed
874 875 876 877 878 879 880 881 882 883 884 885 886 887 888
        const int cost_diff = try_neighbor_level_down_nz(
            nb_coeff_idx, coeff_idx, txb_cache, txb_probs, txb_info);
        if (cost_map)
          cost_map[nb_row - row + COST_MAP_OFFSET]
                  [nb_col - col + COST_MAP_OFFSET] += cost_diff;
        accu_cost_diff += cost_diff;
      }
    }
  }

  if (check_base_neighbor(qc)) {
    for (int i = 0; i < BASE_CONTEXT_POSITION_NUM; ++i) {
      const int nb_row = row - base_ref_offset[i][0];
      const int nb_col = col - base_ref_offset[i][1];
      const int nb_coeff_idx = nb_row * txb_info->stride + nb_col;
889 890 891 892 893

      if (!(nb_row >= 0 && nb_col >= 0 && nb_row < txb_info->height &&
            nb_col < txb_info->stride))
        continue;

Angie Chiang's avatar
Angie Chiang committed
894
      const int nb_scan_idx = iscan[nb_coeff_idx];
895
      if (nb_scan_idx < eob) {
Angie Chiang's avatar
Angie Chiang committed
896 897 898 899 900 901 902 903 904 905 906 907 908 909 910
        const int cost_diff = try_neighbor_level_down_base(
            nb_coeff_idx, coeff_idx, txb_cache, txb_probs, txb_info);
        if (cost_map)
          cost_map[nb_row - row + COST_MAP_OFFSET]
                  [nb_col - col + COST_MAP_OFFSET] += cost_diff;
        accu_cost_diff += cost_diff;
      }
    }
  }

  if (check_br_neighbor(qc)) {
    for (int i = 0; i < BR_CONTEXT_POSITION_NUM; ++i) {
      const int nb_row = row - br_ref_offset[i][0];
      const int nb_col = col - br_ref_offset[i][1];
      const int nb_coeff_idx = nb_row * txb_info->stride + nb_col;
911 912 913 914 915

      if (!(nb_row >= 0 && nb_col >= 0 && nb_row < txb_info->height &&
            nb_col < txb_info->stride))
        continue;

Angie Chiang's avatar
Angie Chiang committed
916
      const int nb_scan_idx = iscan[nb_coeff_idx];
917
      if (nb_scan_idx < eob) {
Angie Chiang's avatar
Angie Chiang committed
918 919 920 921 922 923 924 925 926 927 928 929 930
        const int cost_diff = try_neighbor_level_down_br(
            nb_coeff_idx, coeff_idx, txb_cache, txb_probs, txb_info);
        if (cost_map)
          cost_map[nb_row - row + COST_MAP_OFFSET]
                  [nb_col - col + COST_MAP_OFFSET] += cost_diff;
        accu_cost_diff += cost_diff;
      }
    }
  }

  return accu_cost_diff;
}

Angie Chiang's avatar
Angie Chiang committed
931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950
static int get_low_coeff_cost(int coeff_idx, const TxbCache *txb_cache,
                              const TxbProbs *txb_probs,
                              const TxbInfo *txb_info) {
  const tran_low_t qc = txb_info->qcoeff[coeff_idx];
  const int abs_qc = abs(qc);
  assert(abs_qc <= 1);
  int cost = 0;
  const int scan_idx = txb_info->scan_order->iscan[coeff_idx];
  if (scan_idx < txb_info->seg_eob) {
    const aom_prob level_prob =
        get_level_prob(0, coeff_idx, txb_cache, txb_probs);
    cost += av1_cost_bit(level_prob, qc != 0);
  }

  if (qc != 0) {
    const int base_idx = 0;
    const int ctx = txb_cache->base_ctx_arr[base_idx][coeff_idx][0];
    cost += get_base_cost(abs_qc, ctx, txb_probs->coeff_base, base_idx);
    if (scan_idx < txb_info->seg_eob) {
      const int eob_ctx =
951
          get_eob_ctx(txb_info->qcoeff, coeff_idx, txb_info->txs_ctx);
Angie Chiang's avatar
Angie Chiang committed
952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982
      cost += av1_cost_bit(txb_probs->eob_flag[eob_ctx],
                           scan_idx == (txb_info->eob - 1));
    }
    cost += get_sign_bit_cost(qc, coeff_idx, txb_probs->dc_sign_prob,
                              txb_info->txb_ctx->dc_sign_ctx);
  }
  return cost;
}

static INLINE void set_eob(TxbInfo *txb_info, int eob) {
  txb_info->eob = eob;
  txb_info->seg_eob = AOMMIN(eob, tx_size_2d[txb_info->tx_size] - 1);
}

// TODO(angiebird): add static to this function once it's called
int try_change_eob(int *new_eob, int coeff_idx, const TxbCache *txb_cache,
                   const TxbProbs *txb_probs, TxbInfo *txb_info) {
  assert(txb_info->eob > 0);
  const tran_low_t qc = txb_info->qcoeff[coeff_idx];
  const int abs_qc = abs(qc);
  if (abs_qc != 1) {
    *new_eob = -1;
    return 0;
  }
  const int16_t *iscan = txb_info->scan_order->iscan;
  const int16_t *scan = txb_info->scan_order->scan;
  const int scan_idx = iscan[coeff_idx];
  *new_eob = 0;
  int cost_diff = 0;
  cost_diff -= get_low_coeff_cost(coeff_idx, txb_cache, txb_probs, txb_info);
  // int coeff_cost =
983
  //     get_coeff_cost(qc, scan_idx, txb_info, txb_probs);
Angie Chiang's avatar
Angie Chiang committed
984 985 986
  // if (-cost_diff != coeff_cost) {
  //   printf("-cost_diff %d coeff_cost %d\n", -cost_diff, coeff_cost);
  //   get_low_coeff_cost(coeff_idx, txb_cache, txb_probs, txb_info);
987
  //   get_coeff_cost(qc, scan_idx, txb_info, txb_probs);
Angie Chiang's avatar
Angie Chiang committed
988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007
  // }
  for (int si = scan_idx - 1; si >= 0; --si) {
    const int ci = scan[si];
    if (txb_info->qcoeff[ci] != 0) {
      *new_eob = si + 1;
      break;
    } else {
      cost_diff -= get_low_coeff_cost(ci, txb_cache, txb_probs, txb_info);
    }
  }

  const int org_eob = txb_info->eob;
  set_eob(txb_info, *new_eob);
  cost_diff += try_level_down(coeff_idx, txb_cache, txb_probs, txb_info, NULL);
  set_eob(txb_info, org_eob);

  if (*new_eob > 0) {
    // Note that get_eob_ctx does NOT actually account for qcoeff, so we don't
    // need to lower down the qcoeff here
    const int eob_ctx =
1008
        get_eob_ctx(txb_info->qcoeff, scan[*new_eob - 1], txb_info->txs_ctx);
Angie Chiang's avatar
Angie Chiang committed
1009 1010 1011 1012 1013 1014 1015 1016 1017
    cost_diff -= av1_cost_bit(txb_probs->eob_flag[eob_ctx], 0);
    cost_diff += av1_cost_bit(txb_probs->eob_flag[eob_ctx], 1);
  } else {
    const int txb_skip_ctx = txb_info->txb_ctx->txb_skip_ctx;
    cost_diff -= av1_cost_bit(txb_probs->txb_skip[txb_skip_ctx], 0);
    cost_diff += av1_cost_bit(txb_probs->txb_skip[txb_skip_ctx], 1);
  }
  return cost_diff;
}
Angie Chiang's avatar
Angie Chiang committed
1018 1019 1020 1021 1022 1023

static INLINE tran_low_t qcoeff_to_dqcoeff(tran_low_t qc, int dqv, int shift) {
  int sgn = qc < 0 ? -1 : 1;
  return sgn * ((abs(qc) * dqv) >> shift);
}

Angie Chiang's avatar
Angie Chiang committed
1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041
// TODO(angiebird): add static to this function it's called
void update_level_down(int coeff_idx, TxbCache *txb_cache, TxbInfo *txb_info) {
  const tran_low_t qc = txb_info->qcoeff[coeff_idx];
  const int abs_qc = abs(qc);
  if (qc == 0) return;
  const tran_low_t low_coeff = get_lower_coeff(qc);
  txb_info->qcoeff[coeff_idx] = low_coeff;
  const int dqv = txb_info->dequant[coeff_idx != 0];
  txb_info->dqcoeff[coeff_idx] =
      qcoeff_to_dqcoeff(low_coeff, dqv, txb_info->shift);

  const int row = coeff_idx >> txb_info->bwl;
  const int col = coeff_idx - (row << txb_info->bwl);
  const int eob = txb_info->eob;
  const int16_t *iscan = txb_info->scan_order->iscan;
  for (int i = 0; i < SIG_REF_OFFSET_NUM; ++i) {
    const int nb_row = row - sig_ref_offset[i][0];
    const int nb_col = col - sig_ref_offset[i][1];
1042 1043 1044 1045 1046

    if (!(nb_row >= 0 && nb_col >= 0 && nb_row < txb_info->height &&
          nb_col < txb_info->stride))
      continue;

Angie Chiang's avatar
Angie Chiang committed
1047 1048
    const int nb_coeff_idx = nb_row * txb_info->stride + nb_col;
    const int nb_scan_idx = iscan[nb_coeff_idx];
1049
    if (nb_scan_idx < eob) {
Angie Chiang's avatar
Angie Chiang committed
1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072
      const int scan_idx = iscan[coeff_idx];
      if (scan_idx < nb_scan_idx) {
        const int level = 1;
        if (abs_qc == level) {
          txb_cache->nz_count_arr[nb_coeff_idx] -= 1;
          assert(txb_cache->nz_count_arr[nb_coeff_idx] >= 0);
        }
        const int count = txb_cache->nz_count_arr[nb_coeff_idx];
        txb_cache->nz_ctx_arr[nb_coeff_idx][0] = get_nz_map_ctx_from_count(
            count, txb_info->qcoeff, nb_coeff_idx, txb_info->bwl, iscan);
        // int ref_ctx = get_nz_map_ctx2(txb_info->qcoeff, nb_coeff_idx,
        // txb_info->bwl, iscan);
        // if (ref_ctx != txb_cache->nz_ctx_arr[nb_coeff_idx][0])
        //   printf("nz ctx %d ref_ctx %d\n",
        //   txb_cache->nz_ctx_arr[nb_coeff_idx][0], ref_ctx);
      }
    }
  }

  for (int i = 0; i < BASE_CONTEXT_POSITION_NUM; ++i) {
    const int nb_row = row - base_ref_offset[i][0];
    const int nb_col = col - base_ref_offset[i][1];
    const int nb_coeff_idx = nb_row * txb_info->stride + nb_col;
1073 1074 1075 1076 1077

    if (!(nb_row >= 0 && nb_col >= 0 && nb_row < txb_info->height &&
          nb_col < txb_info->stride))
      continue;

Angie Chiang's avatar
Angie Chiang committed
1078 1079 1080
    const tran_low_t nb_coeff = txb_info->qcoeff[nb_coeff_idx];
    if (!has_base(nb_coeff, 0)) continue;
    const int nb_scan_idx = iscan[nb_coeff_idx];
1081
    if (nb_scan_idx < eob) {
Angie Chiang's avatar
Angie Chiang committed
1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109
      if (row >= nb_row && col >= nb_col)
        update_mag_arr(txb_cache->base_mag_arr[nb_coeff_idx], abs_qc);
      const int mag =
          get_mag_from_mag_arr(txb_cache->base_mag_arr[nb_coeff_idx]);
      for (int base_idx = 0; base_idx < NUM_BASE_LEVELS; ++base_idx) {
        if (!has_base(nb_coeff, base_idx)) continue;
        const int level = base_idx + 1;
        if (abs_qc == level) {
          txb_cache->base_count_arr[base_idx][nb_coeff_idx] -= 1;
          assert(txb_cache->base_count_arr[base_idx][nb_coeff_idx] >= 0);
        }
        const int count = txb_cache->base_count_arr[base_idx][nb_coeff_idx];
        txb_cache->base_ctx_arr[base_idx][nb_coeff_idx][0] =
            get_base_ctx_from_count_mag(nb_row, nb_col, count, mag, level);
        // int ref_ctx = get_base_ctx(txb_info->qcoeff, nb_coeff_idx,
        // txb_info->bwl, level);
        // if (ref_ctx != txb_cache->base_ctx_arr[base_idx][nb_coeff_idx][0]) {
        //   printf("base ctx %d ref_ctx %d\n",
        //   txb_cache->base_ctx_arr[base_idx][nb_coeff_idx][0], ref_ctx);
        // }
      }
    }
  }

  for (int i = 0; i < BR_CONTEXT_POSITION_NUM; ++i) {
    const int nb_row = row - br_ref_offset[i][0];
    const int nb_col = col - br_ref_offset[i][1];
    const int nb_coeff_idx = nb_row * txb_info->stride + nb_col;
1110 1111 1112 1113 1114

    if (!(nb_row >= 0 && nb_col >= 0 && nb_row < txb_info->height &&
          nb_col < txb_info->stride))
      continue;

Angie Chiang's avatar
Angie Chiang committed
1115 1116 1117
    const int nb_scan_idx = iscan[nb_coeff_idx];
    const tran_low_t nb_coeff = txb_info->qcoeff[nb_coeff_idx];
    if (!has_br(nb_coeff)) continue;
1118
    if (nb_scan_idx < eob) {
Angie Chiang's avatar
Angie Chiang committed
1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139
      const int level = 1 + NUM_BASE_LEVELS;
      if (abs_qc == level) {
        txb_cache->br_count_arr[nb_coeff_idx] -= 1;
        assert(txb_cache->br_count_arr[nb_coeff_idx] >= 0);
      }
      if (row >= nb_row && col >= nb_col)
        update_mag_arr(txb_cache->br_mag_arr[nb_coeff_idx], abs_qc);
      const int count = txb_cache->br_count_arr[nb_coeff_idx];
      const int mag = get_mag_from_mag_arr(txb_cache->br_mag_arr[nb_coeff_idx]);
      txb_cache->br_ctx_arr[nb_coeff_idx][0] =
          get_br_ctx_from_count_mag(nb_row, nb_col, count, mag);
      // int ref_ctx = get_level_ctx(txb_info->qcoeff, nb_coeff_idx,
      // txb_info->bwl);
      // if (ref_ctx != txb_cache->br_ctx_arr[nb_coeff_idx][0]) {
      //   printf("base ctx %d ref_ctx %d\n",
      //   txb_cache->br_ctx_arr[nb_coeff_idx][0], ref_ctx);
      // }
    }
  }
}

1140
static int get_coeff_cost(tran_low_t qc, int scan_idx, TxbInfo *txb_info,
1141
                          const TxbProbs *txb_probs) {
1142 1143 1144 1145 1146 1147 1148 1149
  const TXB_CTX *txb_ctx = txb_info->txb_ctx;
  const int is_nz = (qc != 0);
  const tran_low_t abs_qc = abs(qc);
  int cost = 0;
  const int16_t *scan = txb_info->scan_order->scan;
  const int16_t *iscan = txb_info->scan_order->iscan;

  if (scan_idx < txb_info->seg_eob) {
1150 1151
    int coeff_ctx = get_nz_map_ctx2(txb_info->qcoeff, scan[scan_idx],
                                    txb_info->bwl, txb_info->height, iscan);
1152 1153 1154 1155 1156 1157 1158 1159
    cost += av1_cost_bit(txb_probs->nz_map[coeff_ctx], is_nz);
  }

  if (is_nz) {
    cost += get_sign_bit_cost(qc, scan_idx, txb_probs->dc_sign_prob,
                              txb_ctx->dc_sign_ctx);

    int ctx_ls[NUM_BASE_LEVELS] = { 0 };
1160 1161
    get_base_ctx_set(txb_info->qcoeff, scan[scan_idx], txb_info->bwl,
                     txb_info->height, ctx_ls);
1162 1163 1164 1165 1166 1167 1168

    int i;
    for (i = 0; i < NUM_BASE_LEVELS; ++i) {
      cost += get_base_cost(abs_qc, ctx_ls[i], txb_probs->coeff_base, i);
    }

    if (abs_qc > NUM_BASE_LEVELS) {
1169 1170
      int ctx = get_br_ctx(txb_info->qcoeff, scan[scan_idx], txb_info->bwl,
                           txb_info->height);
1171 1172 1173 1174 1175 1176
      cost += get_br_cost(abs_qc, ctx, txb_probs->coeff_lps);
      cost += get_golomb_cost(abs_qc);
    }

    if (scan_idx < txb_info->seg_eob) {
      int eob_ctx =
1177
          get_eob_ctx(txb_info->qcoeff, scan[scan_idx], txb_info->txs_ctx);
1178 1179 1180 1181 1182 1183 1184
      cost += av1_cost_bit(txb_probs->eob_flag[eob_ctx],
                           scan_idx == (txb_info->eob - 1));
    }
  }
  return cost;
}

Angie Chiang's avatar
Angie Chiang committed
1185
#if TEST_OPTIMIZE_TXB
1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241
#define ALL_REF_OFFSET_NUM 17
static int all_ref_offset[ALL_REF_OFFSET_NUM][2] = {
  { 0, 0 },  { -2, -1 }, { -2, 0 }, { -2, 1 }, { -1, -2 }, { -1, -1 },
  { -1, 0 }, { -1, 1 },  { 0, -2 }, { 0, -1 }, { 1, -2 },  { 1, -1 },
  { 1, 0 },  { 2, 0 },   { 0, 1 },  { 0, 2 },  { 1, 1 },
};

static int try_level_down_ref(int coeff_idx, const TxbProbs *txb_probs,
                              TxbInfo *txb_info,
                              int (*cost_map)[COST_MAP_SIZE]) {
  if (cost_map) {
    for (int i = 0; i < COST_MAP_SIZE; ++i) av1_zero(cost_map[i]);
  }
  tran_low_t qc = txb_info->qcoeff[coeff_idx];
  if (qc == 0) return 0;
  int row = coeff_idx >> txb_info->bwl;
  int col = coeff_idx - (row << txb_info->bwl);
  int org_cost = 0;
  for (int i = 0; i < ALL_REF_OFFSET_NUM; ++i) {
    int nb_row = row - all_ref_offset[i][0];
    int nb_col = col - all_ref_offset[i][1];
    int nb_coeff_idx = nb_row * txb_info->stride + nb_col;
    int nb_scan_idx = txb_info->scan_order->iscan[nb_coeff_idx];
    if (nb_scan_idx < txb_info->eob && nb_row >= 0 && nb_col >= 0 &&
        nb_row < txb_info->stride && nb_col < txb_info->stride) {
      tran_low_t nb_coeff = txb_info->qcoeff[nb_coeff_idx];
      int cost = get_coeff_cost(nb_coeff, nb_scan_idx, txb_info, txb_probs);
      if (cost_map)
        cost_map[nb_row - row + COST_MAP_OFFSET]
                [nb_col - col + COST_MAP_OFFSET] -= cost;
      org_cost += cost;
    }
  }
  txb_info->qcoeff[coeff_idx] = get_lower_coeff(qc);
  int new_cost = 0;
  for (int i = 0; i < ALL_REF_OFFSET_NUM; ++i) {
    int nb_row = row - all_ref_offset[i][0];
    int nb_col = col - all_ref_offset[i][1];
    int nb_coeff_idx = nb_row * txb_info->stride + nb_col;
    int nb_scan_idx = txb_info->scan_order->iscan[nb_coeff_idx];
    if (nb_scan_idx < txb_info->eob && nb_row >= 0 && nb_col >= 0 &&
        nb_row < txb_info->stride && nb_col < txb_info->stride) {
      tran_low_t nb_coeff = txb_info->qcoeff[nb_coeff_idx];
      int cost = get_coeff_cost(nb_coeff, nb_scan_idx, txb_info, txb_probs);
      if (cost_map)
        cost_map[nb_row - row + COST_MAP_OFFSET]
                [nb_col - col + COST_MAP_OFFSET] += cost;
      new_cost += cost;
    }
  }
  txb_info->qcoeff[coeff_idx] = qc;
  return new_cost - org_cost;
}

static void test_level_down(int coeff_idx, const TxbCache *txb_cache,
                            const TxbProbs *txb_probs, TxbInfo *txb_info) {
Angie Chiang's avatar
Angie Chiang committed
1242 1243 1244 1245
  int cost_map[COST_MAP_SIZE][COST_MAP_SIZE];
  int ref_cost_map[COST_MAP_SIZE][COST_MAP_SIZE];
  const int cost_diff =
      try_level_down(coeff_idx, txb_cache, txb_probs, txb_info, cost_map);
1246 1247
  const int cost_diff_ref =
      try_level_down_ref(coeff_idx, txb_probs, txb_info, ref_cost_map);
Angie Chiang's avatar
Angie Chiang committed
1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260
  if (cost_diff != cost_diff_ref) {
    printf("qc %d cost_diff %d cost_diff_ref %d\n", txb_info->qcoeff[coeff_idx],
           cost_diff, cost_diff_ref);
    for (int r = 0; r < COST_MAP_SIZE; ++r) {
      for (int c = 0; c < COST_MAP_SIZE; ++c) {
        printf("%d:%d ", cost_map[r][c], ref_cost_map[r][c]);
      }
      printf("\n");
    }
  }
}
#endif

1261
// TODO(angiebird): make this static once it's called
1262
int get_txb_cost(TxbInfo *txb_info, const TxbProbs *txb_probs) {
1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278
  int cost = 0;
  int txb_skip_ctx = txb_info->txb_ctx->txb_skip_ctx;
  const int16_t *scan = txb_info->scan_order->scan;
  if (txb_info->eob == 0) {
    cost = av1_cost_bit(txb_probs->txb_skip[txb_skip_ctx], 1);
    return cost;
  }
  cost = av1_cost_bit(txb_probs->txb_skip[txb_skip_ctx], 0);
  for (int c = 0; c < txb_info->eob; ++c) {
    tran_low_t qc = txb_info->qcoeff[scan[c]];
    int coeff_cost = get_coeff_cost(qc, c, txb_info, txb_probs);
    cost += coeff_cost;
  }
  return cost;
}

1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309
#if TEST_OPTIMIZE_TXB
void test_try_change_eob(TxbInfo *txb_info, TxbProbs *txb_probs,
                         TxbCache *txb_cache) {
  int eob = txb_info->eob;
  const int16_t *scan = txb_info->scan_order->scan;
  if (eob > 0) {
    int last_si = eob - 1;
    int last_ci = scan[last_si];
    int last_coeff = txb_info->qcoeff[last_ci];
    if (abs(last_coeff) == 1) {
      int new_eob;
      int cost_diff =
          try_change_eob(&new_eob, last_ci, txb_cache, txb_probs, txb_info);
      int org_eob = txb_info->eob;
      int cost = get_txb_cost(txb_info, txb_probs);

      txb_info->qcoeff[last_ci] = get_lower_coeff(last_coeff);
      set_eob(txb_info, new_eob);
      int new_cost = get_txb_cost(txb_info, txb_probs);
      set_eob(txb_info, org_eob);
      txb_info->qcoeff[last_ci] = last_coeff;

      int ref_cost_diff = -cost + new_cost;
      if (cost_diff != ref_cost_diff)
        printf("org_eob %d new_eob %d cost_diff %d ref_cost_diff %d\n", org_eob,
               new_eob, cost_diff, ref_cost_diff);
    }
  }
}
#endif

Angie Chiang's avatar
Angie Chiang committed
1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367
static INLINE int64_t get_coeff_dist(tran_low_t tcoeff, tran_low_t dqcoeff,
                                     int shift) {
  const int64_t diff = (tcoeff - dqcoeff) * (1 << shift);
  const int64_t error = diff * diff;
  return error;
}

typedef struct LevelDownStats {
  int update;
  tran_low_t low_qc;
  tran_low_t low_dqc;
  int64_t rd_diff;
  int cost_diff;
  int64_t dist_diff;
  int new_eob;
} LevelDownStats;

void try_level_down_facade(LevelDownStats *stats, int scan_idx,
                           const TxbCache *txb_cache, const TxbProbs *txb_probs,
                           TxbInfo *txb_info) {
  const int16_t *scan = txb_info->scan_order->scan;
  const int coeff_idx = scan[scan_idx];
  const tran_low_t qc = txb_info->qcoeff[coeff_idx];
  stats->new_eob = -1;
  stats->update = 0;
  if (qc == 0) {
    return;
  }

  const tran_low_t tqc = txb_info->tcoeff[coeff_idx];
  const int dqv = txb_info->dequant[coeff_idx != 0];

  const tran_low_t dqc = qcoeff_to_dqcoeff(qc, dqv, txb_info->shift);
  const int64_t dqc_dist = get_coeff_dist(tqc, dqc, txb_info->shift);

  stats->low_qc = get_lower_coeff(qc);
  stats->low_dqc = qcoeff_to_dqcoeff(stats->low_qc, dqv, txb_info->shift);
  const int64_t low_dqc_dist =
      get_coeff_dist(tqc, stats->low_dqc, txb_info->shift);

  stats->dist_diff = -dqc_dist + low_dqc_dist;
  stats->cost_diff = 0;
  stats->new_eob = txb_info->eob;
  if (scan_idx == txb_info->eob - 1 && abs(qc) == 1) {
    stats->cost_diff = try_change_eob(&stats->new_eob, coeff_idx, txb_cache,
                                      txb_probs, txb_info);
  } else {
    stats->cost_diff =
        try_level_down(coeff_idx, txb_cache, txb_probs, txb_info, NULL);
#if TEST_OPTIMIZE_TXB
    test_level_down(coeff_idx, txb_cache, txb_probs, txb_info);
#endif
  }
  stats->rd_diff = RDCOST(txb_info->rdmult, txb_info->rddiv, stats->cost_diff,
                          stats->dist_diff);
  if (stats->rd_diff < 0) stats->update = 1;
  return;
}
1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382

static int optimize_txb(TxbInfo *txb_info, const TxbProbs *txb_probs,
                        TxbCache *txb_cache, int dry_run) {
  int update = 0;
  if (txb_info->eob == 0) return update;
  int cost_diff = 0;
  int64_t dist_diff = 0;
  int64_t rd_diff = 0;
  const int max_eob = tx_size_2d[txb_info->tx_size];

#if TEST_OPTIMIZE_TXB
  int64_t sse;
  int64_t org_dist =
      av1_block_error_c(txb_info->tcoeff, txb_info->dqcoeff, max_eob, &sse) *
      (1 << (2 * txb_info->shift));
1383
  int org_cost = get_txb_cost(txb_info, txb_probs);
1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443
#endif

  tran_low_t *org_qcoeff = txb_info->qcoeff;
  tran_low_t *org_dqcoeff = txb_info->dqcoeff;

  tran_low_t tmp_qcoeff[MAX_TX_SQUARE];
  tran_low_t tmp_dqcoeff[MAX_TX_SQUARE];
  const int org_eob = txb_info->eob;
  if (dry_run) {
    memcpy(tmp_qcoeff, org_qcoeff, sizeof(org_qcoeff[0]) * max_eob);
    memcpy(tmp_dqcoeff, org_dqcoeff, sizeof(org_dqcoeff[0]) * max_eob);
    txb_info->qcoeff = tmp_qcoeff;
    txb_info->dqcoeff = tmp_dqcoeff;
  }

  const int16_t *scan = txb_info->scan_order->scan;

  // forward optimize the nz_map
  const int cur_eob = txb_info->eob;
  for (int si = 0; si < cur_eob; ++si) {
    const int coeff_idx = scan[si];
    tran_low_t qc = txb_info->qcoeff[coeff_idx];
    if (abs(qc) == 1) {
      LevelDownStats stats;
      try_level_down_facade(&stats, si, txb_cache, txb_probs, txb_info);
      if (stats.update) {
        update = 1;
        cost_diff += stats.cost_diff;
        dist_diff += stats.dist_diff;
        rd_diff += stats.rd_diff;
        update_level_down(coeff_idx, txb_cache, txb_info);
        set_eob(txb_info, stats.new_eob);
      }
    }
  }

  // backward optimize the level-k map
  for (int si = txb_info->eob - 1; si >= 0; --si) {
    LevelDownStats stats;
    try_level_down_facade(&stats, si, txb_cache, txb_probs, txb_info);
    const int coeff_idx = scan[si];
    if (stats.update) {
#if TEST_OPTIMIZE_TXB
// printf("si %d low_qc %d cost_diff %d dist_diff %ld rd_diff %ld eob %d new_eob
// %d\n", si, stats.low_qc, stats.cost_diff, stats.dist_diff, stats.rd_diff,
// txb_info->eob, stats.new_eob);
#endif
      update = 1;
      cost_diff += stats.cost_diff;
      dist_diff += stats.dist_diff;
      rd_diff += stats.rd_diff;
      update_level_down(coeff_idx, txb_cache, txb_info);
      set_eob(txb_info, stats.new_eob);
    }
    if (si > txb_info->eob) si = txb_info->eob;
  }
#if TEST_OPTIMIZE_TXB
  int64_t new_dist =
      av1_block_error_c(txb_info->tcoeff, txb_info->dqcoeff, max_eob, &sse) *
      (1 << (2 * txb_info->shift));
1444
  int new_cost = get_txb_cost(txb_info, txb_probs);
1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474
  int64_t ref_dist_diff = new_dist - org_dist;
  int ref_cost_diff = new_cost - org_cost;
  if (cost_diff != ref_cost_diff || dist_diff != ref_dist_diff)
    printf(
        "overall rd_diff %ld\ncost_diff %d ref_cost_diff%d\ndist_diff %ld "
        "ref_dist_diff %ld\neob %d new_eob %d\n\n",
        rd_diff, cost_diff, ref_cost_diff, dist_diff, ref_dist_diff, org_eob,
        txb_info->eob);
#endif
  if (dry_run) {
    txb_info->qcoeff = org_qcoeff;
    txb_info->dqcoeff = org_dqcoeff;
    set_eob(txb_info, org_eob);
  }
  return update;
}

// These numbers are empirically obtained.
static const int plane_rd_mult[REF_TYPES][PLANE_TYPES] = {
#if CONFIG_EC_ADAPT
  { 17, 13 }, { 16, 10 },
#else
  { 20, 12 }, { 16, 12 },
#endif
};

int av1_optimize_txb(const AV1_COMMON *cm, MACROBLOCK *x, int plane, int block,
                     TX_SIZE tx_size, TXB_CTX *txb_ctx) {
  MACROBLOCKD *const xd = &x->e_mbd;
  const PLANE_TYPE plane_type = get_plane_type(plane);
1475
  const TX_SIZE txs_ctx = get_txsize_context(tx_size);
1476 1477 1478 1479 1480 1481 1482 1483 1484 1485
  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
  const MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
  const struct macroblock_plane *p = &x->plane[plane];
  struct macroblockd_plane *pd = &xd->plane[plane];
  const int eob = p->eobs[block];
  tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
  tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
  const tran_low_t *tcoeff = BLOCK_OFFSET(p->coeff, block);
  const int16_t *dequant = pd->dequant;
  const int seg_eob = AOMMIN(eob, tx_size_2d[tx_size] - 1);
1486
  const aom_prob *nz_map = xd->fc->nz_map[txs_ctx][plane_type];
1487 1488 1489

  const int bwl = b_width_log2_lookup[txsize_to_bsize[tx_size]] + 2;
  const int stride = 1 << bwl;
1490
  const int height = tx_size_high[tx_size];
1491
  aom_prob(*coeff_base)[COEFF_BASE_CONTEXTS] =
1492
      xd->fc->coeff_base[txs_ctx][plane_type];
1493

1494
  const aom_prob *coeff_lps = xd->fc->coeff_lps[txs_ctx][plane_type];
1495 1496 1497 1498 1499 1500 1501 1502 1503

  const int is_inter = is_inter_block(mbmi);
  const SCAN_ORDER *const scan_order =
      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));

  const TxbProbs txb_probs = { xd->fc->dc_sign[plane_type],
                               nz_map,
                               coeff_base,
                               coeff_lps,
1504 1505
                               xd->fc->eob_flag[txs_ctx][plane_type],
                               xd->fc->txb_skip[txs_ctx] };
1506 1507 1508 1509 1510 1511

  const int shift = av1_get_tx_scale(tx_size);
  const int64_t rdmult =
      (x->rdmult * plane_rd_mult[is_inter][plane_type] + 2) >> 2;
  const int64_t rddiv = x->rddiv;

1512 1513 1514
  TxbInfo txb_info = { qcoeff,     dqcoeff, tcoeff, dequant, shift, tx_size,
                       txs_ctx,    bwl,     stride, height,  eob,   seg_eob,
                       scan_order, txb_ctx, rdmult, rddiv };
1515 1516 1517 1518 1519 1520 1521
  TxbCache txb_cache;
  gen_txb_cache(&txb_cache, &txb_info);

  const int update = optimize_txb(&txb_info, &txb_probs, &txb_cache, 0);
  if (update) p->eobs[block] = txb_info.eob;
  return txb_info.eob;
}
1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536
int av1_get_txb_entropy_context(const tran_low_t *qcoeff,
                                const SCAN_ORDER *scan_order, int eob) {
  const int16_t *scan = scan_order->scan;
  int cul_level = 0;
  int c;
  for (c = 0; c < eob; ++c) {
    cul_level += abs(qcoeff[scan[c]]);
  }

  cul_level = AOMMIN(COEFF_CONTEXT_MASK, cul_level);
  set_dc_sign(&cul_level, qcoeff[0]);

  return cul_level;
}

1537 1538 1539
void av1_update_txb_context_b(int plane, int block, int blk_row, int blk_col,
                              BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                              void *arg) {
Jingning Han's avatar
Jingning Han committed
1540
  struct tokenize_b_args *const args = arg;
1541 1542
  const AV1_COMP *cpi = args->cpi;
  const AV1_COMMON *cm = &cpi->common;
Angie Chiang's avatar
Angie Chiang committed
1543 1544 1545
  ThreadData *const td = args->td;
  MACROBLOCK *const x = &td->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
1546
  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
Angie Chiang's avatar
Angie Chiang committed
1547 1548
  struct macroblock_plane *p = &x->plane[plane];
  struct macroblockd_plane *pd = &xd->plane[plane];
1549 1550 1551 1552 1553 1554
  const uint16_t eob = p->eobs[block];
  const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
  const PLANE_TYPE plane_type = pd->plane_type;
  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
  const SCAN_ORDER *const scan_order =
      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
Angie Chiang's avatar
Angie Chiang committed
1555
  (void)plane_bsize;
1556

1557
  int cul_level = av1_get_txb_entropy_context(qcoeff, scan_order, eob);
1558
  av1_set_contexts(xd, pd, plane, tx_size, cul_level, blk_col, blk_row);
Angie Chiang's avatar
Angie Chiang committed
1559 1560
}

1561 1562 1563
void av1_update_and_record_txb_context(int plane, int block, int blk_row,
                                       int blk_col, BLOCK_SIZE plane_bsize,
                                       TX_SIZE tx_size, void *arg) {
Jingning Han's avatar
Jingning Han committed
1564
  struct tokenize_b_args *const args = arg;
Angie Chiang's avatar
Angie Chiang committed
1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580
  const AV1_COMP *cpi = args->cpi;
  const AV1_COMMON *cm = &cpi->common;
  ThreadData *const td = args->td;
  MACROBLOCK *const x = &td->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  struct macroblock_plane *p = &x->plane[plane];
  struct macroblockd_plane *pd = &xd->plane[plane];
  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
  int eob = p->eobs[block], update_eob = 0;
  const PLANE_TYPE plane_type = pd->plane_type;
  const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
  tran_low_t *tcoeff = BLOCK_OFFSET(x->mbmi_ext->tcoeff[plane], block);
  const int segment_id = mbmi->segment_id;
  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
  const SCAN_ORDER *const scan_order =
      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
1581
  const int16_t *scan = scan_order->scan;
Angie Chiang's avatar
Angie Chiang committed
1582 1583
  const int seg_eob = get_tx_eob(&cpi->common.seg, segment_id, tx_size);
  int c, i;
1584 1585 1586
  TXB_CTX txb_ctx;
  get_txb_ctx(plane_bsize, tx_size, plane, pd->above_context + blk_col,
              pd->left_context + blk_row, &txb_ctx);
Angie Chiang's avatar
Angie Chiang committed
1587
  const int bwl = b_width_log2_lookup[txsize_to_bsize[tx_size]] + 2;
1588
  const int height = tx_size_high[tx_size];
Angie Chiang's avatar
Angie Chiang committed
1589 1590 1591 1592
  int cul_level = 0;
  unsigned int(*nz_map_count)[SIG_COEF_CONTEXTS][2];
  uint8_t txb_mask[32 * 32] = { 0 };

1593 1594 1595
  TX_SIZE txsize_ctx = get_txsize_context(tx_size);

  nz_map_count = &td->counts->nz_map[txsize_ctx][plane_type];
Angie Chiang's avatar
Angie Chiang committed
1596 1597 1598

  memcpy(tcoeff, qcoeff, sizeof(*tcoeff) * seg_eob);

1599
  ++td->counts->txb_skip[txsize_ctx][txb_ctx.txb_skip_ctx][eob == 0];
1600
  x->mbmi_ext->txb_skip_ctx[plane][block] = txb_ctx.txb_skip_ctx;
Angie Chiang's avatar
Angie Chiang committed
1601 1602 1603 1604 1605 1606 1607 1608

  x->mbmi_ext->eobs[plane][block] = eob;

  if (eob == 0) {
    av1_set_contexts(xd, pd, plane, tx_size, 0, blk_col, blk_row);
    return;
  }

Angie Chiang's avatar
Angie Chiang committed
1609
#if CONFIG_TXK_SEL
1610 1611
  av1_update_tx_type_count(cm, xd, block, plane, mbmi->sb_type, tx_size,
                           td->counts);
Angie Chiang's avatar
Angie Chiang committed
1612
#endif
Angie Chiang's avatar
Angie Chiang committed
1613 1614 1615 1616

  for (c = 0; c < eob; ++c) {
    tran_low_t v = qcoeff[scan[c]];
    int is_nz = (v != 0);
1617
    int coeff_ctx = get_nz_map_ctx(tcoeff, txb_mask, scan[c], bwl, height);
1618
    int eob_ctx = get_eob_ctx(tcoeff, scan[c], txsize_ctx);
Angie Chiang's avatar
Angie Chiang committed
1619 1620 1621 1622 1623 1624

    if (c == seg_eob - 1) break;

    ++(*nz_map_count)[coeff_ctx][is_nz];

    if (is_nz) {
1625
      ++td->counts->eob_flag[txsize_ctx][plane_type][eob_ctx][c == (eob - 1)];
Angie Chiang's avatar
Angie Chiang committed
1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639
    }
    txb_mask[scan[c]] = 1;
  }

  // Reverse process order to handle coefficient level and sign.
  for (i = 0; i < NUM_BASE_LEVELS; ++i) {
    update_eob = 0;
    for (c = eob - 1; c >= 0; --c) {
      tran_low_t v = qcoeff[scan[c]];
      tran_low_t level = abs(v);
      int ctx;

      if (level <= i) continue;

1640
      ctx = get_base_ctx(tcoeff, scan[c], bwl, height, i + 1);
Angie Chiang's avatar
Angie Chiang committed
1641 1642

      if (level == i + 1) {
1643
        ++td->counts->coeff_base[txsize_ctx][plane_type][i][ctx][1];
Angie Chiang's avatar
Angie Chiang committed
1644
        if (c == 0) {
1645
          int dc_sign_ctx = txb_ctx.dc_sign_ctx;
Angie Chiang's avatar
Angie Chiang committed
1646 1647 1648 1649 1650 1651 1652

          ++td->counts->dc_sign[plane_type][dc_sign_ctx][v < 0];
          x->mbmi_ext->dc_sign_ctx[plane][block] = dc_sign_ctx;
        }
        cul_level += level;
        continue;
      }